Photo by Solen Feyissa on Unsplash
Google's latest update to its Gemini API and AI Studio introduces a significant new feature: the ability to ground AI responses using data from Google Search. This development promises more accurate and up-to-date responses for developers building AI-based services and chatbots.
The new grounding feature allows developers to connect their AI models with verifiable data from Google's vast search catalogue. This connection helps reduce hallucinations - those pesky instances where AI models confidently state incorrect information. For example, when asked about recent Emmy winners, an ungrounded model might cite outdated information, whilst a grounded model can provide the correct, current answer with additional context and sources.
Implementing this feature is straightforward. Developers can simply toggle a switch and adjust the 'dynamic retrieval' setting to determine how frequently the API should use grounding. Options range from grounding every prompt to a more nuanced approach that uses a smaller model to evaluate whether a prompt would benefit from search data augmentation.
This grounding capability is available for free trials in AI Studio, Google's playground for developers to test and refine prompts. However, Gemini API users will need to be on the paid tier, with a cost of £35 per 1,000 grounded queries.
Importantly, when Google enriches results with search data, it provides supporting links to the underlying sources. This is a requirement for anyone using this feature, ensuring proper credit to publishers and offering users a way to verify information easily.
While AI Studio began as a prompt tuning tool, it has evolved significantly. Google's aim is for developers to experiment with models in AI Studio, then move on to building real applications using the Gemini API.
This update represents a significant step forward in AI development, potentially improving the accuracy and reliability of AI-powered services across various applications.
Comments