Search grounding is not working - Error: use google_search field instead of google_search_retrieval field

[thenameless7741]

Description of the bug:

I implemented the search grounding feature by following the sample code and it seems to be broken at the moment:

  const modelParams: ModelParams = {
    model: 'gemini-2.0-flash-exp',
    tools: [
      {
        googleSearchRetrieval: {
          dynamicRetrievalConfig: {
            mode: DynamicRetrievalMode.MODE_DYNAMIC,
            dynamicThreshold: 0.7, // default
          },
        },
      },
    ],
  };

Error: [GoogleGenerativeAI Error]: Error fetching from https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash-exp:generateContent: [400 Bad Request] Unable to submit request because Please use google_search field instead of google_search_retrieval field..

A temporary workaround is to replace googleSearchRetrieval with an empty googleSearch:

  const modelParams: ModelParams = {
    model: 'gemini-2.0-flash-exp',
    tools: [
      { googleSearch: {} },
    ],
  };

Simply replacing the field name would result in this error:

Error: [GoogleGenerativeAI Error]: Error fetching from https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash-exp:generateContent: [400 Bad Request] Invalid JSON payload received. Unknown name "dynamicRetrievalConfig" at 'tools[0].google_search': Cannot find field.

[{"@type":"type.googleapis.com/google.rpc.BadRequest","fieldViolations":[{"field":"tools[0].google_search","description":"Invalid JSON payload received. Unknown name "dynamicRetrievalConfig" at 'tools[0].google_search': Cannot find field."}]}]

Actual vs expected behavior:

Expected behavior: the documentation, sample code, and type definition should be updated to use googleSearch field.

Thank you.

[Leask]

The issue is fixed, but no one has merged it: #318

[HunderlineK]

this library is effectively abandoned - no commit merged since about 3 months ago?

[Leask]

this library is effectively abandoned - no commit merged since about 3 months ago?

It seems so, but unfortunately.

[colintoh]

Bumping this. Having the exact same issue.

[Hth4nh]

Bump

[sugoidesune]

I could not make it work...

[hkt74]

Thanks for your feedback.

Starting with Gemini 2.0 models, search grounding tool changed from googleSearchRetrieval to googleSearch, as part of the change, dynamicRetrievalConfig has been removed from the new tool, the model can decide when to use Google Search.

Gemini 1.5 models can still use googleSearchRetrieval

The team is working on updating the api document, thanks @Leask for contributing the samples, I will review your PR shortly.

[Leask]

Thanks for your feedback.

Starting with Gemini 2.0 models, search grounding tool changed from googleSearchRetrieval to googleSearch, as part of the change, dynamicRetrievalConfig has been removed from the new tool, the model can decide when to use Google Search.

Gemini 1.5 models can still use googleSearchRetrieval

The team is working on updating the api document, thanks @Leask for contributing the samples, I will review your PR shortly.

Thank you for the review.
I have already submitted an updated version.
😄

[hkt74]

PR #318 has been merged

[Glutch]

Thanks for your feedback.

Starting with Gemini 2.0 models, search grounding tool changed from googleSearchRetrieval to googleSearch, as part of the change, dynamicRetrievalConfig has been removed from the new tool, the model can decide when to use Google Search.

Gemini 1.5 models can still use googleSearchRetrieval

The team is working on updating the api document, thanks @Leask for contributing the samples, I will review your PR shortly.

I'm using tools: [{google_search: {},},] but the groundingMetadata is almost always empty. "the model can decide when to use Google Search". So enabling this tool, lets the model decide for itself if it needs to use google_search? I need to force it, my use case completely breaks if it doesn't use the grounding. Is there no way to force the grounding?

[Leask]

Thanks for your feedback.
Starting with Gemini 2.0 models, search grounding tool changed from googleSearchRetrieval to googleSearch, as part of the change, dynamicRetrievalConfig has been removed from the new tool, the model can decide when to use Google Search.
Gemini 1.5 models can still use googleSearchRetrieval
The team is working on updating the api document, thanks @Leask for contributing the samples, I will review your PR shortly.

I'm using tools: [{google_search: {},},] but the groundingMetadata is almost always empty. "the model can decide when to use Google Search". So enabling this tool, lets the model decide for itself if it needs to use google_search? I need to force it, my use case completely breaks if it doesn't use the grounding. Is there no way to force the grounding?

Try to ask "tell me the latest movies on IMDB". You will see the result.

[Glutch]

Thanks for your feedback.
Starting with Gemini 2.0 models, search grounding tool changed from googleSearchRetrieval to googleSearch, as part of the change, dynamicRetrievalConfig has been removed from the new tool, the model can decide when to use Google Search.
Gemini 1.5 models can still use googleSearchRetrieval
The team is working on updating the api document, thanks @Leask for contributing the samples, I will review your PR shortly.

I'm using tools: [{google_search: {},},] but the groundingMetadata is almost always empty. "the model can decide when to use Google Search". So enabling this tool, lets the model decide for itself if it needs to use google_search? I need to force it, my use case completely breaks if it doesn't use the grounding. Is there no way to force the grounding?

Try to ask "tell me the latest movies on IMDB". You will see the result.

I have in my prompt "SEARCH THE WEB, use google_search, use grounding. Return in json format information about "Iphone 16" etc etc".

And it returns, without using grounding. "The upcoming iphone 16 is expected to feature blabla".

How do i force it to use grounding when it's obviously needed?

[hkt74]

@Glutch Could you try a more specific prompt?

for example could you return the latest Iphone 16 price?, it returned the results with search sources

I recommend to play around it in AI Studio first to test the prompt and tools, thanks

[HungGithubUser]

Added another PR to use googleSearch in typescript

[Glutch]

@Glutch Could you try a more specific prompt?

for example could you return the latest Iphone 16 price?, it returned the results with search sources

I recommend to play around it in AI Studio first to test the prompt and tools, thanks

Two months later, this issue still feels critical. I've built multiple services that heavily rely on search grounding, and I consistently struggle to get long prompts to trigger a search. We need a boolean option to force a search, similar to how dynamic_threshold worked before. Even better would be to provide the AI with exactly what to search for, for most cases i know exactly what would be the best search query - letting the ai come up with search queries isn't always the best option either.

It feels like the API is developed to work autonomously like a chat client, we're developers, not grandmas, we can handle it

[hkt74]

Hi @Glutch do you have an example prompt to reproduce the issue? thanks