What it does

Vald is a distributed approximate-nearest-neighbor search engine for dense vectors. It runs on Kubernetes, automatically indexes incoming vectors, backs up those indexes, and scales horizontally to handle billions of feature vectors. The project targets anything you can embed—images, audio, text, binary blobs—and exposes it through load-balanced gateways with filtering support.

The interesting bit

Rather than building its own ANN algorithm from scratch, Vald embeds NGT (Neighborhood Graph and Tree), which the project bills as the fastest option on ann-benchmarks.com. The actual engineering work is in the surrounding machinery: Kubernetes-native service discovery, sidecar-based index management, and a Helm operator that treats vector search as just another stateful workload.

Key highlights

• Built explicitly for Kubernetes 1.19+ with Helm-based deployment
• Automatic vector indexing and index backup (not manual reindexing)
• Horizontal scaling for billion-vector datasets
• Multiple gateway types: load-balancing and filter gateways
• Official SDKs available; demo repository provided
• CNCF Landscape listed; used in production by LY Corporation and Japan Search

Caveats

• Requires AVX2 instructions, which limits hardware compatibility
• Default Helm install pulls nightly images, not stable releases
• README claims “fastest” ANN algorithm but doesn’t provide self-run benchmarks

Verdict

Worth evaluating if you’re already running Kubernetes at scale and need managed vector search without leaving the cluster. Skip it if you want a single-binary local solution or need to support non-AVX2 hardware.