Insurance RAG QA Service

An end-to-end Retrieval-Augmented Generation (RAG) system for querying large insurance-domain documents. It ingests PDFs (and other supported formats), semantically indexes them, and answers natural-language questions using context retrieved from both existing and newly uploaded policy documents.

---

🚀 Features

• Document Ingestion

  • Supports PDF (plain-text via pdfplumber or Markdown via pymupdf4llm), DOCX, TXT, and EML.
  • Automatically downloads from any HTTP(s) URL.

• Smart Chunking

  • Uses a character-based recursive splitter (langchain) with custom separators for Markdown semantics.
  • Configurable chunk size and overlap to match your embedding model’s context window.

• Domain-Specific Embeddings

  • Leverages a locally-hosted HuggingFace Sentence-Transformer fine-tuned on insurance data.
  • Fast GPU/CPU inference via sentence-transformers.

• Vector Store with FAISS

  • On-disk FAISS index + pickle metadata for persistent storage.
  • In-memory FAISS for newly uploaded docs, then merges into the main store.
  • Cosine (L2) retrieval of top-k relevant chunks.

• RAG Prompt Assembly

  • Combines “new” document context (if provided) with existing global context.
  • Deduplicates overlapping chunks.
  • Produces clear, instruction-driven prompts that constrain the LLM to only use provided context.

• LLM Integration

  • Async calls to Google Gemini QnA endpoint via httpx.
  • Exponential backoff + retry logic for transient errors (5xx, timeouts).
  • Strips numbering and whitespace from generated answers.

• Secure API

  • FastAPI server with a bearer-token auth dependency.
  • Health-check endpoint.
  • Pydantic request/response models for strict validation.

---

🏗️ Architecture Overview

Client Request
   └── POST /api/v1/hackrx/run
         ├─ Headers: Authorization: Bearer <AUTH_TOKEN>
         └─ Body: { documents: "<URL>", questions: ["Q1", "Q2", …] }

main.py (FastAPI)
 ├─ verify_token()    # checks AUTH_TOKEN
 ├─ generate_prompts() # in rag_system.py
 │    ├─ load persistent FAISS store
 │    ├─ if documents provided:
 │    │     ├─ download & parse (document_parser.py)
 │    │     ├─ chunk (text_chunker.py)
 │    │     ├─ dedupe against persistent .texts
 │    │     ├─ embed_texts() → local BERT model
 │    │     └─ build in-memory FAISS index
 │    ├─ for each question:
 │    │     ├─ embed_query() → local BERT model
 │    │     ├─ search new & existing FAISS stores
 │    │     ├─ dedupe & assemble context
 │    │     └─ build instruction-driven prompt
 │    └─ if new chunks exist & ALLOW_DB_UPDATE:
 │          └─ merge them into persistent FAISS store
 ├─ call_gemini_api()  # in embedder/llm.py
 │    ├─ POST to Gemini QnA endpoint
 │    ├─ retry on 5xx / timeouts with exponential backoff
 │    └─ extract candidates[0].content.parts[0].text
 └─ return JSON { answers: […] }

Persistent Storage:
 ├─ /vector_store/index.faiss
 └─ /vector_store/texts.pkl

---

🛠️ Tech Stack

• Language & Framework: Python 3.11 + FastAPI

• Document Parsing:

  • pdfplumber (plain text)
  • pymupdf4llm (Markdown conversion)
  • python-docx (DOCX)

• Chunking: langchain.text_splitter.RecursiveCharacterTextSplitter

• Embeddings: sentence-transformers (local BERT embed model)

• Vector Search: faiss-cpu for persistent on-disk and in-memory indexing

• LLM Client: httpx.AsyncClient → Google Gemini QnA API

• Configuration & Secrets: pydantic-settings, .env via python-dotenv

• Containerization: Docker + optional Docker Compose

• Authentication: Bearer token via FastAPI dependency

---

⚙️ Configuration

Copy the example .env file and fill in your secrets:

# .env
AUTH_TOKEN=your_custom_bearer_token
GEMINI_API_KEY=your_google_gemini_api_key

All other settings (chunk sizes, model paths, retry counts) live in config.py.

---

📦 Installation & Local Run

1. Clone & enter

   git clone <repo-url>
   cd rag_insurance

2. Create & activate a virtual environment

   python -m venv venv
   source venv/bin/activate   # on Windows: venv\Scripts\activate

3. Install dependencies

   pip install -r requirements.txt

4. Prepare your embedding model & vector store

   • Place your Sentence-Transformer model in insurance_bert_embed/
   • Create an empty vector_store/ folder for FAISS artifacts

5. Run the server

   uvicorn main:app --reload --port 8000

6. Test your health and QA endpoints

   curl http://localhost:8000/health
   # POST /api/v1/hackrx/run with Bearer auth to get answers

---

🚢 Docker Deployment (in progress)

1. Build the image

   docker build -t rag-insurance .

2. Run (mount volumes for persistence/model)

   docker run -d \
     -p 8000:8000 \
     -e AUTH_TOKEN="$AUTH_TOKEN" \
     -e GEMINI_API_KEY="$GEMINI_API_KEY" \
     -v ./vector_store:/app/vector_store \
     -v ./insurance_bert_embed:/app/insurance_bert_embed \
     rag-insurance

---

🔄 Extensibility

• Swap LLM: point generator/llm.py at OpenAI or any other endpoint.
• Alternative Vector Store: swap FAISS for Chroma or Pinecone by implementing the same interface.
• Enhanced Chunking: adjust separators or integrate hierarchical splitting.
• Multi‐Doc Queries: accept arrays of URLs & merge their contexts in generate_prompts().

---

Feel free to explore, customize, and extend!