PaveDB 0.9.3

English

PaveDB 0.9.3 is a patch release focused on raw-vector workflows and operational hardening.

Raw vectors / BYOV. Adds identity, vector-only collections and raw-vector ingest and search; validates vector dimensions and metadata; preserves BYOV vectors across restarts; and exposes vector add and search through the public Python client.
API and input safety. Rejects invalid collection names, overlong document IDs, non-object metadata, invalid search limits, metadata-only multipart ingest, and oversized ingest requests with structured errors. The CLI also returns JSON errors for malformed options.
Recovery, concurrency, and authentication. Preserves per-collection backend and embedder configuration during catalog recovery, keeps search capacity occupied until timed-out or cancelled workers finish, reuses the generated admin key across workers, validates authentication configuration, and protects metrics with admin authentication.
SDK and local-client alignment. Shares the local client surface with pavedb-sdk, scopes query lookup and replay by tenant and collection, aligns collection-move responses and error mapping, and makes BYOV available through the public client API.
Distribution and containers. Moves build-tool configuration to pyproject.toml, generates third-party license data, records a pip freeze snapshot for release tags, builds the CPU container in CI, and includes project metadata and the license in the image.

PaveDB persists per-collection backend and embedder settings in collection.json. When upgrading a legacy custom-model collection, keep its existing catalog.db intact for the first 0.9.3 startup so PaveDB can backfill a missing sidecar before any later catalog recovery.
The documented docker run example now passes PAVEDB_AUTH__GLOBAL_KEY into the container and mounts /root/pavedb, so it uses the operator-selected key and preserves instance data.