Backend clients in CI; account-free live tests for Pinecone, MongoDB Atlas and Qdrant server; 3 adapter fixes

[thorwhalen]

Follow-up to #25, answering two requests: run the qdrant and LanceDB tests in CI, and find a way to test the "account-only" backends without accounts.

Tests

Where	Before (master)	After
GitHub CI, Python 3.10 (test,ci-backends)	210 passed	286 passed, 338 skipped
This VM, every server backend up (dev env, pinecone 9.1)	server backends mostly skipped	631 passed, 13 skipped

The 13 skips are by design: fallback-only tests skip on native-hybrid backends. Dependents are unchanged: ef 111 passed, ir 583 passed, newsmood 192 passed with the same 6 pre-existing config2py failures.

Account-free testing

Backend	How it's tested now
pinecone	Pinecone Local, the official in-memory emulator (ghcr.io/pinecone-io/pinecone-local). It speaks the pre-2026-07 API, so it needs the pinecone SDK < 10; the tests skip with that reason on SDK 10.
mongodb	MongoDB Atlas Local (mongodb/mongodb-atlas-local), with real $vectorSearch.
qdrant (server)	A qdrant/qdrant container, as the new qdrant_server entry next to embedded qdrant.
weaviate, pgvector, redis, elasticsearch	Containers, as before. They are now reachable from the cloud VM too.
turbopuffer	Not possible: no emulator exists, and turbopuffer's docs say to test against production with a throwaway namespace.

tests/docker-compose.yml brings all seven up. DOCKERHUB_MIRROR=mirror.gcr.io pulls Docker Hub images through Google's mirror, avoiding the anonymous rate limit that blocked the previous sweep.

Changes

• ci: new ci-backends extra (qdrant-client, lancedb, pgvector, psycopg[binary]), installed via [tool.wads.ci.install] extras = "test,ci-backends". The Windows job installs these extras as well; it is non-blocking.
• fix(mongodb): recreating a collection under a dropped name hung for 120 s. Atlas briefly lists the old index as DOES_NOT_EXIST.
• fix(pinecone): the adapter failed 19 of 32 contract tests on Pinecone Local; all now pass.
  • A new host= option reaches the emulator, and indexes are opened over plain HTTP for an http:// host.
  • A reversible _→-- name mapping, because Pinecone allows only [a-z0-9-] while Milvus rejects hyphens.
  • Collections created without a dimension are tracked until their first write.
  • IDs are parsed from every Index.list() page shape.
• fix(qdrant): get_collection reads the metric and dimension from the server. A second client used to score l2 collections as cosine, returning raw distances.
• test: new server entries (pinecone, qdrant_server), live regression tests, and a native-async qdrant test against the server.
  • Server probes follow the VD_* overrides.
  • Non-local or unparseable targets are skipped unless VD_ALLOW_REMOTE_TESTS=1, because the fixture deletes every collection it sees.
• docs: tests/README.md, CLAUDE.md, the README contributor section and the vd-setup-backend skill, which now has an account-free Pinecone and Atlas section; its snippets were run.

Issues

• Fixes mongodb: searching a collection recreated under a dropped name hangs 120 s and fails #29, filed here with a reproduction.
• Filed get_collection assumes cosine for collections created by another client (chroma, lancedb, milvus, turbopuffer) #31: the same "assume cosine" bug in the chroma, lancedb, milvus and turbopuffer adapters. It is reproduced for lancedb and chroma and left for a focused fix.
• Commented on Native hybrid_search for the 6 remaining hybrid-capable backends #17 and Native AsyncClient/AsyncCollection for the 10 backends with async SDKs (Phase 2) #20: every remaining item except turbopuffer is now testable without an account.

Adversarial review (Opus subagent)

Two rounds by a separate Opus subagent, briefed to refute the change. It ran everything it suspected, including CI's configuration with the network cut off: 286 passed, 0 failed. Verdict: APPROVE in both rounds, with no blocking findings.

Round 1 (9 non-blocking):

• Fixed:
  • Unreachable Pinecone names like a-_b: names that don't round-trip are now rejected, with unit tests.
  • The test guard could wipe a real account through a VD_* override.
  • Pinecone tests now skip on SDK 10 instead of failing.
  • The milvus wording in CLAUDE.md.
• Accepted:
  • Existing -- index names now list as _; the not-found message explains the mapping.
  • ci-backends is a public extra.
  • A mongodb race path relies on the error text containing "already exists".
  • Named or sparse qdrant collections fall back to cosine, and get_collection makes one extra round trip.
• Corrected in this description: the Windows job also installs the CI extras.

Round 2 (4 non-blocking), all fixed in the last commit:

• Keyword-style Postgres DSNs slipped past the guard.
• Multi-host MongoDB URIs errored instead of skipping.
• The Pinecone version check failed with the legacy package.
• The error wording overstated the name rule.

That last commit touches only test infrastructure and one error string, so it was not re-reviewed. The suites were rerun after it.

For whoever lands it

• Merging publishes a release.
• Pinecone users see collection names mapped (my_docs ↔ index my--docs). Pre-existing hyphenated indexes like my-index are unaffected.
• The pinecone adapter is verified on SDK 9.1 against Pinecone Local. On SDK 10 it imports and its calls exist, but it can't be exercised without a real account.

Fixes #29

🤖 Generated with Claude Code