completely restructured, added comments, and significantly reduced the length of the prompt (~72% shorter and cheaper)

1 files changed, 138 insertions, 0 deletions

diff --git a/src/client/views/nodes/chatbot/tools/RAGTool.ts b/src/client/views/nodes/chatbot/tools/RAGTool.ts
new file mode 100644
index 000000000..c24306dcd
--- /dev/null
+++ b/src/client/views/nodes/chatbot/tools/RAGTool.ts
@@ -0,0 +1,138 @@
+import { BaseTool } from './BaseTool';
+import { Vectorstore } from '../vectorstore/Vectorstore';
+import { RAGChunk } from '../types/types';
+import * as fs from 'fs';
+import { Networking } from '../../../../Network';
+import { file } from 'jszip';
+import { ChatCompletion, ChatCompletionContentPart, ChatCompletionMessageParam } from 'openai/resources';
+
+export class RAGTool extends BaseTool {
+    constructor(private vectorstore: Vectorstore) {
+        super(
+            'rag',
+            'Perform a RAG search on user documents',
+            {
+                hypothetical_document_chunk: {
+                    type: 'string',
+                    description:
+                        "Detailed version of the prompt that is effectively a hypothetical document chunk that would be ideal to embed and compare to the vectors of real document chunks to fetch the most relevant document chunks to answer the user's query",
+                    required: 'true',
+                },
+            },
+            `
+            Your task is to provide a comprehensive response to the user's prompt based on the given chunks and chat history. Follow these structural guidelines meticulously:
+
+            1. Overall Structure:
+            <answer>
+                [Main content with grounded_text tags interspersed with normal plain text (information that is not derived from chunks' information)]
+                <citations>
+                [Individual citation tags]
+                </citations>
+                <follow_up_questions>
+                [Three question tags]
+                </follow_up_questions>
+            </answer>
+
+            2. Grounded Text Tag Structure:
+            - Basic format: 
+                <grounded_text citation_index="[citation index number(s)]">
+                [Your generated text based on information from a subset of a chunk (a citation's direct text)]
+                </grounded_text>
+
+            3. Citation Tag Structure:
+            <citation index="[unique number]" chunk_id="[UUID v4]" type="[text/image/table]">
+                [For text: relevant subset of original chunk]
+                [For image/table: leave empty]
+            </citation>
+
+            4. Detailed Grounded Text Guidelines:
+                a. Wrap all information derived from chunks in grounded_text tags.
+                b. DO NOT PUT ANYTHING THAT IS NOT DIRECTLY DERIVED FROM INFORMATION FROM CHUNKS (EITHER IMAGE, TABLE, OR TEXT) IN GROUNDED_TEXT TAGS.
+                c. Use a single grounded_text tag for suquential and closely related information that references the same citation. If other citations' information are used sequentially, create new grounded_text tags.
+                d. Ensure every grounded_text tag has up to a few corresponding citations (should not be more than 3 and only 1 is fine). Multiple citation indices should be separated by commas.
+                e. Grounded text can be as short as a few words or as long as several sentences.
+                f. Avoid overlapping or nesting grounded_text tags; instead, use sequential tags.
+
+            5. Detailed Citation Guidelines:
+                a. Create a unique citation for each distinct piece of information from the chunks that is used to support grounded_text.
+                b. ALL TEXT CITATIONS must have direct text in its element content (e.g. <citation ...>DIRECT TEXT HERE</citation>) that is a relevant SUBSET of the original text chunk that is being cited specifically.
+                c. DO NOT paraphrase or summarize the text; use the original text as much as possible.
+                d. DO NOT USE THE FULL TEXT CHUNK as the citation content; only use the relevant subset of the text that the grounded_text is base. AS SHORT AS POSSIBLE WHILE PROVIDING INFORMATION (ONE TO TWO SENTENCES USUALLY)!
+                e. Ensure each citation has a unique index number.
+                f. Specify the correct type: "text", "image", or "table".
+                g. For text chunks, the content of the citation should ALWAYS have the relevant subset of the original text that the grounded_text is based on.
+                h. For image/table chunks, leave the citation content empty.
+                i. One citation can be used for multiple grounded_text tags if they are based on the same chunk information.
+                j. !!!DO NOT OVERCITE - only include citations for information that is directly relevant to the grounded_text.
+
+            6. Structural Integrity Checks:
+                a. Ensure all opening tags have corresponding closing tags.
+                b. Verify that all grounded_text tags have valid citation_index attributes (they should be equal to the associated citation(s) index field—not their chunk_id field).
+                c. Check that all cited indices in grounded_text tags have corresponding citations.
+
+            Example of grounded_text usage:
+
+            <answer>
+                <grounded_text citation_index="1,2">
+                Artificial Intelligence (AI) is revolutionizing various sectors, with healthcare experiencing significant transformations in areas such as diagnosis and treatment planning.
+                </grounded_text>
+                <grounded_text citation_index="2,3,4">
+                In the field of medical diagnosis, AI has shown remarkable capabilities, particularly in radiology. For instance, AI systems have drastically improved mammogram analysis, achieving 99% accuracy at a rate 30 times faster than human radiologists.
+                </grounded_text>
+                <grounded_text citation_index="4">
+                This advancement not only enhances the efficiency of healthcare systems but also significantly reduces the occurrence of false positives, leading to fewer unnecessary biopsies and reduced patient stress.
+                </grounded_text>
+            
+                <grounded_text citation_index="5,6">
+                Beyond diagnosis, AI is playing a crucial role in drug discovery and development. By analyzing vast amounts of genetic and molecular data, AI algorithms can identify potential drug candidates much faster than traditional methods.
+                </grounded_text>
+                <grounded_text citation_index="6">
+                    This could potentially reduce the time and cost of bringing new medications to market, especially for rare diseases that have historically received less attention due to limited market potential.
+                </grounded_text>
+
+                [... rest of the content ...]
+
+                <citations>
+                    <citation index="1" chunk_id="123e4567-e89b-12d3-a456-426614174000" type="text">Artificial Intelligence is revolutionizing various industries, with healthcare being one of the most profoundly affected sectors.</citation>
+                    <citation index="2" chunk_id="123e4567-e89b-12d3-a456-426614174001" type="text">AI has shown particular promise in the field of radiology, enhancing the accuracy and speed of image analysis.</citation>
+                    <citation index="3" chunk_id="123e4567-e89b-12d3-a456-426614174002" type="text">According to recent studies, AI systems have achieved 99% accuracy in mammogram analysis, performing the task 30 times faster than human radiologists.</citation>
+                    <citation index="4" chunk_id="123e4567-e89b-12d3-a456-426614174003" type="text">The improvement in mammogram accuracy has led to a significant reduction in false positives, decreasing the need for unnecessary biopsies and reducing patient anxiety.</citation>
+                    <citation index="5" chunk_id="123e4567-e89b-12d3-a456-426614174004" type="text">AI is accelerating the drug discovery process by analyzing complex molecular and genetic data to identify potential drug candidates.</citation>
+                    <citation index="6" chunk_id="123e4567-e89b-12d3-a456-426614174005" type="text">The use of AI in drug discovery could significantly reduce the time and cost associated with bringing new medications to market, particularly for rare diseases.</citation>
+                </citations>
+
+                <follow_up_questions>
+                    <question>How might AI-driven personalized medicine impact the cost and accessibility of healthcare in the future?</question>
+                    <question>What measures can be taken to ensure that AI systems in healthcare are free from biases and equally effective for diverse populations?</question>
+                    <question>How could the role of healthcare professionals evolve as AI becomes more integrated into medical practices?</question>
+                </follow_up_questions>
+            </answer>
+            `,
+
+            `Performs a RAG (Retrieval-Augmented Generation) search on user documents and returns a 
+            set of document chunks (either images or text) that can be used to provide a grounded response based on 
+            user documents`
+        );
+    }
+
+    async execute(args: { hypothetical_document_chunk: string }): Promise {
+        const relevantChunks = await this.vectorstore.retrieve(args.hypothetical_document_chunk);
+        const formatted_chunks = await this.getFormattedChunks(relevantChunks);
+        return formatted_chunks;
+    }
+
+    async getFormattedChunks(relevantChunks: RAGChunk[]): Promise {
+        try {
+            const { formattedChunks } = await Networking.PostToServer('/formatChunks', { relevantChunks });
+
+            if (!formattedChunks) {
+                throw new Error('Failed to format chunks');
+            }
+
+            return formattedChunks;
+        } catch (error) {
+            console.error('Error formatting chunks:', error);
+            throw error;
+        }
+    }
+}