Google Cloud expands AlloyDB with new AI functions

Google Cloud has expanded AlloyDB with new AI functions and faster processing options for model-driven database queries. The update adds summarisation and sentiment analysis tools, while making existing AI functions more broadly available.

The database service now includes three new functions: ai.summarize, ai.agg_summarize and ai.analyze_sentiment. These join four existing functions - ai.generate, ai.rank, ai.if and ai.forecast - which are now generally available.

AlloyDB is Google Cloud's PostgreSQL-compatible database service, and the latest changes focus on using large language models directly inside SQL queries. The new functions are intended to summarise long passages of text, create a single summary from grouped records, and classify text as positive, negative or neutral.

Google gave a retail example in which ai.agg_summarize produced a combined view of customer reviews for gaming consoles. In that example, one console's review summary highlighted praise for 4K graphics and frame rates, while noting complaints about fan noise during long sessions.

Speed claims

Google also outlined two methods intended to reduce the cost and delay of running AI-assisted queries across large datasets.

The first, called smart batching, groups AI function calls so repeated prompt instructions are not sent for every row. It is currently available for ai.if and ai.rank. In internal testing, Google said, it improved performance by as much as 2,400 times, or up to 10,000 rows a second, compared with processing one row at a time.

The second method, described as an optimised mode for ai.if, uses a smaller proxy model trained on a sample of a customer's data and the outputs of a larger language model. Once prepared, the query can run locally in the database, with AlloyDB falling back to the external model if accuracy is too low or no trained model is available.

In internal testing, Google said, this approach delivered throughput of up to 100,000 rows a second, a 23,000-fold improvement, and a 6,000-fold reduction in costs. The optimised mode for ai.if is in preview.

Database workflow

The product changes reflect a wider industry effort to move AI processing closer to operational data, rather than exporting records into separate pipelines. In Google's example, a business could use ai.generate to turn unstructured customer feedback or system logs into structured JSON inside the database.

Sample output shared by Google showed raw log entries converted into fields such as error code, service name and root cause. The examples covered incidents including database connection timeouts, token validation failures, memory errors, missing resources and third-party gateway timeouts.

Google also described a product search example that used ai.if to identify underwater cameras rated for depths of 60 metres or more. The aim was to show how an AI-based filter could apply a condition expressed in natural language, where standard semantic or keyword search might return items that appear relevant but do not meet a numeric requirement.

In that example, the database returned products whose descriptions explicitly referred to 60-metre waterproofing. Google used the same scenario to illustrate both smart batching and the optimised ai.if workflow, in which a proxy model is trained with a PREPARE statement and then used with EXECUTE.

Broader push

The move adds to Google Cloud's effort to position AlloyDB as a database service for applications that combine transactional data with AI-driven search, retrieval and analysis. The company has already added vector search, hybrid search and natural language query tools to the product.

For database buyers, the more immediate question is whether the claimed gains in speed and lower costs will hold up in production environments. Google's figures were based on internal testing, and the largest improvements are tied to preview features and to ai.if rather than the full set of AI functions.

Even so, the additions suggest a clear product direction: Google wants more AI work to happen inside the database engine itself, from summarising grouped reviews to filtering products and extracting structure from raw text, while reducing the need for external processing pipelines.

According to Google, smart batching is available in preview for ai.if and ai.rank, while array-based functions are generally available for all large-language-model-based AI functions.