Merge branch 'master' of https://github.com/brown-dash/Dash-Web into Merge

1 files changed, 339 insertions, 0 deletions

diff --git a/src/client/views/nodes/chatbot/vectorstore/Vectorstore.ts b/src/client/views/nodes/chatbot/vectorstore/Vectorstore.ts
new file mode 100644
index 000000000..afd34f28d
--- /dev/null
+++ b/src/client/views/nodes/chatbot/vectorstore/Vectorstore.ts
@@ -0,0 +1,339 @@
+/**
+ * @file Vectorstore.ts
+ * @description This file defines the Vectorstore class, which integrates with Pinecone for vector-based document indexing and OpenAI text-embedding-3-large for text embeddings.
+ * It manages AI document handling, including adding documents, processing media files, combining document chunks, indexing documents,
+ * and retrieving relevant sections based on user queries.
+ */
+
+import { Index, IndexList, Pinecone, PineconeRecord, QueryResponse, RecordMetadata } from '@pinecone-database/pinecone';
+import dotenv from 'dotenv';
+import path from 'path';
+import { v4 as uuidv4 } from 'uuid';
+import { Doc } from '../../../../../fields/Doc';
+import { AudioCast, CsvCast, PDFCast, StrCast, VideoCast } from '../../../../../fields/Types';
+import { Networking } from '../../../../Network';
+import { AI_Document, CHUNK_TYPE, RAGChunk } from '../types/types';
+import OpenAI from 'openai';
+import { Embedding } from 'openai/resources';
+import { PineconeEnvironmentVarsNotSupportedError } from '@pinecone-database/pinecone/dist/errors';
+
+dotenv.config();
+
+/**
+ * The Vectorstore class integrates with Pinecone for vector-based document indexing and retrieval,
+ * 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 index!: Index; // The specific Pinecone index used for document chunks.
+    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 _doc_ids: () => string[]; // List of document IDs handled by this instance.
+
+    documents: AI_Document[] = []; // Store the documents indexed in the vectorstore.
+
+    /**
+     * Initializes the Pinecone and OpenAI clients, sets up the document ID list,
+     * and initializes the Pinecone index.
+     * @param id The unique identifier for the vectorstore instance.
+     * @param doc_ids A function that returns a list of document IDs.
+     */
+    constructor(id: string, doc_ids: () => string[]) {
+        const pineconeApiKey = process.env.PINECONE_API_KEY;
+        if (!pineconeApiKey) {
+            throw new Error('PINECONE_API_KEY is not defined.');
+        }
+
+        // Initialize Pinecone and OpenAI clients with API keys from the environment.
+        this.pinecone = new Pinecone({ apiKey: pineconeApiKey });
+        this.openai = new OpenAI({ apiKey: process.env.OPENAI_API_KEY, dangerouslyAllowBrowser: true });
+        this._id = id;
+        this._doc_ids = doc_ids;
+        this.initializeIndex();
+    }
+
+    /**
+     * Initializes the Pinecone index by checking if it exists and creating it if necessary.
+     * Sets the index to use cosine similarity for vector similarity calculations.
+     */
+    private async initializeIndex() {
+        const indexList: IndexList = await this.pinecone.listIndexes();
+
+        // Check if the index already exists, otherwise create it.
+        if (!indexList.indexes?.some(index => index.name === this.indexName)) {
+            await this.pinecone.createIndex({
+                name: this.indexName,
+                dimension: 3072,
+                metric: 'cosine',
+                spec: {
+                    serverless: {
+                        cloud: 'aws',
+                        region: 'us-east-1',
+                    },
+                },
+            });
+        }
+
+        // Set the index for future use.
+        this.index = this.pinecone.Index(this.indexName);
+    }
+
+    /**
+     * Adds an AI document to the vectorstore. Handles media file processing for audio/video,
+     * and text embedding for all document types. Updates document metadata during processing.
+     * @param doc The document to add.
+     * @param progressCallback Callback to track the progress of the addition process.
+     */
+    async addAIDoc(doc: Doc, progressCallback: (progress: number, step: string) => void) {
+        const ai_document_status: string = StrCast(doc.ai_document_status);
+
+        // Skip if the document is already in progress or completed.
+        if (ai_document_status !== undefined && ai_document_status.trim() !== '' && ai_document_status !== '{}') {
+            if (ai_document_status === 'PROGRESS') {
+                console.log('Already in progress.');
+                return;
+            } else if (ai_document_status === 'COMPLETED') {
+                console.log('Already completed.');
+                return;
+            }
+        } else {
+            // Start processing the document.
+            doc.ai_document_status = 'PROGRESS';
+            const local_file_path: string = CsvCast(doc.data)?.url?.pathname ?? PDFCast(doc.data)?.url?.pathname ?? VideoCast(doc.data)?.url?.pathname ?? AudioCast(doc.data)?.url?.pathname;
+
+            if (!local_file_path) {
+                console.log('Invalid file path.');
+                return;
+            }
+
+            const isAudioOrVideo = local_file_path.endsWith('.mp3') || local_file_path.endsWith('.mp4');
+            let result: AI_Document & { doc_id: string };
+            if (isAudioOrVideo) {
+                console.log('Processing media file...');
+                const response = await Networking.PostToServer('/processMediaFile', { fileName: path.basename(local_file_path) });
+                const segmentedTranscript = response.condensed;
+                console.log(segmentedTranscript);
+                const summary = response.summary;
+                doc.summary = summary;
+                // Generate embeddings for each chunk
+                const texts = segmentedTranscript.map((chunk: any) => chunk.text);
+
+                try {
+                    const embeddingsResponse = await this.openai.embeddings.create({
+                        model: 'text-embedding-3-large',
+                        input: texts,
+                        encoding_format: 'float',
+                    });
+
+                    doc.original_segments = JSON.stringify(response.full);
+                    doc.ai_type = local_file_path.endsWith('.mp3') ? 'audio' : 'video';
+                    const doc_id = uuidv4();
+
+                    // Add transcript and embeddings to metadata
+                    result = {
+                        doc_id,
+                        purpose: '',
+                        file_name: local_file_path,
+                        num_pages: 0,
+                        summary: '',
+                        chunks: segmentedTranscript.map((chunk: any, index: number) => ({
+                            id: uuidv4(),
+                            values: (embeddingsResponse.data as Embedding[])[index].embedding, // Assign embedding
+                            metadata: {
+                                indexes: chunk.indexes,
+                                original_document: local_file_path,
+                                doc_id: doc_id,
+                                file_path: local_file_path,
+                                start_time: chunk.start,
+                                end_time: chunk.end,
+                                text: chunk.text,
+                                type: CHUNK_TYPE.VIDEO,
+                            },
+                        })),
+                        type: 'media',
+                    };
+                } catch (error) {
+                    console.error('Error generating embeddings:', error);
+                    throw new Error('Embedding generation failed');
+                }
+
+                doc.segmented_transcript = JSON.stringify(segmentedTranscript);
+                // Simplify chunks for storage
+                const simplifiedChunks = result.chunks.map(chunk => ({
+                    chunkId: chunk.id,
+                    start_time: chunk.metadata.start_time,
+                    end_time: chunk.metadata.end_time,
+                    indexes: chunk.metadata.indexes,
+                    chunkType: CHUNK_TYPE.VIDEO,
+                    text: chunk.metadata.text,
+                }));
+                doc.chunk_simpl = JSON.stringify({ chunks: simplifiedChunks });
+            } else {
+                // Existing document processing logic remains unchanged
+                console.log('Processing regular document...');
+                const { jobId } = await Networking.PostToServer('/createDocument', { file_path: local_file_path });
+
+                while (true) {
+                    await new Promise(resolve => setTimeout(resolve, 2000));
+                    const resultResponse = await Networking.FetchFromServer(`/getResult/${jobId}`);
+                    const resultResponseJson = JSON.parse(resultResponse);
+                    if (resultResponseJson.status === 'completed') {
+                        result = resultResponseJson;
+                        break;
+                    }
+                    const progressResponse = await Networking.FetchFromServer(`/getProgress/${jobId}`);
+                    const progressResponseJson = JSON.parse(progressResponse);
+                    if (progressResponseJson) {
+                        progressCallback(progressResponseJson.progress, progressResponseJson.step);
+                    }
+                }
+                if (!doc.chunk_simpl) {
+                    doc.chunk_simpl = JSON.stringify({ chunks: [] });
+                }
+                doc.summary = result.summary;
+                doc.ai_purpose = result.purpose;
+
+                result.chunks.forEach((chunk: RAGChunk) => {
+                    const chunkToAdd = {
+                        chunkId: chunk.id,
+                        startPage: chunk.metadata.start_page,
+                        endPage: chunk.metadata.end_page,
+                        location: chunk.metadata.location,
+                        chunkType: chunk.metadata.type as CHUNK_TYPE,
+                        text: chunk.metadata.text,
+                    };
+                    const new_chunk_simpl = JSON.parse(StrCast(doc.chunk_simpl));
+                    new_chunk_simpl.chunks = new_chunk_simpl.chunks.concat(chunkToAdd);
+                    doc.chunk_simpl = JSON.stringify(new_chunk_simpl);
+                });
+            }
+
+            // Index the document
+            await this.indexDocument(result);
+
+            // Preserve existing metadata updates
+            if (!doc.vectorstore_id) {
+                doc.vectorstore_id = JSON.stringify([this._id]);
+            } else {
+                doc.vectorstore_id = JSON.stringify(JSON.parse(StrCast(doc.vectorstore_id)).concat([this._id]));
+            }
+
+            doc.ai_doc_id = result.doc_id;
+
+            console.log(`Document added: ${result.file_name}`);
+            doc.ai_document_status = 'COMPLETED';
+        }
+    }
+
+    /**
+     * Uploads the document's vector chunks to the Pinecone index.
+     * Prepares the metadata for each chunk and uses Pinecone's upsert operation.
+     * @param document The processed document containing its chunks and metadata.
+     */
+    private async indexDocument(document: AI_Document) {
+        console.log('Uploading vectors to content namespace...');
+
+        // Prepare Pinecone records for each chunk in the document.
+        const pineconeRecords: PineconeRecord[] = (document.chunks as RAGChunk[]).map(chunk => ({
+            id: chunk.id,
+            values: chunk.values,
+            metadata: { ...chunk.metadata } as RecordMetadata,
+        }));
+
+        // Upload the records to Pinecone.
+        await this.index.upsert(pineconeRecords);
+    }
+
+    /**
+     * Combines document chunks until their combined text reaches a minimum word count.
+     * This is used to optimize retrieval and indexing processes.
+     * @param chunks The original chunks to combine.
+     * @returns Combined chunks with updated text and metadata.
+     */
+    private combineChunks(chunks: RAGChunk[]): RAGChunk[] {
+        const combinedChunks: RAGChunk[] = [];
+        let currentChunk: RAGChunk | null = null;
+        let wordCount = 0;
+
+        chunks.forEach(chunk => {
+            const textWords = chunk.metadata.text.split(' ').length;
+
+            if (!currentChunk) {
+                currentChunk = { ...chunk, metadata: { ...chunk.metadata, text: chunk.metadata.text } };
+                wordCount = textWords;
+            } else if (wordCount + textWords >= 500) {
+                combinedChunks.push(currentChunk);
+                currentChunk = { ...chunk, metadata: { ...chunk.metadata, text: chunk.metadata.text } };
+                wordCount = textWords;
+            } else {
+                currentChunk.metadata.text += ` ${chunk.metadata.text}`;
+                wordCount += textWords;
+            }
+        });
+
+        if (currentChunk) {
+            combinedChunks.push(currentChunk);
+        }
+
+        return combinedChunks;
+    }
+
+    /**
+     * Retrieves the most relevant document chunks for a given query.
+     * Uses OpenAI for embedding the query and Pinecone for vector similarity matching.
+     * @param query The search query string.
+     * @param topK The number of top results to return (default is 10).
+     * @returns A list of document chunks that match the query.
+     */
+    async retrieve(query: string, topK: number = 10): Promise<RAGChunk[]> {
+        console.log(`Retrieving chunks for query: ${query}`);
+        try {
+            // Generate an embedding for the query using OpenAI.
+            const queryEmbeddingResponse = await this.openai.embeddings.create({
+                model: 'text-embedding-3-large',
+                input: query,
+                encoding_format: 'float',
+            });
+
+            let queryEmbedding = queryEmbeddingResponse.data[0].embedding;
+
+            // Extract the embedding from the response.
+
+            console.log(this._doc_ids());
+            // Query the Pinecone index using the embedding and filter by document IDs.
+            const queryResponse: QueryResponse = await this.index.query({
+                vector: queryEmbedding,
+                filter: {
+                    doc_id: { $in: this._doc_ids() },
+                },
+                topK,
+                includeValues: true,
+                includeMetadata: true,
+            });
+            console.log(queryResponse);
+
+            // Map the results into RAGChunks and return them.
+            return queryResponse.matches.map(
+                match =>
+                    ({
+                        id: match.id,
+                        values: match.values as number[],
+                        metadata: match.metadata as {
+                            text: string;
+                            type: string;
+                            original_document: string;
+                            file_path: string;
+                            doc_id: string;
+                            location: string;
+                            start_page: number;
+                            end_page: number;
+                        },
+                    }) as RAGChunk
+            );
+        } catch (error) {
+            console.error(`Error retrieving chunks: ${error}`);
+            return [];
+        }
+    }
+}