feat: full-stack RAG pipeline for PDF ingestion and contextual chat capabilities

[Nyumat]

This pull request implements a full-stack retrieval-augmented generation (RAG) ingestion pipeline and querying capability. As the core functionality of Beavs AI, this will be a constant work in progress. Although, we can finally say: We did it 🚀

Demo

beavsai_demo.mov

How It Works

PDF Upload Workflow

1. File Upload: User uploads a PDF via the UI.
2. Presigned URL: Backend generates a Cloudflare R2 presigned URL for secure access.
3. File Parsing: PDF is parsed into text chunks using WebPDFLoader.
4. Chunking: Text is split into chunks with RecursiveCharacterTextSplitter.
5. Embedding Creation: Chunks are embedded with OpenAIEmbeddings.
6. Pinecone Storage: Embeddings are upserted into Pinecone.
7. Database Update: documentIds and isIndexed fields in the course_materials table are updated.

Chat Context Workflow

1. File Selection: User selects a previously uploaded PDF when starting a chat.
2. Message Retrieval: Latest user message is retrieved for querying.
3. Embedding Search: Message is embedded and searched in Pinecone for relevant chunks.
4. Context Preparation: Retrieved chunks are combined into a file context.
5. Chat Session: The chat session starts using from the PDF metadata (fileName).

Key Changes

• Dependencies: Added @langchain/pinecone and removed langchain.
• Schema: Added documentIds and isIndexed fields to course_materials.
• API Updates: Enhanced embeddings and chat routes for PDF processing and context retrieval.
• Components: Updated chat components and actions for file-based context support.
• Utilities: Added PDF parsing utilities and improved Pinecone and OpenAI clients.

Next Steps

• Improve error handling and performance.
• Support multi-tenant use relevant info.
• Support sources for the AI response.
• Allow instant chat to query across all documents.

Note

This foundational RAG pipeline allows Beavs AI to provide document-based contextual chats, enhancing the user interactions and responses.

[Nyumat]

Going to clean up the code a bit to help prepare for our next meeting, where we'll do a deep dive of this implementation!