Gemini 3 tool calls fail on a Google Vertex AI inference endpoint

On Stack 9.2 and 9.3, Elastic Agent Builder tool calls fail when using a Gemini 3.x model through a Google Vertex AI inference endpoint configured in External inference, even though endpoint creation succeeds.

Stack 9.3 The Elastic Stack includes Elastic's core products such as Elasticsearch, Kibana, Logstash, and Beats. Availability Generally available in 9.3 This functionality is generally available and ready for production usage. Unless stated otherwise on the page, this functionality is available when your Elastic Stack is deployed on Elastic Cloud Hosted, Elastic Cloud Enterprise, Elastic Cloud on Kubernetes, and self-managed environments. This documentation corresponds to the latest patch available for each minor version. If you're not using the latest patch, check the release notes for changes. Gemini 3.x models remain supported through the Elastic Inference Service (EIS).

• The agent fails when it tries to call a tool. You get no tool result.

• The error includes status 400 and thought_signature. For example:

  Error: Received an unsuccessful status code for request from inference entity id [<inference_id>] status [400]. Function call is missing a thought_signature in functionCall parts. This is required for tools to work correctly, and missing thought_signature may lead to degraded model performance.
  

  <inference_id> is the ID of the Google Vertex AI endpoint Elastic Agent Builder is using.

You can choose one of the following options:

• Upgrade to Stack 9.5.6 or later.
• Create a new endpoint that uses a Gemini 2.x model.
• Stack 9.3 The Elastic Stack includes Elastic's core products such as Elasticsearch, Kibana, Logstash, and Beats. Availability Generally available in 9.3 This functionality is generally available and ready for production usage. Unless stated otherwise on the page, this functionality is available when your Elastic Stack is deployed on Elastic Cloud Hosted, Elastic Cloud Enterprise, Elastic Cloud on Kubernetes, and self-managed environments. This documentation corresponds to the latest patch available for each minor version. If you're not using the latest patch, check the release notes for changes. Use Gemini 3.x models through EIS.

Upgrade your cluster to Stack 9.5.6 or later. For instructions, refer to Upgrade your deployment or cluster.

Create a new Google Vertex AI endpoint that uses a Gemini 2.x model. The model ID of an existing endpoint cannot be changed. For example:

PUT _inference/chat_completion/gemini_2_5_flash
{
  "service": "googlevertexai",
  "service_settings": {
    "service_account_json": "<service_account_json>",
    "model_id": "gemini-2.5-flash",
    "project_id": "<project_id>",
    "location": "<location>"
  }
}

1. The endpoint ID. Use an ID that is not already in use.
2. The service account JSON key for the Google Vertex AI API.
3. The Google Cloud project ID.
4. The Google Cloud region where Vertex AI is enabled, such as us-central1.

Select this endpoint in the model selector. For the full field list, refer to Create a Google Vertex AI inference endpoint.

Gemini 3.x models are available through the Elastic Inference Service (EIS). EIS is automatically available on Elastic Cloud Serverless projects and Elastic Cloud Hosted deployments. On Elastic Cloud Enterprise, Elastic Cloud on Kubernetes, and self-managed deployments, connect the cluster to EIS first.

To use a Gemini 3.x model on EIS, select it in the model selector.

• Recommended models
• Elastic Inference Service