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_scoreonPOST /query: candidates scoring below it are dropped before the sort andtop_kcut, so every returned item scores at leastmin_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
wherefilter 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/imagefieldsDeveloper 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_kis always greater thank, 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.