Hybrid vs Vector-Only Retrieval for a Private Knowledge Manager

The article argues that hybrid retrieval combining lexical and vector search suits cases where exact identifiers and semantic meaning both matter, while vector-only retrieval fits small, stable, prose-heavy corpora. It uses a logistics example of detecting near-duplicate shipment records, where tracking numbers are exact keys and differently worded descriptions must still be matched. The author stresses that retrieval quality should be judged with a labeled test set and measured using recall, precision, and reciprocal rank rather than model leaderboards. Vector indexes carry ongoing costs such as storage, approximate-nearest-neighbor maintenance, and re-embedding whenever chunking or models change. The piece concludes that the choice depends on an index-cost budget and a relevance test set tailored to the application.