🛡️ Finserv-Insurance-Agent — RAG-Based Insurance Question Answering System

A Retrieval-Augmented Generation (RAG) system built to answer insurance-related queries by extracting relevant information from uploaded policy documents (PDF / DOCX / EML) and generating concise, formal, human-style answers.

---

🚀 What This Project Does

Stage	Component	Description
1️⃣ Data ingestion	preprocessing.py	Reads documents, splits into chunks, generates embeddings, builds FAISS index
2️⃣ Query handling	query_final.py	Reformulates query → retrieves chunks via FAISS → reranks with Cross-Encoder → produces LLM answer
3️⃣ API layer	router.py	Exposes /hackrx/run endpoint to upload a document URL + multiple questions and returns answers

The system behaves like an insurance agent — brief, factual, and strictly based on the documents.

---

📦 Project Structure

├── preprocessing.py # Load docs, chunk, embed, build FAISS index ├── query_final.py # RAG pipeline + reranking + Groq LLM answering ├── router.py # FastAPI routes (/hackrx/run) ├── utils.py # File processing helper ├── faiss_index/ # FAISS index + chunks (auto-generated) ├── all-MiniLM-L6-v2/ # Local SentenceTransformer model ├── local_cross_encoder/ # Local reranking model ├── data/ # Optional document store └── requirements.txt

🔧 Installation

git clone https://github.com/SilentCanary/Finserv-Insurance-Agent.git
cd Finserv-Insurance-Agent
pip install -r requirements.txt

Environment Variables (.env)

GROQ_KEY = your_groq_api_key
VALID_TOKEN = your_fastapi_auth_token

🧰 Running the API

uvicorn main:app --reload

Endpoint

POST /hackrx/run

Example Request

{
  "documents": "https://example.com/policy.pdf",
  "questions": [
    "Is flood damage covered?",
    "What is the waiting period for hospitalization?"
  ]
}

Example Response

  "answers": [
    "Yes, flood damage is covered under Section 3 with exclusions.",
    "The policy requires a 30-day waiting period for hospitalization."
  ]
}

🧠 Core Technology

Component	Purpose
SentenceTransformer	Computes embeddings for document chunks and user queries
FAISS	Performs vector similarity search for retrieval
CrossEncoder	Reranks retrieved chunks based on relevance
Groq LLM	Generates the final answer using the top-ranked chunks
LRU Cache	Speeds up execution by caching repeated queries and responses

👤 Author Advitiya