Vector search understands paraphrase. Ask about "reducing customer churn" and it will find a document titled "keeping subscribers longer". Keyword search does not do that, and vector search has the opposite blind spot: it is unreliable on exact tokens. Search for error code TS-4471 and the nearest neighbours are documents about error codes in general.
Hybrid search runs both. A sparse retriever such as BM25 handles literal terms, product codes, and rare names. A dense retriever handles meaning. The two ranked lists are then fused, most often with reciprocal rank fusion, which scores each document by its position in each list and needs no score normalisation between two very different scales.
For technical documentation, support corpora, and e-commerce catalogues this is usually the default worth starting from. Those corpora are full of identifiers that must match literally and descriptions that users paraphrase freely.
Adding a reranker on top of the fused list is a common third stage, and the combination tends to outperform either retriever tuned alone.