new: QdrantServerless client prototype

[generall]

Stacked on #1394 (CI congruence fixes) to keep integration tests green; merge #1394 first, then retarget this PR to dev.

What

Prototype of a dedicated client for Qdrant Serverless, parallel to the regular QdrantClient, per the serverless client proposal.

from qdrant_client.serverless import QdrantServerless, DenseVectorConfig, Distance

client = QdrantServerless(url="https://serverless....qdrant.io", api_key="...")

client.create_collection(          # serverless-specific input
    "my-collection",
    dense_vectors=DenseVectorConfig(size=1536, distance=Distance.COSINE),
)
client.get_collection("my-collection")   # serverless-specific output
client.query_points("my-collection", query=[...])  # same as regular qdrant

Why

Serverless is not 100% identical to a regular cluster deployment: collection management exposes only a simplified tenant-facing config (no quantization/WAL/segments), and point operations don't support read consistency, shard selection, write ordering, or filtered updates. A separate client keeps those parameters out of the serverless interface instead of forcing one client to paper over the divergence.

How

Everything lives in a dedicated qdrant_client/serverless/ module; nothing is added to the top-level package, so regular users don't pay any import cost.

• gRPC stubs generated from qdrant-cloud-public-api's serverless/collections.proto with the same pinned grpcio-tools 1.62.0 as the existing stubs (tools/generate_serverless_grpc_client.sh regenerates). The proto is renamed to serverless_collections.proto because the protobuf descriptor pool registers files by name and collections.proto is already taken by the regular client — without the rename, importing both clients in one process crashes. Generated types stay internal.
• Models: hand-written pydantic models (v1.10+ compatible) mirroring the tenant-facing config — DenseVectorConfig, SparseVectorConfig, PrecisionTier, the eight payload-index types, CollectionConfig, CollectionInfo, CollectionSummary. The existing Distance / TokenizerType enums are reused so values look familiar. These may later be replaced with models generated from a serverless OpenAPI spec.
• Client: QdrantServerless wraps an internal QdrantRemote(prefer_grpc=True, check_compatibility=False) and delegates point operations to it, reusing all existing conversion machinery; collection methods talk to the new CollectionsService stub on the same channel (one place for TLS, api-key metadata, user-agent). The API divergence is expressed purely in signatures:
  • point methods (query_points, upsert, retrieve, scroll, count, delete, set_payload, delete_payload) are the regular ones minus consistency / shard_key_selector / ordering
  • delete / payload methods take ids only, since serverless rejects filtered updates
  • create_collection accepts a bare config as the unnamed "" vector and returns the service's result string ("created" / "already exists"); get_collection returns CollectionInfo with exists / point_count instead of raising on missing collections
  • default grpc_port is 443, not 6334

Not included (mechanical follow-ups)

query_batch_points / batch_update_points / update_vectors, update_collection (no such RPC yet).

Async client

AsyncQdrantServerless is generated from the sync client the same way the regular async client is: tools/async_client_generator/serverless_generator.py runs as part of tools/generate_async_client.sh, delegates to AsyncQdrantRemote, awaits the CollectionsService stub RPCs on the aio channel, and drops the sync context manager (the regular async client has none either). The generated file is verified by tests/async-client-consistency-check.sh alongside the other generated clients.

Testing

• tests/test_serverless.py: model↔gRPC round-trips, proto3 optional-field presence, offline client construction (headers, port, TLS)
• mypy clean on the new module
• not yet run against a live serverless endpoint

🤖 Generated with Claude Code

[netlify]

✅ Deploy Preview for poetic-froyo-8baba7 ready!

Name	Link
🔨 Latest commit	4254cc6
🔍 Latest deploy log	https://app.netlify.com/projects/poetic-froyo-8baba7/deploys/6aa12f438f5e0200087234a1
😎 Deploy Preview	https://deploy-preview-1393--poetic-froyo-8baba7.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.

---

To edit notification comments on pull requests, go to your Netlify project configuration.

[qdrant-cloud-bot]

Documented the serverless client as in development — do not use yet in package/client/models/conversions docs, the async client (regenerated), and the example.

[joein]

The example code crashes with the following error:

created
[CollectionSummary(collection_name='my-collection', point_count=None)]
exists=True config=CollectionConfig(dense_vectors={'': DenseVectorConfig(size=4, distance=<Distance.COSINE: 'Cosine'>, multivector=False, precision_tier=None)}, sparse_vectors={}, payload_indexes={'color': KeywordIndex(type='keyword')}) point_count=None
Traceback (most recent call last):
  File "../qdrant/qdrant_client/examples/serverless_client.py", line 52, in <module>
    main()
  File "../qdrant/qdrant_client/examples/serverless_client.py", line 45, in main
    print(client.query_points("my-collection", query=[0.1, 0.2, 0.3, 0.4]))
  File "../qdrant/qdrant_client/qdrant_client/serverless/client.py", line 355, in query_points
    return self._remote.query_points(
  File "../qdrant/qdrant_client/qdrant_client/qdrant_remote.py", line 535, in query_points
    res: grpc.QueryResponse = self.grpc_points.Query(
  File "../qdrant/qdrant_client/venv/lib/python3.10/site-packages/grpc/_interceptor.py", line 276, in __call__
    response, ignored_call = self._with_call(
  File "../qdrant/qdrant_client/venv/lib/python3.10/site-packages/grpc/_interceptor.py", line 331, in _with_call
    return call.result(), call
  File "../qdrant/qdrant_client/venv/lib/python3.10/site-packages/grpc/_channel.py", line 447, in result
    raise self
  File "../qdrant/qdrant_client/venv/lib/python3.10/site-packages/grpc/_interceptor.py", line 314, in continuation
    response, call = self._thunk(new_method).with_call(
  File "../qdrant/qdrant_client/venv/lib/python3.10/site-packages/grpc/_channel.py", line 1182, in with_call
    return _end_unary_response_blocking(state, call, True, None)
  File "../qdrant/qdrant_client/venv/lib/python3.10/site-packages/grpc/_channel.py", line 999, in _end_unary_response_blocking
    raise _InactiveRpcError(state)  # pytype: disable=not-instantiable
grpc._channel._InactiveRpcError: <_InactiveRpcError of RPC that terminated with:
	status = StatusCode.INVALID_ARGUMENT
	details = "Not existing vector name error: "
	debug_error_string = "INVALID_ARGUMENT:Not existing vector name error: "
>

In the beginning, it failed just on the first attempt and then worked when I reran it.
I tried running it again, and got the problem with not existing vector name 5 times in a row before it worked.

[qdrant-cloud-bot]

Updated this PR to match the latest serverless collections API in qdrant-cloud-public-api:

• Pagination (ListCollections): get_collections(limit=..., offset_token=...) now returns CollectionsList with next_offset_token
• Keyword prefix + text analysis (stopwords / stemmer / ascii_folding) on models and conversions
• Proto sync script strips server-only buf.validate annotations

Tests: pytest tests/test_serverless.py (4 passed)