Document Q&A with hybrid retrieval and evidence checks

The project

A portfolio/research prototype for document-grounded question answering across mixed input formats.

What I built

• A modular Python pipeline separating ingestion, retrieval, generation, verification and evaluation.
• Dense and BM25 retrieval with reranking to assemble relevant context.
• Citation and evidence checks, with confidence-based retry or abstention paths.
• Evaluation scripts and documentation so individual pipeline stages can be inspected and improved.

Relevant client work

Document search, knowledge assistants, retrieval debugging and a focused RAG proof of concept.

Tools: Python, FAISS, BM25, Hugging Face Transformers and LLM workflows.

[View source code and documentation](https://www.twine.net/signin

This is a portfolio prototype; its documented evaluation includes limitations.