How do dense vectors connect paraphrases?
An encoder may place “cancel my subscription” near a passage about ending a membership despite different wording. This helps natural language questions, though nearby vectors can also represent only loosely related topics. It is a common approach to semantic search.
How are candidates retrieved?
Documents are encoded in advance with source and access metadata. A query is encoded at search time, and a nearest neighbor index finds candidates. A later reranker may refine them. Model choice, chunking, and index settings determine recall. Multimodal retrieval and analytics shows a broader retrieval use case.
When should sparse retrieval be added?
Exact names, codes, and uncommon terms may favor sparse retrieval. Combining dense and sparse candidates often improves mixed enterprise queries. Test candidate recall before tuning generation, because a model cannot reliably cite evidence never retrieved.
FAQ
Is dense retrieval the same as semantic search?
It is a common implementation, not the entire category.
Must the encoder be trained in house?
No. Test existing models before considering domain adaptation.
How are misses found?
Use paraphrase, rare term, and no answer cases in a labeled test set.