An embedding is a model-produced numeric vector representing aspects of an input for similarity tasks. It is derived, versioned data—not meaning, truth, or authorization. Useful vector search requires compatible dimensions, a chosen distance metric, representative evaluation, source lineage, tenant policy, deletion propagation, and a plan for rebuilding when the model or preprocessing changes.
Imagine a model mapping each text chunk to a point in a high-dimensional space. A query maps through the same compatible model; distance retrieves nearby points. Nearness is a learned statistical signal, not a guarantee of relevance or correctness.
An embedding row needs tenant, chunk/version identity, embedding model/version, dimensions, preprocessing/chunker version, vector, creation time, and status. A unique constraint prevents duplicate active output for the same source/generator combination.
Do not mix incompatible vector dimensions or models in one search. Normalize only when required by the metric/model design. Use the operator corresponding to the metric and the matching index operator class later.
It is compelling when relational filters, transactions, access policy, lineage, and vector search belong together and scale fits the extension/provider. Consider specialized systems when measured corpus/query scale, independent scaling, search features, operational separation, or availability needs justify the added data synchronization and governance.
AI-native means embeddings participate in a governed lifecycle: source versioning, evaluation, access, observability, recovery, and deletion. Simply adding a vector column is feature-native, not system-native.
Before installing any extension, create five small 2D points on paper. Calculate Euclidean distance and cosine similarity for one query and show their rankings can differ. Then define an embedding registry row for two model versions and prove a query cannot mix them.
Design embedding_models and chunk_embeddings relations without writing vector-specific SQL yet.
Acceptance criterion: incompatible models cannot be mixed accidentally, source lineage is complete, and every embedding is rebuildable/deletable.