Proposed Changes

Chroma metadata can contain an id key independently of Document.id. A top-level filter on meta.id currently loses its namespace during conversion and is then sent as Chroma's document IDs parameter

This can return the wrong document. Given document-a with meta.id="target", and a different document whose Document.id is "target" but meta.id="other", filtering meta.id == "target" currently returns the second document. Wrapping that same metadata condition in AND happens to return the first document

Only the original, unprefixed id field now uses the IDs parameter. meta.id remains a metadata comparison, so equality, inequality and membership retain their metadata semantics. Existing validation is extracted into a small helper without changing it

Reproduction

from haystack import Document
from haystack_integrations.document_stores.chroma import ChromaDocumentStore

store = ChromaDocumentStore(client_settings={"anonymized_telemetry": False})
store.write_documents([
    Document(id="document-a", content="A", meta={"id": "target"}, embedding=[1.0, 0.0]),
    Document(id="target", content="B", meta={"id": "other"}, embedding=[0.0, 1.0]),
])
result = store.filter_documents({"field": "meta.id", "operator": "==", "value": "target"})
print([document.id for document in result])
store.close()

Before, this prints ["target"]. After, it prints ["document-a"]. Filtering the unprefixed id == "target" still returns ["target"]

This uses a real in-memory Chroma collection and supplied embeddings, with no model download, API key or mocked document-store result

How did you test it?

The 13 added regressions give 9 failures and 4 passes on the base, then all pass with the fix. They cover metadata equality, inequality, in/not in, empty-string metadata, no match, a logical AND, and a separate Document.id control

Two additional boundary cases exercise real Chroma validation for a scalar membership value and a non-string content value. They verify the integration's filter exception, metadata/content category, and original ValueError cause without mocking the validator or matching Chroma's full error message

• hatch run fmt-check passes, including all 16 Chroma package files
• hatch run test:types passes for all 7 source files
• The full Chroma unit suite passes with 68 passed and 152 integration cases deselected by marker
• The existing real in-memory filtering matrix passes with 40 passed and 10 pre-existing null/date limitation skips
• git diff --check passes

The HTTP-server asynchronous integration suite was not run locally. No dependencies, public parameters, version files or generated changelogs were changed

Notes for the reviewer

AI tooling generated the implementation and regression tests. The patch was checked independently against the original conversion logic and the real Chroma results above. No personal human review is claimed

Checklist

• Read this repository's contributor guidelines and code of conduct
• Checked current issues and open PRs, including fix: Chroma - guard empty filter conditions and increase unit tests coverage #3797, which only addresses empty conditions
• Added converter and real document-store regression tests
• Used a conventional fix title