AlloyDB AI is an advanced extension of AlloyDB for PostgreSQL, engineered to integrate high-performance vector processing directly within the enterprise-grade relational database. It enables developers to build sophisticated, intelligent applications that understand the semantic meaning of data across both structured and unstructured formats. By unifying SQL capabilities with cutting-edge vector search technology, AlloyDB AI provides a singular, AI-ready platform for data management and retrieval-augmented generation (RAG) workflows, eliminating the need to deploy and manage a separate vector database.
Core Architecture and Performance
AlloyDB uses the pgvector extension of PostgreSQL for vector functionality. It supports the popular HNSW index, as well as Scalable Nearest Neighbors (ScaNN) index, a state-of-the-art vector search algorithm derived from years of Google Search research. This deep integration into the PostgreSQL environment allows for high-speed, pgvector-compatible operations and advanced semantic querying.
ScaNN Index Performance
The ScaNN index offers substantial speed advantages over the standard PostgreSQL HNSW (Hierarchical Navigable Small Worlds) index, providing rapid vector operations necessary for demanding AI workloads:
| Operation | Performance Gain (vs. standard PostgreSQL HNSW Index) | Features |
|---|---|---|
| Index Creation | Up to 10x faster | Parallel index build |
| Vector Search Queries | Up to 4x faster | Index auto-maintenance |
| Filtered Vector Search Queries | Up to 10x faster | Enterprise-grade observability |
Adaptive Filtering and Query Optimization
The ScaNN index is deeply integrated with the PostgreSQL query planner to ensure optimal performance for hybrid queries involving both vector similarity and standard structured metadata filters.
- Adaptive Filtering: This mechanism ensures optimal performance when filters, joins, and vector indexes are utilized concurrently. It functions by observing query execution statistics at runtime, learning the selectivity of filters, and adaptively adjusting the execution plan to determine the most efficient ordering of filter application and vector search.
- Inline Filtering: Specific to the ScaNN algorithm, inline filtering is an optimization strategy that executes vector search and metadata filter evaluation in tandem. It leverages secondary PostgreSQL indexes (such as B-trees, GIN, or GiST) to create a shortlist (bitmap) of rows that satisfy the filter criteria. The subsequent vector scan only computes distances for vectors contained within that pre-filtered set, drastically reducing computational overhead, especially when filters are moderately selective.
Seamless Integration with Vertex AI
AlloyDB AI is engineered for built-in integration with the Vertex AI ecosystem, enabling the transformation and enrichment of data directly within the database using SQL.
Automated Embedding Generation
The platform supports automated embedding generation using the google_ml_integration extension. This capability allows operational data (text, images, and video) to be converted into vector embeddings using Vertex AI models, such as text-embedding-005.
- Model Registration: Vertex AI models (embedding and generative) are registered within the database using the
google_ml.create_modelprocedure. - Virtual Columns: Embeddings can be automatically generated and persisted in a generated column using the
embedding()function, ensuring the vector data remains synchronized with the source column data:
ALTER TABLE my_table
ADD COLUMN embedding vector(768)
GENERATED ALWAYS AS (embedding('text-embedding-005', source_column)) STORED;
Querying and Generative AI Enrichment
With vector embeddings stored, developers can perform sophisticated semantic searches using the vector similarity operator (<=>) in standard SQL. Furthermore, the results of these searches can be passed directly to generative AI models, such as gemini-2.0-flash-001, using functions like google_ml.generate_content for RAG-based query enrichment and user-friendly output generation.