Azure AI Search retrieval: hybrid vs vector vs full-text vs semantic ranking
Verdict: Hybrid runs full-text and vector queries in parallel and merges them, best when queries mix intent with exact tokens. Vector finds concepts, full-text matches keywords, semantic ranking reranks an existing result set.
| Criterion | Hybrid search | Vector search | Full-text search | Semantic ranking |
|---|---|---|---|---|
| What it does | Full-text and vector in one request, merged | Similarity query over embeddings | BM25 keyword scoring | Reranks an existing result set |
| Strength | Exact tokens plus conceptual intent | Conceptually similar content | Exact codes, names, jargon | Relevance by meaning over an existing set |
| Choose when | Queries mix natural language with exact tokens | Vague, meaning-based queries | Precise keyword lookups | Improve ordering of retrieved results |
Rules
- Hybrid search is a single request with both search and vector parameters, running full-text and vector queries in parallel and merging into one result set.
- Hybrid combines strengths: keyword search finds exact codes and names, while vector search finds conceptually similar content without keyword matches.
- For terminology mismatch, hybrid queries plus semantic ranking and scoring profiles improve retrieval recall and relevance.
Traps
- Semantic ranking is a reranking pass over already-retrieved results, not a way to run two retrieval queries in parallel.
- Vector-only search misses exact tokens like product codes; that is the failing requirement hybrid fixes.