Vald

Vald is a distributed vector search engine for developers building similarity search into applications. It runs on a self-hosted Kubernetes cluster and finds approximate nearest neighbors in dense vectors using NGT. It's open source under the Apache 2.0 license, with an architecture designed to spread large vector collections across multiple machines.

Its automatic indexing lets searches continue while it builds indexes, avoiding the pauses that graph indexing can otherwise cause. Vald distributes indexes across agents and keeps replicas on multiple agents. If an agent goes down, it automatically rebalances those replicas. You can scale memory and CPU capacity horizontally as your collection grows, including workloads with billions of feature vectors.

Vald can search representations of images, audio, video, text or binary data once they're converted into vectors. That makes it relevant to teams building image search, recommendation systems or recognition features. It searches supplied vectors; your application needs to handle the conversion of its data.

For applications with their own processing requirements, Vald supports customizable input and output filters through its gRPC interface. Client libraries cover Go, Java, Node.js and Python. You can adjust vector dimensions and replica counts, while automatic backups to object storage or persistent volumes support disaster recovery. The NGT agent requires a CPU with AVX2 instructions.