fix: avoid NaN cosine scores for zero-norm embeddings in InMemoryDocumentStore

[i-anubhav-anand]

Related Issues

Self-found bug (no existing issue). InMemoryDocumentStore.embedding_retrieval returns NaN similarity scores when a document (or the query) has a zero-norm embedding, silently corrupting ranking.

Proposed Changes:

For cosine similarity, embeddings are normalized by their L2 norm:

query_embedding /= np.linalg.norm(x=query_embedding, axis=1, keepdims=True)
document_embeddings /= np.linalg.norm(x=document_embeddings, axis=1, keepdims=True)

A zero-norm vector (e.g. a zero embedding, which some models emit for empty/whitespace input) makes this divide by zero, producing NaN scores (numpy even emits a RuntimeWarning: invalid value encountered in divide). NaN then sorts unpredictably and silently corrupts the ranking.

This guards the normalization so zero-norm vectors stay zero (denominator forced to 1.0), giving such documents a cosine score of 0.0 instead of NaN. Non-zero embeddings are unaffected.

Reproduction (before the fix):

from haystack import Document
from haystack.document_stores.in_memory import InMemoryDocumentStore

store = InMemoryDocumentStore(embedding_similarity_function="cosine")
store.write_documents([Document(content="zero", embedding=[0.0, 0.0, 0.0])])
print(store.embedding_retrieval(query_embedding=[1.0, 0.0, 0.0])[0].score)  # nan

How did you test it?

Added test_embedding_retrieval_with_zero_vector_does_not_produce_nan in test/document_stores/test_in_memory.py: a zero-embedding document no longer yields a NaN score (it gets 0.0) while a normal document is unaffected. It fails on main (NaN) and passes with this change. Ran hatch run test:unit test/document_stores/test_in_memory.py (148 passed, 4 skipped), hatch run fmt (clean), hatch run test:types haystack/document_stores/in_memory/document_store.py (mypy clean), and added a release note.

Notes for the reviewer

Behavior for non-zero embeddings is unchanged; only the zero-norm edge case is guarded.

Checklist

• I have read the contributors guidelines and the code of conduct.
• I have added unit tests and updated the docstrings.
• I've used a conventional commit type for my PR title (fix:).
• I have added a release note file.
• I have run pre-commit hooks / hatch run fmt and fixed any issue.

This PR was generated with the help of an AI assistant. I have reviewed the changes, reproduced the bug, and run the relevant tests locally.

[vercel]

@i-anubhav-anand is attempting to deploy a commit to the deepset Team on Vercel.

A member of the Team first needs to authorize it.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
haystack-docs	Ignored	Preview	Jun 16, 2026 9:07am

[julian-risch]

Looks good to me! Thank you for your contribution @i-anubhav-anand !

[github-actions]

Coverage report

Click to see where and how coverage changed

File	Statements	Missing	Coverage	Coverage (new stmts)	Lines missing
haystack/document_stores/in_memory
document_store.py					801
Project Total

This report was generated by python-coverage-comment-action