HackRx 6.0 – LLM Document Processing System

An end-to-end, well-documented Python project that demonstrates how to build an LLM-powered query–retrieval and decision engine over large, unstructured documents (PDFs, Word, e-mails).

Features

1. Multi-format ingestion – loaders for PDF, DOCX, and E-mail files extract text + metadata.
2. Semantic vector store – documents are chunked, embedded via sentence-transformers, and indexed with FAISS for fast nearest-neighbour search.
3. Natural-language query parser – rule-based + LLM fallback extracts structured fields (age, procedure, location, policy age, …).
4. Decision engine – pluggable logic evaluates retrieved clauses and returns JSON containing:

   {
     "decision": "approved | rejected",
     "amount": 12345.67,
     "justification": "Text explanation …",
     "clauses": [ {"id": "…", "text": "…"} ]
   }

5. CLI – python -m hackrx_llm --docs /path/to/folder --ask "46M knee surgery Pune 3-month policy".
6. Extensible – swap embeddings, LLM provider, or decision logic.
7. Test-driven – pytest covers parsing and retrieval.

Quick-start

python -m venv .venv && source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt

# Build vector store (runs ingestion automatically)
python -m hackrx_llm ingest --docs sample_docs

# Ask a question
python -m hackrx_llm ask --query "46-year-old male, knee surgery in Pune, 3-month policy" --top_k 5

If an OpenAI key is present (export OPENAI_API_KEY=…), the parser will enrich/validate fields via GPT automatically; otherwise, rule-based extraction is used.

Project Structure

├── hackrx_llm/          ← Library package
│   ├── ingestion/       ← PDF, Word, e-mail loaders
│   ├── parser.py        ← Query → structured data
│   ├── retriever.py     ← Vector store + semantic search
│   ├── decision_engine.py
│   ├── schema.py        ← Pydantic models
│   ├── cli.py           ← Typer CLI
│   └── __init__.py
├── tests/               ← Unit tests (pytest)
├── requirements.txt
└── README.md            ← You are here

Design Diagram

graph TD
    A(User Query) --> B(Parser)
    B --> C[Structured Query]
    C --> D(Retriever)
    D --> E[Relevant Clauses]
    C --> F(Decision Engine)
    E --> F
    F --> G[JSON Response]

Loading

Contributing

1. Fork -> git clone -> create feature branch.
2. Ensure pytest passes & run black / ruff.
3. PR with clear description.

License

MIT © 2025 Bajaj Finserv Health Ltd.