Skip to main content
Use this guide to enable optional meaning-based search in the Studio search modal. Keep Meilisearch setup and text extraction available when comparing results. Studio uses Meilisearch’s native embedding and vector retrieval. PHP does not scan vectors or contact the image-label provider while searching. Configure the embedding model separately from the image-analysis provider:
The example identifies a model, not a guarantee of your installation’s supported languages or relevance. Confirm the model/runtime support in your Meilisearch installation. Studio accepts huggingFace, openAi and rest embedder settings; credentials, REST templates and provider-specific parameters stay server-side. The embedding input is capped at12000bytes, with a default6000-byte document template budget. Use an explicit model/revision and validate your languages. Meilisearch embedding documentation. Run media-studio:search-index --configure, wait for the settings task to succeed, then index/synchronize. A model change can require engine-side re-embedding; do not mix model spaces or assume a live atomic index switch. OCR keeps source-language text. Visual labels/descriptions retain provider output. Multilingual matching is provided by the configured embedding model, without translating each stored file into every language.
lexical is the default. hybrid uses equal lexical/semantic weighting; semantic requests semantic retrieval only. Non-lexical modes are unavailable while disabled. In Spotlight, By meaning searches semantically as you type. Smart uses hybrid retrieval when it expands a search, if semantic search is enabled. The same workspace and file filters apply to indexed results, which open through normal access checks. Expiring cursors bind the query, mode, workspace, actor, index and semantic configuration. Live ranking can change between pages; this is a bounded result window, not a frozen or exhaustive result session. The index receives file names, bounded content, and the latest current-version AI description/non-rejected labels. Results recheck their content and AI projection fingerprints in SQL. Changing/removing a projection therefore blocks stale hits before the next engine synchronization. Search is not a permission database. Real-engine relevance, provider quality/cost and production-scale throughput need testing against your deployment and corpus; automated transport tests alone do not establish those results.