ragleap-vectorstores v0.4.0 release notes

Added

UpstashBackend - fourth vector backend, via Upstash Vector's managed serverless REST API. Available via the upstash optional extra (pip install ragleap-vectorstores[upstash]).
No local/embedded mode at all - the index (fixed dimension, dense-vs-hybrid type) must already exist, created via the Upstash console. init_schema() verifies compatibility via info() rather than creating anything, and raises a clear error if the index is hybrid-type or has the wrong dimension.
Metadata filtering is a genuine strength here: Upstash's filter= is a real SQL-like string over an actual JSON dict, so arbitrary multi-key filters work natively - more flexible than Redis's document_id-only limitation, without Chroma's $and wrapping either.
similarity_score is used directly from Upstash's query() - its score is already a normalized similarity in [0, 1], unlike Chroma/Redis which return a distance requiring conversion.
delete_document() and chunk-counting use Upstash's native prefix= parameter - simpler than Redis's manual scan-then-delete pattern.
14 new tests against a real Upstash Vector index (no mocks); they skip cleanly if credentials aren't set rather than failing.

See CHANGELOG for full details, including the v0.3.2/v0.3.3 fixes (README gaps, a MODULE LIST parsing regression, and a real RedisBackend data-isolation bug) folded into this release cycle.