From a99d8df26d331d87bae4f27cd52ce5ec4d97fa7a Mon Sep 17 00:00:00 2001 From: "A.J. Shulman" Date: Thu, 3 Apr 2025 11:45:02 -0400 Subject: attempt --- src/server/chunker/pdf_chunker.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'src/server/chunker') diff --git a/src/server/chunker/pdf_chunker.py b/src/server/chunker/pdf_chunker.py index 697550f2e..9d4cfb1b2 100644 --- a/src/server/chunker/pdf_chunker.py +++ b/src/server/chunker/pdf_chunker.py @@ -153,7 +153,7 @@ class ElementExtractor: xref = img_info[0] # XREF of the image in the PDF base_image = page.parent.extract_image(xref) # Extract the image by its XREF image_bytes = base_image["image"] - image = Image.open(io.BytesIO(image_bytes)) # Convert bytes to PIL image + image = Image.open(io.BytesIO(image_bytes)).convert("RGB") # Ensure it's RGB before saving as PNG width_ratio = img.width / page.rect.width # Scale factor for width height_ratio = img.height / page.rect.height # Scale factor for height -- cgit v1.2.3-70-g09d2 From 6e8f05495dfcf7d64bdc424503f874fece85e291 Mon Sep 17 00:00:00 2001 From: "A.J. Shulman" Date: Thu, 10 Apr 2025 12:36:07 -0400 Subject: trying again --- .../views/nodes/chatbot/agentsystem/Agent.ts | 14 +- src/client/views/nodes/chatbot/guides/guide.md | 647 +++++++++++++++++++++ .../views/nodes/chatbot/tools/CreateAnyDocTool.ts | 158 ----- .../nodes/chatbot/tools/DocumentMetadataTool.ts | 561 +++++++++--------- src/server/chunker/pdf_chunker.py | 5 +- 5 files changed, 952 insertions(+), 433 deletions(-) create mode 100644 src/client/views/nodes/chatbot/guides/guide.md delete mode 100644 src/client/views/nodes/chatbot/tools/CreateAnyDocTool.ts (limited to 'src/server/chunker') diff --git a/src/client/views/nodes/chatbot/agentsystem/Agent.ts b/src/client/views/nodes/chatbot/agentsystem/Agent.ts index bab2081b3..b166254bb 100644 --- a/src/client/views/nodes/chatbot/agentsystem/Agent.ts +++ b/src/client/views/nodes/chatbot/agentsystem/Agent.ts @@ -8,7 +8,6 @@ import { StreamedAnswerParser } from '../response_parsers/StreamedAnswerParser'; import { BaseTool } from '../tools/BaseTool'; import { CalculateTool } from '../tools/CalculateTool'; //import { CreateAnyDocumentTool } from '../tools/CreateAnyDocTool'; -import { CreateDocTool } from '../tools/CreateDocumentTool'; import { DataAnalysisTool } from '../tools/DataAnalysisTool'; import { DocumentMetadataTool } from '../tools/DocumentMetadataTool'; import { ImageCreationTool } from '../tools/ImageCreationTool'; @@ -83,13 +82,8 @@ export class Agent { dataAnalysis: new DataAnalysisTool(csvData), websiteInfoScraper: new WebsiteInfoScraperTool(addLinkedUrlDoc), searchTool: new SearchTool(addLinkedUrlDoc), - // createCSV: new CreateCSVTool(createCSVInDash), noTool: new NoTool(), - imageCreationTool: new ImageCreationTool(createImage), - // createTextDoc: new CreateTextDocTool(addLinkedDoc), - // createDoc: new CreateDocTool(addLinkedDoc), - // createAnyDocument: new CreateAnyDocumentTool(addLinkedDoc), - // dictionary: new DictionaryTool(), + //imageCreationTool: new ImageCreationTool(createImage), documentMetadata: new DocumentMetadataTool(this), }; } @@ -103,7 +97,7 @@ export class Agent { * @param maxTurns The maximum number of turns to allow in the conversation. * @returns The final response from the assistant. */ - async askAgent(question: string, onProcessingUpdate: (processingUpdate: ProcessingInfo[]) => void, onAnswerUpdate: (answerUpdate: string) => void, maxTurns: number = 30): Promise { + async askAgent(question: string, onProcessingUpdate: (processingUpdate: ProcessingInfo[]) => void, onAnswerUpdate: (answerUpdate: string) => void, maxTurns: number = 50): Promise { console.log(`Starting query: ${question}`); const MAX_QUERY_LENGTH = 1000; // adjust the limit as needed @@ -470,7 +464,7 @@ export class Agent { actionInput.fieldValue = String(actionInput.fieldValue); } } - + // Handle fieldEdits parameter (for multiple field edits) if ('fieldEdits' in actionInput && actionInput.fieldEdits) { try { @@ -521,7 +515,7 @@ export class Agent { /** * Reinitializes the DocumentMetadataTool with a direct reference to the ChatBox instance. * This ensures that the tool can properly access the ChatBox document and find related documents. - * + * * @param chatBox The ChatBox instance to pass to the DocumentMetadataTool */ public reinitializeDocumentMetadataTool(chatBox: any): void { diff --git a/src/client/views/nodes/chatbot/guides/guide.md b/src/client/views/nodes/chatbot/guides/guide.md new file mode 100644 index 000000000..2af76490d --- /dev/null +++ b/src/client/views/nodes/chatbot/guides/guide.md @@ -0,0 +1,647 @@ +# Dash Agent Tool Development Guide + +**Table of Contents** + +1. [Introduction: The Role and Potential of Tools](#1-introduction-the-role-and-potential-of-tools) + - Beyond Information Retrieval: Action and Creation + - The Agent as an Extension of the User within Dash +2. [Core Agent Architecture Deep Dive](#2-core-agent-architecture-deep-dive) + - The ReAct-Inspired Interaction Loop: Rationale and Flow + - XML Structure: Why XML? Parsing and LLM Guidance + - Stages (``) and Roles (`role="..."`): Enforcing Order + - Message Management (`messages`, `interMessages`): Building Context + - State Handling: Agent's Internal State vs. Tool Statelessness + - Key Components Revisited (`Agent.ts`, `prompts.ts`, `BaseTool.ts`, Parsers) + - Role of `prompts.ts`: Template and Dynamic Content Injection + - Limits and Safeguards (`maxTurns`) +3. [Anatomy of a Dash Agent Tool (Detailed Breakdown)](#3-anatomy-of-a-dash-agent-tool-detailed-breakdown) + - The `BaseTool` Abstract Class: Foundation and Contract + - `ToolInfo`: Defining Identity and LLM Instructions + - `name`: Uniqueness and LLM Invocation Trigger + - `description`: The LLM's Primary Guide - _Dynamically Injected into Prompt_ + - `parameterRules`: The Input Contract (In Depth) + - `citationRules`: Controlling Grounding in the Final Answer + - The `execute` Method: Heart of the Tool + - Asynchronous Nature (`async/await`) + - Receiving Arguments (`args: ParametersType

`) + - Performing the Core Logic (API calls, Dash functions) + - Returning `Observation[]`: The Output Contract (In Depth) + - The `inputValidator` Method: Handling Edge Cases +4. [The Agent-Tool Interaction Flow (Annotated XML Trace)](#4-the-agent-tool-interaction-flow-annotated-xml-trace) + - Detailed Step-by-Step with `Agent.ts` actions highlighted +5. [Step-by-Step Guide: Creating a New Tool](#5-step-by-step-guide-creating-a-new-tool) + - Step 1: Define Goal, Scope, Inputs, Outputs, Dash Interactions, Side Effects + - Step 2: Create the Tool Class File (Directory Structure) + - Step 3: Define Parameters (`parameterRules`) - Type Handling, Arrays + - Step 4: Define Tool Information (`ToolInfo`) - Crafting the _Crucial_ `description` + - Step 5: Implement `execute` - Defensive Coding, Using Injected Functions, Error Handling Pattern + - Step 6: Format Output (`Observation[]`) - Chunk Structure, `chunk_type`, IDs + - Step 7: Register Tool in `Agent.ts` - _This makes the tool available to the prompt_ + - Step 8: Verify Prompt Integration (No Manual Editing Needed) + - Step 9: Testing Your Tool - Strategies and What to Look For +6. [Deep Dive: Advanced Concepts & Patterns](#6-deep-dive-advanced-concepts--patterns) + - Handling Complex Data Types (Arrays, Objects) in Parameters/Observations + - Binary Data Handling (e.g., Base64 in Chunks) + - Managing Long-Running Tasks (Beyond simple `await`) + - Tools Needing Dash Context (Passing `this` vs. specific functions) + - The Role of `chunk_id` and `chunk_type` +7. [Best Practices and Advanced Considerations](#7-best-practices-and-advanced-considerations) + - Error Handling & Reporting (Specific Error Chunks) + - Security Considerations (Input Sanitization, API Key Management, Output Filtering) + - Performance Optimization (Minimize `execute` workload) + - Idempotency: Designing for Retries + - Tool Granularity: Single Responsibility Principle + - Context Window Management (Concise Descriptions are Key) + - User Experience (Tool output clarity) + - Maintainability and Code Comments +8. [Debugging Strategies](#8-debugging-strategies) + - Console Logging within `execute` + - Inspecting `interMessages` in `Agent.ts` + - Testing Tool Logic Standalone + - Analyzing LLM Failures (Incorrect tool choice -> Check `description`, bad parameters) +9. [Example: `CreateDashNoteTool`](#9-example-createdashnotetool) +10. [Glossary of Key Terms](#10-glossary-of-key-terms) +11. [Conclusion](#11-conclusion) + +--- + +## 1. Introduction: The Role and Potential of Tools + +Welcome, Dash team member! This guide will walk you through creating new tools for the Dash Agent. The Agent is designed to interact with users, understand their queries, and leverage specialized **Tools** to perform actions or retrieve information that the core Large Language Model (LLM) cannot do on its own. + +Tools extend the Agent's capabilities beyond simple conversation. They allow the Agent to: + +- Interact with external APIs (e.g., web search, calculators, image generation). +- Access and process data specific to the user's Dash environment (e.g., querying document metadata, analyzing linked CSVs). +- Perform actions within Dash (e.g., creating new documents, adding links, modifying metadata). + +By building new tools, you directly enhance the Agent's utility and integration within the Dash ecosystem. + +### Beyond Information Retrieval: Action and Creation + +While tools like `RAGTool` and `SearchTool` retrieve information, others _act_. `CalculateTool` performs computations, `ImageCreationTool` generates content, and importantly, tools like `DocumentMetadataTool` and your custom tools can **modify the Dash environment**, creating documents, adding links, or changing properties. + +### The Agent as an Extension of the User within Dash + +Think of the Agent, equipped with tools, as an intelligent assistant that can perform tasks _on behalf of the user_ directly within their Dash workspace. This deep integration is a key differentiator. + +--- + +## 2. Core Agent Architecture Deep Dive + +Understanding the "why" behind the architecture helps in tool development. + +### The ReAct-Inspired Interaction Loop: Rationale and Flow + +The Agent operates based on a loop inspired by the ReAct (Reason + Act) framework. The LLM alternates between: + +- **Reasoning (``):** Analyzing the query and deciding the next step. +- **Acting (``, ``):** Selecting and preparing to use a tool, or formulating a final answer (``). +- **Observing (``):** Receiving the results from a tool execution. + +This structure (Reason -> Act -> Observe -> Reason...) forces the LLM to break down complex tasks into manageable steps, making the process more reliable and auditable than letting the LLM generate a monolithic plan upfront. + +### XML Structure: Why XML? Parsing and LLM Guidance + +- **Why XML?** LLMs are generally adept at generating well-formed XML. XML's explicit start/end tags make parsing by `Agent.ts` (using libraries like `fast-xml-parser`) more robust and less prone to LLM "hallucinations" breaking the structure compared to formats like JSON in some complex scenarios. +- **LLM Guidance:** The strict XML schema defined in the system prompt provides clear guardrails for the LLM's output, constraining it to valid actions and formats. + +### Stages (``) and Roles (`role="..."`): Enforcing Order + +The `` ensures sequential processing. The `role` attribute indicates the source (e.g., `user`, `assistant`) and dictates control flow. `Agent.ts` _waits_ for a `user` (or `system-error-reporter`) stage after sending an `assistant` stage, enforcing the turn-based nature. The LLM is explicitly told only to generate `assistant` stages. + +### Message Management (`messages`, `interMessages`): Building Context + +- `messages`: The user-facing chat history (persisted in the Dash Doc `data` field). +- `interMessages`: The **internal, complete context** sent to the LLM for each turn. It includes the system prompt, user queries, _all intermediate thoughts, actions, rules, inputs, and observations_. This ensures the LLM has the full history of the current reasoning chain. It grows with each step in the loop. + +### State Handling: Agent's Internal State vs. Tool Statelessness + +- `Agent.ts` manages the conversational state (`interMessages`, current turn number, `processingInfo`, etc.). +- **Tools should be designed to be stateless.** They receive inputs via `args`, perform their action, and return results. Any persistent state relevant to the user's work should reside within Dash Docs/Fobs, accessible perhaps via tools like `DocumentMetadataTool` or specific functions passed to the tool. + +### Key Components Revisited (`Agent.ts`, `prompts.ts`, `BaseTool.ts`, Parsers) + +- `Agent.ts`: The central controller. Parses XML, validates actions, manages the loop, calls `tool.execute`, formats `Observation`s. Handles the streaming updates for the _final_ answer via `StreamedAnswerParser`. Holds the registry of available tools (`this.tools`). +- `prompts.ts` (`getReactPrompt`): Generates the system prompt for the LLM. It acts as a **template** defining the Agent's overall task, rules, and the required XML structure. Crucially, it **dynamically injects the list of available tools** (including their names and descriptions) based on the tools registered in the `Agent.ts` instance at runtime. **_You do not manually add tool descriptions here._** +- `BaseTool.ts`: The abstract class defining the _interface_ all tools must adhere to. Contains properties like `name` and `description` used by `getReactPrompt`. +- Parsers (`AnswerParser`, `StreamedAnswerParser`): Handle the final `` tag, extracting structured content, citations, etc., for UI display (`ChatBox.tsx`). + +### Limits and Safeguards (`maxTurns`) + +`Agent.ts` includes a `maxTurns` limit (default 30) to prevent infinite loops if the LLM gets stuck or fails to reach an `` stage. + +--- + +## 3. Anatomy of a Dash Agent Tool (Detailed Breakdown) + +All tools inherit from the abstract class `BaseTool`. + +### The `BaseTool` Abstract Class: Foundation and Contract + +- Located in `src/components/views/nodes/chatbot/agentsystem/tools/BaseTool.ts`. +- Generic `BaseTool

>`: `P` represents the specific, readonly array of `Parameter` definitions for _your_ tool, ensuring type safety for the `args` in `execute`. +- Defines the public properties (`name`, `description`, `parameterRules`, `citationRules`) and the abstract `execute` method that all tools must implement. + +### `ToolInfo`: Defining Identity and LLM Instructions + +- A configuration object (`{ name: string; description: string; parameterRules: P; citationRules: string; }`) passed to the `BaseTool` constructor. +- `name: string`: + - The **unique identifier** for your tool (e.g., `dictionaryLookup`, `createDashNote`). + - Must match the key used when registering the tool in `Agent.ts`'s `this.tools` map. + - This is the string the LLM will output in the `` tag to invoke your tool. + - Keep it concise and descriptive (camelCase recommended). +- `description: string`: The LLM's Primary Guide - _Dynamically Injected into Prompt_. + - This text is extracted from your `ToolInfo` object when `getReactPrompt` is called. + - It's **the text the LLM sees** to understand your tool's purpose and when to use it. + - **Crafting this is critical.** Make it extremely clear, concise, and accurate. Explicitly state: + - What the tool _does_. + - What _inputs_ it needs (briefly). + - What _output_ it provides. + - _Crucially_, under what circumstances the Agent should _choose_ this tool over others. (e.g., "Use this tool to create _new_ Dash notes, not for editing existing ones.") +- `parameterRules: P` (where `P extends ReadonlyArray`): + - The readonly array defining the **exact inputs** your `execute` method expects. + - Each element is a `Parameter` object (`{ name: string; type: 'string' | ... ; required: boolean; description: string; max_inputs?: number }`): + - `name`: Name of the parameter (e.g., `wordToDefine`, `noteContent`). Used as the key in the `args` object passed to `execute`. + - `type`: `'string' | 'number' | 'boolean' | 'string[]' | 'number[]' | 'boolean[]'`. `Agent.ts` uses this for basic validation and parsing (specifically for arrays). + - `required`: `true` if the LLM _must_ provide this parameter for the tool to function. `Agent.ts` checks this before calling `execute` (unless `inputValidator` overrides). + - `description`: Explanation of the parameter _for the LLM_. Guides the LLM on _what value_ to provide. Be specific (e.g., "The exact URL to scrape", "A search query suitable for web search"). + - `max_inputs?`: Optional. For array types (`string[]`, etc.), suggests a limit to the LLM on the number of items to provide. +- `citationRules: string`: + - Instructions for the LLM on how to construct the `` block within the final `` tag _when information obtained from this specific tool is used_. + - Directly influences the grounding and verifiability of the Agent's final response. + - Be explicit about the format: "Cite using `Optional text snippet`". + - Specify what goes into `chunk_id` (e.g., "Use the ID provided in the observation chunk"), `type` (a constant string representing your tool's output type), and whether the text snippet should be included (often empty for URLs, calculations). + - If no citation is appropriate (e.g., calculator, ephemeral action), state clearly: "No citation needed for this tool's output." + +### The `execute` Method: Heart of the Tool + +- `abstract execute(args: ParametersType

): Promise;` +- **Asynchronous Nature (`async/await`):** Must be `async` because tool actions often involve I/O (network requests, database access via Dash functions, filesystem). Must return a `Promise`. +- **Receiving Arguments (`args: ParametersType

`):** + - Receives a single argument `args`. This object's keys are your defined parameter names (from `parameterRules`), and the values are those provided by the LLM in the `` block. + - The type `ParametersType

` infers the structure of `args` based on your specific `parameterRules` definition (`P`), providing TypeScript type safety. + - `Agent.ts` performs basic validation (required fields) and type coercion (for arrays) before calling `execute`. However, **always perform defensive checks** within `execute` (e.g., check if required args are truly present and not empty strings, check types if crucial). +- **Performing the Core Logic:** This is where your tool does its work. Examples: + - Call external APIs (using `axios`, `fetch`, or specific SDKs). + - Call Dash functions passed via the constructor (e.g., `this._createDocInDash(...)`, `this._addLinkedUrlDoc(...)`). + - Perform calculations or data transformations. + - Interact with other backend systems if necessary. +- **Returning `Observation[]`: The Output Contract (In Depth)** + - The method **must** resolve to an array of `Observation` objects. Even if there's only one piece of output, return it in an array: `[observation]`. + - Each `Observation` object usually has the structure `{ type: 'text', text: string }`. Other types might be possible but `text` is standard for LLM interaction. + - The `text` string **must** contain the tool's output formatted within one or more `` tags. This is how the Agent passes structured results back to the LLM. + - Format: `OUTPUT_DATA` + - `chunk_id`: A unique identifier for this specific piece of output (use `uuidv4()`). Essential for linking citations back to observations. + - `chunk_type`: A string literal describing the _semantic type_ of the data (e.g., `'search_result_url'`, `'calculation_result'`, `'note_creation_status'`, `'error'`, `'metadata_info'`). Helps the LLM interpret the result. Choose consistent and descriptive names. + - `OUTPUT_DATA`: The actual result from your tool. Can be simple text, JSON stringified data, etc. Keep it concise if possible for the LLM context. + - **Return errors** using the same format, but with `chunk_type="error"` and a descriptive error message inside the chunk tag. This allows the Agent loop to continue gracefully and potentially inform the LLM or user. + +### The `inputValidator` Method: Handling Edge Cases + +- `inputValidator(inputParam: ParametersType) { return false; }` (Default implementation in `BaseTool`). +- Override this method _only_ if your tool needs complex input validation logic beyond simple `required` checks (e.g., dependencies between parameters). +- If you override it to return `true`, `Agent.ts` will skip its standard check for missing _required_ parameters. Your `execute` method becomes fully responsible for validating the `args` object. +- Use case example: `DocumentMetadataTool` uses it to allow either `fieldName`/`fieldValue` OR `fieldEdits` to be provided for the "edit" action. + +--- + +## 4. The Agent-Tool Interaction Flow (Annotated XML Trace) + +Let's trace the `dictionaryLookup` example with `Agent.ts` actions: + +1. **User Input:** User types "What is hypermedia?" and submits. + + - `// ChatBox.tsx calls agent.askAgent("What is hypermedia?")` + - `// Agent.ts adds stage 1 to interMessages:` + + ```xml + + What is hypermedia? + + ``` + + - `// Agent.ts calls LLM with interMessages.` + +2. **LLM Thought & Action:** LLM processes the query and system prompt (which includes the dynamically generated description for `dictionaryLookup`). + + - `// LLM responds with stage 2:` + + ```xml + + The user is asking for a definition. The dictionaryLookup tool is appropriate for this. + dictionaryLookup + + ``` + + - `// Agent.ts parses stage 2. Finds dictionaryLookup.` + - `// Agent.ts retrieves the dictionaryTool instance: tool = this.tools['dictionaryLookup'].` + - `// Agent.ts gets parameter rules: rules = tool.getActionRule().` + +3. **Agent Provides Rules:** `Agent.ts` formats the rules into XML. + + - `// Agent.ts adds stage 3 to interMessages:` + + ```xml + + + dictionaryLookup + Looks up the definition of a given English word. + Cite the definition using the provided chunk_id and type="dictionary_definition". Leave citation content empty. + + + string + The word to define. + true + + + + + ``` + + - `// Agent.ts calls LLM with updated interMessages.` + +4. **LLM Provides Inputs:** LLM uses the rules to formulate the required inputs. + + - `// LLM responds with stage 4:` + + ```xml + + + Looking up the definition for the word 'hypermedia'. + + hypermedia + + + + ``` + + - `// Agent.ts parses stage 4. Finds . Extracts inputs: { word: 'hypermedia' }.` + - `// Agent.ts validates required params (finds 'word'). Checks tool.inputValidator (returns false). OK.` + - `// Agent.ts calls: const observations = await dictionaryTool.execute({ word: 'hypermedia' });` + +5. **Tool Execution -> Observation:** The `dictionaryTool.execute` method runs, calls an API, gets the definition, and formats the `Observation`. + + - `// dictionaryTool.execute resolves Promise with:` + + ```javascript + [{ type: 'text', text: 'Hypermedia is an extension of hypertext allowing...' }]; + ``` + + - `// Agent.ts receives the resolved Observation[].` + +6. **Agent Provides Observation:** `Agent.ts` formats the observation(s) into the next stage. + + - `// Agent.ts adds stage 5 to interMessages:` + + ```xml + + + Hypermedia is an extension of hypertext allowing... + + + ``` + + - `// Agent.ts calls LLM with updated interMessages.` + +7. **LLM Generates Final Answer:** LLM uses the full context (query, thoughts, actions, observation) and the `citationRules` from the `dictionaryLookup` tool. + - `// LLM responds with stage 6:` + ```xml + + I have received the definition from the dictionary tool. I will now format the final answer according to the specified structure, including the citation. + + Hypermedia is an extension of hypertext allowing graphics, audio, video, plain text and hyperlinks to intertwine to create a generally non-linear medium of information. + + + + + Can you give an example of a hypermedia system? + How does hypermedia differ from multimedia? + Who coined the term hypermedia? + + Used dictionaryLookup tool to define 'hypermedia'. + + + ``` + - `// Agent.ts parses stage 6. Finds . Calls AnswerParser.` + - `// Agent.ts updates ChatBox UI (`\_history.push(...)`). Loop ends.` + +--- + +## 5. Step-by-Step Guide: Creating a New Tool + +Let's use the example of creating a `CreateDashNoteTool`. + +### Step 1: Define Goal, Scope, Inputs, Outputs, Dash Interactions, Side Effects + +- **Goal:** Allow Agent to create a new text note (`DocumentType.TEXT` or equivalent) in Dash. +- **Scope:** Creates a _simple_ note with title and text content. Does not handle complex formatting, linking (beyond default linking to ChatBox if handled by the creation function), or specific placement beyond a potential default offset. +- **Inputs:** `noteTitle` (string, required), `noteContent` (string, required). +- **Outputs (Observation):** Confirmation message with new note's Dash Document ID, or an error message. +- **Dash Interactions:** Calls a function capable of creating Dash documents (e.g., `createDocInDash` passed via constructor). +- **Side Effects:** A new Dash text document is created in the user's space and potentially linked to the ChatBox. + +### Step 2: Create the Tool Class File (Directory Structure) + +- Create file: `src/components/views/nodes/chatbot/agentsystem/tools/CreateDashNoteTool.ts` +- Ensure it's within the `tools` subdirectory. + +### Step 3: Define Parameters (`parameterRules`) - Type Handling, Arrays + +- Use `as const` for the array to allow TypeScript to infer literal types, which aids `ParametersType`. +- Define `noteTitle` and `noteContent` as required strings. + +```typescript +import { Parameter } from '../types/tool_types'; + +const createDashNoteToolParams = [ + { + name: 'noteTitle', + type: 'string', + required: true, + description: 'The title for the new Dash note document. Cannot be empty.', + }, + { + name: 'noteContent', + type: 'string', + required: true, + description: 'The text content for the new Dash note. Can be an empty string.', // Specify if empty content is allowed + }, +] as const; // Use 'as const' for precise typing + +// Infer the type for args object in execute +type CreateDashNoteToolParamsType = typeof createDashNoteToolParams; +``` + +### Step 4: Define Tool Information (`ToolInfo`) - Crafting the _Crucial_ `description` + +- This object's `description` is key for the LLM. + +```typescript +import { ToolInfo, ParametersType } from '../types/tool_types'; + +// Assuming createDashNoteToolParams and CreateDashNoteToolParamsType are defined above + +const createDashNoteToolInfo: ToolInfo = { + name: 'createDashNote', // Must match registration key in Agent.ts + description: + 'Creates a *new*, simple text note document within the current Dash view. Requires a title and text content. The note will be linked to the ChatBox and placed nearby with default dimensions. Use this when the user asks to create a new note, save information, or write something down persistently in Dash.', + parameterRules: createDashNoteToolParams, + citationRules: 'This tool creates a document. The observation confirms success and provides the new document ID. No citation is typically needed in the final answer unless confirming the action.', +}; +``` + +### Step 5: Implement `execute` - Defensive Coding, Using Injected Functions, Error Handling Pattern + +- Implement the `execute` method within your class. +- Wrap logic in `try...catch`. +- Validate inputs defensively. +- Check injected dependencies (`this._createDocInDash`). +- Call the Dash function. +- Handle the return value. +- Format success or error `Observation`. + +```typescript +import { BaseTool } from './BaseTool'; +import { Observation } from '../types/types'; +import { supportedDocTypes } from '../types/tool_types'; +import { parsedDoc } from '../chatboxcomponents/ChatBox'; // May need adjustment based on actual path +import { Doc } from '../../../../../../fields/Doc'; // Adjust path as needed +import { v4 as uuidv4 } from 'uuid'; +import { RTFCast } from '../../../../../../fields/Types'; // Adjust path as needed + +// Assuming createDashNoteToolParams, CreateDashNoteToolParamsType, createDashNoteToolInfo are defined above + +export class CreateDashNoteTool extends BaseTool { + // Dependency: Function to create a document in Dash + private _createDocInDash: (doc: parsedDoc) => Doc | undefined; + + // Constructor to inject dependencies + constructor(createDocInDash: (doc: parsedDoc) => Doc | undefined) { + super(createDashNoteToolInfo); + if (typeof createDocInDash !== 'function') { + console.error('CreateDashNoteTool Error: createDocInDash function dependency not provided during instantiation!'); + // Consider throwing an error or setting a flag to prevent execution + } + this._createDocInDash = createDocInDash; + } + + async execute(args: ParametersType): Promise { + const chunkId = uuidv4(); // Unique ID for this observation + const { noteTitle, noteContent } = args; + + // --- Input Validation --- + if (typeof noteTitle !== 'string' || !noteTitle.trim()) { + return [{ type: 'text', text: `Invalid input: Note title must be a non-empty string.` }]; + } + if (typeof noteContent !== 'string') { + // Assuming empty content IS allowed based on description + // If not allowed, return error here. + return [{ type: 'text', text: `Invalid input: Note content must be a string.` }]; + } + if (!this._createDocInDash) { + return [{ type: 'text', text: `Tool Configuration Error: Document creation function not available.` }]; + } + // --- End Validation --- + + try { + const trimmedTitle = noteTitle.trim(); + + // Prepare the document object for the creation function + const noteDoc: parsedDoc = { + doc_type: supportedDocTypes.note, // Use the correct type for a text note + title: trimmedTitle, + data: RTFCast(noteContent) as unknown as string, // Ensure data is correctly formatted if needed + // Example default properties: + _width: 300, + _layout_fitWidth: false, + _layout_autoHeight: true, + backgroundColor: '#FFFFE0', // Light yellow background + // Add x, y coordinates if desired, potentially relative to ChatBox if context is available + }; + + console.log(`CreateDashNoteTool: Attempting to create doc:`, { title: noteDoc.title, type: noteDoc.doc_type }); // Avoid logging full content + + // Call the injected Dash function + const createdDoc = this._createDocInDash(noteDoc); + + // Check the result + if (createdDoc && createdDoc.id) { + const successMessage = `Successfully created note titled "${trimmedTitle}" with ID: ${createdDoc.id}. It has been added to your current view.`; + console.log(`CreateDashNoteTool: Success - ${successMessage}`); + // Return observation confirming success + return [{ type: 'text', text: `${successMessage}` }]; + } else { + console.error('CreateDashNoteTool Error: _createDocInDash returned undefined or document without an ID.'); + throw new Error('Dash document creation failed or did not return a valid document ID.'); + } + } catch (error) { + console.error(`CreateDashNoteTool: Error creating note titled "${noteTitle.trim()}":`, error); + const errorMessage = error instanceof Error ? error.message : 'An unknown error occurred during note creation.'; + // Return observation indicating error + return [{ type: 'text', text: `Error creating note: ${errorMessage}` }]; + } + } +} +``` + +### Step 6: Format Output (`Observation[]`) - Chunk Structure, `chunk_type`, IDs + +- Ensure the `text` field within the returned `Observation` contains `...`. +- Use a specific `chunk_type` (e.g., `note_creation_status`, `error`). +- Generate a unique `chunk_id` using `uuidv4()`. +- The text inside the chunk should be informative for the LLM and potentially for debugging. + +### Step 7: Register Tool in `Agent.ts` - _This makes the tool available to the prompt_ + +- Import your tool class at the top of `Agent.ts`: + ```typescript + import { CreateDashNoteTool } from '../tools/CreateDashNoteTool'; + ``` +- In the `Agent` constructor, instantiate your tool within the `this.tools = { ... };` block. Ensure the key matches `ToolInfo.name` and pass any required dependencies (like the `createDocInDash` function). + + ```typescript + constructor( + _vectorstore: Vectorstore, + summaries: () => string, + history: () => string, + csvData: () => { filename: string; id: string; text: string }[], + addLinkedUrlDoc: (url: string, id: string) => void, + createImage: (result: any, options: any) => void, // Use specific types if known + createDocInDashFunc: (doc: parsedDoc) => Doc | undefined, // Renamed for clarity + createCSVInDash: (url: string, title: string, id: string, data: string) => void + ) { + // ... existing initializations (OpenAI client, vectorstore, etc.) ... + this.vectorstore = _vectorstore; + this._summaries = summaries; + this._history = history; + this._csvData = csvData; + + this.tools = { + calculate: new CalculateTool(), + rag: new RAGTool(this.vectorstore), + dataAnalysis: new DataAnalysisTool(csvData), + websiteInfoScraper: new WebsiteInfoScraperTool(addLinkedUrlDoc), + searchTool: new SearchTool(addLinkedUrlDoc), + noTool: new NoTool(), + imageCreationTool: new ImageCreationTool(createImage), + documentMetadata: new DocumentMetadataTool(this), // Pass ChatBox instance if needed by tool + // Register the new tool here: + createDashNote: new CreateDashNoteTool(createDocInDashFunc), // Pass the required function + }; + // ... rest of constructor + } + ``` + +- **Verify Dependencies:** Ensure that the `createDocInDashFunc` parameter (or however you name it) is actually being passed into the `Agent` constructor when it's instantiated (likely within `ChatBox.tsx`). Trace the dependency chain. + +### Step 8: Verify Prompt Integration (No Manual Editing Needed) + +- **No manual changes are needed in `prompts.ts`**. The `getReactPrompt` function dynamically builds the `` section from `this.tools`. +- **Verify (Recommended):** Temporarily add `console.log(systemPrompt)` in `Agent.ts` right after `const systemPrompt = getReactPrompt(...)` within the `askAgent` method. Run a query. Examine the console output to confirm the system prompt includes your tool's `` and `<description>` within the `<tools>` block. Remove the log afterward. + +### Step 9: Testing Your Tool - Strategies and What to Look For + +- **Functional Tests:** Use specific prompts like "Create a note called 'Ideas' with content 'Test 1'." Check the Dash UI for the note and the chat for the success message/ID. +- **Edge Case Tests:** Test empty titles (should fail validation), empty content (should succeed if allowed), titles/content with special characters or excessive length. +- **LLM Interaction Tests:** Use less direct prompts like "Save this thought: Remember to buy milk." Does the LLM correctly identify the need for your tool and extract/request the title and content? +- **Failure Tests:** If possible, simulate failure in the dependency (`createDocInDash`) to ensure the `error` chunk is returned correctly. +- **Console/Debugging:** Use `console.log` within `execute` and inspect `interMessages` in `Agent.ts` to trace the flow and identify issues. + +--- + +## 6. Deep Dive: Advanced Concepts & Patterns + +### Handling Complex Data Types (Arrays, Objects) in Parameters/Observations + +- **Parameters:** For complex inputs, define the parameter `type` as `string` in `parameterRules`. In the `description`, instruct the LLM to provide a **valid JSON string**. Inside your `execute` method, use `JSON.parse()` within a `try...catch` block to parse this string. Handle potential parsing errors gracefully (return an `error` chunk). +- **Observations:** To return structured data, `JSON.stringify` your object/array and embed this string _inside_ the `<chunk>` tag. Use a specific `chunk_type` (e.g., `json_data_analysis`). The LLM might need guidance (via prompt engineering) on how to interpret and use this JSON data effectively in its final response. + +### Binary Data Handling (e.g., Base64 in Chunks) + +- **Avoid large binary data in observations.** Context windows are limited. +- **Preferred:** Save binary data server-side (e.g., using `DashUploadUtils` or similar) or reference existing Dash media docs. Return a **reference** (URL, Doc ID, file path accessible by Dash) within the `<chunk>`. +- **If absolutely necessary:** For small images/data needed _directly_ by the LLM, Base64 encode it inside the chunk: `<chunk chunk_id="..." chunk_type="base64_image_png">BASE64_STRING</chunk>`. + +### Managing Long-Running Tasks (Beyond simple `await`) + +- The agent's `askAgent` loop `await`s `tool.execute()`. Tasks taking more than ~5-10 seconds degrade user experience. Very long tasks risk timeouts. +- **Limitation:** The current architecture doesn't have built-in support for asynchronous background jobs with status polling. +- **Possible (Complex) Workaround:** + 1. Tool `execute` initiates a long-running _external_ process (like the Python PDF chunker) or backend job. + 2. `execute` _immediately_ returns an `Observation` like `<chunk chunk_type="task_initiated" job_id="JOB123">Processing started. Use status check tool with ID JOB123.</chunk>`. + 3. Requires a _separate_ `StatusCheckTool` that takes a `job_id` and queries the external process/backend for status. + 4. This adds significant complexity to the agent's reasoning flow. Use only if absolutely necessary. + +### Tools Needing Dash Context (Passing `this` vs. specific functions) + +- **Specific Functions (Preferred):** Pass only necessary functions from `ChatBox`/Dash utilities. Promotes modularity and testability. Requires updating constructors if needs change. +- **`ChatBox` Instance (`this`) (As in `DocumentMetadataTool`):** Provides broad access to `ChatBox` state (`Document`, `layoutDoc`, computed properties) and methods. Easier for tools with complex Dash interactions but increases coupling and makes testing harder. +- **Decision:** Start with specific functions. Escalate to passing `this` only if the tool's requirements become extensive and unmanageable via individual function injection. + +### The Role of `chunk_id` and `chunk_type` + +- `chunk_id` (e.g., `uuidv4()`): **Traceability & Citation.** Uniquely identifies a piece of data returned by a tool. Allows the final `<answer>`'s `<citation>` tag to precisely reference the source observation via this ID. Essential for debugging and grounding. +- `chunk_type`: **Semantic Meaning.** Tells the LLM _what kind_ of information the chunk contains (e.g., `url`, `calculation_result`, `error`, `note_creation_status`). Guides the LLM in processing the observation and formatting the final answer appropriately. Use consistent and descriptive type names. + +--- + +## 7. Best Practices and Advanced Considerations + +- **Error Handling & Reporting:** Return errors in structured `<chunk chunk_type="error">...</chunk>` format. Include context in the message (e.g., "API call failed for URL: [url]", "Invalid value for parameter: [param_name]"). +- **Security:** + - **Input Sanitization:** **Crucial.** If tool inputs influence API calls, file paths, database queries, etc., validate and sanitize them rigorously. Do not trust LLM output implicitly. + - **API Keys:** Use server-side environment variables (`process.env`) for keys used in backend routes called by tools. Avoid exposing keys directly in client-side tool code if possible. + - **Output Filtering:** Be mindful of sensitive data. Don't leak PII or internal details in observations or error messages. +- **Performance Optimization:** Keep `execute` logic efficient. Minimize blocking operations. Use asynchronous patterns correctly. +- **Idempotency:** Design tools (especially those causing side effects like creation/modification) to be safe if run multiple times with the same input, if possible. +- **Tool Granularity (SRP):** Aim for tools that do one thing well. Complex workflows can be achieved by the LLM chaining multiple focused tools. +- **Context Window Management:** Write concise but clear tool `description`s. Keep `Observation` data relevant and succinct. +- **User Experience:** Tool output (via observations) influences the final answer. Ensure returned data is clear and `citationRules` guide the LLM to produce understandable results. +- **Maintainability:** Use clear code, comments for complex logic, TypeScript types, and follow project conventions. + +--- + +## 8. Debugging Strategies + +1. **`console.log`:** Liberally use `console.log` inside your tool's `execute` method to inspect `args`, intermediate variables, API responses, and the `Observation[]` object just before returning. +2. **Inspect `interMessages`:** Temporarily modify `Agent.ts` (e.g., in the `askAgent` `while` loop) to `console.log(JSON.stringify(this.interMessages, null, 2))` before each LLM call. This shows the exact XML context the LLM sees and its raw XML response. Pinpoint where the conversation deviates or breaks. +3. **Test Standalone:** Create a simple test script (`.ts` file run with `ts-node` or similar). Import your tool. Create mock objects/functions for its dependencies (e.g., `const mockCreateDoc = (doc) => ({ id: 'mock-doc-123', ...doc });`). Instantiate your tool with mocks: `const tool = new YourTool(mockCreateDoc);`. Call `await tool.execute(testArgs);` and assert the output. This isolates tool logic. +4. **Analyzing LLM Failures:** Use the `interMessages` log: + - **Wrong Tool Chosen:** LLM's `<thought>` selects the wrong tool, or uses `<action>noTool</action>` inappropriately. -> **Refine your tool's `description`** in `ToolInfo` for clarity and better differentiation. + - **Missing/Incorrect Parameters:** LLM fails to provide required parameters in `<inputs>`, or provides wrong values. -> **Refine parameter `description`s** in `parameterRules`. Check the `<action_input>` stage in the log. + - **Ignoring Observation/Bad Answer:** LLM gets the correct `<observation>` but generates a poor `<answer>` (ignores data, bad citation). -> Check `chunk_type`, data format inside the chunk, and **clarify `citationRules`**. Simplify observation data if needed. + - **XML Formatting Errors:** LLM returns malformed XML. -> This might require adjusting the system prompt's structure rules or adding more robust parsing/error handling in `Agent.ts`. + +--- + +## 9. Example: `CreateDashNoteTool` + +The code provided in Step 5 serves as a practical example, demonstrating dependency injection, input validation, calling a Dash function, and formatting success/error observations within the required `<chunk>` structure. Ensure the dependency (`createDocInDashFunc`) is correctly passed during `Agent` instantiation in `ChatBox.tsx`. + +--- + +## 10. Glossary of Key Terms + +- **Agent (`Agent.ts`):** The orchestrator class managing the LLM interaction loop and tool usage. +- **Tool (`BaseTool.ts`):** A class extending `BaseTool` to provide specific functionality (API calls, Dash actions). +- **LLM (Large Language Model):** The AI model providing reasoning and text generation (e.g., GPT-4o). +- **ReAct Loop:** The core interaction pattern: Reason -> Act -> Observe. +- **XML Structure:** The tag-based format (`<stage>`, `<thought>`, etc.) for LLM communication. +- **`interMessages`:** The internal, complete conversational context sent to the LLM. +- **`ToolInfo`:** Configuration object (`name`, `description`, `parameterRules`, `citationRules`) defining a tool. **Source of dynamic prompt content for the tool list.** +- **`parameterRules`:** Array defining a tool's expected input parameters. +- **`citationRules`:** Instructions for the LLM on citing a tool's output. +- **`execute`:** The primary asynchronous method within a tool containing its core logic. +- **`Observation`:** The structured object (`{ type: 'text', text: '<chunk>...' }`) returned by `execute`. +- **`<chunk>`:** The required XML-like wrapper within an Observation's `text`, containing `chunk_id` and `chunk_type`. +- **`chunk_type`:** Semantic identifier for the data type within a `<chunk>`. +- **System Prompt (`getReactPrompt`):** The foundational instructions for the LLM, acting as a **template dynamically populated** with registered tool descriptions. +- **Dash Functions:** Capabilities from the Dash environment (e.g., `createDocInDash`) injected into tools. +- **Stateless Tool:** A tool whose output depends solely on current inputs, not past interactions. + +--- + +## 11. Conclusion + +This guide provides a detailed framework for extending the Dash Agent with custom tools. By adhering to the `BaseTool` structure, understanding the agent's interaction flow, crafting clear `ToolInfo` descriptions, implementing robust `execute` methods, and correctly registering your tool in `Agent.ts`, you can build powerful integrations that leverage both AI and the unique capabilities of the Dash hypermedia environment. Remember that testing and careful consideration of dependencies, errors, and security are crucial for creating reliable tools. diff --git a/src/client/views/nodes/chatbot/tools/CreateAnyDocTool.ts b/src/client/views/nodes/chatbot/tools/CreateAnyDocTool.ts deleted file mode 100644 index 754d230c8..000000000 --- a/src/client/views/nodes/chatbot/tools/CreateAnyDocTool.ts +++ /dev/null @@ -1,158 +0,0 @@ -import { toLower } from 'lodash'; -import { Doc } from '../../../../../fields/Doc'; -import { Id } from '../../../../../fields/FieldSymbols'; -import { DocumentOptions } from '../../../../documents/Documents'; -import { parsedDoc } from '../chatboxcomponents/ChatBox'; -import { ParametersType, ToolInfo } from '../types/tool_types'; -import { Observation } from '../types/types'; -import { BaseTool } from './BaseTool'; -import { supportedDocTypes } from './CreateDocumentTool'; - -const standardOptions = ['title', 'backgroundColor']; -/** - * Description of document options and data field for each type. - */ -const documentTypesInfo: { [key in supportedDocTypes]: { options: string[]; dataDescription: string } } = { - [supportedDocTypes.flashcard]: { - options: [...standardOptions, 'fontColor', 'text_align'], - dataDescription: 'an array of two strings. the first string contains a question, and the second string contains an answer', - }, - [supportedDocTypes.text]: { - options: [...standardOptions, 'fontColor', 'text_align'], - dataDescription: 'The text content of the document.', - }, - [supportedDocTypes.html]: { - options: [], - dataDescription: 'The HTML-formatted text content of the document.', - }, - [supportedDocTypes.equation]: { - options: [...standardOptions, 'fontColor'], - dataDescription: 'The equation content as a string.', - }, - [supportedDocTypes.functionplot]: { - options: [...standardOptions, 'function_definition'], - dataDescription: 'The function definition(s) for plotting. Provide as a string or array of function definitions.', - }, - [supportedDocTypes.dataviz]: { - options: [...standardOptions, 'chartType'], - dataDescription: 'A string of comma-separated values representing the CSV data.', - }, - [supportedDocTypes.notetaking]: { - options: standardOptions, - dataDescription: 'The initial content or structure for note-taking.', - }, - [supportedDocTypes.rtf]: { - options: standardOptions, - dataDescription: 'The rich text content in RTF format.', - }, - [supportedDocTypes.image]: { - options: standardOptions, - dataDescription: 'The image content as an image file URL.', - }, - [supportedDocTypes.pdf]: { - options: standardOptions, - dataDescription: 'the pdf content as a PDF file url.', - }, - [supportedDocTypes.audio]: { - options: standardOptions, - dataDescription: 'The audio content as a file url.', - }, - [supportedDocTypes.video]: { - options: standardOptions, - dataDescription: 'The video content as a file url.', - }, - [supportedDocTypes.message]: { - options: standardOptions, - dataDescription: 'The message content of the document.', - }, - [supportedDocTypes.diagram]: { - options: ['title', 'backgroundColor'], - dataDescription: 'diagram content as a text string in Mermaid format.', - }, - [supportedDocTypes.script]: { - options: ['title', 'backgroundColor'], - dataDescription: 'The compilable JavaScript code. Use this for creating scripts.', - }, -}; - -const createAnyDocumentToolParams = [ - { - name: 'document_type', - type: 'string', - description: `The type of the document to create. Supported types are: ${Object.values(supportedDocTypes).join(', ')}`, - required: true, - }, - { - name: 'data', - type: 'string', - description: 'The content or data of the document. The exact format depends on the document type.', - required: true, - }, - { - name: 'options', - type: 'string', - required: false, - description: `A JSON string representing the document options. Available options depend on the document type. For example: - ${Object.entries(documentTypesInfo).map( ([doc_type, info]) => ` -- For '${doc_type}' documents, options include: ${info.options.join(', ')}`) - .join('\n')}`, // prettier-ignore - }, -] as const; - -type CreateAnyDocumentToolParamsType = typeof createAnyDocumentToolParams; - -const createAnyDocToolInfo: ToolInfo<CreateAnyDocumentToolParamsType> = { - name: 'createAnyDocument', - description: - `Creates any type of document with the provided options and data. - Supported document types are: ${Object.values(supportedDocTypes).join(', ')}. - dataviz is a csv table tool, so for CSVs, use dataviz. Here are the options for each type: - <supported_document_types>` + - Object.entries(documentTypesInfo) - .map( - ([doc_type, info]) => - `<document_type name="${doc_type}"> - <data_description>${info.dataDescription}</data_description> - <options>` + - info.options.map(option => `<option>${option}</option>`).join('\n') + - `</options> - </document_type>` - ) - .join('\n') + - `</supported_document_types>`, - parameterRules: createAnyDocumentToolParams, - citationRules: 'No citation needed.', -}; - -export class CreateAnyDocumentTool extends BaseTool<CreateAnyDocumentToolParamsType> { - private _addLinkedDoc: (doc: parsedDoc) => Doc | undefined; - - constructor(addLinkedDoc: (doc: parsedDoc) => Doc | undefined) { - super(createAnyDocToolInfo); - this._addLinkedDoc = addLinkedDoc; - } - - async execute(args: ParametersType<CreateAnyDocumentToolParamsType>): Promise<Observation[]> { - try { - const documentType = toLower(args.document_type) as unknown as supportedDocTypes; - const info = documentTypesInfo[documentType]; - - if (info === undefined) { - throw new Error(`Unsupported document type: ${documentType}. Supported types are: ${Object.values(supportedDocTypes).join(', ')}.`); - } - - if (!args.data) { - throw new Error(`Data is required for ${documentType} documents. ${info.dataDescription}`); - } - - const options: DocumentOptions = !args.options ? {} : JSON.parse(args.options); - - // Call the function to add the linked document (add default title that can be overriden if set in options) - const doc = this._addLinkedDoc({ doc_type: documentType, data: args.data, title: `New ${documentType.charAt(0).toUpperCase() + documentType.slice(1)} Document`, ...options }); - - return [{ type: 'text', text: `Created ${documentType} document with ID ${doc?.[Id]}.` }]; - } catch (error) { - return [{ type: 'text', text: 'Error creating document: ' + (error as Error).message }]; - } - } -} diff --git a/src/client/views/nodes/chatbot/tools/DocumentMetadataTool.ts b/src/client/views/nodes/chatbot/tools/DocumentMetadataTool.ts index a9fb45b5a..a3d86287d 100644 --- a/src/client/views/nodes/chatbot/tools/DocumentMetadataTool.ts +++ b/src/client/views/nodes/chatbot/tools/DocumentMetadataTool.ts @@ -11,7 +11,6 @@ import { DocCast, StrCast } from '../../../../../fields/Types'; import { supportedDocTypes } from '../types/tool_types'; import { parsedDoc } from '../chatboxcomponents/ChatBox'; - // Define the parameters for the DocumentMetadataTool const parameterDefinitions: ReadonlyArray<Parameter> = [ { @@ -61,7 +60,7 @@ const parameterDefinitions: ReadonlyArray<Parameter> = [ type: 'string', required: false, description: `The type of document to create. Required for "create" action. Options: ${Object.keys(supportedDocTypes).join(',')}`, - } + }, ] as const; type DocumentMetadataToolParamsType = typeof parameterDefinitions; @@ -105,14 +104,14 @@ IMPORTANT: Some fields have dependencies that must be handled for edits to work - You can edit dependent fields in a single operation using the fieldEdits parameter Example: To change document height, first disable auto-height: -1. { action: "edit", documentId: "doc123", fieldName: "layout_autoHeight", fieldValue: false } -2. { action: "edit", documentId: "doc123", fieldName: "height", fieldValue: 300 } +1. {... inputs: { action: "edit", documentId: "doc123", fieldName: "layout_autoHeight", fieldValue: false }} +2. {... inputs: { action: "edit", documentId: "doc123", fieldName: "height", fieldValue: 300 }} OR using multi-field edit (recommended for dependent fields): -{ action: "edit", documentId: "doc123", fieldEdits: [ +{... inputs: { action: "edit", documentId: "doc123", fieldEdits: [ { fieldName: "layout_autoHeight", fieldValue: false }, { fieldName: "height", fieldValue: 300 } -]}`; +]}}`; // Extensive usage guidelines for the tool const citationRules = `USAGE GUIDELINES: @@ -132,7 +131,7 @@ To CREATE a new document: - title: The title of the document to create - data: The content data for the document (text content, URL, etc.) - doc_type: The type of document to create (text, web, image, etc.) -- Example: { action: "create", title: "My Notes", data: "This is the content", doc_type: "text" } +- Example: {...inputs: { action: "create", title: "My Notes", data: "This is the content", doc_type: "text" }} - After creation, you can edit the document with more specific properties To EDIT document metadata: @@ -147,7 +146,7 @@ To EDIT document metadata: SPECIAL FIELD HANDLING: - Text fields: When editing the 'text' field, provide simple plain text - Example: { action: "edit", documentId: "doc123", fieldName: "text", fieldValue: "Hello world" } + Example: {...inputs: { action: "edit", documentId: "doc123", fieldName: "text", fieldValue: "Hello world" }} The tool will automatically convert your text to the proper RichTextField format - Width/Height: Set layout_autoHeight/layout_autoWidth to false before editing @@ -165,10 +164,10 @@ HANDLING DEPENDENT FIELDS: - When editing some fields, you may need to update related dependent fields - For example, when changing "height", you should also set "layout_autoHeight" to false - Use the fieldEdits parameter to update dependent fields in a single operation (recommended): - { action: "edit", documentId: "doc123", fieldEdits: [ + {...inputs: { action: "edit", documentId: "doc123", fieldEdits: [ { fieldName: "layout_autoHeight", fieldValue: false }, { fieldName: "height", fieldValue: 300 } - ]} +]}} - Always check for dependent fields that might affect your edits, such as: - height → layout_autoHeight (set to false to allow manual height) - width → layout_autoWidth (set to false to allow manual width) @@ -200,8 +199,8 @@ Examples: { action: "edit", documentId: "doc123", fieldEdits: [ { fieldName: "layout_autoHeight", fieldValue: false }, { fieldName: "height", fieldValue: 200 } - ]}`; - + ]} +- IMPORTANT: MULTI STEP WORKFLOWS ARE NOT ONLY ALLOWED BUT ENCOURAGED. TAKE THINGS 1 STEP AT A TIME.`; const documentMetadataToolInfo: ToolInfo<DocumentMetadataToolParamsType> = { name: 'documentMetadata', description: toolDescription, @@ -227,7 +226,7 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp constructor(chatBox: any) { super(documentMetadataToolInfo); this.chatBox = chatBox; - + // Store a direct reference to the ChatBox document if (chatBox && chatBox.Document) { this.chatBoxDocument = chatBox.Document; @@ -246,7 +245,7 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp } else { console.warn('DocumentMetadataTool initialized without valid ChatBox Document reference'); } - + this.initializeFieldMetadata(); } @@ -273,12 +272,12 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp // Check if fieldInfo has description property (it's likely a FInfo instance) if (fieldInfo && typeof fieldInfo === 'object' && 'description' in fieldInfo) { fieldData.description = fieldInfo.description; - + // Extract field type if available if ('fieldType' in fieldInfo) { fieldData.type = fieldInfo.fieldType; } - + // Extract possible values if available if ('values' in fieldInfo && Array.isArray(fieldInfo.values)) { fieldData.possibleValues = fieldInfo.values; @@ -303,30 +302,30 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp // Use the LinkManager approach which is proven to work in ChatBox if (this.chatBoxDocument) { console.log('Finding documents linked to ChatBox document with ID:', this.chatBoxDocument.id); - + // Get directly linked documents via LinkManager const linkedDocs = LinkManager.Instance.getAllRelatedLinks(this.chatBoxDocument) .map(d => DocCast(LinkManager.getOppositeAnchor(d, this.chatBoxDocument!))) .map(d => DocCast(d?.annotationOn, d)) .filter(d => d); - + console.log(`Found ${linkedDocs.length} linked documents via LinkManager`); - + // Process the linked documents linkedDocs.forEach((doc: Doc) => { if (doc) { this.processDocument(doc); } }); - + // Include the ChatBox document itself this.processDocument(this.chatBoxDocument); - + // If we have access to the Document's parent, try to find sibling documents if (this.chatBoxDocument.parent) { const parent = this.chatBoxDocument.parent; console.log('Found parent document, checking for siblings'); - + // Check if parent is a Doc type and has a childDocs function if (parent && typeof parent === 'object' && 'childDocs' in parent && typeof parent.childDocs === 'function') { try { @@ -356,7 +355,7 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp } }); } - + // Process the ChatBox document if available if (this.chatBox.Document) { this.processDocument(this.chatBox.Document); @@ -366,11 +365,11 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp } console.log(`DocumentMetadataTool found ${this.documentsById.size} total documents`); - + // If we didn't find any documents, try a fallback method if (this.documentsById.size === 0 && this.chatBox) { console.log('No documents found, trying fallback method'); - + // Try to access any field that might contain documents if (this.chatBox.props && this.chatBox.props.documents) { const documents = this.chatBox.props.documents; @@ -396,17 +395,17 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp private processDocument(doc: Doc) { // Ensure document has a persistent ID const docId = this.ensureDocumentId(doc); - + // Only add if we haven't already processed this document if (!this.documentsById.has(docId)) { this.documentsById.set(docId, doc); - + // Get layout doc (the document itself or its layout) const layoutDoc = Doc.Layout(doc); if (layoutDoc) { this.layoutDocsById.set(docId, layoutDoc); } - + // Get data doc const dataDoc = doc[DocData]; if (dataDoc) { @@ -422,15 +421,15 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp */ private ensureDocumentId(doc: Doc): string { let docId: string | undefined; - + // First try to get the ID from our custom field if (doc[this.DOCUMENT_ID_FIELD]) { docId = String(doc[this.DOCUMENT_ID_FIELD]); return docId; } - + // Try different ways to get a document ID - + // 1. Try the direct id property if it exists if (doc.id && typeof doc.id === 'string') { docId = doc.id; @@ -448,14 +447,14 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp docId = uuidv4(); console.log(`Generated new UUID for document with title: ${doc.title || 'Untitled'}`); } - + // Store the ID in the document's metadata so it persists try { doc[this.DOCUMENT_ID_FIELD] = docId; } catch (e) { console.warn(`Could not assign ID to document property`, e); } - + return docId; } @@ -472,7 +471,7 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp const layoutDoc = this.layoutDocsById.get(docId); const dataDoc = this.dataDocsById.get(docId); - + const metadata: Record<string, any> = { id: docId, title: doc.title || '', @@ -488,7 +487,7 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp Object.keys(this.fieldMetadata).forEach(fieldName => { const fieldDef = this.fieldMetadata[fieldName]; const strippedName = fieldName.startsWith('_') ? fieldName.substring(1) : fieldName; - + // Check if field exists on layout document let layoutValue = undefined; if (layoutDoc) { @@ -499,7 +498,7 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp metadata.fieldLocationMap[strippedName] = 'layout'; } } - + // Check if field exists on data document let dataValue = undefined; if (dataDoc) { @@ -512,12 +511,12 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp } } } - + // For fields with stripped names (without leading underscore), // also check if they exist on documents without the underscore if (fieldName.startsWith('_')) { const nonUnderscoreFieldName = fieldName.substring(1); - + if (layoutDoc) { const nonUnderscoreLayoutValue = layoutDoc[nonUnderscoreFieldName]; if (nonUnderscoreLayoutValue !== undefined) { @@ -525,7 +524,7 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp metadata.fieldLocationMap[nonUnderscoreFieldName] = 'layout'; } } - + if (dataDoc) { const nonUnderscoreDataValue = dataDoc[nonUnderscoreFieldName]; if (nonUnderscoreDataValue !== undefined) { @@ -544,7 +543,7 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp metadata.fields.layout.width = metadata.fields.layout._width; metadata.fieldLocationMap.width = 'layout'; } - + if (metadata.fields.layout._height !== undefined && metadata.fields.layout.height === undefined) { metadata.fields.layout.height = metadata.fields.layout._height; metadata.fieldLocationMap.height = 'layout'; @@ -560,18 +559,22 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp * @param fieldValue The new value for the field (string, number, or boolean) * @returns Object with success status, message, and additional information */ - private editDocumentField(docId: string, fieldName: string, fieldValue: string | number | boolean): { - success: boolean; - message: string; - fieldName?: string; - originalFieldName?: string; + private editDocumentField( + docId: string, + fieldName: string, + fieldValue: string | number | boolean + ): { + success: boolean; + message: string; + fieldName?: string; + originalFieldName?: string; newValue?: any; warning?: string; } { // Normalize field name (handle with/without underscore) let normalizedFieldName = fieldName.startsWith('_') ? fieldName : fieldName; const strippedFieldName = fieldName.startsWith('_') ? fieldName.substring(1) : fieldName; - + // Handle common field name aliases (width → _width, height → _height) // Many document fields use '_' prefix for layout properties if (fieldName === 'width') { @@ -579,36 +582,36 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp } else if (fieldName === 'height') { normalizedFieldName = '_height'; } - + // Get the documents const doc = this.documentsById.get(docId); if (!doc) { return { success: false, message: `Document with ID ${docId} not found` }; } - + const layoutDoc = this.layoutDocsById.get(docId); const dataDoc = this.dataDocsById.get(docId); - + if (!layoutDoc && !dataDoc) { return { success: false, message: `Could not find layout or data document for document with ID ${docId}` }; } - + try { // Convert the field value to the appropriate type based on field metadata const convertedValue = this.convertFieldValue(normalizedFieldName, fieldValue); - + let targetDoc: Doc | undefined; let targetLocation: string; - + // First, check if field exists on layout document using Doc.Get if (layoutDoc) { const fieldExistsOnLayout = Doc.Get(layoutDoc, normalizedFieldName, true) !== undefined; - + // If it exists on layout document, update it there if (fieldExistsOnLayout) { targetDoc = layoutDoc; targetLocation = 'layout'; - } + } // If it has an underscore prefix, it's likely a layout property even if not yet set else if (normalizedFieldName.startsWith('_')) { targetDoc = layoutDoc; @@ -618,13 +621,13 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp else if (dataDoc) { targetDoc = dataDoc; targetLocation = 'data'; - } + } // If no data document available, default to layout else { targetDoc = layoutDoc; targetLocation = 'layout'; } - } + } // If no layout document, use data document else if (dataDoc) { targetDoc = dataDoc; @@ -632,26 +635,26 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp } else { return { success: false, message: `No valid document found for editing` }; } - + if (!targetDoc) { return { success: false, message: `Target document not available` }; } - + // Set the field value on the target document targetDoc[normalizedFieldName] = convertedValue; - - return { - success: true, + + return { + success: true, message: `Successfully updated field '${normalizedFieldName}' on ${targetLocation} document (ID: ${docId})`, fieldName: normalizedFieldName, originalFieldName: fieldName, - newValue: convertedValue + newValue: convertedValue, }; } catch (error) { console.error('Error editing document field:', error); - return { - success: false, - message: `Error updating field: ${error instanceof Error ? error.message : String(error)}` + return { + success: false, + message: `Error updating field: ${error instanceof Error ? error.message : String(error)}`, }; } } @@ -667,7 +670,7 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp if (typeof fieldValue === 'number' || typeof fieldValue === 'boolean') { return fieldValue; } - + // If fieldValue is a string "true" or "false", convert to boolean if (typeof fieldValue === 'string') { if (fieldValue.toLowerCase() === 'true') { @@ -677,12 +680,12 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp return false; } } - + // If fieldValue is not a string (and not a number or boolean), convert it to string if (typeof fieldValue !== 'string') { fieldValue = String(fieldValue); } - + // Special handling for text field - convert to proper RichTextField format if (fieldName === 'text') { try { @@ -710,28 +713,28 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp return JSON.stringify(rtf); } } - + // Get field metadata const normalizedFieldName = fieldName.startsWith('_') ? fieldName : `_${fieldName}`; const strippedFieldName = fieldName.startsWith('_') ? fieldName.substring(1) : fieldName; - + // Check both versions of the field name in metadata const fieldMeta = this.fieldMetadata[normalizedFieldName] || this.fieldMetadata[strippedFieldName]; - + // Special handling for width and height without metadata if (!fieldMeta && (fieldName === '_width' || fieldName === '_height' || fieldName === 'width' || fieldName === 'height')) { const num = Number(fieldValue); return isNaN(num) ? fieldValue : num; } - + if (!fieldMeta) { // If no metadata found, just return the string value return fieldValue; } - + // Convert based on field type const fieldType = fieldMeta.type; - + if (fieldType === 'boolean') { // Convert to boolean return fieldValue.toLowerCase() === 'true'; @@ -761,7 +764,7 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp return fieldValue; } } - + // Default to string return fieldValue; } @@ -775,7 +778,7 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp if (value === undefined || value === null) { return null; } - + // Handle Doc objects if (value instanceof Doc) { return { @@ -785,7 +788,7 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp docType: value.type || '', }; } - + // Handle RichTextField (try to extract plain text) if (typeof value === 'string' && value.includes('"type":"doc"') && value.includes('"content":')) { try { @@ -802,9 +805,9 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp node.content.forEach((child: any) => extractText(child)); } }; - + extractText(rtfObj.doc); - + // If we successfully extracted text, show it, but also preserve the original value if (plainText) { return { @@ -819,14 +822,14 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp // If parsing fails, just treat as a regular string } } - + // Handle arrays and complex objects if (typeof value === 'object') { // If the object has a toString method, use it if (value.toString && value.toString !== Object.prototype.toString) { return value.toString(); } - + try { // Try to convert to JSON string return JSON.stringify(value); @@ -834,7 +837,7 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp return '[Complex Object]'; } } - + // Return primitive values as is return value; } @@ -856,7 +859,7 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp list: [], date: [], enumeration: [], - other: [] + other: [], }, fieldNameMappings: {}, commonFields: { @@ -865,19 +868,19 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp size: [], content: [], behavior: [], - layout: [] - } + layout: [], + }, }; // Process each field in the metadata Object.entries(this.fieldMetadata).forEach(([fieldName, fieldInfo]) => { const strippedName = fieldName.startsWith('_') ? fieldName.substring(1) : fieldName; - + // Add to fieldNameMappings if (fieldName.startsWith('_')) { result.fieldNameMappings[strippedName] = fieldName; } - + // Create structured field metadata const fieldData: Record<string, any> = { name: fieldName, @@ -886,10 +889,10 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp type: fieldInfo.fieldType || 'unknown', possibleValues: fieldInfo.values || [], }; - + // Add field to fields collection result.fields[fieldName] = fieldData; - + // Categorize by field type const type = fieldInfo.fieldType?.toLowerCase() || 'unknown'; if (type === 'string') { @@ -909,7 +912,7 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp } else { result.fieldsByType.other.push(fieldName); } - + // Categorize by field purpose if (fieldName.includes('width') || fieldName.includes('height') || fieldName.includes('size')) { result.commonFields.size.push(fieldName); @@ -925,23 +928,23 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp result.commonFields.layout.push(fieldName); } }); - + // Add special section for auto-sizing related fields result.autoSizingFields = { height: { autoHeightField: '_layout_autoHeight', heightField: '_height', displayName: 'height', - usage: 'To manually set height, first set layout_autoHeight to false' + usage: 'To manually set height, first set layout_autoHeight to false', }, width: { autoWidthField: '_layout_autoWidth', widthField: '_width', displayName: 'width', - usage: 'To manually set width, first set layout_autoWidth to false' - } + usage: 'To manually set width, first set layout_autoWidth to false', + }, }; - + // Add special section for text field format result.specialFields = { text: { @@ -949,10 +952,10 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp description: 'Document text content', format: 'RichTextField', note: 'When setting text, provide plain text - it will be automatically converted to the correct format', - example: 'For setting: "Hello world" (plain text); For getting: Will be converted to plaintext for display' - } + example: 'For setting: "Hello world" (plain text); For getting: Will be converted to plaintext for display', + }, }; - + return result; } @@ -963,121 +966,129 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp */ async execute(args: ParametersType<DocumentMetadataToolParamsType>): Promise<Observation[]> { console.log('DocumentMetadataTool: Executing with args:', args); - + // Find all documents in the Freeform view this.findDocumentsInFreeformView(); - + try { // Validate required input parameters based on action if (!this.inputValidator(args)) { - return [{ - type: 'text', - text: `Error: Invalid or missing parameters for action "${args.action}". ${this.getParameterRequirementsByAction(String(args.action))}` - }]; + return [ + { + type: 'text', + text: `Error: Invalid or missing parameters for action "${args.action}". ${this.getParameterRequirementsByAction(String(args.action))}`, + }, + ]; } - + // Ensure the action is valid and convert to string const action = String(args.action); if (!['get', 'edit', 'list', 'getFieldOptions', 'create'].includes(action)) { - return [{ - type: 'text', - text: 'Error: Invalid action. Valid actions are "get", "edit", "list", "getFieldOptions", or "create".' - }]; + return [ + { + type: 'text', + text: 'Error: Invalid action. Valid actions are "get", "edit", "list", "getFieldOptions", or "create".', + }, + ]; } // Safely convert documentId to string or undefined const documentId = args.documentId ? String(args.documentId) : undefined; - + // Perform the specified action switch (action) { case 'get': { // Get metadata for a specific document or all documents const result = this.getDocumentMetadata(documentId); console.log('DocumentMetadataTool: Get metadata result:', result); - return [{ - type: 'text', - text: `Document metadata ${documentId ? 'for document ' + documentId : ''} retrieved successfully:\n${JSON.stringify(result, null, 2)}` - }]; + return [ + { + type: 'text', + text: `Document metadata ${documentId ? 'for document ' + documentId : ''} retrieved successfully:\n${JSON.stringify(result, null, 2)}`, + }, + ]; } - + case 'edit': { // Edit a specific field on a document if (!documentId) { - return [{ - type: 'text', - text: 'Error: Document ID is required for edit actions.' - }]; + return [ + { + type: 'text', + text: 'Error: Document ID is required for edit actions.', + }, + ]; } - + // Ensure document exists if (!this.documentsById.has(documentId)) { - return [{ - type: 'text', - text: `Error: Document with ID ${documentId} not found.` - }]; + return [ + { + type: 'text', + text: `Error: Document with ID ${documentId} not found.`, + }, + ]; } - + // Check if we're doing a multi-field edit or a single field edit if (args.fieldEdits) { try { // Parse fieldEdits array const edits = JSON.parse(String(args.fieldEdits)); if (!Array.isArray(edits) || edits.length === 0) { - return [{ - type: 'text', - text: 'Error: fieldEdits must be a non-empty array of field edits.' - }]; + return [ + { + type: 'text', + text: 'Error: fieldEdits must be a non-empty array of field edits.', + }, + ]; } - + // Track results for all edits - const results: { - success: boolean; - message: string; - fieldName?: string; - originalFieldName?: string; + const results: { + success: boolean; + message: string; + fieldName?: string; + originalFieldName?: string; newValue?: any; warning?: string; }[] = []; - + let allSuccessful = true; - + // Process each edit for (const edit of edits) { // Get fieldValue in its original form let fieldValue = edit.fieldValue; - + // Only convert to string if it's neither boolean nor number if (typeof fieldValue !== 'boolean' && typeof fieldValue !== 'number') { fieldValue = String(fieldValue); } - + const fieldName = String(edit.fieldName); - + // Edit the field - const result = this.editDocumentField( - documentId, - fieldName, - fieldValue - ); - + const result = this.editDocumentField(documentId, fieldName, fieldValue); + console.log(`DocumentMetadataTool: Edit field result for ${fieldName}:`, result); - + // Add to results results.push(result); - + // Update success status if (!result.success) { allSuccessful = false; } } - + // Format response based on results let responseText = ''; if (allSuccessful) { responseText = `Successfully edited ${results.length} fields on document ${documentId}:\n`; results.forEach(result => { responseText += `- Field '${result.originalFieldName}': updated to ${JSON.stringify(result.newValue)}\n`; - + // Add any warnings if (result.warning) { responseText += ` Warning: ${result.warning}\n`; @@ -1088,7 +1099,7 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp results.forEach(result => { if (result.success) { responseText += `- Field '${result.originalFieldName}': updated to ${JSON.stringify(result.newValue)}\n`; - + // Add any warnings if (result.warning) { responseText += ` Warning: ${result.warning}\n`; @@ -1098,120 +1109,134 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp } }); } - + // Get the updated metadata to return const updatedMetadata = this.getDocumentMetadata(documentId); - - return [{ - type: 'text', - text: `${responseText}\nUpdated metadata:\n${JSON.stringify(updatedMetadata, null, 2)}` - }]; + + return [ + { + type: 'text', + text: `${responseText}\nUpdated metadata:\n${JSON.stringify(updatedMetadata, null, 2)}`, + }, + ]; } catch (error) { - return [{ - type: 'text', - text: `Error processing fieldEdits: ${error instanceof Error ? error.message : String(error)}` - }]; + return [ + { + type: 'text', + text: `Error processing fieldEdits: ${error instanceof Error ? error.message : String(error)}`, + }, + ]; } } else { // Single field edit (original behavior) if (!args.fieldName) { - return [{ - type: 'text', - text: 'Error: Field name and field value are required for edit actions.' - }]; + return [ + { + type: 'text', + text: 'Error: Field name and field value are required for edit actions.', + }, + ]; } - + // Get fieldValue in its original form - we'll handle conversion in editDocumentField let fieldValue = args.fieldValue; - + // Only convert to string if it's neither boolean nor number if (typeof fieldValue !== 'boolean' && typeof fieldValue !== 'number') { fieldValue = String(fieldValue); } - + const fieldName = String(args.fieldName); - + // Edit the field - const result = this.editDocumentField( - documentId, - fieldName, - fieldValue - ); - + const result = this.editDocumentField(documentId, fieldName, fieldValue); + console.log('DocumentMetadataTool: Edit field result:', result); - + if (!result.success) { return [{ type: 'text', text: result.message }]; } - + // Include warning if present let responseText = result.message; if (result.warning) { responseText += `\n\n${result.warning}`; } - + // Get the updated metadata to return const updatedMetadata = this.getDocumentMetadata(documentId); - - return [{ - type: 'text', - text: `${responseText}\nUpdated metadata:\n${JSON.stringify(updatedMetadata, null, 2)}` - }]; + + return [ + { + type: 'text', + text: `${responseText}\nUpdated metadata:\n${JSON.stringify(updatedMetadata, null, 2)}`, + }, + ]; } } - + case 'list': { // List all available documents in simple format const docs = Array.from(this.documentsById.entries()).map(([id, doc]) => ({ id, title: doc.title || 'Untitled Document', - type: doc.type || 'Unknown Type' + type: doc.type || 'Unknown Type', })); - + if (docs.length === 0) { - return [{ - type: 'text', - text: 'No documents found in the current view.' - }]; + return [ + { + type: 'text', + text: 'No documents found in the current view.', + }, + ]; } - - return [{ - type: 'text', - text: `Found ${docs.length} document(s) in the current view:\n${JSON.stringify(docs, null, 2)}` - }]; + + return [ + { + type: 'text', + text: `Found ${docs.length} document(s) in the current view:\n${JSON.stringify(docs, null, 2)}`, + }, + ]; } - + case 'getFieldOptions': { // Get all available field options with metadata const fieldOptions = this.getAllFieldMetadata(); - - return [{ - type: 'text', - text: `Document field options retrieved successfully.\nThis information should be consulted before editing document fields to understand available options and dependencies:\n${JSON.stringify(fieldOptions, null, 2)}` - }]; + + return [ + { + type: 'text', + text: `Document field options retrieved successfully.\nThis information should be consulted before editing document fields to understand available options and dependencies:\n${JSON.stringify(fieldOptions, null, 2)}`, + }, + ]; } - + case 'create': { // Create a new document if (!args.title || !args.data || !args.doc_type) { - return [{ - type: 'text', - text: 'Error: Title, data, and doc_type are required for create action.' - }]; + return [ + { + type: 'text', + text: 'Error: Title, data, and doc_type are required for create action.', + }, + ]; } - + const docType = String(args.doc_type); const title = String(args.title); const data = String(args.data); - + // Validate doc_type if (!this.isValidDocType(docType)) { - return [{ - type: 'text', - text: `Error: Invalid doc_type. Valid options are: ${Object.keys(supportedDocTypes).join(',')}` - }]; + return [ + { + type: 'text', + text: `Error: Invalid doc_type. Valid options are: ${Object.keys(supportedDocTypes).join(',')}`, + }, + ]; } - + try { // Create simple document with just title and data const simpleDoc: parsedDoc = { @@ -1223,35 +1248,40 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp _width: 300, _height: 300, _layout_fitWidth: false, - _layout_autoHeight: true + _layout_autoHeight: true, }; - + // Use the chatBox's createDocInDash method to create and link the document if (!this.chatBox || !this.chatBox.createDocInDash) { - return [{ - type: 'text', - text: 'Error: Could not access document creation functionality.' - }]; + return [ + { + type: 'text', + text: 'Error: Could not access document creation functionality.', + }, + ]; } - + const createdDoc = this.chatBox.createDocInDash(simpleDoc); - + if (!createdDoc) { - return [{ - type: 'text', - text: 'Error: Failed to create document.' - }]; + return [ + { + type: 'text', + text: 'Error: Failed to create document.', + }, + ]; } - + // Update our local document maps with the new document this.processDocument(createdDoc); - + // Get the created document's metadata const createdMetadata = this.getDocumentMetadata(createdDoc.id); - - return [{ - type: 'text', - text: `Document created successfully. + + return [ + { + type: 'text', + text: `Document created successfully. Document ID: ${createdDoc.id} Type: ${docType} Title: "${title}" @@ -1266,36 +1296,43 @@ Next steps: 4. For text documents, you can edit the content with: { action: "edit", documentId: "${createdDoc.id}", fieldName: "text", fieldValue: "New content" } Full metadata for the created document: -${JSON.stringify(createdMetadata, null, 2)}` - }]; +${JSON.stringify(createdMetadata, null, 2)}`, + }, + ]; } catch (error) { - return [{ - type: 'text', - text: `Error creating document: ${error instanceof Error ? error.message : String(error)}` - }]; + return [ + { + type: 'text', + text: `Error creating document: ${error instanceof Error ? error.message : String(error)}`, + }, + ]; } } - + default: - return [{ - type: 'text', - text: 'Error: Unknown action. Valid actions are "get", "edit", "list", "getFieldOptions", or "create".' - }]; + return [ + { + type: 'text', + text: 'Error: Unknown action. Valid actions are "get", "edit", "list", "getFieldOptions", or "create".', + }, + ]; } } catch (error) { console.error('DocumentMetadataTool execution error:', error); - return [{ - type: 'text', - text: `Error executing DocumentMetadataTool: ${error instanceof Error ? error.message : String(error)}` - }]; + return [ + { + type: 'text', + text: `Error executing DocumentMetadataTool: ${error instanceof Error ? error.message : String(error)}`, + }, + ]; } } /** * Validates the input parameters for the DocumentMetadataTool - * This custom validator allows numbers and booleans to be passed for fieldValue + * This custom validator allows numbers and booleans to be passed for fieldValue * while maintaining compatibility with the standard validation - * + * * @param params The parameters to validate * @returns True if the parameters are valid, false otherwise */ @@ -1304,12 +1341,12 @@ ${JSON.stringify(createdMetadata, null, 2)}` if (params.action === undefined) { return false; } - + // For create action, validate required parameters if (params.action === 'create') { return !!(params.title && params.data && params.doc_type); } - + // For edit action, validate either single field edit or multiple field edits if (params.action === 'edit') { // If fieldEdits is provided, it must be valid and we'll ignore fieldName/fieldValue @@ -1317,13 +1354,13 @@ ${JSON.stringify(createdMetadata, null, 2)}` try { // Parse fieldEdits and validate its structure const edits = JSON.parse(String(params.fieldEdits)); - + // Ensure it's an array if (!Array.isArray(edits)) { console.log('fieldEdits is not an array'); return false; } - + // Ensure each item has fieldName and fieldValue for (const edit of edits) { if (!edit.fieldName) { @@ -1335,7 +1372,7 @@ ${JSON.stringify(createdMetadata, null, 2)}` return false; } } - + // Everything looks good with fieldEdits return !!params.documentId; // Just ensure documentId is provided } catch (error) { @@ -1349,22 +1386,22 @@ ${JSON.stringify(createdMetadata, null, 2)}` } } } - + // For get action with documentId, documentId is required if (params.action === 'get' && params.documentId === '') { return false; } - + // getFieldOptions action doesn't require any additional parameters if (params.action === 'getFieldOptions') { return true; } - + // list action doesn't require any additional parameters if (params.action === 'list') { return true; } - + // Allow for numeric or boolean fieldValue even though the type is defined as string if (params.fieldValue !== undefined) { if (typeof params.fieldValue === 'number') { @@ -1372,14 +1409,14 @@ ${JSON.stringify(createdMetadata, null, 2)}` // We'll convert it later, so don't fail validation return true; } - + if (typeof params.fieldValue === 'boolean') { console.log('Boolean fieldValue detected, will be converted appropriately'); // We'll handle boolean conversion in the execute method return true; } } - + return true; } @@ -1424,11 +1461,11 @@ ${JSON.stringify(createdMetadata, null, 2)}` for (const docId of this.documentsById.keys()) { documentsMetadata[docId] = this.extractDocumentMetadata(docId); } - + return { documentCount: this.documentsById.size, documents: documentsMetadata, - fieldDefinitions: this.fieldMetadata + fieldDefinitions: this.fieldMetadata, }; } } @@ -1441,4 +1478,4 @@ ${JSON.stringify(createdMetadata, null, 2)}` private isValidDocType(docType: string): boolean { return Object.values(supportedDocTypes).includes(docType as supportedDocTypes); } -} \ No newline at end of file +} diff --git a/src/server/chunker/pdf_chunker.py b/src/server/chunker/pdf_chunker.py index 9d4cfb1b2..feb437f1f 100644 --- a/src/server/chunker/pdf_chunker.py +++ b/src/server/chunker/pdf_chunker.py @@ -602,12 +602,11 @@ class PDFChunker: try: response = self.client.messages.create( - model='claude-3-5-sonnet-20240620', + model='claude-3-7-sonnet-20250219', system=prompt, max_tokens=400 * len(images), # Increased token limit for more detailed summaries messages=messages, temperature=0, - extra_headers={"anthropic-beta": "max-tokens-3-5-sonnet-2024-07-15"} ) # Parse the response @@ -767,7 +766,7 @@ class Document: client = OpenAI() # Initialize OpenAI client for text generation completion = client.chat.completions.create( - model="gpt-3.5-turbo", # Specify the language model + model="gpt-4o", # Specify the language model messages=[ {"role": "system", "content": "You are an AI assistant tasked with summarizing a document. You are provided with important chunks from the document and provide a summary, as best you can, of what the document will contain overall. Be concise and brief with your response."}, -- cgit v1.2.3-70-g09d2 From 8bc8bd13293e64a99b68408ec3d24a50a5dfe4bc Mon Sep 17 00:00:00 2001 From: "A.J. Shulman" <Shulman.aj@gmail.com> Date: Mon, 21 Apr 2025 14:23:38 -0400 Subject: feat & fix: citations now scroll to correct page and using OpenAI 4o with structured outputs --- .../nodes/chatbot/chatboxcomponents/ChatBox.tsx | 28 ++-- src/server/chunker/pdf_chunker.py | 169 ++++++++------------- 2 files changed, 77 insertions(+), 120 deletions(-) (limited to 'src/server/chunker') diff --git a/src/client/views/nodes/chatbot/chatboxcomponents/ChatBox.tsx b/src/client/views/nodes/chatbot/chatboxcomponents/ChatBox.tsx index 3152fe923..33419e559 100644 --- a/src/client/views/nodes/chatbot/chatboxcomponents/ChatBox.tsx +++ b/src/client/views/nodes/chatbot/chatboxcomponents/ChatBox.tsx @@ -107,21 +107,12 @@ export class ChatBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { this.vectorstore_id = StrCast(this.dataDoc.vectorstore_id); } this.vectorstore = new Vectorstore(this.vectorstore_id, this.retrieveDocIds); - this.agent = new Agent( - this.vectorstore, - this.retrieveSummaries, - this.retrieveFormattedHistory, - this.retrieveCSVData, - this.addLinkedUrlDoc, - this.createImageInDash, - this.createDocInDash, - this.createCSVInDash - ); - + this.agent = new Agent(this.vectorstore, this.retrieveSummaries, this.retrieveFormattedHistory, this.retrieveCSVData, this.addLinkedUrlDoc, this.createImageInDash, this.createDocInDash, this.createCSVInDash); + // Reinitialize the DocumentMetadataTool with a direct reference to this ChatBox instance // This ensures the tool can properly access documents in the same Freeform view this.agent.reinitializeDocumentMetadataTool(this); - + this.messagesRef = React.createRef<HTMLDivElement>(); // Reaction to update dataDoc when chat history changes @@ -743,6 +734,7 @@ export class ChatBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { DocumentManager.Instance.showDocument(doc, { willZoomCentered: true }, () => {}); return; } + const x1 = parseFloat(values[0]) * Doc.NativeWidth(doc); const y1 = parseFloat(values[1]) * Doc.NativeHeight(doc) + foundChunk.startPage * Doc.NativeHeight(doc); const x2 = parseFloat(values[2]) * Doc.NativeWidth(doc); @@ -751,9 +743,18 @@ export class ChatBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { const annotationKey = Doc.LayoutFieldKey(doc) + '_annotations'; const existingDoc = DocListCast(doc[DocData][annotationKey]).find(d => d.citation_id === citation.citation_id); + if (existingDoc) { + existingDoc.x = x1; + existingDoc.y = y1; + existingDoc._width = x2 - x1; + existingDoc._height = y2 - y1; + } const highlightDoc = existingDoc ?? this.createImageCitationHighlight(x1, y1, x2, y2, citation, annotationKey, doc); - DocumentManager.Instance.showDocument(highlightDoc, { willZoomCentered: true }, () => {}); + //doc.layout_scroll = y1; + doc._layout_curPage = foundChunk.startPage + 1; + DocumentManager.Instance.showDocument(doc, { willZoomCentered: true }, () => {}); + //DocumentManager.Instance.showDocument(highlightDoc, { willZoomCentered: true }, () => {}); } break; case CHUNK_TYPE.TEXT: @@ -795,6 +796,7 @@ export class ChatBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { backgroundColor: 'rgba(255, 255, 0, 0.5)', }); highlight_doc[DocData].citation_id = citation.citation_id; + highlight_doc.freeform_scale = 1; Doc.AddDocToList(pdfDoc[DocData], annotationKey, highlight_doc); highlight_doc.annotationOn = pdfDoc; Doc.SetContainer(highlight_doc, pdfDoc); diff --git a/src/server/chunker/pdf_chunker.py b/src/server/chunker/pdf_chunker.py index feb437f1f..e9b9ef2b3 100644 --- a/src/server/chunker/pdf_chunker.py +++ b/src/server/chunker/pdf_chunker.py @@ -276,12 +276,13 @@ class PDFChunker: :param output_folder: Folder to store the output files (extracted tables/images). :param image_batch_size: The batch size for processing visual elements. """ - self.client = Anthropic(api_key=os.getenv("ANTHROPIC_API_KEY")) # Initialize the Anthropic API client + self.client = OpenAI() # ← replaces Anthropic() self.output_folder = output_folder self.image_batch_size = image_batch_size # Batch size for image processing self.doc_id = doc_id # Add doc_id self.element_extractor = ElementExtractor(output_folder, doc_id) + async def chunk_pdf(self, file_data: bytes, file_name: str, doc_id: str, job_id: str) -> List[Dict[str, Any]]: """ Processes a PDF file, extracting text and visual elements, and returning structured chunks. @@ -518,123 +519,77 @@ class PDFChunker: def batch_summarize_images(self, images: Dict[int, str]) -> Dict[int, str]: """ - Summarize images or tables by generating descriptive text. - - :param images: A dictionary mapping image numbers to base64-encoded image data. - :return: A dictionary mapping image numbers to their generated summaries. - """ - # Prompt for the AI model to summarize images and tables - prompt = f"""<instruction> - <task> - You are tasked with summarizing a series of {len(images)} images and tables for use in a RAG (Retrieval-Augmented Generation) system. - Your goal is to create concise, informative summaries that capture the essential content of each image or table. - These summaries will be used for embedding, so they should be descriptive and relevant. The image or table will be outlined in red on an image of the full page that it is on. Where necessary, use the context of the full page to heklp with the summary but don't summarize other content on the page. - </task> - - <steps> - <step>Identify whether it's an image or a table.</step> - <step>Examine its content carefully.</step> - <step> - Write a detailed summary that captures the main points or visual elements: - <details> - <table>After summarizing what the table is about, include the column headers, a detailed summary of the data, and any notable data trends.</table> - <image>Describe the main subjects, actions, or notable features.</image> - </details> - </step> - <step>Focus on writing summaries that would make it easy to retrieve the content if compared to a user query using vector similarity search.</step> - <step>Keep summaries concise and include important words that may help with retrieval (but do not include numbers and numerical data).</step> - </steps> - - <important_notes> - <note>Avoid using special characters like &, <, >, ", ', $, %, etc. Instead, use their word equivalents:</note> - <note>Use "and" instead of &.</note> - <note>Use "dollars" instead of $.</note> - <note>Use "percent" instead of %.</note> - <note>Refrain from using quotation marks " or apostrophes ' unless absolutely necessary.</note> - <note>Ensure your output is in valid XML format.</note> - </important_notes> - - <formatting> - <note>Enclose all summaries within a root element called <summaries>.</note> - <note>Use <summary> tags to enclose each individual summary.</note> - <note>Include an attribute 'number' in each <summary> tag to indicate the sequence, matching the provided image numbers.</note> - <note>Start each summary by indicating whether it's an image or a table (e.g., "This image shows..." or "The table presents...").</note> - <note>If an image is completely blank, leave the summary blank (e.g., <summary number="3"></summary>).</note> - </formatting> - - <example> - <note>Do not replicate the example below—stay grounded to the content of the table or image and describe it completely and accurately.</note> - <output> - <summaries> - <summary number="1"> - The image shows two men shaking hands on stage at a formal event. The man on the left, in a dark suit and glasses, has a professional appearance, possibly an academic or business figure. The man on the right, Tim Cook, CEO of Apple, is recognizable by his silver hair and dark blue blazer. Cook holds a document titled "Tsinghua SEM EMBA," suggesting a link to Tsinghua University’s Executive MBA program. The backdrop displays English and Chinese text about business management and education, with the event dated October 23, 2014. - </summary> - <summary number="2"> - The table compares the company's assets between December 30, 2023, and September 30, 2023. Key changes include an increase in cash and cash equivalents, while marketable securities had a slight rise. Accounts receivable and vendor non-trade receivables decreased. Inventories and other current assets saw minor fluctuations. Non-current assets like marketable securities slightly declined, while property, plant, and equipment remained stable. Total assets showed minimal change, holding steady at around three hundred fifty-three billion dollars. - </summary> - <summary number="3"> - The table outlines the company's shareholders' equity as of December 30, 2023, versus September 30, 2023. Common stock and additional paid-in capital increased, and retained earnings shifted from a deficit to a positive figure. Accumulated other comprehensive loss decreased. Overall, total shareholders' equity rose significantly, while total liabilities and equity remained nearly unchanged at about three hundred fifty-three billion dollars. - </summary> - <summary number="4"> - The table details the company's liabilities as of December 30, 2023, compared to September 30, 2023. Current liabilities decreased due to lower accounts payable and other current liabilities, while deferred revenue slightly increased. Commercial paper significantly decreased, and term debt rose modestly. Non-current liabilities were stable, with minimal changes in term debt and other non-current liabilities. Total liabilities dropped from two hundred ninety billion dollars to two hundred seventy-nine billion dollars. - </summary> - <summary number="5"> - </summary> - </summaries> - </output> - </example> - - <final_notes> - <note>Process each image or table in the order provided.</note> - <note>Maintain consistent formatting throughout your response.</note> - <note>Ensure the output is in full, valid XML format with the root <summaries> element and each summary being within a <summary> element with the summary number specified as well.</note> - </final_notes> -</instruction> - """ - content = [] - for number, img in images.items(): - content.append({"type": "text", "text": f"\nImage {number}:\n"}) - content.append({"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": img}}) + Summarise a batch of images/tables with GPT‑4o using Structured Outputs. + :param images: {image_number: base64_png} + :return: {image_number: summary_text} + """ + # -------- 1. Build the prompt ----------- + content: list[dict] = [] + for n, b64 in images.items(): + content.append({"type": "text", + "text": f"\nImage {n} (outlined in red on the page):"}) + content.append({"type": "image_url", + "image_url": {"url": f"data:image/png;base64,{b64}"}}) messages = [ - {"role": "user", "content": content} + { + "role": "system", + "content": ( + "You are generating retrieval‑ready summaries for each highlighted " + "image or table. Start by identifying whether the element is an " + "image or a table, then write one informative sentence that a vector " + "search would find useful. Provide detail but limit to a couple of paragraphs per image." + ), + }, + {"role": "user", "content": content}, ] + schema = { + "type": "object", + "properties": { + "summaries": { + "type": "array", + "items": { + "type": "object", + "properties": { + "number": {"type": "integer"}, + "type": {"type": "string", "enum": ["image", "table"]}, + "summary": {"type": "string"} + }, + "required": ["number", "type", "summary"], + "additionalProperties": False + } + } + }, + "required": ["summaries"], + "additionalProperties": False + } + + # ---------- OpenAI call ----------------------------------------------------- try: - response = self.client.messages.create( - model='claude-3-7-sonnet-20250219', - system=prompt, - max_tokens=400 * len(images), # Increased token limit for more detailed summaries + resp = self.client.chat.completions.create( + model="gpt-4o", messages=messages, + max_tokens=400 * len(images), temperature=0, + response_format={ + "type": "json_schema", + "json_schema": { + "name": "image_batch_summaries", # ← REQUIRED + "schema": schema, # ← REQUIRED + "strict": True # ← strongly recommended + }, + }, ) - # Parse the response - text = response.content[0].text - #print(text) - # Attempt to parse and fix the XML if necessary - parser = etree.XMLParser(recover=True) - root = etree.fromstring(text, parser=parser) - # Check if there were errors corrected - # if parser.error_log: - # #print("XML Parsing Errors:") - # for error in parser.error_log: - # #print(error) - # Extract summaries - summaries = {} - for summary in root.findall('summary'): - number = int(summary.get('number')) - content = summary.text.strip() if summary.text else "" - if content: # Only include non-empty summaries - summaries[number] = content - - return summaries + parsed = json.loads(resp.choices[0].message.content) # schema‑safe + return {item["number"]: item["summary"] + for item in parsed["summaries"]} except Exception as e: - # Print errors to stderr so they don't interfere with JSON output - print(json.dumps({"error": str(e)}), file=sys.stderr) - sys.stderr.flush() - + # Log and fall back gracefully + print(json.dumps({"error": str(e)}), file=sys.stderr, flush=True) + return {} class DocumentType(Enum): """ -- cgit v1.2.3-70-g09d2 From a5d7f5c38192b91b7df3bd6ecace5ba7365449a6 Mon Sep 17 00:00:00 2001 From: "A.J. Shulman" <Shulman.aj@gmail.com> Date: Sun, 11 May 2025 13:42:00 -0400 Subject: Made it so chunk Ids are seperately managed and made sure the doc id is sonsistent and not created in python spawn --- src/client/views/nodes/WebBox.scss | 241 ++++---- src/client/views/nodes/WebBox.tsx | 605 +++++++-------------- src/client/views/nodes/WebBoxRenderer.js | 103 ++++ .../views/nodes/chatbot/agentsystem/prompts.ts | 4 +- .../nodes/chatbot/chatboxcomponents/ChatBox.tsx | 36 +- .../nodes/chatbot/tools/DocumentMetadataTool.ts | 38 +- src/client/views/nodes/chatbot/tools/RAGTool.ts | 5 +- .../nodes/chatbot/utils/AgentDocumentManager.ts | 213 +++----- .../views/nodes/chatbot/vectorstore/Vectorstore.ts | 33 +- src/server/ApiManagers/AssistantManager.ts | 8 +- src/server/chunker/pdf_chunker.py | 13 +- 11 files changed, 548 insertions(+), 751 deletions(-) (limited to 'src/server/chunker') diff --git a/src/client/views/nodes/WebBox.scss b/src/client/views/nodes/WebBox.scss index a1991d1d0..77d7716f4 100644 --- a/src/client/views/nodes/WebBox.scss +++ b/src/client/views/nodes/WebBox.scss @@ -1,9 +1,13 @@ @use '../global/globalCssVariables.module.scss' as global; .webBox { + height: 100%; + width: 100%; + top: 0; + left: 0; position: relative; + display: flex; overflow: hidden; - aspect-ratio: 1 / 1; // Explicitly enforce square aspect ratio .webBox-sideResizer { position: absolute; @@ -16,119 +20,6 @@ .webBox-background { width: 100%; height: 100%; - position: absolute; - top: 0; - left: 0; - } - - // Simple container for screenshot - .webBox-screenshot-container { - width: 100%; - } - - .webBox-screenshot { - width: 100%; - height: auto; // Maintain aspect ratio - display: block; - pointer-events: none; - } - - .webBox-loading { - padding: 20px; - text-align: center; - color: #666; - background-color: #f5f5f5; - min-height: 200px; - display: flex; - flex-direction: column; - align-items: center; - justify-content: center; - } - - .webBox-loading-spinner { - margin-top: 15px; - color: #1976d2; - font-size: 24px; - } - - .webBox-error { - padding: 20px; - color: #d32f2f; - text-align: center; - background-color: #ffebee; - min-height: 200px; - display: flex; - flex-direction: column; - align-items: center; - justify-content: center; - gap: 15px; - } - - .webBox-placeholder { - padding: 20px; - text-align: center; - color: #757575; - background-color: #fafafa; - min-height: 200px; - display: flex; - align-items: center; - justify-content: center; - } - - // Basic container layout - .webBox-container { - width: 100%; - height: 100%; - position: relative; - } - - // Simple scrollable container - vertical only - .webBox-outerContent { - width: 100%; - position: relative; - overflow-y: auto; - overflow-x: hidden; - background-color: #f5f5f5; - - // Improve scrollbar styling - &::-webkit-scrollbar-thumb { - background-color: #888; - border-radius: 6px; - } - - &::-webkit-scrollbar { - width: 8px; - background-color: #f5f5f5; - } - } - - .webBox-innerContent { - width: 100%; - background-color: #f5f5f5; - } - - .webBox-htmlSpan { - position: absolute; - top: 0; - left: 0; - cursor: text; - padding: 15px; - width: 100%; - height: 100%; - } - - .webBox-annotationLayer { - position: absolute; - transform-origin: left top; - top: 0; - width: 100%; - pointer-events: none; - mix-blend-mode: multiply; - } - - .webBox-annotationBox { - position: absolute; - background-color: rgba(245, 230, 95, 0.616); } .webBox-ui { @@ -177,14 +68,14 @@ } } - .webBox-refreshButton { + .webBox-nextIcon, + .webBox-prevIcon { background: #121721; + color: white; height: 20px; width: 25px; display: flex; - position: absolute; - bottom: 0; - right: 40px; + position: relative; align-items: center; justify-content: center; border-radius: 3px; @@ -192,6 +83,10 @@ padding: 0px; } + .webBox-overlayButton:hover { + background: none; + } + .webBox-overlayCont { position: absolute; width: calc(100% - 40px); @@ -223,7 +118,8 @@ justify-content: center; border-radius: 3px; pointer-events: all; - z-index: 1; + z-index: 1; // so it appears on top of the document's title, if shown + box-shadow: global.$standard-box-shadow; transition: 0.2s; @@ -238,6 +134,89 @@ opacity: 0.1; } + .webBox-annotationLayer { + position: absolute; + transform-origin: left top; + top: 0; + width: 100%; + pointer-events: none; + mix-blend-mode: multiply; // bcz: makes text fuzzy! + } + + .webBox-annotationBox { + position: absolute; + background-color: rgba(245, 230, 95, 0.616); + } + + .webBox-container { + transform-origin: top left; + width: 100%; + height: 100%; + position: absolute; + + .webBox-htmlSpan { + position: absolute; + top: 0; + left: 0; + cursor: text; + padding: 15px; + height: 100%; + } + + .webBox-cont { + pointer-events: none; + } + + .webBox-cont, + .webBox-cont-interactive { + padding: 0vw; + position: absolute; + top: 0; + left: 0; + width: 100%; + height: 100%; + transform-origin: top left; + + .webBox-iframe { + width: 100%; + height: 100%; + position: absolute; + top: 0; + left: 0; + body { + ::selection { + color: white; + background: orange; + } + } + } + } + + .webBox-cont-interactive { + span { + user-select: text !important; + } + } + + .webBox-outerContent { + width: 100%; + height: 100%; + position: absolute; + transform-origin: top left; + top: 0; + left: 0; + overflow: auto; + + .webBox-innerContent { + position: relative; + } + } + + div.webBox-outerContent::-webkit-scrollbar-thumb { + cursor: nw-resize; + } + } + .webBox-overlay { width: 100%; height: 100%; @@ -277,13 +256,37 @@ width: 100%; height: 100%; position: absolute; + pointer-events: all; .indicator { position: absolute; + transition: background-color 0.2s ease; + border-radius: 2px; &.active { background-color: rgba(0, 0, 0, 0.1); + box-shadow: 0 0 2px rgba(0, 0, 0, 0.2); } } } + + // Add styles to hide font errors and improve user experience + .font-error-hidden { + font-family: + system-ui, + -apple-system, + BlinkMacSystemFont, + 'Segoe UI', + Roboto, + Arial, + sans-serif !important; + } + + // Change iframe behavior when resource loading errors occur + iframe.webBox-iframe { + &.loading-error { + // Make full content accessible when external resources fail + pointer-events: all !important; + } + } } diff --git a/src/client/views/nodes/WebBox.tsx b/src/client/views/nodes/WebBox.tsx index 045af7ecd..1e158f484 100644 --- a/src/client/views/nodes/WebBox.tsx +++ b/src/client/views/nodes/WebBox.tsx @@ -4,7 +4,6 @@ import { htmlToText } from 'html-to-text'; import { action, computed, IReactionDisposer, makeObservable, observable, ObservableMap, reaction, runInAction } from 'mobx'; import { observer } from 'mobx-react'; import * as React from 'react'; -import axios from 'axios'; import * as WebRequest from 'web-request'; import { addStyleSheet, addStyleSheetRule, clearStyleSheetRules, ClientUtils, DivHeight, getWordAtPoint, lightOrDark, returnFalse, returnOne, returnZero, setupMoveUpEvents, smoothScroll } from '../../../ClientUtils'; import { Doc, DocListCast, Field, FieldType, Opt, StrListCast } from '../../../fields/Doc'; @@ -70,20 +69,23 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { private _scrollTimer: NodeJS.Timeout | undefined; private _getAnchor: (savedAnnotations: Opt<ObservableMap<number, HTMLDivElement[]>>, addAsAnnotation: boolean) => Opt<Doc> = () => undefined; - @observable private _webUrl = ''; // url of the page we want to display - @observable private _hackHide = false; + @observable private _webUrl = ''; // url of the src parameter of the embedded iframe but not necessarily the rendered page - eg, when following a link, the rendered page changes but we don't want the src parameter to also change as that would cause an unnecessary re-render. + @observable private _hackHide = false; // apparently changing the value of the 'sandbox' prop doesn't necessarily apply it to the active iframe. so thisforces the ifrmae to be rebuilt when allowScripts is toggled @observable private _searching: boolean = false; @observable private _showSidebar = false; @observable private _webPageHasBeenRendered = false; @observable private _marqueeing: number[] | undefined = undefined; - @observable private _screenshotUrl: string | null = null; // URL to the screenshot image - @observable private _fullHeight: number = 0; // Full height of the webpage screenshot - @observable private _isLoadingScreenshot: boolean = false; // Loading state for the screenshot + get marqueeing() { + return this._marqueeing; + } + set marqueeing(val) { + val && this._marqueeref.current?.onInitiateSelection(val); + !val && this._marqueeref.current?.onTerminateSelection(); + this._marqueeing = val; + } @observable private _iframe: HTMLIFrameElement | null = null; @observable private _savedAnnotations = new ObservableMap<number, (HTMLDivElement & { marqueeing?: boolean })[]>(); @observable private _scrollHeight = NumCast(this.layoutDoc.scrollHeight); - @observable private _screenshotError: string | null = null; // Error message if screenshot fails - @observable private _loadingFromCache: boolean = false; @computed get _url() { return this.webField?.toString() || ''; } @@ -143,38 +145,31 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { }; updateIcon = async () => { - if (!this._screenshotUrl) { - // If we don't have a screenshot yet, capture one first - await this.captureWebScreenshot(); - } - + if (!this._iframe) return new Promise<void>(res => res()); const scrollTop = NumCast(this.layoutDoc._layout_scrollTop); const nativeWidth = NumCast(this.layoutDoc.nativeWidth); const nativeHeight = (nativeWidth * this._props.PanelHeight()) / this._props.PanelWidth(); - + let htmlString = this._iframe.contentDocument && new XMLSerializer().serializeToString(this._iframe.contentDocument); + if (!htmlString) { + htmlString = await fetch(ClientUtils.CorsProxy(this.webField!.href)).then(response => response.text()); + } this.layoutDoc.thumb = undefined; this.Document.thumbLockout = true; // lock to prevent multiple thumb updates. - - try { - // If we have a screenshot, use it directly for the thumbnail - if (this._screenshotUrl) { - return ClientUtils.convertDataUri(this._screenshotUrl, this.layoutDoc[Id] + '_icon_' + new Date().getTime(), true, this.layoutDoc[Id] + '_icon_').then(returnedfilename => { + return (CreateImage(this._webUrl.endsWith('/') ? this._webUrl.substring(0, this._webUrl.length - 1) : this._webUrl, this._iframe.contentDocument?.styleSheets ?? [], htmlString, nativeWidth, nativeHeight, scrollTop) as Promise<string>) + .then((dataUrl: string) => { + if (dataUrl.includes('<!DOCTYPE')) { + console.log('BAD DATA IN THUMB CREATION'); + return; + } + return ClientUtils.convertDataUri(dataUrl, this.layoutDoc[Id] + '_icon_' + new Date().getTime(), true, this.layoutDoc[Id] + '_icon_').then(returnedfilename => { this.Document.thumbLockout = false; this.layoutDoc.thumb = new ImageField(returnedfilename); this.layoutDoc.thumbScrollTop = scrollTop; this.layoutDoc.thumbNativeWidth = nativeWidth; this.layoutDoc.thumbNativeHeight = nativeHeight; }); - } else { - console.log('No screenshot available for thumbnail generation'); - this.Document.thumbLockout = false; - return Promise.resolve(); - } - } catch (error) { - console.error('Error creating thumbnail:', error); - this.Document.thumbLockout = false; - return Promise.reject(error); - } + }) + .catch((error: object) => console.error('oops, something went wrong!', error)); }; componentDidMount() { @@ -243,64 +238,13 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { }, { fireImmediately: true } ); - - // Check if we have a cached screenshot URL in metadata - if (this._url) { - this._webUrl = this._url; - const cachedScreenshotUrl = StrCast(this.dataDoc[this.fieldKey + '_screenshotUrl']); - const cachedHeight = NumCast(this.dataDoc[this.fieldKey + '_screenshotHeight']); - - if (cachedScreenshotUrl && cachedHeight) { - // Use cached screenshot - this._loadingFromCache = true; - this._isLoadingScreenshot = true; - - // Verify the cached screenshot exists by loading the image - const img = new Image(); - img.onload = action(() => { - this._screenshotUrl = cachedScreenshotUrl; - this._fullHeight = cachedHeight; - this._scrollHeight = cachedHeight; - this._webPageHasBeenRendered = true; - this._isLoadingScreenshot = false; - this._loadingFromCache = false; - - // Apply dimensions and initial scroll - if (this.layoutDoc._layout_autoHeight) { - this.layoutDoc._nativeHeight = this._fullHeight; - this._props.setHeight?.(this._fullHeight * (this._props.NativeDimScaling?.() || 1)); - } - - if (this._initialScroll !== undefined) { - this.setScrollPos(this._initialScroll); - } - - console.log(`Loaded cached screenshot: ${this._screenshotUrl}`); - }); - - img.onerror = action(() => { - // If image fails to load, capture a new screenshot - console.log('Cached screenshot not found, capturing new one'); - this._loadingFromCache = false; - this.captureWebScreenshot(); - }); - - img.src = cachedScreenshotUrl; - } else { - // No cached screenshot, capture a new one - this.captureWebScreenshot(); - } - } } componentWillUnmount() { - // Clean up timers - if (this._scrollTimer) { - clearTimeout(this._scrollTimer); - this._scrollTimer = undefined; - } - - // Clean up reaction disposers + this._iframetimeout && clearTimeout(this._iframetimeout); + this._iframetimeout = undefined; Object.values(this._disposers).forEach(disposer => disposer?.()); + // this._iframe?.removeEventListener('wheel', this.iframeWheel, true); + // this._iframe?.contentDocument?.removeEventListener("pointerup", this.iframeUp); } private _selectionText: string = ''; @@ -414,6 +358,59 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { _textAnnotationCreator: (() => ObservableMap<number, (HTMLDivElement & { marqueeing?: boolean })[]>) | undefined; savedAnnotationsCreator: () => ObservableMap<number, (HTMLDivElement & { marqueeing?: boolean })[]> = () => this._textAnnotationCreator?.() || this._savedAnnotations; + @action + iframeMove = (e: PointerEvent) => { + const theclick = this.props + .ScreenToLocalTransform() + .inverse() + .transformPoint(e.clientX, e.clientY - NumCast(this.layoutDoc.layout_scrollTop)); + this._marqueeref.current?.onMove(theclick); + }; + @action + iframeUp = (e: PointerEvent) => { + this._iframe?.contentDocument?.removeEventListener('pointermove', this.iframeMove); + this.marqueeing = undefined; + this._getAnchor = AnchorMenu.Instance?.GetAnchor; // need to save AnchorMenu's getAnchor since a subsequent selection on another doc will overwrite this value + this._textAnnotationCreator = undefined; + this.DocumentView?.()?.cleanupPointerEvents(); // pointerup events aren't generated on containing document view, so we have to invoke it here. + if (this._iframe?.contentWindow && this._iframe.contentDocument && !this._iframe.contentWindow.getSelection()?.isCollapsed) { + const mainContBounds = ClientUtils.GetScreenTransform(this._mainCont.current!); + const scale = (this._props.NativeDimScaling?.() || 1) * mainContBounds.scale; + const sel = this._iframe.contentWindow.getSelection(); + if (sel) { + this._selectionText = sel.toString(); + AnchorMenu.Instance.setSelectedText(sel.toString()); + this._textAnnotationCreator = () => this.createTextAnnotation(sel, !sel.isCollapsed ? sel.getRangeAt(0) : undefined); + AnchorMenu.Instance.jumpTo(e.clientX * scale + mainContBounds.translateX, e.clientY * scale + mainContBounds.translateY - NumCast(this.layoutDoc._layout_scrollTop) * scale); + // Changing which document to add the annotation to (the currently selected WebBox) + GPTPopup.Instance.setSidebarFieldKey(`${this._props.fieldKey}_${this._urlHash ? this._urlHash + '_' : ''}sidebar`); + GPTPopup.Instance.addDoc = this.sidebarAddDocument; + } + } else { + const theclick = this.props + .ScreenToLocalTransform() + .inverse() + .transformPoint(e.clientX, e.clientY - NumCast(this.layoutDoc.layout_scrollTop)); + if (!this._marqueeref.current?.isEmpty) this._marqueeref.current?.onEnd(theclick[0], theclick[1]); + else { + if (!(e.target as HTMLElement)?.tagName?.includes('INPUT')) this.finishMarquee(theclick[0], theclick[1]); + this._getAnchor = AnchorMenu.Instance?.GetAnchor; + this.marqueeing = undefined; + } + + ContextMenu.Instance.closeMenu(); + ContextMenu.Instance.setIgnoreEvents(false); + if (e?.button === 2 || e?.altKey) { + e?.preventDefault(); + e?.stopPropagation(); + setTimeout(() => { + // if menu comes up right away, the down event can still be active causing a menu item to be selected + this.specificContextMenu(); + this.DocumentView?.().onContextMenu(undefined, theclick[0], theclick[1]); + }); + } + } + }; @action webClipDown = (e: React.PointerEvent) => { e.stopPropagation(); @@ -508,6 +505,98 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { this._scrollHeight = this._iframe?.contentDocument?.body?.scrollHeight ?? 0; this.addWebStyleSheetRule(this.addWebStyleSheet(this._iframe?.contentDocument), '::selection', { color: 'white', background: 'orange' }, ''); + // Add error handler to suppress font CORS errors + if (this._iframe?.contentWindow) { + try { + // Track if any resource errors occurred + let hasResourceErrors = false; + + // Override the console.error to filter out font CORS errors + const win = this._iframe.contentWindow as Window & { console: Console }; + const originalConsoleError = win.console.error; + win.console.error = (...args: unknown[]) => { + const errorMsg = args.map(arg => String(arg)).join(' '); + if (errorMsg.includes('Access to font') && errorMsg.includes('has been blocked by CORS policy')) { + // Mark that we have font errors + hasResourceErrors = true; + // Ignore font CORS errors + return; + } + // Also catch other resource loading errors + if (errorMsg.includes('ERR_FAILED') || errorMsg.includes('ERR_BLOCKED_BY_CLIENT')) { + hasResourceErrors = true; + } + originalConsoleError.apply(win.console, args); + }; + + // Listen for resource loading errors + this._iframe.contentWindow.addEventListener( + 'error', + (e: Event) => { + const target = e.target as HTMLElement; + if (target instanceof HTMLElement) { + // If it's a resource that failed to load + if (target.tagName === 'LINK' || target.tagName === 'IMG' || target.tagName === 'SCRIPT') { + hasResourceErrors = true; + // Apply error class after a short delay to allow initial content to load + setTimeout(() => { + if (this._iframe && hasResourceErrors) { + this._iframe.classList.add('loading-error'); + } + }, 1000); + } + } + }, + true + ); + + // Add fallback CSS for fonts that fail to load + const style = this._iframe.contentDocument?.createElement('style'); + if (style) { + style.textContent = ` + @font-face { + font-family: 'CORS-fallback-serif'; + src: local('Times New Roman'), local('Georgia'), serif; + } + @font-face { + font-family: 'CORS-fallback-sans'; + src: local('Arial'), local('Helvetica'), sans-serif; + } + /* Fallback for all fonts that fail to load */ + @font-face { + font-display: swap !important; + } + + /* Add a script to find and fix elements with failed fonts */ + @font-face { + font-family: '__failed_font__'; + src: local('Arial'); + unicode-range: U+0000; + } + `; + this._iframe.contentDocument?.head.appendChild(style); + + // Add a script to detect and fix font loading issues + const script = this._iframe.contentDocument?.createElement('script'); + if (script) { + script.textContent = ` + // Fix font loading issues with fallbacks + setTimeout(function() { + document.querySelectorAll('*').forEach(function(el) { + if (window.getComputedStyle(el).fontFamily.includes('__failed_font__')) { + el.classList.add('font-error-hidden'); + } + }); + }, 1000); + `; + this._iframe.contentDocument?.head.appendChild(script); + } + } + } catch (e) { + console.log('Error setting up font error handling:', e); + } + } + let href: Opt<string>; try { href = iframe?.contentWindow?.location.href; @@ -658,23 +747,15 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { this.dataDoc[this.fieldKey + '_history'] = new List<string>([...history, this._url]); this.dataDoc[this.fieldKey] = new WebField(new URL(future.pop()!)); this._scrollHeight = 0; - - // Reset screenshot state for new URL - this._screenshotUrl = null; - this._fullHeight = 0; - this._isLoadingScreenshot = false; - if (this._webUrl === this._url) { this._webUrl = curUrl; setTimeout( action(() => { this._webUrl = this._url; - this.captureWebScreenshot(); // Capture screenshot for new URL }) ); } else { this._webUrl = this._url; - this.captureWebScreenshot(); // Capture screenshot for new URL } return true; } @@ -694,18 +775,11 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { else this.dataDoc[this.fieldKey + '_future'] = new List<string>([...future, this._url]); this.dataDoc[this.fieldKey] = new WebField(new URL(history.pop()!)); this._scrollHeight = 0; - - // Reset screenshot state for new URL - this._screenshotUrl = null; - this._fullHeight = 0; - this._isLoadingScreenshot = false; - if (this._webUrl === this._url) { this._webUrl = curUrl; setTimeout(action(() => (this._webUrl = this._url))); } else { this._webUrl = this._url; - this.captureWebScreenshot(); // Capture screenshot for new URL } return true; } @@ -724,11 +798,10 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { this.layoutDoc.thumbNativeWidth = undefined; this.layoutDoc.thumbNativeHeight = undefined; } - + } + if (!preview) { if (!dontUpdateIframe) { this._webUrl = this._url; - // Capture screenshot when URL changes - this.captureWebScreenshot(); } } } catch { @@ -737,85 +810,6 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { return true; }; - @action - captureWebScreenshot = async () => { - if (!this._url || this._loadingFromCache) return; - - try { - this._isLoadingScreenshot = true; - this._screenshotError = null; - - console.log(`Capturing screenshot for URL: ${this._url}`); - - try { - const response = await axios.post('/captureWebScreenshot', { - url: this._url, - width: NumCast(this.Document.nativeWidth, 1200), - height: NumCast(this.Document.nativeHeight, 800), - fullPage: true, // Request a full page screenshot - }); - - runInAction(() => { - this._screenshotUrl = response.data.screenshotUrl; - this._fullHeight = response.data.fullHeight; - this._scrollHeight = response.data.fullHeight; - this._webPageHasBeenRendered = true; - this._isLoadingScreenshot = false; - - // Store screenshot URL and height in document metadata - this.dataDoc[this.fieldKey + '_screenshotUrl'] = response.data.screenshotUrl; - this.dataDoc[this.fieldKey + '_screenshotHeight'] = response.data.fullHeight; - - // Update native dimensions to match the screenshot - if (!this.dataDoc[this.fieldKey + '_nativeWidth']) { - this.dataDoc[this.fieldKey + '_nativeWidth'] = 1200; // Default width - } - - if (!this.dataDoc[this.fieldKey + '_nativeHeight']) { - this.dataDoc[this.fieldKey + '_nativeHeight'] = this._fullHeight; - } - - // Set document height if needed - if (this.layoutDoc._layout_autoHeight) { - this.layoutDoc._nativeHeight = this._fullHeight; - this._props.setHeight?.(this._fullHeight * (this._props.NativeDimScaling?.() || 1)); - } - - // Apply initial scroll if needed - if (this._initialScroll !== undefined) { - this.setScrollPos(this._initialScroll); - } - - console.log(`Screenshot captured successfully: ${this._screenshotUrl} with height: ${this._fullHeight}px`); - }); - } catch (error: any) { - // Handle error from the API - console.error('Error capturing screenshot:', error); - let errorMessage = 'Failed to capture webpage screenshot'; - - // Try to extract detailed error message from response - if (error.response && error.response.data && error.response.data.error) { - errorMessage = error.response.data.error; - } else if (error.message) { - errorMessage = error.message; - } - - runInAction(() => { - this._screenshotError = errorMessage; - this._isLoadingScreenshot = false; - }); - } - } catch (error: any) { - // Handle unexpected errors - runInAction(() => { - console.error('Unexpected error in captureWebScreenshot:', error); - this._screenshotError = 'An unexpected error occurred'; - this._isLoadingScreenshot = false; - }); - } - }; - - @action onWebUrlDrop = (e: React.DragEvent) => { const { dataTransfer } = e; const html = dataTransfer.getData('text/html'); @@ -830,28 +824,13 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { setData = (data: FieldType | Promise<RefField | undefined>) => { if (!(typeof data === 'string') && !(data instanceof WebField)) return false; if (Field.toString(data) === this._url) return false; - - // Reset state for new URL this._scrollHeight = 0; - this._screenshotUrl = null; - this._fullHeight = 0; - this._isLoadingScreenshot = false; - - // Clear stored screenshot metadata for the previous URL - this.dataDoc[this.fieldKey + '_screenshotUrl'] = undefined; - this.dataDoc[this.fieldKey + '_screenshotHeight'] = undefined; - const oldUrl = this._url; const history = Cast(this.dataDoc[this.fieldKey + '_history'], listSpec('string'), []); const weburl = new WebField(Field.toString(data)); this.dataDoc[this.fieldKey + '_future'] = new List<string>([]); this.dataDoc[this.fieldKey + '_history'] = new List<string>([...(history || []), oldUrl]); this.dataDoc[this.fieldKey] = weburl; - - // Capture screenshot for the new URL - this._webUrl = weburl.toString(); - this.captureWebScreenshot(); - return true; }; onWebUrlValueKeyDown = (e: React.KeyboardEvent) => { @@ -868,14 +847,26 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { description: (this.layoutDoc[this.fieldKey + '_useCors'] ? "Don't Use" : 'Use') + ' Cors', event: () => { this.layoutDoc[this.fieldKey + '_useCors'] = !this.layoutDoc[this.fieldKey + '_useCors']; - // Re-capture screenshot with the new setting - this.captureWebScreenshot(); }, icon: 'snowflake', }); - - // Remove the "Allow Scripts" option since it's not relevant for screenshots - + funcs.push({ + description: (this.dataDoc[this.fieldKey + '_allowScripts'] ? 'Prevent' : 'Allow') + ' Scripts', + event: () => { + this.dataDoc[this.fieldKey + '_allowScripts'] = !this.dataDoc[this.fieldKey + '_allowScripts']; + if (this._iframe) { + runInAction(() => { + this._hackHide = true; + }); + setTimeout( + action(() => { + this._hackHide = false; + }) + ); + } + }, + icon: 'snowflake', + }); funcs.push({ description: (!this.layoutDoc.layout_reflowHorizontal ? 'Force' : 'Prevent') + ' Reflow', event: () => { @@ -887,21 +878,7 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { }, icon: 'snowflake', }); - - // Add a refresh option to re-capture the screenshot - funcs.push({ - description: 'Refresh Screenshot', - event: () => this.captureWebScreenshot(), - icon: 'sync-alt', - }); - - !Doc.noviceMode && - funcs.push({ - description: 'Update Icon', - event: () => this.updateIcon(), - icon: 'portrait', - }); - + !Doc.noviceMode && funcs.push({ description: 'Update Icon', event: () => this.updateIcon(), icon: 'portrait' }); cm.addItem({ description: 'Options...', subitems: funcs, icon: 'asterisk' }); } }; @@ -913,7 +890,7 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { */ @action onMarqueeDown = (e: React.PointerEvent) => { - const sel = window.document.getSelection(); + const sel = this._url ? this._iframe?.contentDocument?.getSelection() : window.document.getSelection(); this._textAnnotationCreator = undefined; if (sel?.empty) sel.empty(); // Chrome @@ -948,7 +925,6 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { @computed get urlContent() { if (this.ScreenToLocalBoxXf().Scale > 25) return <div />; - setTimeout( action(() => { if (this._initialScroll === undefined && !this._webPageHasBeenRendered) { @@ -957,10 +933,7 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { this._webPageHasBeenRendered = true; }) ); - const field = this.dataDoc[this._props.fieldKey]; - - // Handle HTML field (text content) if (field instanceof HtmlField) { return ( <span @@ -977,8 +950,6 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { /> ); } - - // Handle WebField (screenshot of webpage) if (field instanceof WebField) { const url = this.layoutDoc[this.fieldKey + '_useCors'] ? '/corsproxy/' + this._webUrl : this._webUrl; const scripts = this.dataDoc[this.fieldKey + '_allowScripts'] || this._webUrl.includes('wikipedia.org') || this._webUrl.includes('google.com') || this._webUrl.startsWith('https://bing'); @@ -1198,7 +1169,7 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { TraceMobx(); // const previewScale = this._previewNativeWidth ? 1 - this.sidebarWidth() / this._previewNativeWidth : 1; const pointerEvents = this.layoutDoc._lockedPosition ? 'none' : (this._props.pointerEvents?.() as Property.PointerEvents | undefined); - const scale = this._props.NativeDimScaling?.() || 1; + // const scale = previewScale * (this._props.NativeDimScaling?.() || 1); return ( <div className="webBox-outerContent" @@ -1207,16 +1178,11 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { height: '100%', //`${100 / scale}%`, pointerEvents, }} + // when active, block wheel events from propagating since they're handled by the iframe onWheel={this.onZoomWheel} onScroll={() => this.setDashScrollTop(this._outerRef.current?.scrollTop || 0)} onPointerDown={this.onMarqueeDown}> - <div - className="webBox-innerContent" - style={{ - width: '100%', - pointerEvents, - backgroundColor: '#f5f5f5', - }}> + <div className="webBox-innerContent" style={{ height: (this._webPageHasBeenRendered && this._scrollHeight > this._props.PanelHeight() && this._scrollHeight) || '100%', pointerEvents }}> {this.content} <div style={{ display: SnappingManager.CanEmbed ? 'none' : undefined, mixBlendMode: 'multiply' }}>{this.renderTransparentAnnotations}</div> {this.renderOpaqueAnnotations} @@ -1258,13 +1224,6 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { <FontAwesomeIcon icon={this._searching ? 'times' : 'search'} size="lg" /> </div> </button> - - {/* Refresh button */} - <button type="button" className="webBox-overlayButton webBox-refreshButton" title="Refresh webpage" onClick={() => this.captureWebScreenshot()}> - <div className="webBox-overlayButton-iconCont" onPointerDown={e => e.stopPropagation()}> - <FontAwesomeIcon icon="sync" size="lg" /> - </div> - </button> </div> ); } @@ -1293,25 +1252,16 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { annotationPointerEvents = () => (this._props.isContentActive() && (SnappingManager.IsDragging || Doc.ActiveTool !== InkTool.None) ? 'all' : 'none'); render() { TraceMobx(); - const containerWidth = NumCast(this.layoutDoc._width) || this._props.PanelWidth(); + const previewScale = this._previewNativeWidth ? 1 - this.sidebarWidth() / this._previewNativeWidth : 1; const pointerEvents = this.layoutDoc._lockedPosition ? 'none' : (this._props.pointerEvents?.() as Property.PointerEvents); - const scale = this._props.NativeDimScaling?.() || 1; - - // Force the component to be square - this.layoutDoc._height = containerWidth; - this.layoutDoc._width = containerWidth; - this.layoutDoc._forceActive = true; - + const scale = previewScale * (this._props.NativeDimScaling?.() || 1); return ( <div className="webBox" ref={this._mainCont} style={{ - pointerEvents: this.pointerEvents(), + pointerEvents: this.pointerEvents(), // position: SnappingManager.IsDragging ? 'absolute' : undefined, - width: `${containerWidth}px`, - height: `${containerWidth}px`, - aspectRatio: '1 / 1', // Explicitly enforce square aspect ratio }}> <div className="webBox-background" style={{ backgroundColor: this._props.styleProvider?.(this.layoutDoc, this._props, StyleProp.BackgroundColor) as string }} /> <div @@ -1376,15 +1326,6 @@ export class WebBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { </div> ); } - - get marqueeing() { - return this._marqueeing; - } - set marqueeing(val) { - val && this._marqueeref.current?.onInitiateSelection(val); - !val && this._marqueeref.current?.onTerminateSelection(); - this._marqueeing = val; - } } // eslint-disable-next-line prefer-arrow-callback ScriptingGlobals.add(function urlHash(url: string) { @@ -1395,149 +1336,3 @@ Docs.Prototypes.TemplateMap.set(DocumentType.WEB, { layout: { view: WebBox, dataField: 'data' }, options: { acl: '', _height: 300, _layout_fitWidth: true, _layout_nativeDimEditable: true, _layout_reflowVertical: true, waitForDoubleClickToClick: 'always', systemIcon: 'BsGlobe' }, }); - -// Add CSS styles for screenshot mode -const webBoxStyles = ` -.webBox-screenshot-container { - width: 100%; - position: relative; - overflow: visible; - display: flex; - align-items: flex-start; - justify-content: center; - background-color: #f5f5f5; -} - -.webBox-screenshot { - width: 100%; - pointer-events: none; - display: block; - user-select: none; - object-fit: contain; - transition: opacity 0.3s ease; -} - -.webBox-loading { - padding: 20px; - text-align: center; - color: #666; - background-color: #f5f5f5; - border-radius: 4px; - min-height: 200px; - display: flex; - flex-direction: column; - align-items: center; - justify-content: center; -} - -.webBox-loading-message { - font-size: 16px; - margin-bottom: 15px; - color: #555; -} - -.webBox-loading-spinner { - margin-top: 10px; - color: #1976d2; -} - -.webBox-error { - padding: 20px; - color: #d32f2f; - text-align: center; - background-color: #ffebee; - border-radius: 4px; - min-height: 200px; - display: flex; - flex-direction: column; - align-items: center; - justify-content: center; - gap: 15px; -} - -.webBox-error-icon { - color: #d32f2f; - margin-bottom: 10px; -} - -.webBox-error-message { - color: #d32f2f; - font-size: 14px; - max-width: 80%; - line-height: 1.5; -} - -.webBox-error-actions { - margin-top: 10px; -} - -.webBox-retry-button { - background-color: #f44336; - color: white; - border: none; - padding: 8px 16px; - border-radius: 4px; - cursor: pointer; - font-size: 14px; - transition: background-color 0.3s; -} - -.webBox-retry-button:hover { - background-color: #d32f2f; -} - -.webBox-placeholder { - padding: 20px; - text-align: center; - color: #757575; - background-color: #fafafa; - border-radius: 4px; - min-height: 200px; - display: flex; - align-items: center; - justify-content: center; -} - -.webBox-refreshButton { - margin-right: 5px; -} - -.webBox-innerContent { - position: relative; - width: 100%; - background-color: #f5f5f5; - overflow: visible; -} - -.webBox-outerContent { - overflow: auto; - width: 100%; - background-color: #f5f5f5; - position: relative; -} - -.webBox-container { - position: relative; - display: flex; - flex-direction: column; - height: 100%; - background-color: white; - border-radius: 4px; - overflow: hidden; -} - -.webBox { - position: relative; - height: 100%; - width: 100%; - overflow: hidden; - background-color: white; - border-radius: 4px; - box-shadow: 0 1px 3px rgba(0, 0, 0, 0.12), 0 1px 2px rgba(0, 0, 0, 0.24); -} -`; - -// Add the styles to the document -const styleEl = document.createElement('style'); -styleEl.textContent = webBoxStyles; -document.head.appendChild(styleEl); diff --git a/src/client/views/nodes/WebBoxRenderer.js b/src/client/views/nodes/WebBoxRenderer.js index ef465c453..31e0ef5e4 100644 --- a/src/client/views/nodes/WebBoxRenderer.js +++ b/src/client/views/nodes/WebBoxRenderer.js @@ -145,6 +145,29 @@ const ForeignHtmlRenderer = function (styleSheets) { return urlsFound; }; + /** + * Extracts font-face URLs from CSS rules + * @param {String} cssRuleStr + * @returns {String[]} + */ + const getFontFaceUrlsFromCss = function (cssRuleStr) { + const fontFaceUrls = []; + // Find @font-face blocks + const fontFaceBlocks = cssRuleStr.match(/@font-face\s*{[^}]*}/g) || []; + + fontFaceBlocks.forEach(block => { + // Extract URLs from src properties + const urls = block.match(/src\s*:\s*[^;]*/g) || []; + urls.forEach(srcDeclaration => { + // Find all url() references in the src declaration + const fontUrls = getUrlsFromCssString(srcDeclaration); + fontFaceUrls.push(...fontUrls); + }); + }); + + return fontFaceUrls; + }; + /** * * @param {String} html @@ -158,6 +181,61 @@ const ForeignHtmlRenderer = function (styleSheets) { return string.replace(/[.*+?^${}()|[\]\\]/g, '\\$&'); // $& means the whole matched string }; + /** + * Create a fallback font-face rule for handling CORS errors + * @returns {String} + */ + const createFallbackFontFaceRules = function () { + return ` + @font-face { + font-family: 'CORS-fallback-serif'; + src: local('Times New Roman'), local('Georgia'), serif; + } + @font-face { + font-family: 'CORS-fallback-sans'; + src: local('Arial'), local('Helvetica'), sans-serif; + } + /* Add fallback font handling */ + [data-font-error] { + font-family: 'CORS-fallback-sans', sans-serif !important; + } + [data-font-error="serif"] { + font-family: 'CORS-fallback-serif', serif !important; + } + `; + }; + + /** + * Clean up and optimize CSS for better rendering + * @param {String} cssStyles + * @returns {String} + */ + const optimizeCssForRendering = function (cssStyles) { + // Add fallback font-face rules + const enhanced = cssStyles + createFallbackFontFaceRules(); + + // Replace problematic font-face declarations with proxied versions + let optimized = enhanced.replace(/(url\(['"]?)(https?:\/\/[^)'"]+)(['"]?\))/gi, (match, prefix, url, suffix) => { + // If it's a font file, proxy it + if (url.match(/\.(woff2?|ttf|eot|otf)(\?.*)?$/i)) { + return `${prefix}${CorsProxy(url)}${suffix}`; + } + return match; + }); + + // Add error handling for fonts + optimized += ` + /* Suppress font CORS errors in console */ + @supports (font-display: swap) { + @font-face { + font-display: swap !important; + } + } + `; + + return optimized; + }; + /** * * @param {String} contentHtml @@ -175,6 +253,7 @@ const ForeignHtmlRenderer = function (styleSheets) { // copy styles let cssStyles = ''; const urlsFoundInCss = []; + const fontUrlsInCss = []; for (let i = 0; i < styleSheets.length; i += 1) { try { @@ -182,6 +261,7 @@ const ForeignHtmlRenderer = function (styleSheets) { for (let j = 0; j < rules.length; j += 1) { const cssRuleStr = rules[j].cssText; urlsFoundInCss.push(...getUrlsFromCssString(cssRuleStr)); + fontUrlsInCss.push(...getFontFaceUrlsFromCss(cssRuleStr)); cssStyles += cssRuleStr; } } catch (e) { @@ -189,6 +269,9 @@ const ForeignHtmlRenderer = function (styleSheets) { } } + // Optimize and enhance CSS + cssStyles = optimizeCssForRendering(cssStyles); + // const fetchedResourcesFromStylesheets = await getMultipleResourcesAsBase64(webUrl, urlsFoundInCss); // for (let i = 0; i < fetchedResourcesFromStylesheets.length; i++) { // const r = fetchedResourcesFromStylesheets[i]; @@ -203,6 +286,26 @@ const ForeignHtmlRenderer = function (styleSheets) { .replace(/<div class="mediaset"><\/div>/g, '') // when scripting isn't available (ie, rendering web pages here), <noscript> tags should become <div>'s. But for Brown CS, there's a layout problem if you leave the empty <mediaset> tag .replace(/<link[^>]*>/g, '') // don't need to keep any linked style sheets because we've already processed all style sheets above .replace(/srcset="([^ "]*)[^"]*"/g, 'src="$1"'); // instead of converting each item in the srcset to a data url, just convert the first one and use that + + // Add script to handle font loading errors + contentHtml += ` + <script> + // Handle font loading errors with fallbacks + document.addEventListener('DOMContentLoaded', function() { + // Mark elements with font issues + document.querySelectorAll('*').forEach(function(el) { + const style = window.getComputedStyle(el); + const fontFamily = style.getPropertyValue('font-family'); + if (fontFamily && !fontFamily.includes('serif') && !fontFamily.includes('sans')) { + el.setAttribute('data-font-error', 'sans'); + } else if (fontFamily && fontFamily.includes('serif')) { + el.setAttribute('data-font-error', 'serif'); + } + }); + }); + </script> + `; + const urlsFoundInHtml = getImageUrlsFromFromHtml(contentHtml).filter(url => !url.startsWith('data:')); return getMultipleResourcesAsBase64(webUrl, urlsFoundInHtml).then(fetchedResources => { for (let i = 0; i < fetchedResources.length; i += 1) { diff --git a/src/client/views/nodes/chatbot/agentsystem/prompts.ts b/src/client/views/nodes/chatbot/agentsystem/prompts.ts index e551ef830..fcb4ab450 100644 --- a/src/client/views/nodes/chatbot/agentsystem/prompts.ts +++ b/src/client/views/nodes/chatbot/agentsystem/prompts.ts @@ -103,9 +103,9 @@ export function getReactPrompt(tools: BaseTool<ReadonlyArray<Parameter>>[], summ <note>If no external tool is required, use 'no_tool', but if there might be relevant external information, use the appropriate tool.</note> </tools> - <summaries> + <available_documents> ${summaries()} - </summaries> + </available_documents> <chat_history> ${chatHistory} diff --git a/src/client/views/nodes/chatbot/chatboxcomponents/ChatBox.tsx b/src/client/views/nodes/chatbot/chatboxcomponents/ChatBox.tsx index d919b5f7f..34a1ade2e 100644 --- a/src/client/views/nodes/chatbot/chatboxcomponents/ChatBox.tsx +++ b/src/client/views/nodes/chatbot/chatboxcomponents/ChatBox.tsx @@ -650,16 +650,8 @@ export class ChatBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { citation: JSON.stringify(citation, null, 2), }); - // Try to find the document - let doc: Doc | undefined; - // First try to find the document using the document manager's chunk ID lookup - const parentDocId = this.docManager.getDocIdByChunkId(chunkId); - if (parentDocId) { - doc = this.docManager.getDocument(parentDocId); - console.log(`Found document by chunk ID lookup: ${parentDocId}`); - } - + const doc: Doc | undefined = this.docManager.getDocByChunkId(chunkId); if (!doc) { console.warn(`Document not found for citation with chunk_id: ${chunkId}`); return; @@ -989,32 +981,13 @@ export class ChatBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { componentWillUnmount() { this.removeScrollListener(); } - - /** - * Getter that retrieves all linked documents for the current document. - */ - @computed get linkedDocs(): Doc[] { - const docIds = this.docManager.listDocs(); - const docs: Doc[] = []; - - // Get documents from the document manager using the getDocument method - docIds.forEach(id => { - const doc = this.docManager.getDocument(id); - if (doc) { - docs.push(doc); - } - }); - - return docs; - } - /** * Getter that retrieves document IDs of linked documents that have PDF_chunker–parsed content. */ @computed get docIds(): string[] { // Use the document manager to get all document IDs - return Array.from(this.docManager.listDocs()); + return Array.from(this.docManager.listDocs); } /** @@ -1023,7 +996,8 @@ export class ChatBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { @computed get summaries(): string { // Use the document manager to get all summaries - return this.docManager.getAllDocumentSummaries(); + console.log(this.docManager.listDocs); + return JSON.stringify(this.docManager.listDocs); } /** @@ -1064,7 +1038,7 @@ export class ChatBox extends ViewBoxAnnotatableComponent<FieldViewProps>() { }; retrieveDocIds = (): string[] => { - return Array.from(this.docManager.listDocs()); + return Array.from(this.docManager.docIds); }; /** diff --git a/src/client/views/nodes/chatbot/tools/DocumentMetadataTool.ts b/src/client/views/nodes/chatbot/tools/DocumentMetadataTool.ts index e6c2421e5..5297292bf 100644 --- a/src/client/views/nodes/chatbot/tools/DocumentMetadataTool.ts +++ b/src/client/views/nodes/chatbot/tools/DocumentMetadataTool.ts @@ -18,13 +18,13 @@ const parameterDefinitions: ReadonlyArray<Parameter> = [ name: 'action', type: 'string', required: true, - description: 'The action to perform: "get" to retrieve metadata, "edit" to modify metadata, "list" to enumerate documents, "getFieldOptions" to retrieve all available field options, or "create" to create a new document', + description: 'The action to perform: "get" to retrieve metadata, "edit" to modify metadata, "getFieldOptions" to retrieve all available field options, or "create" to create a new document', }, { name: 'documentId', type: 'string', required: false, - description: 'The ID of the document to get or edit metadata for. Required for "edit", optional for "get", ignored for "list", "getFieldOptions", and "create"', + description: 'The ID of the document to get or edit metadata for. Required for "edit", optional for "get", ignored for "getFieldOptions", and "create"', }, { name: 'fieldEdits', @@ -68,7 +68,6 @@ This tool provides the following capabilities: - Get metadata from a specific document - Edit metadata fields on documents (in either layout or data documents) - Edit multiple fields at once (useful for updating dependent fields together) -- List all available documents in the current view - Retrieve all available field options with metadata (IMPORTANT: always call this before editing) - Understand which fields are stored where (layout vs data document) - Get detailed information about all available document fields @@ -137,8 +136,8 @@ SPECIAL FIELD HANDLING: - Width/Height: Set layout_autoHeight/layout_autoWidth to false before editing RECOMMENDED WORKFLOW: -1. First call action="list" to identify available documents -2. Then call action="getFieldOptions" to understand available fields +0. Understand the currently available documents that were provided as <available_documents> in the prompt +1. Call action="getFieldOptions" to understand available fields 3. Get document metadata with action="get" to see current values 4. Edit fields with action="edit" using proper dependencies OR @@ -159,10 +158,6 @@ HANDLING DEPENDENT FIELDS: - width → layout_autoWidth (set to false to allow manual width) - Other auto-sizing related properties -To LIST available documents: -- Use action="list" to get a simple list of all documents in the current view -- This is useful when you need to identify documents before getting details or editing them - Editing fields follows these rules: 1. First checks if the field exists on the layout document using Doc.Get 2. If it exists on the layout document, it's updated there @@ -172,7 +167,6 @@ Editing fields follows these rules: Examples: - To get field options: { action: "getFieldOptions" } -- To list all documents: { action: "list" } - To get all document metadata: { action: "get" } - To get metadata for a specific document: { action: "get", documentId: "doc123" } - To edit a single field: { action: "edit", documentId: "doc123", fieldEdits: [{ fieldName: "backgroundColor", fieldValue: "#ff0000" }] } @@ -186,7 +180,8 @@ Examples: { fieldName: "layout_autoHeight", fieldValue: false }, { fieldName: "height", fieldValue: 200 } ]} -- IMPORTANT: MULTI STEP WORKFLOWS ARE NOT ONLY ALLOWED BUT ENCOURAGED. TAKE THINGS 1 STEP AT A TIME.`; +- IMPORTANT: MULTI STEP WORKFLOWS ARE NOT ONLY ALLOWED BUT ENCOURAGED. TAKE THINGS 1 STEP AT A TIME. +- IMPORTANT: WHEN CITING A DOCUMENT, MAKE THE CHUNK ID THE DOCUMENT ID. WHENEVER YOU CITE A DOCUMENT, ALWAYS MAKE THE CITATION TYPE "text", THE "direct_text" FIELD BLANK, AND THE "chunk_id" FIELD THE DOCUMENT ID.`; const documentMetadataToolInfo: ToolInfo<DocumentMetadataToolParamsType> = { name: 'documentMetadata', description: toolDescription, @@ -232,11 +227,11 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp // Ensure the action is valid and convert to string const action = String(args.action); - if (!['get', 'edit', 'list', 'getFieldOptions', 'create'].includes(action)) { + if (!['get', 'edit', 'getFieldOptions', 'create'].includes(action)) { return [ { type: 'text', - text: 'Error: Invalid action. Valid actions are "get", "edit", "list", "getFieldOptions", or "create".', + text: 'Error: Invalid action. Valid actions are "get", "edit", "getFieldOptions", or "create".', }, ]; } @@ -386,10 +381,6 @@ export class DocumentMetadataTool extends BaseTool<DocumentMetadataToolParamsTyp } } - case 'list': { - this._docManager.listDocs(); - } - case 'getFieldOptions': { // Get all available field options with metadata const fieldOptions = this._docManager.getAllFieldMetadata(); @@ -457,7 +448,7 @@ ${JSON.stringify(createdMetadata, null, 2)}`, return [ { type: 'text', - text: 'Error: Unknown action. Valid actions are "get", "edit", "list", "getFieldOptions", or "create".', + text: 'Error: Unknown action. Valid actions are "get", "edit", "getFieldOptions", or "create".', }, ]; } @@ -537,11 +528,6 @@ ${JSON.stringify(createdMetadata, null, 2)}`, return true; } - // list action doesn't require any additional parameters - if (params.action === 'list') { - return true; - } - return true; } @@ -552,7 +538,7 @@ ${JSON.stringify(createdMetadata, null, 2)}`, */ private getParameterRequirementsByAction(action?: string): string { if (!action) { - return 'Please specify an action: "get", "edit", "list", "getFieldOptions", or "create".'; + return 'Please specify an action: "get", "edit", "getFieldOptions", or "create".'; } switch (action.toLowerCase()) { @@ -560,14 +546,12 @@ ${JSON.stringify(createdMetadata, null, 2)}`, return 'The "get" action accepts an optional documentId parameter.'; case 'edit': return 'The "edit" action requires documentId and fieldEdits parameters. fieldEdits must be a JSON array of field edits.'; - case 'list': - return 'The "list" action does not require any additional parameters.'; case 'getFieldOptions': return 'The "getFieldOptions" action does not require any additional parameters. It returns metadata about all available document fields.'; case 'create': return 'The "create" action requires title, data, and doc_type parameters.'; default: - return `Unknown action "${action}". Valid actions are "get", "edit", "list", "getFieldOptions", or "create".`; + return `Unknown action "${action}". Valid actions are "get", "edit", "getFieldOptions", or "create".`; } } } diff --git a/src/client/views/nodes/chatbot/tools/RAGTool.ts b/src/client/views/nodes/chatbot/tools/RAGTool.ts index ef374ed22..90b803d21 100644 --- a/src/client/views/nodes/chatbot/tools/RAGTool.ts +++ b/src/client/views/nodes/chatbot/tools/RAGTool.ts @@ -3,6 +3,7 @@ import { Observation, RAGChunk } from '../types/types'; import { ParametersType, ToolInfo } from '../types/tool_types'; import { Vectorstore } from '../vectorstore/Vectorstore'; import { BaseTool } from './BaseTool'; +import { DocumentMetadataTool } from './DocumentMetadataTool'; const ragToolParams = [ { @@ -17,7 +18,7 @@ type RAGToolParamsType = typeof ragToolParams; const ragToolInfo: ToolInfo<RAGToolParamsType> = { name: 'rag', - description: 'Performs a RAG (Retrieval-Augmented Generation) search on user documents and returns a set of document chunks (text or images) to provide a grounded response based on user documents.', + description: `Performs a RAG (Retrieval-Augmented Generation) search on user documents (only PDF, audio, and video are supported—for information about other document types, use the ${DocumentMetadataTool.name} tool) and returns a set of document chunks (text or images) to provide a grounded response based on user documents.`, citationRules: `When using the RAG tool, the structure must adhere to the format described in the ReAct prompt. Below are additional guidelines specifically for RAG-based responses: 1. **Grounded Text Guidelines**: @@ -75,7 +76,7 @@ export class RAGTool extends BaseTool<RAGToolParamsType> { async getFormattedChunks(relevantChunks: RAGChunk[]): Promise<Observation[]> { try { - const { formattedChunks } = await Networking.PostToServer('/formatChunks', { relevantChunks }) as { formattedChunks: Observation[]} + const { formattedChunks } = (await Networking.PostToServer('/formatChunks', { relevantChunks })) as { formattedChunks: Observation[] }; if (!formattedChunks) { throw new Error('Failed to format chunks'); diff --git a/src/client/views/nodes/chatbot/utils/AgentDocumentManager.ts b/src/client/views/nodes/chatbot/utils/AgentDocumentManager.ts index 14cffcb70..c8a6bb16b 100644 --- a/src/client/views/nodes/chatbot/utils/AgentDocumentManager.ts +++ b/src/client/views/nodes/chatbot/utils/AgentDocumentManager.ts @@ -1,4 +1,4 @@ -import { action, makeObservable, observable, ObservableMap, reaction, runInAction } from 'mobx'; +import { action, computed, makeObservable, observable, ObservableMap, reaction, runInAction } from 'mobx'; import { observer } from 'mobx-react'; import { v4 as uuidv4 } from 'uuid'; import { Doc, StrListCast } from '../../../../../fields/Doc'; @@ -31,7 +31,7 @@ export class AgentDocumentManager { private chatBox: ChatBox; private chatBoxDocument: Doc | null = null; private fieldMetadata: Record<string, any> = {}; - private readonly DOCUMENT_ID_FIELD = '_dash_document_id'; + @observable private documentIdsFromChunkIds: ObservableMap<string, string>; /** * Creates a new DocumentManager @@ -40,8 +40,17 @@ export class AgentDocumentManager { constructor(chatBox: ChatBox) { makeObservable(this); const agentDoc = DocCast(chatBox.Document.agentDocument) ?? new Doc(); + const chunkIds = DocCast(agentDoc.chunkIds) ?? new Doc(); + agentDoc.title = chatBox.Document.title + '_agentDocument'; + chunkIds.title = '_chunkIds'; chatBox.Document.agentDocument = agentDoc; + DocCast(chatBox.Document.agentDocument)!.chunkIds = chunkIds; + this.documentIdsFromChunkIds = StrListCast(chunkIds.mapping).reduce((mapping, content) => { + const [chunkId, docId] = content.split(':'); + mapping.set(chunkId, docId); + return mapping; + }, new ObservableMap<string, string>()); this.documentsById = StrListCast(agentDoc.mapping).reduce((mapping, content) => { const [id, layoutId, docId] = content.split(':'); const layoutDoc = DocServer.GetCachedRefField(layoutId); @@ -66,6 +75,19 @@ export class AgentDocumentManager { } //{ fireImmediately: true } ); + reaction( + () => this.documentIdsFromChunkIds.values(), + () => { + if (this.chatBoxDocument && DocCast(this.chatBoxDocument.agentDocument)) { + // Store the mapping with chunkId:docId format for consistency + const chunkIdsDoc = DocCast(DocCast(this.chatBoxDocument.agentDocument)!.chunkIds); + if (chunkIdsDoc) { + chunkIdsDoc.mapping = new List<string>(Array.from(this.documentIdsFromChunkIds.entries()).map(([chunkId, docId]) => `${chunkId}:${docId}`)); + } + } + } + //{ fireImmediately: true } + ); this.processDocument(this.chatBoxDocument); this.initializeFieldMetadata(); } @@ -120,7 +142,7 @@ export class AgentDocumentManager { try { // Use the LinkManager approach which is proven to work in ChatBox if (this.chatBoxDocument) { - console.log('Finding documents linked to ChatBox document with ID:', this.chatBoxDocument.id); + console.log('Finding documents linked to ChatBox document with ID:', this.chatBoxDocument[Id]); // Get directly linked documents via LinkManager const linkedDocs = LinkManager.Instance.getAllRelatedLinks(this.chatBoxDocument) @@ -134,57 +156,10 @@ export class AgentDocumentManager { linkedDocs.forEach((doc: Doc | undefined) => { if (doc) { this.processDocument(doc); - console.log('Processed linked document:', doc.id, doc.title, doc.type); + console.log('Processed linked document:', doc[Id], doc.title, doc.type); } }); - - // Include the ChatBox document itself - this.processDocument(this.chatBoxDocument); - - // If we have access to the Document's parent, try to find sibling documents - if (this.chatBoxDocument.parent) { - const parent = this.chatBoxDocument.parent; - console.log('Found parent document, checking for siblings'); - - // Check if parent is a Doc type and has a childDocs function - if (parent && typeof parent === 'object' && 'childDocs' in parent && typeof parent.childDocs === 'function') { - try { - const siblingDocs = parent.childDocs(); - if (Array.isArray(siblingDocs)) { - console.log(`Found ${siblingDocs.length} sibling documents via parent.childDocs()`); - siblingDocs.forEach((doc: Doc) => { - if (doc) { - this.processDocument(doc); - } - }); - } - } catch (e) { - console.warn('Error accessing parent.childDocs:', e); - } - } - } - } else if (this.chatBox && this.chatBox.linkedDocs) { - // If we have direct access to the linkedDocs computed property from ChatBox - console.log('Using ChatBox.linkedDocs directly'); - const linkedDocs = this.chatBox.linkedDocs; - if (Array.isArray(linkedDocs)) { - console.log(`Found ${linkedDocs.length} documents via ChatBox.linkedDocs`); - linkedDocs.forEach((doc: Doc) => { - if (doc) { - this.processDocument(doc); - } - }); - } - - // Process the ChatBox document if available - if (this.chatBox.Document) { - this.processDocument(this.chatBox.Document); - } - } else { - console.warn('No ChatBox document reference available for finding linked documents'); } - - console.log(`DocumentMetadataTool found ${this.documentsById.size} total documents`); } catch (error) { console.error('Error finding documents in Freeform view:', error); } @@ -201,6 +176,7 @@ export class AgentDocumentManager { // Only add if we haven't already processed this document if (!this.documentsById.has(docId)) { this.documentsById.set(docId, { layoutDoc: doc, dataDoc: doc[DocData] }); + console.log('Added document to documentsById:', doc[Id], docId, doc[Id], doc[DocData][Id]); } return docId; } @@ -213,37 +189,12 @@ export class AgentDocumentManager { private ensureDocumentId(doc: Doc): string { let docId: string | undefined; - // First try to get the ID from our custom field - if (doc[this.DOCUMENT_ID_FIELD]) { - docId = String(doc[this.DOCUMENT_ID_FIELD]); - return docId; - } - - // Try different ways to get a document ID - // 1. Try the direct id property if it exists - if (doc.id && typeof doc.id === 'string') { - docId = doc.id; - } - // 2. Try doc._id if it exists - else if (doc._id && typeof doc._id === 'string') { - docId = doc._id; - } - // 3. Try doc.data?.id if it exists - else if (doc.data && typeof doc.data === 'object' && 'id' in doc.data && typeof doc.data.id === 'string') { - docId = doc.data.id; - } - // 4. If none of the above work, generate a UUID - else { - docId = uuidv4(); - console.log(`Generated new UUID for document with title: ${doc.title || 'Untitled'}`); - } - - // Store the ID in the document's metadata so it persists - try { - doc[this.DOCUMENT_ID_FIELD] = docId; - } catch (e) { - console.warn(`Could not assign ID to document property`, e); + if (doc[Id]) { + console.log('Found document ID (normal):', doc[Id]); + docId = doc[Id]; + } else { + throw new Error('No document ID found'); } return docId; @@ -256,13 +207,13 @@ export class AgentDocumentManager { */ public extractDocumentMetadata(id: string) { if (!id) return null; - const doc = this.documentsById.get(id); - if (!doc) return null; - const layoutDoc = doc.layoutDoc; - const dataDoc = doc.dataDoc; + const agentDoc = this.documentsById.get(id); + if (!agentDoc) return null; + const layoutDoc = agentDoc.layoutDoc; + const dataDoc = agentDoc.dataDoc; const metadata: Record<string, any> = { - id: layoutDoc.dash_document_id || layoutDoc.id || '', + id: layoutDoc[Id] || dataDoc[Id] || '', title: layoutDoc.title || '', type: layoutDoc.type || '', fields: { @@ -355,7 +306,7 @@ export class AgentDocumentManager { if (value instanceof Doc) { return { type: 'Doc', - id: value.id || this.ensureDocumentId(value), + id: value[Id] || this.ensureDocumentId(value), title: value.title || '', docType: value.type || '', }; @@ -1011,33 +962,17 @@ export class AgentDocumentManager { * Returns a list of all document IDs in the manager. * @returns An array of document IDs (strings). */ - public listDocs(): string[] { - return Array.from(this.documentsById.keys()); + @computed + public get listDocs(): string[] { + console.log( + Array.from(this.documentsById.entries()).map(([id, agentDoc]) => JSON.stringify({ id, title: agentDoc.layoutDoc.title, type: agentDoc.layoutDoc.type, summary: agentDoc.layoutDoc.summary || 'No summary available for this document.' })) + ); + return Array.from(this.documentsById.entries()).map(([id, agentDoc]) => JSON.stringify({ id, title: agentDoc.layoutDoc.title, type: agentDoc.layoutDoc.type, summary: agentDoc.layoutDoc.summary || 'No summary available for this document.' })); } - /** - * Adds a document with a custom ID to the manager - * @param doc The document to add - * @param customId The custom ID to assign to the document - * @returns The customId that was assigned - */ - @action - public addCustomId(doc: Doc, customId: string): string { - if (!doc) { - console.error('Cannot add null document with custom ID'); - return ''; - } - - // Set the custom ID in the document's metadata - doc[this.DOCUMENT_ID_FIELD] = customId; - - // Store the document in our map - this.documentsById.set(customId, { - layoutDoc: doc, - dataDoc: doc, - }); - - return customId; + @computed + public get docIds(): string[] { + return Array.from(this.documentsById.keys()); } /** @@ -1078,11 +1013,8 @@ export class AgentDocumentManager { // Ensure each chunk ID can be linked back to its parent document // Store a mapping from chunk ID to parent document ID // This allows us to easily find a document by any of its chunk IDs - if (!this.documentsById.has(chunkId)) { - this.documentsById.set(chunkId, { - layoutDoc: doc, - dataDoc: docInfo.dataDoc, - }); + if (!this.documentIdsFromChunkIds.has(chunkId) && doc) { + this.documentIdsFromChunkIds.set(chunkId, doc[Id]); } } } @@ -1092,11 +1024,25 @@ export class AgentDocumentManager { * @param chunkId The chunk ID to look up * @returns The parent document ID if found */ - public getDocIdByChunkId(chunkId: string): string | undefined { - const docInfo = this.documentsById.get(chunkId); + public getDocByChunkId(chunkId: string): Doc | undefined { + // First, look up the document ID using the chunk ID mapping + const docId = this.documentIdsFromChunkIds.get(chunkId); + console.log('this.documentIdsFromChunkIds', this.documentIdsFromChunkIds); + console.log('docId', docId); + if (!docId) { + if (this.documentsById.has(chunkId)) { + return this.documentsById.get(chunkId)?.layoutDoc; + } else { + console.error('No document found for chunkId and docId', chunkId); + return undefined; + } + } + // Then get the document using the document ID + const docInfo = this.documentsById.get(docId); if (docInfo) { - return docInfo.layoutDoc[this.DOCUMENT_ID_FIELD] as string; + return docInfo.layoutDoc; } + console.error('No document found for docId', docId); return undefined; } @@ -1157,7 +1103,7 @@ export class AgentDocumentManager { return baseChunk; } }); - + console.log('simplifiedChunks', simplifiedChunks); // Update the document with all simplified chunks at once doc.chunk_simpl = JSON.stringify({ chunks: simplifiedChunks }); @@ -1165,32 +1111,25 @@ export class AgentDocumentManager { } /** - * Gets the simplified chunks from a document - * @param doc The document to get simplified chunks from - * @returns Array of simplified chunks or empty array if none exist + * Gets a specific simplified chunk by ID + * @param doc The document containing chunks + * @param chunkId The ID of the chunk to retrieve + * @returns The simplified chunk if found, undefined otherwise */ - public getSimplifiedChunks(doc: Doc): any[] { + public getSimplifiedChunkById(doc: Doc, chunkId: string): any | undefined { + let chunks: any[] = []; if (!doc || !doc.chunk_simpl) { + chunks = []; + console.warn('No chunk found for chunkId', chunkId, '. Checking if document exists in documentsById.'); return []; } - try { const parsed = JSON.parse(StrCast(doc.chunk_simpl)); - return parsed.chunks || []; + chunks = parsed.chunks || []; } catch (e) { console.error('Error parsing simplified chunks:', e); return []; } - } - - /** - * Gets a specific simplified chunk by ID - * @param doc The document containing chunks - * @param chunkId The ID of the chunk to retrieve - * @returns The simplified chunk if found, undefined otherwise - */ - public getSimplifiedChunkById(doc: Doc, chunkId: string): any | undefined { - const chunks = this.getSimplifiedChunks(doc); return chunks.find(chunk => chunk.chunkId === chunkId); } diff --git a/src/client/views/nodes/chatbot/vectorstore/Vectorstore.ts b/src/client/views/nodes/chatbot/vectorstore/Vectorstore.ts index 3df1294e9..1349df483 100644 --- a/src/client/views/nodes/chatbot/vectorstore/Vectorstore.ts +++ b/src/client/views/nodes/chatbot/vectorstore/Vectorstore.ts @@ -16,6 +16,7 @@ import { AI_Document, CHUNK_TYPE, RAGChunk } from '../types/types'; import OpenAI from 'openai'; import { Embedding } from 'openai/resources'; import { AgentDocumentManager } from '../utils/AgentDocumentManager'; +import { Id } from '../../../../../fields/FieldSymbols'; dotenv.config(); @@ -24,13 +25,12 @@ dotenv.config(); * and OpenAI text-embedding-3-large for text embedding. It handles AI document management, uploads, and query-based retrieval. */ export class Vectorstore { - private pinecone: Pinecone; // Pinecone client for managing the vector index. + private pinecone!: Pinecone; // Pinecone client for managing the vector index. private index!: Index; // The specific Pinecone index used for document chunks. - private openai: OpenAI; // OpenAI client for generating embeddings. + private openai!: OpenAI; // OpenAI client for generating embeddings. private indexName: string = 'pdf-chatbot'; // Default name for the index. - private _id: string; // Unique ID for the Vectorstore instance. - private docManager: AgentDocumentManager; // Document manager for handling documents - + private _id!: string; // Unique ID for the Vectorstore instance. + private docManager!: AgentDocumentManager; // Document manager for handling documents documents: AI_Document[] = []; // Store the documents indexed in the vectorstore. /** @@ -143,10 +143,8 @@ export class Vectorstore { progressCallback(85, 'Embeddings generated. Finalizing document...'); doc.original_segments = JSON.stringify(typedResponse.full); - const doc_id = uuidv4(); - - // Register the document with the AgentDocumentManager - this.docManager.addCustomId(doc, doc_id); + const doc_id = doc[Id]; + console.log('doc_id in vectorstore', doc_id); // Generate chunk IDs upfront so we can register them const chunkIds = segmentedTranscript.map(() => uuidv4()); @@ -191,7 +189,7 @@ export class Vectorstore { } else { // Process regular document console.log('Processing regular document...'); - const createDocumentResponse = await Networking.PostToServer('/createDocument', { file_path: local_file_path }); + const createDocumentResponse = await Networking.PostToServer('/createDocument', { file_path: local_file_path, doc_id: doc[Id] }); // Type assertion for the response const { jobId } = createDocumentResponse as { jobId: string }; @@ -211,12 +209,13 @@ export class Vectorstore { } } - // Register the document with the AgentDocumentManager - this.docManager.addCustomId(doc, result.doc_id); - // Collect all chunk IDs const chunkIds = result.chunks.map(chunk => chunk.id); + if (result.doc_id !== doc[Id]) { + console.log('doc_id in vectorstore', result.doc_id, 'does not match doc_id in doc', doc[Id]); + } + // Register chunks with the document manager this.docManager.registerChunkIds(result.doc_id, chunkIds); @@ -319,16 +318,14 @@ export class Vectorstore { const queryEmbedding = queryEmbeddingResponse.data[0].embedding; - // Get document IDs from the AgentDocumentManager - const docIds = Array.from(this.docManager.listDocs()); - console.log('Using document IDs for retrieval:', docIds); + console.log('Using document IDs for retrieval:', this.docManager.docIds); // Query the Pinecone index using the embedding and filter by document IDs. // We'll query based on document IDs that are registered in the document manager const queryResponse: QueryResponse = await this.index.query({ vector: queryEmbedding, filter: { - doc_id: { $in: docIds }, + doc_id: { $in: this.docManager.docIds }, }, topK, includeValues: true, @@ -356,7 +353,7 @@ export class Vectorstore { // Ensure the document manager knows about this chunk // This is important for maintaining backwards compatibility - if (chunk.id && !this.docManager.getDocIdByChunkId(chunk.id)) { + if (chunk.id && !this.docManager.getDocByChunkId(chunk.id)) { // If the chunk ID isn't registered but we have a doc_id in metadata if (chunk.metadata.doc_id && this.docManager.has(chunk.metadata.doc_id)) { // Register the chunk with its parent document diff --git a/src/server/ApiManagers/AssistantManager.ts b/src/server/ApiManagers/AssistantManager.ts index 6d2779163..378f14094 100644 --- a/src/server/ApiManagers/AssistantManager.ts +++ b/src/server/ApiManagers/AssistantManager.ts @@ -559,7 +559,7 @@ export default class AssistantManager extends ApiManager { method: Method.POST, subscription: '/createDocument', secureHandler: async ({ req, res }) => { - const { file_path } = req.body; + const { file_path, doc_id } = req.body; const public_path = path.join(publicDirectory, file_path); // Resolve the file path in the public directory const file_name = path.basename(file_path); // Extract the file name from the path @@ -572,7 +572,7 @@ export default class AssistantManager extends ApiManager { // Spawn the Python process and track its progress/output // eslint-disable-next-line no-use-before-define - spawnPythonProcess(jobId, public_path); + spawnPythonProcess(jobId, public_path, doc_id); // Send the job ID back to the client for tracking res.send({ jobId }); @@ -850,7 +850,7 @@ export default class AssistantManager extends ApiManager { * @param file_name The name of the file to process. * @param file_path The filepath of the file to process. */ -function spawnPythonProcess(jobId: string, file_path: string) { +function spawnPythonProcess(jobId: string, file_path: string, doc_id: string) { const venvPath = path.join(__dirname, '../chunker/venv'); const requirementsPath = path.join(__dirname, '../chunker/requirements.txt'); const pythonScriptPath = path.join(__dirname, '../chunker/pdf_chunker.py'); @@ -860,7 +860,7 @@ function spawnPythonProcess(jobId: string, file_path: string) { function runPythonScript() { const pythonPath = process.platform === 'win32' ? path.join(venvPath, 'Scripts', 'python') : path.join(venvPath, 'bin', 'python3'); - const pythonProcess = spawn(pythonPath, [pythonScriptPath, jobId, file_path, outputDirectory]); + const pythonProcess = spawn(pythonPath, [pythonScriptPath, jobId, file_path, outputDirectory, doc_id]); let pythonOutput = ''; let stderrOutput = ''; diff --git a/src/server/chunker/pdf_chunker.py b/src/server/chunker/pdf_chunker.py index e9b9ef2b3..e34753176 100644 --- a/src/server/chunker/pdf_chunker.py +++ b/src/server/chunker/pdf_chunker.py @@ -622,7 +622,7 @@ class Document: Represents a document being processed, such as a PDF, handling chunking, embedding, and summarization. """ - def __init__(self, file_path: str, file_name: str, job_id: str, output_folder: str): + def __init__(self, file_path: str, file_name: str, job_id: str, output_folder: str, doc_id: str): """ Initialize the Document with file data, file name, and job ID. @@ -635,7 +635,7 @@ class Document: self.file_path = file_path self.job_id = job_id self.type = self._get_document_type(file_name) # Determine the document type (PDF, CSV, etc.) - self.doc_id = job_id # Use the job ID as the document ID + self.doc_id = doc_id # Use the job ID as the document ID self.chunks = [] # List to hold text and visual chunks self.num_pages = 0 # Number of pages in the document (if applicable) self.summary = "" # The generated summary for the document @@ -755,7 +755,7 @@ class Document: "doc_id": self.doc_id }, indent=2) # Convert the document's attributes to JSON format -def process_document(file_path, job_id, output_folder): +def process_document(file_path, job_id, output_folder, doc_id): """ Top-level function to process a document and return the JSON output. @@ -763,26 +763,27 @@ def process_document(file_path, job_id, output_folder): :param job_id: The job ID for this document processing task. :return: The processed document's data in JSON format. """ - new_document = Document(file_path, file_path, job_id, output_folder) + new_document = Document(file_path, file_path, job_id, output_folder, doc_id) return new_document.to_json() def main(): """ Main entry point for the script, called with arguments from Node.js. """ - if len(sys.argv) != 4: + if len(sys.argv) != 5: print(json.dumps({"error": "Invalid arguments"}), file=sys.stderr) return job_id = sys.argv[1] file_path = sys.argv[2] output_folder = sys.argv[3] # Get the output folder from arguments + doc_id = sys.argv[4] try: os.makedirs(output_folder, exist_ok=True) # Process the document - document_result = process_document(file_path, job_id, output_folder) # Pass output_folder + document_result = process_document(file_path, job_id, output_folder,doc_id) # Pass output_folder # Output the final result as JSON to stdout print(document_result) -- cgit v1.2.3-70-g09d2