Version: Next

Embeddings (Experimental)

caution

This is an experimental feature. Configuration, data format, index layout, and APIs may still change in breaking ways between releases.

Embeddings add meaning-based capabilities on top of the Generic Data Index:

Semantic search — queries match by meaning instead of exact words ("British roadster" finds an MG even if no field contains "roadster"), including multilingual queries.
Image search — text queries against image assets ("yellow van"), no tagging required.
Duplicate detection — exact duplicates via a content hash, near-duplicates via vector and MinHash similarity, for both objects and images.

The feature is off by default and calls an external inference service you provide — the bundle ships no model. See Getting Started to turn it on.

How it works

When an element is saved or re-indexed, its content becomes an embedding input — a composed text for data objects, a thumbnail for image assets. An inference service you configure turns that input into a vector, and the vector is stored in the element's regular search index document. A search embeds the query with the same (or a paired cross-modal) model and runs a kNN query against those fields.

Two properties are worth knowing up front because they shape day-to-day operations:

Vectors are reused, not recomputed. Each stored vector carries an identity key derived from the model and the exact input, so re-indexing unchanged content costs no inference calls — see Caching.
Generation is asynchronous and never blocks. It runs through Symfony Messenger; if the inference service is unavailable, the element is indexed without vectors, marked for backfill, and repaired later. Saving and indexing should never fail because of embeddings.

How it works​

Further reading​

How it works

Further reading