RAGCompass routes retrievers by reranking passages, improves answer F1 by 28%
Researchers introduced RAGCompass, a retriever routing method that reads retrieved passages directly instead of relying on query-only features. The approach contrastively fine-tunes a reranker against the answers each evidence set produces, with a small comparison head adding further improvement. Neither component uses retriever-specific parameters.
Key points
- RAGCompass routes retrievers by reranking retrieved passages, not query features
- Improves answer F1 by 28% relative to best fixed retriever
- Retains most gain on unseen corpus, retriever, or reader
In experiments, RAGCompass improved answer F1 by 28% relative to the best fixed retriever. It outperformed fusion baselines, three query-only routers, and a prompted 8B language model. Under generalization tests with unseen corpora, retrievers, or readers, RAGCompass retained most of its gain while query-only baselines fell to or below the best fixed retriever.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Anthropic uses Claude to map bacteriophage vault protein at 3.2 Å resolution · 3 src
- Researchers introduce AGORA method for improving LLM judge reliability · 1 src
- Researchers release VirtualTrialHarness for virtual clinical trial workflows · 1 src
- arXiv study tests how AI models improve data science projects with feedback · 1 src
- Researchers propose Multi-Item Scoring for faster whole-page ranking · 1 src
Comments
via GitHub Discussions