Release Notes: Vector Retriever#

These release notes describe the Vector Retriever microservice. For sample-application integration context, see the Video Search and Summarization release notes.

Version 2026.3.0-rc1#

Release Date: November 6, 2026

New:

  • Optional min_score on POST /query: candidates scoring below it are dropped before the sort and top_k cut, so every returned item scores at least min_score. Omitting it keeps the previous behaviour. Scores are compared as returned by the backend, so the filter assumes higher is better (inner product or cosine), the same assumption as the existing descending sort.

Version 2026.2.0#

Release Date: August 4, 2026

New (initial release):

  • Multi-backend retriever support (vdms, milvus, pgvector, faiss)

  • Backend-centric folder structure under src/retriever/backends/

  • Registry-based backend dispatch and filter translation

  • Primary where filter grammar with compatibility aliases (tags, time_filter, filters)

  • Filter capability discovery endpoint (GET /capabilities/filters)

  • Batch query API with partial error handling

  • Image query modality: send an image (base64 or URL) instead of text for visual similarity search, with mutually exclusive query/image fields

  • Developer template and guide for adding new backends

  • Docker Compose overlays for running the retriever against a selected backend

Improved:

  • Hardened the VDMS over-fetch floor so fetch_k is always greater than k, improving recall when filters discard candidates.

  • Backend registry, middleware, and VDMS backend fixes with expanded test coverage.

  • Refreshed the OpenAPI specification.

  • Resolved a copyleft licensing issue in the service dependencies.

  • Documentation fixes and corrections.