Vector search — often called semantic search — finds the items most relevant to a query by comparing embeddings, not by matching literal words. The query is embedded the same way the stored content was, and the system finds the stored vectors closest to the query vector.

Comparing a query against every stored vector one by one — an exact nearest-neighbor search — becomes too slow as a collection grows into the millions. In practice, vector databases use approximate nearest neighbor (ANN) algorithms that index vectors for fast lookup, trading a small amount of accuracy for a large gain in speed.

• Exact search — compares the query against every stored vector; accurate but doesn't scale to large collections.
• Approximate nearest neighbor (ANN) — indexes vectors so most queries can skip comparing against everything, dramatically faster at large scale with a small accuracy tradeoff.
• Top-k retrieval — vector search typically returns the k closest items, not a single best match, since RAG and search applications usually want several candidates to work with.

Neither approach strictly dominates the other, which is exactly why hybrid search — combining both — often outperforms either alone.

• Assuming vector search always outperforms keyword search

  Exact identifiers, product codes, or rare technical terms are often matched better by keyword search.

• Ignoring the accuracy tradeoff of approximate nearest neighbor search

  ANN indexes trade a small amount of accuracy for large speed gains — this is usually the right tradeoff at scale, but it is a real tradeoff.

• Retrieving only the single top result

  A single closest match can be wrong or incomplete — retrieving several top-k candidates and letting a later step (like reranking) narrow them down is usually more robust.

• Forgetting that the query and stored content must use the same embedding model

  Embeddings from different models generally aren't comparable to each other.