Policy-RAG-BOT is a Retrieval-Augmented Generation (RAG) based chatbot that allows users to query government policy documents through a simple conversational interface, instead of navigating complex official websites. The project currently demonstrates this idea using the Year-End Review 2024: Ministry of Environment, Forest and Climate Change (India), and is designed to be easily extensible to any government policy document (PDFs, reports, press releases, etc.).

🎯 Project Motivation

Government policy information is often:

• Scattered across multiple web pages.

• Written in dense, formal language

• Hard to search for specific answers

• This project shows how modern LLM-based systems and vector search can be used to build transparent, document-grounded policy assistants that:

• Answer only from official sources

• Refuse gracefully when information is missing

• Remain auditable and extensible

🧠 Key Features

• Retrieval-Augmented Generation (RAG) using LangChain
• Strict grounding: answers only from retrieved context
• Graceful refusal when information is not found
• Conversation memory support (chat history preserved)
• Multi-LLM backend support
• Groq
• OpenRouter (OpenAI-compatible)
• Hugging Face
• Pluggable document ingestion (any PDF / policy document)
  -Streamlit-based UI for nice interface while asking questions

🏗️ Architecture Overview

Policy documents are embedded using Hugging Face sentence transformers.
Embeddings are stored in a Chroma vector database.
User queries are:

• Embedded
• Retrieved against the vector store
• Passed to an LLM with retrieved context.

The LLM is instructed to:

• Answer only from context
• Refuse politely if context is insufficient
• Suggest related information when refusing
• This design ensures faithful, non-hallucinatory answers.

📦 Tech Stack

• Language: Python
• Frameworks:
• LangChain
• Streamlit
• Vector Store: Chroma
• Embeddings: Hugging Face (all-MiniLM-L6-v2)
• LLMs:
  Groq (default)
  OpenRouter (OpenAI-compatible)
  Hugging Face Inference API
  Environment Management: python-doten