What causes "failed to search for nearest neighbors" in weaviate/weaviate?

Inspect the wrapped inner error from Centroids.Search for the root cause.

failed to search for nearest neighbors

Error message

failed to search for nearest neighbors

What it means

HFresh.RNGSelect performs an RNG-based neighbor selection starting from a centroid search (h.Centroids.Search) for the query vector; this error wraps a failure of that underlying nearest-centroid search. Without the candidate centroid list, RNG selection cannot proceed, so add, ensureInitialPosting, and doReassign all fail.

Source

Thrown at adapters/repos/db/vector/hfresh/rng.go:28

//

package hfresh

import (
	"github.com/pkg/errors"
)

// RNGSelect performs a Relative Neighborhood Graph selection to determine
// the postings where the vector should be assigned to.
// It performs a search to find candidate postings, then iteratively selects
// postings that are not too close to already selected ones based on the RNGFactor.
// If `reassignedFromID` is non-zero, the function will abort and return false
// if one of the selected postings is equal to `reassignedFromID`.
func (h *HFresh) RNGSelect(query []float32, reassignedFromID uint64) (*ResultSet, bool, error) {
	replicas := NewResultSet(int(h.replicas))
	candidates, err := h.Centroids.Search(query, h.config.InternalPostingCandidates, nil)
	if err != nil {
		return nil, false, errors.Wrap(err, "failed to search for nearest neighbors")
	}

	for cID, cDistance := range candidates.Iter() {
		cCenter, err := h.Centroids.Get(cID)
		if err != nil {
			return nil, false, errors.Wrap(err, "failed to get candidate centroid")
		}

		tooClose := false
		for _, r := range replicas.data {
			rCenter, err := h.Centroids.Get(r.ID)
			if err != nil {
				return nil, false, errors.Wrap(err, "failed to get replica centroid")
			}
			centerDist, err := h.distancer.DistanceBetweenVectors(cCenter.Uncompressed, rCenter.Uncompressed)
			if err != nil {
				return nil, false, errors.Wrapf(err, "failed to compute distance for edge %d -> %d", cID, r.ID)
			}

View on GitHub (pinned to 75aa4b6d11)

Solutions

Inspect the wrapped inner error from Centroids.Search for the root cause.
Verify centroid store integrity and that vector dimensions match the collection config.
Retry after confirming the shard is healthy; restore from backup if centroids are corrupt.
Avoid shutting down the node while background add/reassign operations run.

Defensive patterns

Strategy: retry

Validate before calling

// validate query dimensionality against collection config before search
if len(query) != expectedDims {
    return fmt.Errorf("query dims %d != index dims %d", len(query), expectedDims)
}

Try / catch

rs, ok, err := h.RNGSelect(query, reassignedFromID)
if err != nil {
    if errors.Is(err, context.Canceled) { return } // shutdown race
    logger.Warnf("RNG select failed, will retry: %v", err)
    retryOrQueueForLater(query)
}

Prevention

Keep vector dimensions consistent with collection schema.
Avoid shutdown while background adds/reassigns run.
Verify centroid store integrity after crashes/restores.
Back up shards; rebuild index if centroids are corrupt.

When it happens

Trigger: RNGSelect(query, reassignedFromID) where h.Centroids.Search(query, h.config.InternalPostingCandidates, nil) returns an error — e.g. corrupted or unreadable centroid store, shutdown during search, or an empty/failed search backend.

Common situations: Shard corruption affecting centroids, searches during node shutdown, dimension mismatches or internal centroid index failures after restore.

Related errors

AI-assisted analysis of weaviate/weaviate@75aa4b6d11 (2026-09-04). Data as JSON: /api/errors/499d773b905ea744. Report an issue: GitHub.