Vald

What Vald is

Vald is a free, open-source engine for distributed approximate nearest-neighbor search across dense vectors. It uses the NGT algorithm by default and says users can switch to another vector search engine. Vald can index data asynchronously while continuing to answer search requests. Its Kubernetes and cloud-native design supports horizontal scaling for memory and CPU. For resilience, it supports index replication and can rebalance replicas automatically when an agent goes down. Index backups can use Persistent Volumes, S3-compatible object storage, or both; saved indexes can be restored when an agent pod starts. Ingress filters can vectorize or preprocess data, and egress filters can post-process results, provided each filter implements Vald’s gRPC interface. Official client libraries are available for Go, Java, Python, and Node.js, with gRPC protobufs published for other languages. Deployment documentation covers Kubernetes with Helm or the Helm operator, plus a tutorial for running an agent on Docker. Getting started requires a Kubernetes cluster, Helm v3 or later, and Go v1.15 or later. Kubernetes network policies can control cluster traffic when the required prerequisites are met.

Who it suits

Vald is for teams building dense-vector search into systems deployed on Kubernetes or Docker. Its listed client libraries and gRPC protobufs accommodate development in several languages.

Limits

The free Vald plan is an open-source, self-deployed project. Its getting-started guide requires a Kubernetes cluster, Helm v3 or later, and Go v1.15 or later.

Getting started

1. Prepare a Kubernetes cluster.
2. Install Helm v3 or later and Go v1.15 or later.
3. Deploy Vald on Kubernetes using Helm or the Helm operator.
4. Alternatively, follow the Docker tutorial to run a Vald agent.
5. Use an official Go, Java, Python, or Node.js client library, or the published gRPC protobufs.

Verdict

Vald offers distributed dense-vector search with asynchronous indexing, scaling, and index backup options. Deployment requires Kubernetes, Helm, and Go according to its getting-started requirements.