Insurance Aware RAG

AI-powered RAG system for insurance policy analysis.

This RAG system is an advanced Retrieval-Augmented Generation (RAG) system designed to extract and reason over complex insurance policy clauses. It utilizes a two-stage retrieval pipeline (Dense Retrieval + Cross-Encoder Reranking) combined with strictly enforced structural validation to accurately answer policy-related queries with high-confidence citations.

Key Features & Architectural Specialties

• Deterministic Clause Splitting: Abandons naive text chunking. Policies are parsed into atomic structural clauses with deterministic, stable IDs (e.g., ICICILombard_p8_Grace_Period_1) to ensure perfect evaluation traceability and prevent vector overwrites.
• Two-Stage Retrieval Pipeline:
  • Candidate Generation: High-recall dense retrieval using FAISS (BAAI/bge-base-en-v1.5).
  • Cross-Encoder Reranking: Precision reranking of the top candidates using BAAI/bge-reranker-base to surface the most relevant evidence.
• Strictly Validated Reasoning: Leverages exact Pydantic schemas (RAGResponse, Citation) to strictly enforce LLM output. The system acts as a structural firewall that rejects ungrounded hallucinated citations and logically contradictory responses.
• Fail-Fast Engineering: Designed with strict internal verification. Duplicate clause IDs raise errors during ingestion, and retrieval/reasoning layers fail fast on missing context to avoid silent degradation.
• Validation-Driven Evaluation Framework: Includes a built-in evaluator (evaluate_stagewise, evaluate_multi_clause) mapping ground-truth clause IDs to measure exact Retrieval Recall@K and MRR.

Architecture Stack

• Framework: LangChain & LangGraph, with Pydantic for strict schema validation
• Vector Store: FAISS (CPU)
• Embeddings: BAAI/bge-base-en-v1.5 (Dense) & BAAI/bge-reranker-base (Cross-Encoder)
• LLM Engine: Groq (Llama-based inference)

Getting Started

Prerequisites

• Python 3.10+
• A valid Groq API Key

Installation

1. Clone the repository:

   git clone https://github.com/kyunbhaii/Insurance-Aware-RAG.git
   cd Insurance-Aware-RAG

2. Set up a virtual environment (recommended):

   python -m venv venv
   source venv/bin/activate  # On Windows use `venv\Scripts\activate`

3. Install dependencies:

   pip install -r requirements.txt

4. Environment Variables: Create a .env file in the root directory and add your Groq API key:

   GROQ_API_KEY="your_actual_groq_api_key"

Usage

1. Place Your Data

Place your insurance policy PDFs in the data/ directory. For example, data/icici_complete_health.pdf.

2. Run the Full Pipeline

The main script processes the document, generates the embeddings, builds the retriever, and executes a sample query.

python -m scripts.run_pipeline

3. Review the Output

The system handles queries (e.g., "Is maternity covered?") and returns a structured response:

Answer: [Detailed reasoning paragraph]
Found: True/False
Citations:
  - Clause 4.2 (page 12)
Confidence: High

Advanced Scripts

• Testing Retrieval: python -m scripts.test_retrieval to benchmark document retrieval performance.
• Data Ingestion Preview: python -m scripts.main to preview how your PDF is split into clauses.
• Evaluation: Explore app/evaluation and scripts/run_evaluation.py for comprehensive RAG testing.

Project Structure

ClaimLens/
├── app/
│   ├── config.py
│   ├── ingestion/
│   ├── retrieval/
│   ├── reasoning/
│   └── pipeline.py
├── data/
├── docs/
├── scripts/
├── indexes/
└── requirements.txt

Contributing

Contributions, issues, and feature requests are welcome! Feel free to check the issues page.

License

This project is licensed under the MIT License - see the LICENSE file for details.