Docquery

RAG-powered PDF Q&A app. Upload a document, ask questions, get answers with source citations.

Live Demo: https://docquery-app.netlify.app

---

How it works

1. Upload a PDF — text is extracted, chunked into 500-word segments, embedded via Voyage AI, and stored in ChromaDB
2. Ask a question — the question is embedded, top 4 matching chunks retrieved via vector search, passed to Claude Haiku as context
3. Get an answer with source citations — collapsible sources show exactly which chunks were used

Stack

Frontend: React 19 + Vite → Netlify Backend: FastAPI + ChromaDB + Voyage AI + Anthropic Claude → Render

Architecture

Browser → Netlify (React) → Render (FastAPI)

The FastAPI backend calls three external services:

• Voyage AI — generates embeddings for chunks and questions
• ChromaDB — stores and retrieves vectors
• Anthropic Claude — generates answers from retrieved context

Run locally

Backend:

cd backend
pip install -r requirements.txt
cp .env.example .env   # add ANTHROPIC_API_KEY and VOYAGE_API_KEY
uvicorn main:app --reload

Frontend:

cd frontend
npm install
npm run dev

API

Method	Endpoint	Description
GET	/documents	List uploaded documents
POST	/upload	Upload a PDF
POST	/ask	Ask a question
DELETE	/document/{doc_id}	Delete a document

Design decisions

• Voyage AI embeddings over local sentence-transformers — API-based, no RAM overhead, deployable on Render free tier
• ChromaDB PersistentClient — document vectors survive server restarts
• Filename stored in collection metadata — document registry rebuilds correctly after restart
• Session-based chat history — in-memory only, no database required for Phase 1
• 10MB file size limit — prevents oversized uploads crashing the free tier server

Security

• CORS restricted to Netlify frontend URL
• 10MB file size limit on uploads
• 1000 character limit on questions
• Rate limiting: 5 uploads/hour and 20 questions/hour per IP
• API keys stored as environment variables, never in code

Evaluation

RAG pipeline evaluated using a custom LLM-as-judge eval suite.

Metric	Score
Faithfulness	1.0
Answer Relevancy	1.0

• Faithfulness — measures whether answers are grounded in retrieved context, not hallucinated
• Answer Relevancy — measures whether answers address the question asked
• Judge model: Claude Haiku
• Test cases: 10 question/answer pairs with ground truth

Run evals:

cd backend/evals
python eval.py