Multi-route Recall
An engineering strategy in the retrieval stage of RAG that simultaneously uses multiple recall paths, such as vector retrieval and keyword retrieval, and merges the results from each path to improve recall rates.
Core Facts
Timeline (last 90 days)
Multi-way retrieval combines vector retrieval, keyword retrieval (such as BM25), knowledge graphs and other methods, complementing the blind spots of different retrieval approaches.
Multi-route recall means simultaneously using multiple retrieval strategies (such as vector retrieval, BM25 keyword retrieval, knowledge graph retrieval, etc.) to recall candidate sets separately, then merging and deduplicating them for unified re-ranking, to compensate for the blind spots of any single retrieval method.
A common fusion approach is parallel multi-channel recall: a CF channel + a content vector channel + a popular-rules channel, with candidates from each channel uniformly scored at the ranking layer.
Multi-channel retrieval simultaneously fetches candidate results through multiple different retrieval channels and then performs fusion ranking to maximize recall rate.
All Facts (5)
In actual production, hybrid retrieval combining both vector similarity and keyword matching is typically used.
65%UnverifiedMulti-channel retrieval simultaneously fetches candidate results through multiple different retrieval channels and then performs fusion ranking to maximize recall rate.
60%UnverifiedMulti-way retrieval combines vector retrieval, keyword retrieval (such as BM25), knowledge graphs and other methods, complementing the blind spots of different retrieval approaches.
50%UnverifiedMulti-route recall means simultaneously using multiple retrieval strategies (such as vector retrieval, BM25 keyword retrieval, knowledge graph retrieval, etc.) to recall candidate sets separately, then merging and deduplicating them for unified re-ranking, to compensate for the blind spots of any single retrieval method.
50%UnverifiedA common fusion approach is parallel multi-channel recall: a CF channel + a content vector channel + a popular-rules channel, with candidates from each channel uniformly scored at the ranking layer.
50%