Listen: Introducing Grounding Classifier
An analysis of Gemini 2.5 Pro's search grounding capabilities and the development of a prompt grounding classifier trained on 10,000 collected prompts.
Transcript
We wanted to understand how Google decides when a search query needs real-time grounding. To find out, we analyzed ten thousand prompts using Google's Gemini Pro model with search grounding enabled.
Our test showed that Gemini’s grounding behavior matches Google's official developer documentation, relying on a default retrieval threshold to decide when to search the live web. For instance, a query about India's current population gets grounded to real-time search results, while a simple request for a computer joke does not.
Using the results from these ten thousand prompts, we built a robust training dataset. We also generated synthetic data to balance out the examples. With this dataset, we fine-tuned Microsoft’s DeBERTa model to replicate Google's internal classifier.
This has given us a commercial-grade model that we now use in our own machine learning toolkit and data processing pipelines to predict whether any given query deserves search grounding.