AlloyDB AI query engine empowers developers to create smarter apps with fast, intelligent data handling and deep insights.
Deep changes are being sparked by the revolutionary potential of AI and intelligent agents, which allow software to comprehend orders and questions in normal language and even act on behalf on its own. The “AI-ready” enterprise database, a dynamic, intelligent engine that comprehends the semantics of both structured and unstructured data and leverages foundation models to build a platform that enables you to access previously unheard-of possibilities from enterprise data, is at the center of this transformation.
Google cloud revealing a number of new AlloyDB AI features this week at Google Cloud Next to speed up the creation of intelligent agents and applications. These include autonomous vector index management, high-performance filtered vector search with enhanced semantic search, and a significant improvement in search quality through the use of the recently released Vertex AI Ranking API and AlloyDB AI query engine. Additionally, the AI query engine introduces AI-powered operators to filter SQL queries.
In order to give users and bots profound insights from natural language queries, They are also introducing natural language capabilities. When combined, these developments establish AlloyDB as the cornerstone of agentic AI, transforming the database from a place for storing data and performing standard SQL queries to one where intelligent agents may interact with the data and do independent research on behalf.
High-performance, high-quality, and easy semantic search
Smart data retrieval that blends organized and unstructured, multimodal data like text and images is necessary for modern apps. Semantic queries over unstructured data were previously made possible by AlloyDB AI, which also deeply integrated vector search with PostgreSQL to ensure that search results are constantly current. Google cloud next AlloyDB AI features respond to user requests for improved search result quality, increased performance, and inexpensive automatic maintenance.
Adaptive filtering: This cutting-edge method, which is currently in preview, can guarantee that filters, joins, and vector indexes work as efficiently as possible when combined. After learning the true filter selectivity as it accesses data, adaptive filtering optimizes the query plan and can then transition between filtered vector search algorithms appropriately.
Vector index auto-maintenance: vector index auto-maintenance minimizes the frequency of vector index rebuilds while guaranteeing that vector indexes maintain their accuracy and performance even when data changes. When creating an index or making changes to an existing index, you can activate vector index auto-maintenance.
Reranking: Using the new Vertex AI cross-attention Ranking API, the recently launched AlloyDB AI query engine may improve semantic search by fusing vector search with high-accuracy AI reranking. In order to reliably select the ultimate best results (like Top 10) from those candidates, reranking capability applies the high quality cross-attention Ranking API after effectively generating initial candidates (like Top N) using vector search. AlloyDB AI can interface with any third-party ranking API, including custom ones, to provide you with the greatest amount of flexibility.
Recall evaluator: Currently widely accessible, this feature offers the transparency required to control and improve the caliber of vector search results. You may assess end-to-end recall for any query, including intricate ones including filters, joins, and reranking, using a straightforward stored procedure.
Parallel index build: Previously several times that amount, index build parallelization is now widely available and enables developers to create indexes with up to 1 billion rows in a matter of hours. AlloyDB AI supports this feature by launching parallel processes to divide the burden and produce indexes more quickly.
These enhancements result in noticeably quicker performance and are made possible by the thorough integration of the Scalable Nearest Neighbors (ScaNN) vector index from AlloyDB AI with the PostgreSQL query planner:
- 10 times faster filtered vector search than ordinary PostgreSQL’s hierarchical navigable small world (HNSW) index.
- Compared to the HNSW index in regular PostgreSQL, index construction is ten times faster.
- Compared to the HNSW index in ordinary PostgreSQL, the vector search is four times faster.
AlloyDB AI natural language
In 2024, natural language interfaces on databases made significant strides, supported by AI technology that converts inquiries whether from agents or end users into SQL queries that yield responses.
A quantum leap was required to increase accuracy even further. Its launching new features to let you create interactive natural language user interfaces that accurately interpret user intent and can create extremely accurate mappings from user queries to SQL queries that provide answers, building on the natural language support that was revealed last year.
Disambiguation: Ambiguity is inherent in natural language. When the AlloyDB AI natural language interface requires additional information about the user’s intent, it will pose follow-up queries. The database is the best at resolving ambiguity since it is frequently deeply ingrained in the data.
For instance, a question might mention “John Smith,” but the database might have two John Smiths, or there might be a “Jon Smith” whose initial name was misspelled or spelled differently. Finding the pertinent entities and their ideas when they are not immediately apparent from the question is made possible by AlloyDB concept types and the AlloyDB values index.
High accuracy and intent explanation: AlloyDB AI natural language provides extremely accurate, virtually certified responses to significant and predictable classes of questions using faceted templates and plain templates that match parameterized SQL queries.
Theoretically, a screen-based faceted search interface could not handle the thousands of product properties found on a retailer’s product search page. A faceted search template, on the other hand, can respond to any query that directly or indirectly raises any combination of property requirements, even with just one basic search field. To increase query coverage, you can supply extra templates in addition to the ones that AlloyDB can automatically generate from query logs. AlloyDB provides a clear description of how it interprets user queries to guarantee confidence in the results.
High accuracy and flexibility: In situations where questions are unpredictable but response must be flexible, AlloyDB allows the user to increase accuracy by automatically adding rich data from the schema, the data (including sample data, which can significantly improve accuracy), and the query logs to the context used when mapping the question to SQL.
Parameterized secure views: To assist defend against rapid injection attacks, AlloyDB provides parameterized secure views, a novel type of database view that restricts access to end-user data at the database level.
Beyond AlloyDB with Agentspace: Google Agentspace offers AlloyDB AI natural language if you want to create your own agents that, for instance, may respond to queries by fusing AlloyDB data with information from other databases or the internet.
AlloyDB AI query engine
Through AI-powered SQL operators, the AlloyDB AI query engine can extract deep semantic insights from enterprise data, enabling you to create user-friendly and potent AI applications. Model Endpoint Management, a way to call any AI model on any platform, is used by AI query engines.
Let’s examine the query engine in AlloyDB AI as well as additional features made possible by new AI models:
AlloyDB is an AI query engine: Simple yet effective AI operators, AI.IF() for filters and joins and AI.RANK() for ordering, are now available in SQL. These operators express the ranking criteria and filtering conditions in SQL queries using natural language. They can employ cross-attention models, which also derive their strength from foundation models and their real-world knowledge, and they can use foundation models to provide logic and practical knowledge to SQL queries. Specifically, the Vertex AI Ranking API can be used by AI.RANK() to retrieve the most pertinent results.
Multimodal embedding generation: Previously, AlloyDB AI made it simple for a SQL developer to create multimodal embeddings from text in SQL statements. To enable you to search using any modality, it have extended this feature to create embeddings for any modality (text, photos, and videos).
Updated text embedding generation: The Google DeepMind text-embedding generation model is integrated out-of-the-box by the AlloyDB AI query engine.
Beginning
Google cloud think that the AlloyDB AI query engine, next-generation natural language support, and improved filtered vector search announced by AlloyDB today lay the groundwork for databases of the future. With the use of AI-ready data, they offer proactive insights for agents that anticipate and take decisive action. The database revolution being built by AlloyDB AI will enable you to confidently enter this intelligent future and unleash the limitless possibilities of your data.
