Manage evaluation metrics

Before you begin

Before you manage evaluation metrics, ensure you have the following:

• A Google Cloud project with the Agent Platform APIenabled.
• (Optional) If using the Agent Platform SDK, initialize the client as described in Evaluate your agents .

The Metric Registryallows you to define, store, and manage reusable configurations for how your agents are evaluated. Instead of configuring criteria for every test run, you can save standardized metrics—such as a custom LLM-based rubric for safety or a Python function for execution accuracy—and apply them consistently to both offline assessments and continuous online monitors.

Metric types

Agent Platform supports three types of metrics in the registry:

• Predefined Metrics:Managed metrics provided by Google, including multi-turn raters for task success, tool use quality, and trajectory compliance.
• Custom LLM Metrics:Natural language rubrics where a "Judge LLM" evaluates an agent's response based on your specific criteria and rating scales.
• Custom Code Metrics:Python functions that programmatically validate agent behavior, such as checking for a specific output format or verifying a tool response.

In addition to managed metrics provided by Google, you can use customized registered metrics for evaluation.

Predefined metrics

Agent Platform provides a set of predefined metrics for evaluating agents. These metrics are managed by Google and cover common evaluation dimensions for both single-turn and multi-turn agent interactions.

You can access predefined metrics in the SDK using types.RubricMetric.METRIC_NAME . For the full details of all managed rubric-based metrics, including input requirements and output formats, see Details for managed rubric-based metrics .

Single-turn agent metrics

The following metrics evaluate a single agent interaction (one prompt and one response, potentially with intermediate tool calls):

Multi-turn agent metrics

The following metrics evaluate multi-turn agent conversations. They analyze the full conversation context to assess overall agent performance across multiple turns:

Use predefined metrics in the SDK

  from 
  
 vertexai 
  
 import 
 evals 
 , 
 types 
 # Run an evaluation with predefined metrics 
 result 
 = 
 client 
 . 
 evals 
 . 
 evaluate 
 ( 
 dataset 
 = 
 eval_dataset 
 , 
 metrics 
 = 
 [ 
 types 
 . 
 RubricMetric 
 . 
 FINAL_RESPONSE_QUALITY 
 , 
 types 
 . 
 RubricMetric 
 . 
 TOOL_USE_QUALITY 
 , 
 types 
 . 
 RubricMetric 
 . 
 HALLUCINATION 
 , 
 types 
 . 
 RubricMetric 
 . 
 SAFETY 
 , 
 ], 
 ) 
 # Visualize results in Colab 
 result 
 . 
 show 
 () 
 

Manage metrics in the console

1. In the Google Cloud console, navigate to the Agent Platform > Agents > Evaluationpage.

   Go to Evaluation

2. Click the Metricstab to view the registry.

3. Create a metric:Click New metricand select Custom LLM metricor Custom code metric.

4. Define rubrics:For LLM metrics, use the Samplebuttons to quickly populate instructions, criteria (for example, Clarityor Excitement), and rating scores.

5. View and edit:Click any metric name to view its definition in read-only mode, or use the More options more_vert icon to Duplicateor Deletethe resource.

Manage metrics with the SDK

You can programmatically register and use metrics using the Agent Platform SDK.

Register a Custom LLM Metric

  from 
  
 vertexai 
  
 import 
 evals 
 , 
 types 
 # Define a metric with a specific rubric 
 tone_check_metric 
 = 
 types 
 . 
 LLMMetric 
 ( 
 name 
 = 
 "tone_check" 
 , 
 prompt_template 
 = 
 "Analyze the tone of the response ..." 
 , 
 result_parsing_function 
 = 
 """ 
 import json, re 
 def parse_results(responses): 
 response = json.loads(responses[0]) 
 return {"score": response.get("score", 0.0), 
 "explanation": response.get("explanation", "default explanation")} 
 """ 
 ) 
 # Register the custom metric 
 tone_check_metric_path 
 = 
 client 
 . 
 evals 
 . 
 create_evaluation_metric 
 ( 
 metric 
 = 
 tone_check_metric 
 ) 
 

Register a Custom Code Metric

  from 
  
 vertexai 
  
 import 
 evals 
 , 
 types 
 # Define a metric with custom python code 
 accuracy_metric_code 
 = 
 """ 
 def evaluate(instance: dict) -> float: 
 agent_data = instance.get('agent_eval_data', 
 {} 
 ) 
 turns = agent_data.get('turns', []) 
 for turn in turns: 
 ... 
 """ 
 accuracy_metric 
 = 
 types 
 . 
 CodeExecutionMetric 
 ( 
 name 
 = 
 "multi_turn_accuracy" 
 , 
 custom_function 
 = 
 accuracy_metric_code 
 ) 
 # Register the custom metric 
 accuracy_metric_path 
 = 
 client 
 . 
 evals 
 . 
 create_evaluation_metric 
 ( 
 metric 
 = 
 accuracy_metric 
 )