Elasticsearch.

Where Elasticsearch fits on a long engagement.

Relevance is the hard part of search and the part teams underestimate. Getting results back is trivial; getting the right result first requires analysers matched to the language, field boosting, synonym handling and an understanding of how scoring actually works. Default relevance is rarely good enough for a product where search is the interface.

Mappings are effectively permanent. Field types and analyser choices are set when the index is created and cannot be changed without reindexing, so a mapping created without thought becomes a constraint that a large index makes expensive to revisit.

What an assigned team does with Elasticsearch.

Search quality needs measurement rather than opinion. Click-through on results, queries that return nothing, and queries where users refine immediately are all signals that say more than anyone's assessment of whether the results look right.

Building that feedback loop is what turns search from a subjective argument into something improvable, and it sits inside the standards held under managed ai services where search is part of a retrieval stack.

What we use Elasticsearch for.

• Relevance tuned rather than accepted Analysers, boosting and synonyms configured for the actual corpus and language.
• Mappings designed before indexing Field types chosen deliberately, because changing them means a reindex.
• Search quality measured Zero-result and refinement rates tracked, so improvement is evidenced.

How Elasticsearch capacity is assigned.

Search capacity is assigned inside the discipline that owns the corpus, with relevance measurement treated as part of delivery.