Agent Retrieval Vector Store

Agent Retrieval (formerly Vector Search 2.0) is Google's managed vector database on the Gemini Enterprise Agent Platform, generally available at vectorsearch.googleapis.com. The agent-retrieval feature of adk-rag implements the [VectorStore] trait against it, so it slots into the existing RAG pipeline exactly like the pgvector, Qdrant, LanceDB, and SurrealDB backends.

[dependencies]
adk-rag = { version = "2.1.0", features = ["agent-retrieval"] }

Quick start

use adk_rag::VectorStore;
use adk_rag::agent_retrieval::{AgentRetrievalConfig, AgentRetrievalStore};

async fn run() -> adk_rag::Result<()> {
    // Application Default Credentials; project/location from the environment.
    let store = AgentRetrievalStore::new_with_adc(AgentRetrievalConfig::from_env()?)?;

    store.create_collection("docs", 768).await?;
    // upsert / search / delete exactly as with any other VectorStore backend
    Ok(())
}

Mapping onto the platform

Trait operation	Agent Retrieval call
`create_collection`	`POST {parent}/collections` (long-running operation; `ALREADY_EXISTS` is a no-op)
`delete_collection`	`DELETE {collection}` (long-running operation; `NOT_FOUND` is a no-op)
`upsert`	`dataObjects:batchCreate` (≤1000, atomic), falling back to per-object create-else-patch for last-write-wins
`delete`	`dataObjects:batchDelete`, falling back to per-object delete when IDs are missing
`search`	`dataObjects:search` with a `vectorSearch` query

Chunk text and metadata are stored in the Data Object alongside the vector — no companion store is needed. The store runs in BYOE (bring-your-own-embeddings) mode: embeddings come from adk-rag's embedder pipeline, keeping behavior identical across backends. Scores are DOT_PRODUCT distances, which equal cosine similarity for normalized embeddings (higher is more relevant).

Configuration

Option	Effect
`with_collection_prefix("rag-")`	Prefixes every collection ID, isolating this store's collections
`with_existing_collections_only(true)`	`create_collection` validates the collection exists instead of creating it — for deployments that pre-provision with platform tooling
`with_endpoint(...)`	Custom API origin (loopback HTTP allowed for tests)

Collection names are sanitized into RFC 1035 labels (lowercased; invalid characters become hyphens), so my_docs becomes collection ID my-docs. Distinct names that sanitize identically collide — choose names that differ in more than case or punctuation.

Beyond the trait

For callers holding the concrete AgentRetrievalStore, hybrid_search runs a semantic and a keyword query in one dataObjects:batchSearch call and returns the reciprocal-rank-fusion ranked list:

# use adk_rag::agent_retrieval::{AgentRetrievalConfig, AgentRetrievalStore};
# async fn run(store: AgentRetrievalStore, embedding: Vec<f32>) -> adk_rag::Result<()> {
let results = store
    .hybrid_search("docs", &embedding, "quarterly revenue", 10, (0.6, 0.4))
    .await?;
# Ok(())
# }

Platform notes

Regions: Agent Retrieval is regional; supported regions include us-central1, us-east4, us-west1, europe-north1, europe-west2, europe-west4, asia-east1, asia-northeast1, and asia-southeast1.
CMEK: standalone Agent Retrieval supports customer-managed encryption keys via the collection's encryptionSpec (immutable at creation). Pre-provision CMEK collections with platform tooling and use with_existing_collections_only(true).
Vector Search 1.0 is deliberately not implemented. Its index/endpoint infrastructure model fits the VectorStore trait poorly, and Google positions Agent Retrieval as its successor.
Quotas: batch writes are capped at 1000 objects per request and are atomic; the store chunks larger upserts automatically.

Testing

The backend passes the shared VectorStore contract and property suite that all adk-rag backends run, so it is behaviorally interchangeable with the other stores. The #[ignore] live test creates and deletes its own uniquely-named collection (data-plane, like the pgvector and Qdrant live tests):

GOOGLE_CLOUD_PROJECT=my-project GOOGLE_CLOUD_LOCATION=us-central1 \
  cargo test -p adk-rag --features agent-retrieval -- --ignored