weaviate/weaviate · error

inconsistent results: input=%d neighbors=%d

Error message

inconsistent results: input=%d neighbors=%d

What it means

After fetching nearest words for every input vector, Multi verifies the search engine returned exactly one neighbor set per input; a mismatch means the contextionary search silently dropped or duplicated results. The error protects downstream per-item assignment from desyncing (wrong neighbors attached to wrong objects).

Source

Thrown at modules/text2vec-contextionary/additional/nearestneighbors/extender.go:87

	if in == nil {
		return nil, nil
	}

	vectors := make([][]float32, len(in))
	for i, res := range in {
		if len(res.Vector) == 0 {
			return nil, fmt.Errorf("item %d has no vector", i)
		}
		vectors[i] = res.Vector
	}

	neighbors, err := e.searcher.MultiNearestWordsByVector(ctx, vectors, DefaultK, limitOrDefault(limit))
	if err != nil {
		return nil, errors.Wrap(err, "get neighbors for search results")
	}

	if len(neighbors) != len(in) {
		return nil, fmt.Errorf("inconsistent results: input=%d neighbors=%d", len(in), len(neighbors))
	}

	for i, res := range in {
		up := res.AdditionalProperties
		if up == nil {
			up = models.AdditionalProperties{}
		}

		up["nearestNeighbors"] = removeDollarElements(neighbors[i])
		in[i].AdditionalProperties = up
	}

	return in, nil
}

func NewExtender(searcher contextionary) *Extender {
	return &Extender{searcher: searcher}
}

View on GitHub (pinned to 75aa4b6d11)

Solutions

1. Retry the query — transient index state may resolve it
2. Check contextionary index health/rebuild it if inconsistent results persist
3. Report/inspect the searcher implementation to ensure it returns one result group per input vector

out, err := extender.Multi(ctx, in, limit)
if err != nil && strings.Contains(err.Error(), "inconsistent results") {
  out, err = extender.Multi(ctx, in, limit) // retry once
}

When it happens

Trigger: MultiNearestWordsByVector returns a different number of neighbor groups than the len(in) input results — e.g. internal contextionary index inconsistency, partial failures swallowed upstream, or a searcher implementation not preserving order/count.

Common situations: Degraded or reindexed contextionary state, custom searcher implementations, concurrent mutations during large multi-searches.

Related errors

• item %d has no vector
• replicate deletion: shard=%q %w
• after DocPointerWithScoreList: %w
• after createBlockTerm: %w
• sortedDocPointerWithScoreMerger do: %w

AI-assisted analysis of weaviate/weaviate@75aa4b6d11 (2026-09-04). Data as JSON: /api/errors/9e39ddcc2898a855. Report an issue: GitHub.