answer-stack

answer-stack is a full-stack Retrieval-Augmented Generation (RAG) chatbot application that lets users upload documents and ask questions grounded in those files. It combines a React chat interface, an Express API, document parsing, chunking, vector retrieval, and LLM-based answering into one Dockerized project.

The project is designed to work in two modes: a local demo mode that runs without external API keys, and a managed mode that can use OpenAI for embeddings/answers and Pinecone for vector search. This makes it useful both as a portfolio-ready GitHub project and as a foundation for building document Q&A, internal knowledge base, support assistant, or research assistant workflows.

Features

• RAG-based document Q&A with source snippets
• User authentication with JWT
• Multi-user document isolation
• PostgreSQL-backed user storage
• Persistent document metadata and chunks with a Docker volume
• File upload and text extraction for common document formats
• Document chunking and retrieval before answer generation
• Demo mode with in-memory indexing and local keyword search
• Managed mode support for OpenAI and Pinecone
• Redis + BullMQ ingestion queue support
• React frontend for upload and chat workflows
• WhatsApp webhook endpoint
• Dockerized frontend, backend, and Redis services

Run Locally

docker compose up --build

Frontend: http://localhost:5173 Backend health check: http://localhost:3000/health

Demo Mode

Set DEMO_MODE=true to run the full flow without any external secret. In demo mode:

• uploads are chunked and saved to a local JSON-backed document store
• retrieval is done with a local keyword scorer
• chat responses include grounded snippets from uploaded documents

Set DEMO_MODE=false and provide OpenAI + Pinecone credentials to use managed embeddings and LLM responses.

Persistence

answer-stack stores registered users in PostgreSQL. New user passwords are saved as salted hashes.

Uploaded document records and chunks are stored in a JSON file controlled by DOCUMENT_STORE_PATH.

With Docker Compose, the backend mounts a named volume at /app/data, so the default path below survives container restarts and rebuilds:

DOCUMENT_STORE_PATH=/app/data/documents.json
DATABASE_URL=postgres://answerstack:answerstack@postgres:5432/answerstack

With Docker Compose, Postgres uses a named postgres-data volume, so registered users survive container restarts and rebuilds.

When DEMO_MODE=false, vector search uses Pinecone while document records and chunk metadata remain persisted locally. For production deployments, you can move the document store into Postgres too.

How It Works

1. A user registers or logs in with JWT authentication.
2. The user uploads a supported document.
3. The backend extracts readable text from the file.
4. The text is split into chunks and indexed.
5. When the user asks a question, answer-stack retrieves relevant chunks.
6. The chatbot returns an answer grounded in the retrieved document context.

Supported Uploads

• .txt
• .md
• .json
• .csv
• .pdf
• .docx

The frontend now sends real files with multipart upload, and the backend extracts text for indexing. The default backend upload limit is 50 MB, configurable with MAX_UPLOAD_SIZE_MB in .env.

APIs

• POST /auth/register
• POST /auth/login
• POST /upload
• GET /upload
• POST /chat
• POST /whatsapp

Tech Stack

• Frontend: React, Vite
• Backend: Node.js, Express
• Auth: JWT, PostgreSQL
• Queue: Redis, BullMQ
• AI/RAG: OpenAI, Pinecone, local demo retrieval
• Deployment: Docker, Docker Compose

Project Structure

backend/   Express API, auth middleware, document upload, parsing, RAG services
frontend/  Vite + React chat UI

Notes

• No real API key is bundled in source code.
• Use .env for your own keys, or stay in DEMO_MODE=true for a working local demo without secrets.