MultiplexRAG

The question

One fixed retrieval method does not handle every query equally well. Routing is useful only if it improves on a strong fixed alternative under a clear evaluation protocol.

How it works

• Derive training labels from retrieval quality and fit a lightweight logistic-regression router.
• Use query/retrieval features to choose sparse, dense or hybrid search, falling back when confidence is low.
• Compare fixed retrievers, routing variants and ablations with query-level splits and ranking metrics.

Evidence & results

• Stored Phase 4 NFCorpus validation report: MRR@10 of 0.5270 versus 0.5110 for the strongest fixed hybrid baseline, over 324 queries.
• The repository includes evaluation reports and experiments across four datasets.

Scope & next steps

• The quoted NFCorpus comparison is validation evidence, not a held-out test gain.
• The Phase 4 configuration is distinct from the default one-command pipeline.
• No production latency improvement or universal advantage across RAG systems is claimed.