HackRx – LLM Intelligent Query Retrieval System

📌 Project Overview

This project is an LLM-powered Intelligent Query Retrieval System built during HackRx 5.0. It is designed to process large, unstructured documents such as PDFs, Word files, and emails, and deliver fast, context-aware answers.

The system combines Pinecone for vector storage, Gemini for ultra-fast inference, and LLMs for context generation to provide precise and structured responses.

✨ Key Features

• 🔍 Semantic Search with Pinecone – Stores document embeddings for efficient and relevant retrieval.
• ⚡ Gemini-powered Inference – High-speed, low-latency query handling.
• 🧠 Context Generation with LLMs – Delivers detailed, structured answers with supporting context.
• 📑 Multi-format Document Support – Works with PDFs, Word documents, and email data.
• 🔐 Secure API Layer – Built with FastAPI and token-based authentication.
• 📊 Explainable Responses – Includes sources, metadata, and confidence levels.

🛠️ Tech Stack

• Vector Store: Pinecone
• LLM Models: Gemini + supporting LLMs for context generation
• Backend Framework: FastAPI (with async support)
• Authentication: Bearer Token Security
• Document Processing: Python (text extraction, chunking, embedding)

⚙️ How It Works

1. Upload Documents → PDFs, Word docs, or emails are ingested.
2. Preprocessing → Text is extracted, cleaned, and split into chunks.
3. Embedding Generation → Chunks are embedded and stored in Pinecone.
4. Query Input → User submits a query through the API/UI.
5. Vector Search → Relevant chunks are retrieved from Pinecone.
6. LLM Context Generation → Gemini + LLMs generate structured, context-aware answers.
7. Response Delivery → User receives a JSON response with answer, metadata, and confidence score.

🚀 Getting Started

Prerequisites

• Python 3.9+
• Pinecone API Key
• Google Gemini API Key
• FastAPI & Uvicorn

Installation

# Clone the repo
git clone https://github.com/WHitE-TITaN/HackRx-LLM-IntelligentQueryRetrievalSystem.git
cd HackRx-LLM-IntelligentQueryRetrievalSystem

# Create a virtual environment
python -m venv venv
source venv/bin/activate   # On Windows use venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Environment Variables

Create a .env file in the root directory with:

#your PineCorn api key command if for windows
setx API_KEY_PineCone "your_secret_key"
# >SUCCESS: Specified value was saved.
setx PINECONE_ENV "your_hosted_region"
# >SUCCESS: Specified value was saved.

setx API_KEY_GeminiAi "your Gemini api key"
# >SUCCESS: Specified value was saved.

setx  mongo_auth "your api key"
# >SUCCESS: Specified value was saved.

setx default_token "your key for normal users"
# >SUCCESS: Specified value was saved.

setx API_KEY_Cohere "your_secret_key" 
# >SUCCESS: Specified value was saved.
#this set the enviromewnt variable that can be use later.

Run the Server

uvicorn app.main:app --reload

API Endpoints

• POST /upload → Upload a document
• POST /query → Submit a query and get context-aware response
• GET /health → Health check endpoint

📚 Example Usage

Upload Document

curl -X POST "http://127.0.0.1:8000/upload"      -H "Authorization: Bearer <your_token>"      -F "file=@document.pdf"

Query the System

curl -X POST "http://127.0.0.1:8000/query"      -H "Authorization: Bearer <your_token>"      -H "Content-Type: application/json"      -d '{"query": "What are the key clauses in this contract?"}'

Response Example:

{
  "answer": "The contract includes key clauses on payment, termination, and liability.",
  "sources": ["document.pdf - page 3", "document.pdf - page 7"],
  "confidence": 0.92
}

📈 Future Improvements

• Add support for real-time streaming responses.
• Expand multi-language support.
• Integrate a simple web dashboard for uploads & queries.