ManoharLegion/legal-document-assistant

Legal Document Assistant

An agentic RAG assistant that answers questions over Indian case law and lets a user upload their own legal document (PDF/.txt) for summarization and session-scoped follow-up questions.

• —Embeddings: sentence-transformers/all-MiniLM-L6-v2 (local, free)
• —Vector store: ChromaDB (local, file-based)
• —Generation: Gemini/Gemma via Google AI Studio's free API
• —Backend: FastAPI (backend/main.py), also serves the static frontend
• —Frontend: plain HTML/CSS/JS (frontend/)

A legacy Streamlit version of the UI is still present at app.py for reference.

Local setup

1. 1.Create a virtual environment and install dependencies:

   python -m venv .venv
   .venv\Scripts\activate   # Windows
   pip install -r requirements.txt

1. 1.Get a free API key at https://aistudio.google.com/apikey and copy it into .env:

   copy .env.example .env
   # then edit .env and paste your key as GEMINI_API_KEY=...

1. 1.Build the vector index (downloads a subset of the dataset, embeds, and persists to chroma_db/):

   python scripts/build_index.py

This indexes the first 300 documents by default. Override with MAX_DOCS=1000 env var if desired.

1. 1.Run the app (backend serves the frontend too — one server, one URL):

   uvicorn backend.main:app --port 8011

Open http://localhost:8011

Deploying publicly (free, no card required)

This repo ships a Dockerfile for Hugging Face Spaces:

1. 1.Create a free Space at huggingface.co/new-space, SDK: Docker.
2. 2.Push this repo to the Space's git remote (or link your GitHub repo in the Space settings).
3. 3.In the Space's Settings → Repository secrets, add GEMINI_API_KEY with your key.
4. 4.The Space builds the Docker image, runs scripts/build_index.py then starts the server — you get a permanent public URL.

Storage on the free tier is ephemeral, so the index rebuilds on every restart (a few minutes); no persistent disk is required.

Usage

• —Ask: type a question about the indexed corpus — the agent retrieves relevant chunks via similarity search and answers using only that retrieved context, streaming the answer as it generates, with sources shown for transparency.
• —Upload a document: upload a PDF/.txt, click Summarize. If it's recognized as a legal document, it's summarized and added to your session's search scope so you can ask follow-up questions about it specifically (in addition to the base indexed corpus).

Notes

• —The free Gemini API tier has request-per-minute limits; src/llm.py retries with backoff on transient errors.
• —Session-uploaded documents are held in memory only (never written to disk) and are discarded when the server process restarts.