orion-search/meilisearch

Orion’s engine adapter for Meilisearch 1.13 and later, on the official meilisearch/meilisearch-php client over Guzzle.

use Orion\Meilisearch\MeilisearchEngine;
use Orion\Registry\EngineRegistry;

$engines = new EngineRegistry(meilisearch: MeilisearchEngine::connect($url, $apiKey));

• Meilisearch is schemaless: the neutral schema becomes index settings plus a declaration document Orion keeps in the index (orion-schema, a tombstone no search returns). inspect() reads the declaration and keeps each field only where the settings still serve it; a field they no longer serve is unknown and fails a plan closed.
• Documents are validated against the declaration before they are sent, so an undeclared field, a wrong type, a keyword over 250 bytes or an id the engine cannot hold fails alone, as a strict mapping would; a task the engine refuses as a whole fails every document it carried, with the engine’s reason. Every write is a task; BulkResult::await() waits for it, flush($index, await: true) waits for every task of the index.
• Versions are Orion’s: a write or delete is compared with the stored version first (stale when not newer), after the engine’s own queued write from this instance has settled, and a delete leaves a tombstone every search excludes. One writer per index, which the outbox is: two processes writing the same index at once can race.
• Lexical clauses are filters on shadow attributes (a text field’s tokens as hex-encoded, letter-delimited strings of at most 500 bytes, chunked with an overlap longer than any phrase; a keyword’s value hex-encoded; hex because the engine’s filters are case-insensitive), never q: Meilisearch’s q requires every word or drops some, tolerates typos and completes the last word, none of which is the portable match. q stays available through the escape hatch, with the host’s own typo, synonym and ranking settings.
• Vectors are user-provided embeddings, ranked by cosine similarity (the one metric available); hybrid fusion is core’s. Facets count within the k nearest, as on every engine, read from the window itself.
• No native aliases: core keeps the indirection. No nested documents. One geo field (_geo). A prefix clause on a number is refused.
• IndexSchema::native('meilisearch', ['settings' => [...]]) merges over the generated settings, except what Orion owns (filterable, sortable and displayed attributes, embedders, pagination, faceting, distinct).
• raw() takes callable(): array returning a search body; scope filters and the tombstone exclusion are ANDed into its filter. Nothing in a search body reads past its filter, so no key is refused.
• The containsFilter experimental feature is turned on at ensure(): Orion’s clauses are CONTAINS and STARTS WITH filters.

Proven by Orion\Testing\EngineConformanceTestCase against Meilisearch 1.56 and 1.13, with the tenancy isolation, determinism, fault-injection and endpoint-fixture suites on both. The property-based query equivalence suite, which holds a vector search to the exact neighbours within a slack, runs on the current line only: 1.13’s approximate index strays further and differently per build, so on that line a vector search returns neighbours the engine ranks nearest, not provably the k nearest.