ad
ad

Grounding with Google Search now in Google AI Studio and the Gemini API

Science & Technology

Introduction

Hey everyone, Stephanie Wong here! Have you ever wished that your AI experience could tap into the vast power of Google Search? Well, now you can! With the recent launch of grounding with Google Search in Gemini models, users will benefit from more accurate and up-to-date results. This feature is now available in the Gemini API and Google AI Studio. Grounding with Google Search enables developers to leverage data from Google’s search engine to enhance the accuracy and relevance of responses generated by Gemini's AI models.

To provide more insights about this feature, we welcomed Shista, the Group Product Manager leading the initiative. The key motivation behind grounding with Google Search is to empower developers to obtain better, richer, and more current responses, particularly in applications that demand factual accuracy. Developers have been asking for this capability for quite some time, reflecting its growing necessity in modern AI applications.

Shista showcased a demo using the Google AI Studio interface. In the demo, queries were run first without search grounding enabled. The model incorrectly stated that "Ted Lasso" won an award that it actually hadn’t, highlighting limitations stemming from knowledge cut-off dates often seen in such models.

When Shista activated the search grounding feature, the query provided accurate and rich details about the award, explaining the rival nomination and offering links to original sources. In addition, the interface showcased Google search suggestions at the end, leading users to related queries.

How It Works

At a high level, the grounding feature operates as follows:

  1. Toggle Activation: The first step is deciding whether search grounding should be active.

  2. Classifier Prediction Score: If grounding is enabled, the original query undergoes classification to determine its suitability for grounding, yielding a score between 0 and 1.

  3. Dynamic Retrieval Threshold: This score is then compared to a dynamic retrieval threshold defined by the developer. Based on these parameters, the system determines whether to ground the information by crafting more pertinent queries.

  4. Information Extraction: The system retrieves results from Google Search, ranks them, and blends them into a coherent context to formulate a response in the AI model.

Use Cases

The primary use cases for grounding include applications where accuracy and factual correctness are vital. It especially benefits research tools by improving answer quality with the added depth and detail retrieved from search. Other notable use cases include translation applications and technical troubleshooting tools, aptly utilizing Google search for enhanced context and data access.

Given that only Google can offer this unique integration with Google Search, the development process involved collaboration across multiple Google sectors, including Google Cloud, Google DeepMind, and Google Search itself.

Best Practices for Developers

To maximize the effectiveness of grounding with Google Search, developers should consider several best practices:

  1. Dynamic Retrieval Threshold: Test different threshold values to find the best fit for their specific applications.

  2. Cost Considerations: Be aware of cost implications associated with grounding, experimenting with dynamic retrieval settings for a beneficial balance.

  3. Supporting Links: Display the citation sources provided in the responses to encourage credibility and allow users to explore more in-depth information.

  4. Search Suggestions: Use Google search suggestions in applications for enhanced user experience.

Conclusion

To exemplify the fun use cases, Shista also asked for trending Halloween costumes for 2024. When search grounding was turned on, the AI yielded relevant and current suggestions, showcasing its ability to generate specific and timely content.

For developers eager to explore grounding with Google Search, visiting Google AI Studio or accessing it via the API is the next step.

Thank you for reading, and I hope this new feature helps you enhance your applications!

Keywords

  • Grounding
  • Google Search
  • Gemini API
  • AI Studio
  • Accurate Responses
  • Dynamic Retrieval
  • Use Cases

FAQ

Q: What is grounding with Google Search?
A: Grounding with Google Search allows AI models to access Google Search results to improve the accuracy and freshness of their responses.

Q: How does the grounding feature work?
A: It involves toggling grounding on, using a classifier to predict query suitability, applying a dynamic retrieval threshold, and then extracting and re-ranking information through Google Search.

Q: What are some use cases for grounding?
A: Grounding is beneficial for applications requiring high accuracy, research tools, translation, and technical troubleshooting.

Q: What best practices should developers follow when using grounding?
A: Developers should test dynamic retrieval thresholds, keep cost considerations in mind, display citation sources, and utilize Google search suggestions.

Q: Where can developers try out grounding with Google Search?
A: Developers can experiment with the feature in Google AI Studio or integrate it via the Gemini API.