Google Cloud is expanding the grounding capabilities for AI models in its Vertex AI platform. These new features are meant to make it easier to build more accurate and reliable AI applications.
The company has announced several new features for Vertex AI to make it easier to build AI agents and applications that provide more accurate and helpful responses.
A key enhancement is "Grounding with Google Search." This feature, which is now generally available, includes dynamic grounding, which means that the Gemini model independently decides whether to use Google search results for a query or rely on its own training knowledge.
Google has also introduced a "high-fidelity grounding" mode in Experimental Preview. This feature of the Grounded Generation API aims to further reduce hallucinations by relying solely on the context provided when generating answers, with each sentence linked to a source. Google is using a fine-tuned Gemini 1.5 Flash model for this purpose.
In the third quarter of 2024, Google plans to enable AI models to be based on third-party data sets to improve factual accuracy. The company is working with specialist providers such as Moody's, MSCI, Thomson Reuters and Zoominfo.
In addition, Google is expanding its vector search capabilities with hybrid search, now available in public preview. This combines vector and keyword-based search techniques to improve results.
These new grounding features are designed to address widespread and valid concerns in the business market that AI applications are outdated and error-prone. Google's approach of referencing proprietary or verified data in prompts and databases, as well as incorporating up-to-date content, aims to mitigate these concerns.
