meilisearch

Executes simultaneous full-text and vector similarity searches, then combines results using a configurable semanticRatio parameter that weights keyword relevance against semantic similarity. The milli crate maintains separate inverted indexes (word_docids, word_pair_proximity_docids) for keyword matching and arroy vector stores for embedding-based retrieval, with fusion logic that merges ranked result sets at query time. This dual-index approach enables applications to balance exact-match precision with semantic understanding without requiring separate search infrastructure.

All write operations (document additions, deletions, index creation, settings changes) are enqueued as tasks in the IndexScheduler, which batches and processes them asynchronously in the background. The scheduler implements intelligent batching logic that groups related operations (e.g., multiple document upserts) into single indexing jobs, reducing overhead and improving throughput. Documents flow through a parallel extraction pipeline in the milli crate that tokenizes text via charabia, builds inverted indexes, and creates vector indexes using arroy, with progress tracked via task status endpoints.

Exposes all search, indexing, and administrative functionality through a RESTful HTTP API built on actix-web, with complete OpenAPI 3.0 specification for API documentation and client generation. The API follows REST conventions for resource management (indexes, documents, tasks) with standard HTTP methods (GET, POST, PUT, DELETE) and status codes. The OpenAPI spec is automatically validated and published, enabling API-first development and integration with API documentation tools.

Implements a task queue system where all write operations are enqueued and processed asynchronously, with webhook support for notifying external systems when tasks complete. The IndexScheduler manages the task queue, persisting task state to LMDB and processing tasks in batches. Applications can poll task status endpoints or subscribe to webhooks to receive completion notifications, enabling event-driven architectures where indexing completion triggers downstream processes (e.g., cache invalidation, analytics updates).

Provides dump and export endpoints that serialize the entire index state (documents, settings, tasks) to a portable format that can be restored on another Meilisearch instance. Dumps include all index metadata, documents, and task history, enabling point-in-time backups and zero-downtime migrations between servers. The dump format is version-aware, allowing upgrades between Meilisearch versions with automatic schema migration.

Allows customization of document ranking through a configurable ranking rules system that applies multiple ranking criteria in sequence (e.g., exact match, word proximity, attribute position, typo count, sort order). Rules are evaluated in order, with earlier rules taking precedence, enabling fine-grained control over relevance without modifying the search algorithm. The ranking system supports both built-in rules and custom sort expressions, allowing applications to tune relevance based on business logic (e.g., boosting bestsellers, deprioritizing out-of-stock items).

Provides InstantSearch.js library that integrates with Meilisearch to enable rapid development of search-as-you-type interfaces with minimal code. The SDK handles query execution, result rendering, facet management, and pagination, with support for popular UI frameworks (React, Vue, Angular). The library abstracts away HTTP request management and provides reactive components that automatically update as users interact with search filters and input.

Implements typo tolerance through the charabia tokenization library, which handles misspellings and character variations during both indexing and query processing. The system builds inverted indexes that support fuzzy matching with configurable Levenshtein distance thresholds (typoTolerance setting), allowing queries like 'speling' to match 'spelling'. The tolerance is applied at the token level during query expansion, where the search engine generates candidate tokens within the distance threshold and retrieves documents containing any of those variants.

Pre-computes facet distributions at indexing time by maintaining facet_id_*_docids databases in LMDB for each faceted attribute, enabling instant facet counts without scanning the entire result set. When a search query is executed, the filter system intersects the result set with pre-computed facet buckets to return accurate counts for each facet value. This approach trades indexing overhead for sub-millisecond facet computation, making it ideal for real-time faceted navigation interfaces.

Supports location-based search through a special _geo attribute that stores latitude/longitude pairs for each document. The filter system can evaluate geographic distance expressions (e.g., 'distance(lat, lng) < 10km') to filter documents within a radius, and results can be sorted by proximity to a reference point. The implementation uses LMDB-backed storage for coordinate data and applies distance calculations during filter evaluation, enabling location-aware search without requiring a separate geospatial database.

Parses complex filter expressions into a FilterCondition abstract syntax tree (AST) using the filter-parser crate, enabling boolean logic (AND, OR, NOT), comparison operators (=, !=, <, >, <=, >=), range queries, and nested conditions. The AST is evaluated during search to determine which documents match the filter criteria, with support for filtering on any indexed attribute. This approach separates filter parsing from evaluation, allowing for query optimization and reuse of parsed filter trees across multiple searches.

The milli crate implements a parallel extraction architecture that processes documents through multiple stages: tokenization via charabia, field extraction, inverted index construction, and vector index building using arroy. Documents are processed in parallel batches using Rayon thread pool, with each stage operating on independent document chunks to maximize CPU utilization. The pipeline outputs LMDB-backed indexes that are atomically committed, ensuring consistency and enabling zero-downtime index updates.

Allows modification of index settings (searchable attributes, filterable attributes, sortable attributes, ranking rules) through API calls that trigger automatic reindexing via the IndexScheduler. The system builds a new index in the background while the old index remains queryable, then atomically swaps them upon completion. Settings changes are enqueued as tasks with status tracking, enabling applications to update index configuration without downtime or manual index rebuilding.

Supports querying multiple indexes simultaneously through federated search endpoints that execute the same query across multiple indexes and merge results using configurable weighting. The system executes searches in parallel across indexes, collects ranked result sets, and applies a merge strategy (e.g., round-robin, weighted scoring) to produce a unified result list. This enables applications to search across logically separate indexes (e.g., products, articles, users) in a single request without client-side result aggregation.

Optimizes for real-time search feedback by returning results with minimal latency (sub-50ms target) as users type each character. The system leverages LMDB's memory-mapped I/O and pre-computed indexes to serve results from cache, with query processing optimized for short, incomplete queries. Prefix matching is built into the inverted index structure, enabling efficient retrieval of documents matching partial tokens without scanning the entire index.