Vertex AI Vector Search

What Vertex AI Vector Search is

Vertex AI Vector Search is a cloud vector search engine for search, recommendations and generative AI applications. It uses Google Research’s ScaNN algorithm and supports dense embeddings, sparse embeddings, and hybrid search that combines semantic and keyword matching. Numeric and text attribute filters can narrow searches. Google describes uses that include semantic search, retrieval-augmented generation, chatbots and multimodal search.

An index can be deployed to a public or private endpoint, including with Private Service Connect for private access to serving backends from multiple user VPC networks. Documented integrations include RAG Engine, Cloud Storage, Cloud Run, TensorFlow Recommenders and LlamaIndex. Developers can use Python and Java clients, the Python SDK or REST API workflows. Security controls include VPC Service Controls and customer-managed encryption keys. The service is paid and billed by usage: serving is charged per node hour, with separate charges for index builds, updates and streaming updates. Build and update data processing costs $3.00/GiB in all regions; streaming inserts cost $0.45/GiB. Serving rates vary by machine type and region. New Google Cloud customers can receive $300 in free credits to try Google Cloud products.

Who it suits

It suits teams building semantic or hybrid search, recommendations, retrieval-augmented generation, chatbots or multimodal search on Google Cloud. Developers working in Python or Java, or through REST workflows, can connect it with the documented Google Cloud and LlamaIndex integrations.

Limits

Per region and project, quotas allow 5 concurrent index creations, 5 concurrent index updates, 50 deployed index nodes and 100 indexes. The service is usage billed: serving is charged per node hour, build and update data processed costs $3.00/GiB in all regions, and streaming inserts cost $0.45/GiB.

Getting started

1. Open the Vertex AI Vector Search documentation website.
2. Use the documented Python or Java client, Python SDK, or REST API workflow.
3. Choose dense, sparse, or hybrid search for the application.
4. Apply numeric or text attribute filters as needed.
5. Deploy the index to a public or private endpoint.

Verdict

Choose Vertex AI Vector Search for cloud search, recommendations or generative AI retrieval when its embedding modes, filters and endpoint choices fit your application. Consider the usage-based billing and per-project quotas when estimating deployment needs.