In-Session Document Processing RAG

A FastAPI-based document ingestion and retrieval-augmented generation (RAG) app. It lets users upload documents, vectorizes them with Google Gemini embeddings, stores them in ChromaDB, and answers session-specific questions with Nemoguardrails safety checks.

Key features

• Upload documents via /upload and build a Chroma vector store
• Query documents by session_id using /chat
• Streams ingestion logs through /logs/stream
• Enforces input/output safety with Nemoguardrails
• Uses Google Generative AI via langchain-google-genai

Project structure

• app.py — FastAPI server and endpoints
• src/ingest.py — document ingestion, splitting, embedding, and Chroma persistence
• src/query.py — session-aware retrieval and Gemini-based answer generation
• src/query_guardrails.py — safety checks before and after answer generation
• src/loader.py — document loader abstraction for PDF, Word, text, and Excel files
• guardrails/query_policy.yml — Nemoguardrails safety policy
• UI/index.html — front-end interface for the chat UI

Environment

Create a .env file in the repository root with:

GOOGLE_API_KEY=your-google-api-key
DATA_DIR=./data

.env is ignored by .gitignore and should not be committed to git.

Install

python -m pip install -r requirements.txt

Run locally

uvicorn app:app --reload

Then open UI/index.html in a browser or connect a client to the API.

API endpoints

• POST /upload — upload a file and index it for a session
• POST /chat — ask a question for a session
• GET /logs/stream — stream log events for a session

Notes

• .env is not committed and should remain private. Replace .env.example to .env and update with your credentials.
• The application requires a valid GOOGLE_API_KEY.

Thanks ! @author - Sanjaykumar Venkatesan