01Hybrid-Search RAG Platform
NotebookRAG
A hybrid-search RAG pipeline with the evaluation built in. It fuses dense and BM25 retrieval with weighted RRF, reranks with an LLM, checks every claim against its citations, and refuses to answer instead of making something up. A cross-model LLM-as-judge harness grades the results.
- Problem
- RAG systems fail silently: when retrieval misses, the model invents an answer and nobody notices.
- Approach
- Dense + BM25 fused in one Qdrant collection with weighted RRF, reranked by an LLM. A judge model verifies every claim against its citation, and a confidence gate refuses instead of guessing.
- Proof
- 0.919 correctness · 0.992 faithfulness · 1.00 retrieval hit rate
- 7/7 unanswerable trap questions refused, dense-only answered all 7
- Hybrid keyword-match 1.00 vs 0.79 dense-only, on a 36-question golden set
