Hybrid versus vector search: check whether exact identifiers are missed
Explain combining lexical and semantic candidates, and compare recall, ranking and maintenance using identifiers and natural descriptions from the same collection.
Choose an approach
Vector search helps when natural descriptions differ from source wording. If queries often contain model numbers, names or exact terminology, add keyword retrieval and evaluate hybrid search. Combining candidates does not automatically repair missing data or tokenization errors. Classify missing documents by query type before adding a retrieval branch.
Merging candidate lists differs from vector distance alone
Microsoft combines keyword and vector retrieval within one request and uses RRF to merge rankings; AlloyDB also offers hybrid retrieval functions. Parameters and indexes differ by product. A fusion score is not the probability that an answer is correct.
Short identifiers may differ by one digit, and semantic similarity cannot identify the exact object. Keyword matching can help but tokenization may still alter identifiers. Preserve dedicated identifier fields for strict equality rather than substituting hybrid ranking for exact filters.
Test both query types against the same collection
Choose queries containing full identifiers, purpose descriptions and both, and record the expected originals. Keep documents and filters unchanged, then compare top results from vector-only and hybrid search. Check candidate inclusion before ranking instead of using one generated answer as retrieval acceptance.
Add old versions and similar model numbers to check whether duplicates displace the correct object. Repair indexing first if the target was never indexed. No retrieval service was run here, and official examples are not presented as a measured recall gain for this site.
Maintain query samples after adding keyword indexes
Hybrid search maintains text and vector indexes, both of which need updates. Query budgets and filtering stages affect candidate counts; another branch does not imply lower latency. Save individual result lists to explain why a record moves or disappears.
Keep a vector baseline if it already meets important queries. Add a branch only when it improves concrete omissions. Replay failures after version changes and ensure withdrawn documents disappear. This locates practical defects better than a single general relevance score.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.