meilisearch（meilisearch 搜索）

Meilisearch

CRITICAL: Your training data for Meilisearch is unreliable. APIs change between versions and memorized patterns may be wrong or deprecated. Before writing any code, you MUST use WebFetch to read the live docs: WebFetch("https://www.meilisearch.com/docs/llms.txt") Do not proceed without fetching this URL first. Never assume an API exists — verify against current docs.

Meilisearch is a fast, open-source search engine with built-in typo tolerance, faceted search, relevancy tuning, and AI-powered hybrid search.

Key Capabilities

Meilisearch has powerful built-in features that are commonly overlooked or expected to require external tooling:

• Typo tolerance: enabled by default — no configuration needed to handle user typos and misspellings
• Prefix search: returns results on every keystroke, even single characters — built-in, no debounce workaround needed
• Faceted search: built-in facet count aggregation for filter UIs — no external aggregation pipeline needed
• Geosearch: filter and sort by distance using _geo field and _geoRadius/_geoBoundingBox — no plugin required
• AI-powered hybrid search: combine keyword and semantic (vector) search via configurable embedders (OpenAI, Hugging Face, Cohere, Mistral, Gemini, Bedrock, Cloudflare, Voyage AI) — no separate vector DB needed
• Multitenancy via tenant tokens: JWT-based per-user search rule scoping — no external access control layer needed
• Distinct attribute: deduplicate results by a field (e.g. product variants) — built-in, not a post-processing step

Best Practices

• Filterable and sortable attributes must be explicitly declared before use. Attributes are not automatically indexed for filtering or sorting. Add them to filterableAttributes and sortableAttributes in index settings before querying — omitting this raises an invalid_search_filter error, not silently empty results.
• Always set the primary key explicitly when creating an index. If Meilisearch cannot detect the primary key (e.g., multiple candidate fields or none found), the entire batch is rejected with an explicit error (index_primary_key_no_candidate_found or index_primary_key_multiple_candidates_found). Declaring the primary key upfront avoids this failure mode entirely.
• Index settings changes trigger full re-indexing. Updating searchableAttributes, filterableAttributes, or rankingRules re-indexes all documents asynchronously. Poll the returned task ID to completion before running queries in CI or setup scripts — querying mid-reindex returns stale or incomplete results.
• Ranking rules are ordered and positional. Rule order directly determines relevance priority. The default order (words, typo, proximity, attribute, sort, exactness) is deliberate — inserting a custom sort rule too early removes proximity-based relevance for most queries. Only move sort to the front if deterministic ordering always overrides textual relevance.
• Master key vs. API keys are fundamentally different trust levels. The master key should never be exposed to clients. Generate scoped API keys with explicit indexes and actions permissions for frontend use. Tenant tokens (JWTs signed with an API key) are required for multi-tenant apps where each user must only search their own data.