Hey everyone,
For those who couldn’t make this session last week, I wanted to share some exciting news from the Vertex AI team. Vertex AI just launched the private preview of the new Vertex AI Vector Search 2.0 (VS2.0)!
It is focused on creating a fully managed vector database that simplifies the process of building AI applications, addresses the limitations of previous versions, and improves integration with the open-source ecosystem.
Let’s dive into what’s new.
What’s New in Vector Search 2.0?
Here are five key features in this release:
1. Introducing Collections
The biggest change in VS2.0 is the shift from indexes to Collections. A Collection is a container that stores your vectors and the original JSON data together. This means you no longer need a separate database (like Cloud Storage or BigQuery) to store your source data for retrieval—it’s all in one place.
You can even configure the schema to automatically generate embeddings from your text fields upon insertion. Here’s how simple it is to get started:
# Create a collection for movies and define a schema
# that automatically embeds the "plot" field.
movies_collection = vectorsearch.create_collection(
collection_id="movies",
schema={
"type": "object",
"properties": {
"plot": {"type": "string"},
"year": {"type": "number"},
"plot_embedding": {
"x-dense-vector": {
"dims": 768,
"embedding-config": {
"modelId": "text-embedding-005",
"sourceProperties": ["plot"]
}
}
}
}
}
)
# Add a movie to the collection.
# The `plot_embedding` is populated automatically.
movies_collection.create_data_object(
data_object_id="shawshank-redemption-1994",
data={
"plot": "Two imprisoned men bond over years, finding solace and redemption.",
"year": 1994
}
)
2. Hybrid Search and Ranking
VS2.0 lets you run hybrid queries that combine vector search, full-text search, and semantic search in a single request. We’ve also built in a Semantic Ranker and support for Reciprocal Rank Fusion (RRF) to blend the results into a single, highly relevant list. Querying your data is also extended, with the rich filtering language you need.
# Use natural language and metadata filters to search your data.
results = movies_collection.search_data_objects(
SemanticSearch(
query="a story about hope in prison",
search_vectors=["plot_embedding"],
filter={"year": {"$gt": 1990}}
)
)
3. Zero Infrastructure Management
VS2.0 is designed to be completely serverless. The underlying infrastructure is abstracted away—no more managing VMs. The system automatically tunes and scales to maintain high performance and cost-effectiveness, so you can focus on building your application, not managing infrastructure.
4. New Integrations
VS2.0 is built to connect with the tools you use every day, with first-class support for LangChain, LlamaIndex, and the Agent Development Kit (aka ADK). It’s also deeply integrated with Google Cloud services like BigQuery, allowing you to generate embeddings with minimal data movement. More connectors and integrations are on the way.
5. Performance at Scale
While VS2.0 prioritized ease of use, it didn’t compromise on performance. Under the hood, VS2.0 is powered by Google’s industry-leading ScaNN (Scalable Nearest Neighbors) technology to deliver fast and massively scalable approximate nearest neighbor (ANN) retrieval.
6. Storage-Optimized Search for Cost-Efficiency
VS2.0 introduces a new storage-optimized tier where your ANN index is backed by SSDs rather than more expensive RAM. This makes it significantly more cost-effective, especially for applications with very large datasets and lower traffic patterns.
What’s Next
The private preview is currently available through a REST API. A tutorial notebook to help you get started will be available soon. I will share it later in this thread. Looking ahead, full, language-specific client libraries will be rolled out for the public preview.
While you will explore the Vertex AI Vector Search 2.0, we’d love for you to share your thoughts, questions, and feedback right here.
Stay tuned!