microsoft/semantic-kernel · error · ValueError

Invalid vectors, cannot compute cosine similarity scoresfor

Error message

Invalid vectors, cannot compute cosine similarity scoresfor zero vectors{embedding_array} or {embedding}

What it means

Raised as a ValueError in chroma_compute_similarity_scores (chroma/utils.py) when cosine similarity cannot be computed for ANY vector in the batch. The function precomputes valid_indices = (query_norm != 0) & (collection_norm != 0); when none are valid it raises. This happens when the query embedding is a zero vector (making all collection vectors invalid) or when every collection embedding is a zero vector. Non-zero vectors mixed with some zero vectors only trigger a warning, not this error.

Source

Thrown at python/semantic_kernel/connectors/memory_stores/chroma/utils.py:121

    # Compute indices for which the similarity scores can be computed
    valid_indices = (query_norm != 0) & (collection_norm != 0)

    # Initialize the similarity scores with -1 to distinguish the cases
    # between zero similarity from orthogonal vectors and invalid similarity
    similarity_scores = array([-1.0] * embedding_array.shape[0])

    if valid_indices.any():
        similarity_scores[valid_indices] = embedding.dot(embedding_array[valid_indices].T) / (
            query_norm * collection_norm[valid_indices]
        )
        if not valid_indices.all():
            logger.warning(
                "Some vectors in the embedding collection are zero vectors."
                "Ignoring cosine similarity score computation for those vectors."
            )
    else:
        raise ValueError(
            f"Invalid vectors, cannot compute cosine similarity scoresfor zero vectors{embedding_array} or {embedding}"
        )
    return similarity_scores

View on GitHub (pinned to c028a0c7dc)

Solutions

Validate the query embedding is non-zero before searching: if numpy.linalg.norm(embedding) == 0, skip or regenerate the embedding.
Investigate why the embedding model produced a zero vector (API failure, wrong input, unloaded model).
If stored embeddings are zero vectors, re-embed the affected records.
Catch ValueError around the search call and fall back to a different retrieval strategy.

Example fix

// before
matches = await store.get_nearest_matches('docs', embedding, limit=5)  # ValueError if embedding is zero
// after
import numpy as np
if np.linalg.norm(embedding) == 0:
    matches = []
else:
    matches = await store.get_nearest_matches('docs', embedding, limit=5)

Defensive patterns

Strategy: validation

Validate before calling

import numpy as np

def is_nonzero_embedding(emb: np.ndarray) -> bool:
    return float(np.linalg.norm(emb)) != 0.0

if is_nonzero_embedding(embedding):
    matches = await store.get_nearest_matches('docs', embedding, limit=5)
else:
    matches = []

Type guard

import numpy as np

def is_valid_query_embedding(emb) -> bool:
    return (
        isinstance(emb, np.ndarray)
        and emb.size > 0
        and float(np.linalg.norm(emb)) != 0.0
    )

Try / catch

try:
    matches = await store.get_nearest_matches('docs', embedding, limit=5)
except ValueError as e:
    if 'zero vectors' in str(e):
        logging.warning('Zero-vector embedding; regenerating')
        embedding = await regenerate_embedding()
    else:
        raise

Prevention

Validate embedding norm is non-zero before searching.
Investigate embedding models that silently return zero vectors on failure.
Re-embed stored records that are zero vectors.
Catch ValueError around similarity search and fall back gracefully.

When it happens

Trigger: Calling get_nearest_matches with an embedding that is all zeros (e.g. a failed embedding model call returning zeros). Or querying against a collection where every stored embedding is a zero vector. The function is called internally by the store's similarity search.

Common situations: An embedding service returning a zero vector on error/timeout without raising. Sentinel/debug embeddings set to zeros. Dimension mismatch causing a degenerate embedding. A model not yet loaded producing zeros. Note the message has a typo ('scoresfor', no space before the array dump) but the condition is legitimate.

Related errors

AI-assisted analysis of microsoft/semantic-kernel@c028a0c7dc (2026-08-13). Data as JSON: /api/errors/856c69c93db79f2c. Report an issue: GitHub.