Hybrid retrieval-augmented generation
Keyword search and embeddings, fused so each covers the other's blind spots, with a local open-source model doing the talking.
Problem
Domain-specific factual question answering where neither keyword search nor embeddings alone are reliable.
Approach
- Convert heterogeneous HTML and PDF sources into chunked embeddings and searchable FAISS and BM25 indexes.
- Hybrid retrieval: run both retrievers, fuse with reciprocal rank fusion, then generate with a locally served open-source LLM.
- Automated evaluation across retrieval relevance and answer quality.
Result
A pipeline where each component can be swapped and measured independently, with the fused retriever beating either retriever alone on the evaluation set.