Google Cloud debuts agent builder to ease GenAI adoption

Google Cloud is making it easier to create advanced chatbots and other conversational artificial intelligence (AI) capabilities through a new agent-building tool in its Vertex AI platform.

Dubbed Vertex AI Agent Builder, it helps developers with different levels of expertise build and deploy AI agents, through a no-code console for novices and open-source frameworks like LlangChain for expert users.

Google Cloud CEO Thomas Kurian said the tool will make it easy for organisations to harness search and conversational capabilities in a variety of use cases. “It allows you to use natural language – English, Chinese, Spanish – and a variety of other languages to guide the creation of agents and to connect agents to your existing systems,” he added.

At Google Cloud Next ’24 in Las Vegas this week, company executives demonstrated the capabilities of a conversational commerce agent built with Vertex AI Agent Builder that was able to recommend items for a customer based on her preferences and process transactions within the same chat widget.

Vertex AI Agent Builder is already being adopted by the likes of ADT, a supplier of security systems, to build an agent that helps customers set up their home security systems. The Intercontinental Hotels Group is also expected to launch an AI agent to help guests plan their vacations.

The performance of AI agents is only as good as the data that goes into training the AI models that power the agents. As models can only work with information they were trained on out of the box, the use of retrieval augmented generation (RAG) is required to ground their outputs with relevant or additional information.

Google Cloud said in a blog post that developers can use a variety of RAG application programming interfaces (APIs) in Vertex AI Agent Builder for tasks like ranking and retrieval to perform checks on grounding outputs.

“For more complex implementations, Vertex AI Agent Builder also offers powerful vector search to build custom embeddings-based RAG systems. Vector search can scale to billions of vectors, find the nearest neighbors in a few milliseconds, and combine with keyword-based search techniques to ensure the most relevant and accurate responses for users,” it added.

To improve the relevance and accuracy of an agent’s outputs, developers can ground model outputs in Google Search, combining the capability of Google’s large language models (LLMs) with access to the latest information on the internet.

Data connectors can also be used to ingest data from business applications like ServiceNow and Salesforce, enabling an agent’s outputs to be grounded with contextual data. For complex use cases spanning multiple workflows, multiple agents can be deployed, with one agent functioning as the main agent and others as subagents.

For now, the LLMs supported by Vertex AI Agent Builder include four variants of Google’s PaLM 2 for Text (text-bison) model and Gemini 1.0 Pro, which is designed for higher input and output scale, and provides better result quality across a wide range of tasks like text summarisation and sentiment analysis.

Developers can also add a telephone interface to their agents, enabling users to interact with agents through voice calls that make use of the speech-to-text and text-to-speech capabilities in Google Cloud.

In a recent interview with Computer Weekly, Boomi CEO Steve Lucas noted the rise of the agent economy where organisations will have armies of AI agents performing different tasks and automating decisions as they look to build next-generation applications, reduce cost and transform their business.