GCP – AlloyDB AI drives innovation for application developers

The transformative power of AI and intelligent agents is driving profound changes, where software can understand natural language questions and commands — and even autonomously act on our behalf. At the heart of this revolution is the “AI-ready” enterprise database, an active, intelligent engine that understands the semantics of structured and unstructured data, and uses the power of foundation models to create a platform where you can unlock unprecedented opportunities from enterprise data.

This week at Google Cloud Next, we’re announcing several new capabilities in AlloyDB AI to accelerate intelligent agent and application development. These include advanced semantic search with high-performance filtered vector search, automatic vector index maintenance, and a major increase in the quality of searches using the newly launched AlloyDB AI query engine and the Vertex AI Ranking API. The AI query engine brings AI-powered operators to SQL queries for filtering, as well.

We’re also launching natural language capabilities to provide users and agents with deep insights from natural language questions. Taken together, these innovations position AlloyDB as the foundation for agentic AI, evolving the database beyond data storage and conventional SQL querying to a future where intelligent agents can converse with the data and autonomously explore it on our behalf.

High-performance, high-quality, and easy semantic search

Modern apps require smart data retrieval that combines structured data with unstructured, multimodal data such as text and images. Previously, AlloyDB AI enabled semantic queries over unstructured data, deeply integrating vector search with PostgreSQL so search results are always up to date. Our next set of AlloyDB AI capabilities addresses customer requests for higher performance, better search result quality, and low-cost automated maintenance.

1. Adaptive filtering: This innovative technique, now in preview, can help ensure that filters, joins, and vector indexes deliver optimal performance when used together. Adaptive filtering optimizes the query plan once it learns the actual filter selectivity as it access data, and then can appropriately switch between filtered vector search methods.

2. Vector index auto-maintenance, also in preview, reduces how often you need to rebuild your vector indexes, while ensuring that vector indexes remain accurate and performant even as data changes. You can enable vector index auto-maintenance during index creation or when altering the index. 

3. Reranking: The newly-released AlloyDB AI query engine can enhance semantic search by combining vector search with high-accuracy AI reranking, through the new Vertex AI cross-attention Ranking API. Our reranking capability uses vector search to efficiently generate initial candidates (such as Top N) and then apply the high quality cross-attention Ranking API to accurately determine the final best results (such as Top 10) from those candidates. To give you as much flexibility as possible, AlloyDB AI can connect with any third-party ranking API, including custom ones.

4. Recall evaluator: Now generally available, this capability provides the transparency you need for managing and tuning the quality of vector search results. With a simple stored procedure, you can evaluate end-to-end recall for any query, including complex ones with filters, joins, and reranking.

5. Parallel index build: Now generally available, index build parallelization allows developers to build indexes of up to 1 billion rows in just hours, down from several times that number. To support this capability, AlloyDB AI spins up parallel processes to distribute the workload and create indexes faster.

These improvements are made possible by the deep integration of AlloyDB AI’s Scalable Nearest Neighbors (ScaNN) vector index with the PostgreSQL query planner, and they lead to notably faster performance:

• 10x faster filtered vector search when compared to the hierarchical navigable small world (HNSW) index in standard PostgreSQL.

• 10x faster index creation when compared to the HNSW index in standard PostgreSQL.

• 4x faster vector search when compared to the HNSW index in standard PostgreSQL.

AlloyDB AI natural language

Natural language interfaces on databases showed great progress in 2024, backed by AI technology that turns questions, posed either by end users or by agents, into SQL queries that provide answers. 

To further improve accuracy, a quantum leap was needed. Building on the natural language support announced last year, we’re introducing new capabilities to help you build interactive natural language user interfaces that decipher user intent accurately, and can build highly-accurate mappings of user questions to SQL queries that answer them.

• Disambiguation: Natural language is inherently ambiguous. The AlloyDB AI natural language interface will ask follow-up questions when it needs more information about user intent. Since ambiguity is often rooted deep in the data, the database is the best at solving it.

  For example, a question may refer to “John Smith,” but there may be two John Smiths in the database, or perhaps there’s a “Jon Smith,” whose first name was spelled differently, or even misspelled. AlloyDB concept types and the AlloyDB values index enable finding the relevant entities and their concepts when they’re not obvious from the question.

• High accuracy and intent explanation: AlloyDB AI natural language uses plain templates, which correspond to parameterized SQL queries, and faceted templates for providing highly-accurate, virtually-certified answers to predictable and important classes of questions.

  For example, a retailer’s product search page could theoretically include dozens of product properties — far too daunting for a screen-based faceted search interface. In contrast, a faceted search template, even with one simple search field, can answer any question that directly or indirectly poses any combination of property requirements. AlloyDB can automatically produce templates from query logs, and you can provide additional templates to boost query coverage. To ensure confidence in results, AlloyDB offers a transparent explanation of its interpretation of user inquiries.

• High accuracy and flexibility: For cases where questions are not predictable but question answering must provide flexibility, AlloyDB enables the user to raise accuracy by automatically enriching the context that is used in the mapping of the question to SQL with the rich data found in the schema, the data (such as sample data that can greatly enhance accuracy), and the query logs.

• Parameterized secure views: AlloyDB offers parameterized secure views, a new kind of database view that locks down access to end-user data at the database level, to help protect against prompt injection attacks.

• Beyond AlloyDB with Agentspace: AlloyDB AI natural language is available in Google Agentspace for building your own agents that, for example, may answer questions by combining AlloyDB data with data from other sources, such as the web or another database.

AlloyDB AI query engine

To empower you to build intuitive and powerful AI applications, AlloyDB AI query engine can unlock deep semantic insights from enterprise data through AI-powered SQL operators. AI query engine leverages Model Endpoint Management, a mechanism for calling any AI model on any platform.

Let’s review AlloyDB AI query engine and other capabilities newly available in AlloyDB AI via new AI models:

• AI query engine: AlloyDB SQL now features simple but powerful AI operators — AI.IF() for filters and joins, and AI.RANK() for ordering. These operators use natural language in SQL queries to express the filtering conditions and the ranking criteria. They can use foundation models to bring reasoning and real-world knowledge to SQL queries, and they can use cross-attention models, which also draw their power from foundation models and their real-world knowledge. In particular, AI.RANK() can use the Vertex AI Ranking API to find the most relevant results.

• Multimodal embedding generation: Previously, AlloyDB AI enabled a SQL developer to easily generate embeddings from text in SQL statements. We’ve expanded this capability to generate embeddings for any modality (text, images, and videos) so you can search using any modality.

• Updated text embedding generation: AlloyDB AI query engine provides out-of-the-box integration with the text-embedding generation model from Google DeepMind.

Getting started

We believe today’s AlloyDB AI announcements — enhanced filtered vector search, next-generation natural language support, and the AI query engine — are the foundation for the future of databases. They provide proactive insights for agents that anticipate and act decisively, powered by AI-ready data. AlloyDB AI is building a database revolution, empowering you to step boldly into this intelligent future and unlock your data’s boundless potential.

Start with a simple vector search on AlloyDB with ScaNN today. You can also get started with our newly-launched AI capabilities from the AlloyDB AI signup form.