Skip to main content
Version: Next

Vector Engines

🎓 Learn how it works with the Amazon S3 Vectors with Spice engineering blog post.

Data sourced by Data Connectors, or views built atop them with vector embedding columns can be indexed and efficiently searched using a vector engine.

A vector engine will store all vector embeddings associated with columns in a dataset/view, provide efficient vector search operations and avoid unnecessary recomputation of embeddings.

A vector engine is configured by setting the vectors configuration. E.g.

datasets:
  - name: dataset_with_embeddings
    vectors:
      enabled: true

For the complete reference specification see datasets.

Supported Vector engines:

NameDescription
s3_vectorsAWS S3 vectors
elasticsearchElasticsearch (Spice.ai Enterprise)

Choosing an engine​

The engines differ in where the index lives:

  • External engines, for high-QPS serving where the index lives outside the Spice process and survives a restart of it. S3 Vectors stores vectors in S3 and pushes predicates on columns marked metadata.vectors: filterable into the index query. Elasticsearch keeps the kNN index in the cluster (and can combine it with full-text in that cluster).
  • No vector engine, when the dataset or view is accelerated with Spice Cayenne. Cayenne indexes embedding columns for vector_search itself, without a vectors block. Its index is exact: it scores every candidate row, so query time grows with the rows searched.

The DuckDB vector engine (vectors.engine: duckdb) is removed in v2.5. See Migrating the DuckDB vector engine.

Add vector search only when semantic retrieval is required. Hybrid retrieval (rrf over vector_search and text_search) is for queries that need both semantic and literal matches. Otherwise stay on SQL, regex, or full-text search. See Where indexes are built.

Limitations
  • A dataset or view must be accelerated (i.e. datasets[].acceleration.enabled: true, see docs) for a vector engine to be provided the appropriate data to ingest.

Vector Engine Docs​