Vertex AI Search vs Vertex AI Vector Search: Two Different Products

If you are using Google Cloud for enterprise search or AI retrieval, you will encounter two products with similar names and very different capabilities: Vertex AI Vector Search and Vertex AI Search. They share a brand name and a GCP project, but they use different APIs, have different capability profiles, and serve different use cases.

Vertex AI Vector Search (formerly Matching Engine)

Vertex AI Vector Search is a managed approximate nearest neighbor (ANN) index. You upload embeddings, deploy an index to an endpoint, and query via the findNeighbors API on a per-endpoint public hostname — not on the standard aiplatform.googleapis.com address. It is a pure vector similarity service: no text is stored alongside the vectors, no keyword search, and no listing API.

In Gateco, the vertex_ai_vector_search connector type connects to this service. It supports vector search only. Because no text is stored in the index, the text field in every result is always an empty string — all readable content must come from sidecar metadata resolution (Gateco's own registry). Retroactive registration is not available for this connector type because there is no listing API to enumerate existing vectors. When to use it: you have built your own embedding pipeline, you need the lowest possible vector search latency, and you are already managing text storage separately.

Vertex AI Search (Discovery Engine API)

Vertex AI Search is a full document search service built on Google's Discovery Engine. It handles ingestion, indexing, keyword search, semantic search, and hybrid (RRF-fused) retrieval. Unlike Vector Search, it stores document content and supports both keyword and vector queries against the same index. It also has a listing API, so Gateco can enumerate indexed documents for retroactive resource registration.

The Gateco vertex_ai_search connector type connects to the Discovery Engine API at discoveryengine.googleapis.com — not aiplatform.googleapis.com. The key configuration fields are data_store_id, location (usually "global"), and optionally serving_config_id (defaults to "default_serving_config"). If you have a regional data store for EU data residency, set location to "eu" rather than "global". Vertex AI Search supports vector, keyword, and hybrid search. Grep is not supported. The alpha parameter for hybrid weighting is accepted but currently ignored — Gateco uses Google's native RRF fusion, since Vertex's internal weighting model does not map cleanly to a simple alpha knob. When to use it: you want Google to manage the full ingestion and indexing pipeline, or you need keyword and hybrid search against Google-hosted content.

Capability comparison

Search modes: Vector Search supports vector only. Vertex AI Search supports vector, keyword, and hybrid. Neither supports grep. Text in results: Vector Search returns empty strings (vectors only). Vertex AI Search returns full document content. Retroactive registration: Vector Search — not available. Vertex AI Search — available. Metadata resolution: both use sidecar by default. Vector Search cannot use inline or SQL view modes. Both use the same service account credentials and the same GCP auth helper in Gateco.

Using both simultaneously

A common architecture uses Vertex AI Vector Search for real-time embedding similarity (where you control the embedding model and need fine-grained latency control) alongside Vertex AI Search for document-level retrieval where Google's ingestion and ranking pipeline is preferable. Both can be connected to Gateco simultaneously as separate connectors, with policies applied independently. The MCP server and SDK both support querying multiple connectors in a single request.