What it means
Keyword search matches strings. Semantic search compares embeddings, so it retrieves by meaning — a search for "how do I cancel" finds a document about "terminating your subscription" even with no shared vocabulary.
The weakness is the mirror image of the strength. Because it matches on meaning rather than characters, semantic search is unreliable on exact identifiers: part numbers, error codes, proper names, version strings. It will happily return something semantically adjacent when you needed an exact match.
This is why serious systems use hybrid search — running keyword and semantic retrieval together and merging the results. It is not a compromise so much as an acknowledgment that the two fail in opposite directions.
Why it matters
Semantic search is what makes retrieval over messy internal documents work at all, since real users never phrase questions the way documents phrase answers. But shipping it alone, without keyword search, produces a system that feels magical in demos and fails on the first support ticket containing an order number.
In practice
Default to hybrid rather than pure vector search. If you must pick one, consider what your users actually type — support and technical domains lean keyword-heavy more than teams expect.
Where this shows up
Tools and models in our catalog.
PineconeThe leading managed vector database for AI applications. Serverless pricing, 99.99% SLA, and billions of vectors at millisecond query speeds. Widely used in production RAG systems.
Elastic AI SearchESRE (Elasticsearch Relevance Engine) for retrieval-augmented generation. Hybrid vector + keyword search combining traditional full-text with semantic vector search in a single platform.
Supabase VectorPostgreSQL-based vector storage using the pgvector extension. Seamlessly combines traditional relational data with vector search in a single database.