AI Document Assistant

A full-stack RAG application that answers questions about a document using only that document with source citations, a confidence level, and a human-review escalation path for low-confidence answers.

What it does

• Paste in any document (contract, policy, report)
• Ask a question in plain language
• The app retrieves the most relevant passages, answers using only those passages, and shows:
  • the answer, grounded in the document
  • the source snippets it used (so the answer is verifiable)
  • a confidence level, with an escalation prompt when confidence is low

Screenshots

Answer with grounded sources and confidence for a sample NDA Contract:

Source citations: showing which passages the answer used

How it works (RAG pipeline)

1. Chunk the document into passages
2. Embed the passages and the question (Gemini text-embedding-004)
3. Retrieve the top passages by cosine similarity
4. Answer using only the retrieved passages, with citations and a confidence estimate

Stack

• Frontend: React (Vite)
• Backend: FastAPI, NumPy for similarity search
• Model: Google Gemini (generation + embeddings)

Why it's designed this way

• Citations for trust. Every answer shows the exact source passages, because in a real setting people can't act on an answer they can't verify.
• Grounded, not open-ended. The model answers only from the document and says so when the answer isn't there — reducing hallucination.
• Confidence routes to humans. Low-confidence answers surface a review prompt instead of being presented as certain.

Running locally

Backend (from the project root):

cd backend
pip install -r ../requirements.txt
uvicorn app:app --reload --port 8001

Frontend:

cd doc-qa-ui
npm install
npm run dev

Requires a GEMINI_API_KEY environment variable set where the backend runs.