What problem does it solve? Pure dense vector search often misses exact keyword matches, and combining multiple representations of the same item (title, abstract, chunks, languages) in one query is hard to design correctly. This Skill guides the construction of prefetch queries in Qdrant's Query API so hybrid retrieval works as intended. ## Core Features & Use Cases - Sparse Vector Selection: Compares BM25, BM42, miniCOIL, SPLADE++, and external learned sparse embeddings (e.g. BGE-M3) with configuration guidance on tokenization, stemming, stopwords, and avg_len calibration. - Multi-Representation Search: Designs multiple named-vector prefetches for items embedded in several ways, with rules for when a representation earns its own prefetch and when to prefer sparse vectors for lexical fields. - Grouping and Multivector Guidance: Covers grouped search with per-prefetch limit sizing, lookup-in-groups to avoid payload duplication, and the tradeoffs of multivectors in prefetch stages. - Use Case: A search engineer building a document search system needs to combine a dense chunk embedding, a BM25 sparse vector over titles, and document-level grouping in a single Qdrant Query API request with correct prefetch limits. ## Quick Start Ask how to combine dense and sparse vectors in one Qdrant search request for a collection with title and chunk fields.
