Local mode: writing a sparse vector to a dense field leaves the collection unreadable

[HuaTNA]

Reproduced on upstream dev at
589a87a10679680921571828d800a120425fa807, Python 3.13.13, on 2026-09-20.

Reproduction

from qdrant_client import QdrantClient, models

client = QdrantClient(":memory:")
client.create_collection(
    "example",
    vectors_config={"dense": models.VectorParams(size=2, distance=models.Distance.DOT)},
)
client.upsert("example", [models.PointStruct(id=1, vector={"dense": [1.0, 2.0]})])

try:
    client.upsert("example", [models.PointStruct(
        id=2,
        vector={"dense": models.SparseVector(indices=[0], values=[3.0])},
    )])
except Exception as exc:
    print(type(exc).__name__, str(exc))

print(client.count("example").count)
print(client.retrieve("example", [1, 2], with_vectors=True))
client.close()

Before the proposed fix, the upsert raises TypeError, but the count is now 2.
Retrieving the points then raises IndexError: the failed insertion has left
the point IDs and vector storage out of sync.

Expected: reject the incompatible vector before inserting the invalid point,
and leave the collection usable.

Cause and proposed fix

_validate_point and _validate_named_vectors check that a vector name exists,
but do not check whether it designates sparse storage before dispatching on the
input's type. A SparseVector can therefore pass validation for a dense or
multivector field. Conversely, a dense list can pass validation for a sparse
field. Failures occur later in the mutation path.

The local patch adds a shared sparse-versus-dense kind check in both validation
paths. It raises ValueError before collection state is changed. Batch updates
use these validation paths too. Dense-versus-multivector shape validation is
outside this patch's scope.

Validation

• 18 new regression cases fail before the fix and pass after it.
• Cases cover upsert, update_vectors and batch updates; sparse inputs for dense
  and multivector fields and dense inputs for sparse fields; memory and disk
  storage. They check original data, count, later queries and writes, and
  reopening persistent storage.
• 32 tests pass together across the new test file, test_in_memory.py and
  test_local_persistence.py.
• Ruff formatting, lint of the new test file, and git diff whitespace checks pass.
• A temporary Qdrant v1.19.1 server rejects the nine corresponding single-invalid-
  operation cases with HTTP 400 and preserves the original point and count.

Server comparison limitation: a mixed request containing a valid point before
an invalid one can apply the valid point before returning an error. The local
regression tests preserve local mode's existing prevalidation behavior; they
do not assert that whole-request atomicity matches the server.

Duplicate check and disclosure

GitHub issues and PRs were searched on 2026-09-20 for sparse/dense validation,
wrong vector kinds, SparseVector TypeError, corruption, and
_validate_named_vectors. No exact duplicate was found in those searches.
Related PR #1445 concerns query errors; #1450 concerns non-finite vector values.

AI assistance was used to investigate, implement and test this local patch.
The reproduction and test results above were executed locally with AI assistance.

A focused fix and regression tests are prepared locally. Would this validation approach be welcome? I can submit the patch against dev after confirmation.

[joein]

fixed in #1476
thanks for reporting this!