Starting today, developers building AI-based services and bots using Google's Gemini API and Google AI Studio will be able to base their prompt results on data from Google Search. This should allow for more accurate responses based on more recent data.
As always, developers can try grounding for free in AI Studio. AI Studio is essentially Google's playground for developers to test and adjust their prompts and access the latest large-scale language models (LLMs). Gemini API users must subscribe to a paid tier and pay $35 for every 1,000 grounded queries.
AI Studio recently released a built-in comparison mode that makes it easy to see how the results of well-founded queries differ from those of queries that rely solely on data in the model itself.
The core of grounding is connecting your model to verifiable data, whether it's your company's internal data or, in this case, Google's entire search catalog. This also helps the system avoid hallucinations. In an example Google showed me ahead of today's announcement, in a prompt asking who won the 2024 Emmy Award for Best Comedy Series, the model gratuitously answered “Ted Lasso.” But it was an illusion. “Ted Lasso” won the award, but in 2022. If evidenced, the model provided the correct result (a “hack”), additional context was included, and the source was cited.
Image credit: Google
Turning on grounding is as simple as turning on a switch and determining how often the API uses grounding by changing the “Dynamic Retrieval” setting. You can choose to enable this for all prompts, or use a smaller model to evaluate your prompts to determine whether they would benefit from enriching them with data from Google Search. , it's as simple as choosing a more subtle setting.
“Grounding is useful when asking very recent questions that are beyond the model's knowledge limits, but it's also useful for less recent questions, but where more detailed information is needed,” says Google. said Shrestha Basu Mallick, Group Product Manager. We covered Gemini API and AI Studio. “Some developers may say we're just basing it on recent facts, and they'll set this [dynamic retrieval value] Higher. And some developers might say, “No, I want detailed Google search information for everything.”
Image credit: Google
When Google enriches its results with data from Google Search, it also provides supporting links to the underlying sources. Logan Kilpatrick, who joined Google earlier this year after previously leading developer relations at OpenAI, told me that displaying these links is a requirement of the Gemini license for anyone using this feature. Ta.
“This is very important to us for two reasons. One is that we want to make sure that publishers can earn trust and visibility,” added Basu Mallick. “But secondly, it's also something that users like. When I receive an LLM answer, I often go to Google search to see what the answer is. We've created an easy way to do this. This is highly appreciated by our users.”
In this context, it's worth noting that while AI Studio started out as more of a prompt tuning tool, it's now much more than that.
“Success at AI Studio looks like this: When you come in and try one of our Gemini models, you see that it's actually very powerful and works well for your use case.” Kilpatrick said. “While we do a lot of work to expose developers to potentially interesting use cases on the UI front, our ultimate goal is to keep users in AI Studio and simply create models. The goal is to get the code in the top right corner.[コードを取得]Click to start building something, and you may be able to return to AI Studio and experiment with future models. ”