You can trace Google Gemini models with LangWatch by using the OpenAI-compatible endpoint provided by Google Cloud Vertex AI. This requires setting up an authenticated endpoint in your Google Cloud project that serves the Gemini model.
Before you begin, ensure you have configured the LangWatch SDK by following the Go setup guide.

​
Setup

This setup is more involved than other providers and requires Google Cloud authentication.

​
1. Enable Vertex AI

Ensure the Vertex AI API is enabled in your Google Cloud project.

​
2. Get an Authentication Token

The Vertex AI API uses a Google Cloud access token, not a static API key. You must generate this token using the gcloud CLI. The token is short-lived (typically 1 hour) and needs to be refreshed. 
# Log in to gcloud
gcloud auth login

# Get a token and set it as an environment variable
export GOOGLE_ACCESS_TOKEN=$(gcloud auth print-access-token)
In a production application, you should use a service account and the Google Cloud client libraries for Go to programmatically generate access tokens instead of using gcloud directly.

​
3. Construct the Endpoint URL

Your Vertex AI endpoint URL will follow this format: https://<region>-aiplatform.googleapis.com/v1/projects/<project-id>/locations/<region>/publishers/google/models/<model-name> For example: https://us-central1-aiplatform.googleapis.com/v1/projects/my-gcp-project/locations/us-central1/publishers/google/models/gemini-1.5-pro-001

​
Example

Configure your openai.Client with the constructed Vertex AI URL and the temporary access token. The gen_ai.system attribute should be set to "google". 
package main

import (
	"context"
	"log"
	"os"

	"github.com/langwatch/langwatch-go/instrumentation/openai"
	"github.com/sashabaranov/go-openai"
)

func main() {
	ctx := context.Background()

	// Assumes LangWatch is already set up.

	// Your Google Cloud details
	projectID := "your-gcp-project-id"
	region := "us-central1"
	modelName := "gemini-1.5-pro-001" // Or another supported Gemini model

	// Construct the Vertex AI endpoint URL
	baseURL := "https://" + region + "-aiplatform.googleapis.com/v1/projects/" + projectID + "/locations/" + region + "/publishers/google/models/" + modelName

	config := openai.DefaultConfig(os.Getenv("GOOGLE_ACCESS_TOKEN"))
	config.BaseURL = baseURL
	// Add the middleware, identifying the system as "google"
	config.HTTPClient = langwatch_openai.Instrument(
		config.HTTPClient,
		"my-gemini-app",
		langwatch_openai.WithGenAISystem("google"),
		langwatch_openai.WithCaptureInput(),
		langwatch_openai.WithCaptureOutput(),
	)

	client := openai.NewClientWithConfig(config)

	// Make a call to the Gemini model
	response, err := client.CreateChatCompletion(ctx, openai.ChatCompletionRequest{
		// The model parameter is not used in the path but can be set
		Model: modelName,
		Messages: []openai.ChatCompletionMessage{
			{
				Role:    openai.ChatMessageRoleUser,
				Content: "Hello, Gemini! Explain the concept of multimodal models.",
			},
		},
	})

	if err != nil {
		log.Fatalf("Google Gemini API call failed: %v", err)
	}

	log.Printf("Response from Gemini: %s", response.Choices[0].Message.Content)
}
The model name is part of the URL itself. While the Model parameter in the request body is less critical, it’s good practice to set it to the model you are targeting.