Meilisearch 1.55 fixes ranking-threshold slowdowns and blocks a vector-index corruption path

Meilisearch 1.55.0, released October 12, 2026, changes rankingScoreThreshold processing to stop after the first below-threshold document, avoiding potentially evaluating every document. Maintainers warn relevance can change. The API now rejects binaryQuantization:null when updating embedder settings because turning off binary quantization after discarding original vectors could corrupt embeddings. It also adds Enterprise-only experimental inlineRule previews and fixes a Unicode highlighting panic.

Teams using Meilisearch for ecommerce, documentation and vector-backed search should test result quality and latency during upgrades. Settings clients that use null to reset binaryQuantization may now receive a deliberate error. Enterprise rule-preview functionality is useful for merchandising experiments but is not stable or generally available.

rankingScoreThreshold formerly could evaluate far too many documents, sometimes the whole index. Version 1.55 stops at the first document below the score threshold. Maintainers explicitly warn that result relevance may change in some circumstances; compare real queries before rollout.

Binary quantization discards exact original embedding values. The old API accidentally allowed binaryQuantization:null, effectively disabling it and risking corrupted embeddings. Version 1.55 rejects this update. Check settings-reset code and handle the new validation error.

inlineRule lets Enterprise Edition users with experimental dynamicSearchRules test a temporary rule in search or multi-search without persisting it. The Meili-Include-Metadata header can explain condition outcomes. This is not a broadly stable open-source feature.

Certain escaped Unicode code points could trigger a panic during highlight formatting with some segmenters. The patch fixes that path; the affected scope depends on index content and configuration.

Benchmark representative threshold queries for both latency and result order, remove unsupported binaryQuantization resets, and check edition/feature flags before using inlineRule. No independent production performance numbers were published.