OpenSearch Query Building Review (Query Review)

Grounding principles are in reference/principles.md. For facts, prefer reference/kb/ (official docs).

Review order (proceed exactly like this)

1. What the query is asking — result set, filter selectivity, frequency, sort/paging pattern.
2. Dynamic bool check — reference/kb/query-dsl-basics.md (filter context, term/match, builder safety).
3. Relevance·paging — reference/kb/relevance-scoring.md (BM25/boost/min_score) · reference/kb/pagination-search-after.md.
4. Vector/hybrid check — reference/kb/vector-hybrid-query.md (k vs ef_search, efficient filtering, normalization/RRF). If hybrid/neural grep returns 0, state "not applicable".
5. Slowness diagnosis — reference/kb/profiling-explain.md (Profile API). Provide a fix (query diff) with severity + re-measurement (recall@k·p99).

Quick checklist

• Is a no-score condition in must? → filter/must_not (skip scoring + cache).
• Does the dynamic builder enforce a field whitelist, a terms size cap, and avoid script/wildcard?
• Is deep paging using from+size? → search_after/PIT.
• In k-NN, is ef_search ≥ k? A variable k with a fixed ef_search means a ceiling mismatch.
• When a filter is present, is it efficient k-NN filtering, and does a selective filter under-fetch (over-fetch/exact fallback)?
• (hybrid) Is there a normalization/RRF processor? Are the neural ingest↔query model_id the same?
• Did a type-based client work around terms via bool.should? → replace with the native form.

KB (read and cite first)

From reference/kb/INDEX.md: reference/kb/query-dsl-basics.md · reference/kb/relevance-scoring.md · reference/kb/vector-hybrid-query.md · reference/kb/pagination-search-after.md · reference/kb/profiling-explain.md

Deliverable

For each finding: problem → principle/KB source → query diff → re-measurement (recall@k sweep·p99). "It got faster" is confirmed only by measurement.