NyaySetu LLM RAG

A legal document question-answering system built with RAG (Retrieval-Augmented Generation) using LangChain, FAISS, and Google Gemini 1.5 Flash (with lightweight local fallbacks for offline use).

Features

• Extract text from PDF legal documents
• Build vector store for efficient document retrieval
• FastAPI web server for question answering
• Hosted Google Gemini 1.5 Flash with automatic fallback to lightweight local responders
• CORS enabled for frontend integration

Setup

1. Install dependencies:

pip install -r requirements.txt

2. Extract PDF texts:

python extract_pdf_to_json.py

3. Build vector store:

python build_vectorstore.py

4. Configure the LLM provider via environment variables:

   # Use Google Gemini 1.5 Flash
   export LLM_PROVIDER=gemini
   export GEMINI_API_KEY=your_google_api_key
   export LLM_MODEL=gemini-1.5-flash

   If these variables are not provided the server falls back to the lightweight keyword-based responder.

5. Run the FastAPI server:

uvicorn app:app --reload

API Endpoints

• GET / - Welcome page
• POST /ask - Submit legal questions

Requirements

• Python 3.11+
• PDF documents in the pdfs/ folder
• If using Gemini: Google Cloud account with Generative Language API enabled and a GEMINI_API_KEY

LLM Configuration

The backend reads the following environment variables to decide which LLM to use:

Variable	Description	Default
LLM_PROVIDER	gemini or fallback	gemini
LLM_MODEL	Model name for the provider (e.g., gemini-1.5-flash)	Provider specific
GEMINI_API_KEY	Google API key for Gemini access	required when LLM_PROVIDER=gemini
LLM_TEMPERATURE	Sampling temperature	0.2
LLM_MAX_OUTPUT_TOKENS	Max tokens generated per response	256

If no provider is configured or initialization fails, the app falls back to the lightweight keyword-based QA system.