> ## Documentation Index
> Fetch the complete documentation index at: https://docs.galtea.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# AI Metric Generation
> Automatically generate evaluation metrics from your product specifications using AI
## Overview
AI Metric Generation lets you automatically create evaluation metrics from your product's [specifications](/concepts/product/specification). Instead of manually crafting judge prompts and configuring evaluation parameters, the AI analyzes your specifications and generates ready-to-use metrics.
Evaluation parameters are automatically selected based on each specification's description and test type. The generated judge prompt follows a format optimized for reliable LLM-based evaluation across different evaluator models.
## Requirements
* A [product](/concepts/product) with a description
* At least one [specification](/concepts/product/specification) of type **POLICY** with a test type assigned (Accuracy, Security, or Behavior)
CAPABILITY and INABILITY specifications cannot be used for AI metric generation because they do not have a test type.
## How to Generate Metrics
There are two ways to trigger AI metric generation from the dashboard:
### From the Specifications Page
1. Navigate to your product's **Specifications** tab
2. Open the dropdown menu on the specification you want to generate metrics for
3. Click **Generate Metrics** — this takes you to the generation page with that specification pre-selected
4. Click **Generate** and wait for the AI to process
5. Review the generated candidates — edit, save, or discard each one
### From the Metrics Page
1. Navigate to your product's **Metrics** tab
2. Click **Generate Metrics with AI**
3. Select the specifications you want to generate metrics for
4. Click **Generate** and wait for the AI to process
5. Review the generated candidates — edit, save, or discard each one
In both cases, the AI analyzes your product name, description, and the selected specifications to generate tailored metrics.
## Complete a Metric with AI
When you already know **what** you want to measure but don't want to write the judge prompt by hand, use **Complete with AI**. You provide the metric's name, description, and evaluation type as the source of truth, and the AI fills in the rest: the judge prompt and the evaluation parameters (plus a recommended evaluator model for AI Evaluation).
This is available in two places:
* **The manual metric form** — when creating a metric from scratch (**Metrics** tab → **New Metric** → blank), fill in the name, description, and evaluation type, then click **Complete with AI**.
* **The AI generation review step** — when editing a generated candidate, click **Complete with AI** to regenerate its judge prompt and evaluation parameters from the candidate's current name, description, and tags.
Your name, description, and tags are never modified — they are the inputs the AI completes from. If you have already written a judge prompt, the dashboard asks for confirmation before replacing it.
Complete with AI is available for **AI Evaluation** and **Human Evaluation** metrics:
* For **AI Evaluation**, it produces a binary LLM-judge prompt and recommends an evaluator model.
* For **Human Evaluation**, it produces annotator guidelines on a 0–100 scoring scale (no evaluator model, since a human assigns the score).
**Self-Hosted** metrics are not supported because they have no judge prompt. See [Evaluation Types](/concepts/metric/evaluation-types) for the differences between these scoring methods.
## Evaluation Parameter Selection
In both spec-driven generation and completion mode, the AI automatically selects the evaluation parameters each metric needs. In spec mode these are inferred from the specification description; in completion mode they are inferred from the metric description you provided. For example:
* A specification about **citation accuracy** or **knowledge-grounded answers** will include `retrieval_context` so the judge can verify answers against retrieved source material.
* A specification about **internal processes, workflows, or tool orchestration** will include `traces` (and often `tools_used`) so the judge can inspect the execution path — not just the final output.
* A specification about **refusal or safety boundaries** typically only needs `input` and `actual_output`, since compliance is fully observable from what was asked and answered.
If a metric includes `traces` or `retrieval_context` as evaluation parameters, the product under test must capture that data during test execution (e.g., via tracing integrations). If the data is not available at evaluation time, the evaluation will fail.
## Generated Metric Properties
Each AI-generated metric candidate includes:
| Property | Description |
| ------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Name** | A descriptive name for the metric |
| **Description** | What the metric evaluates |
| **Judge Prompt** | The evaluation prompt with placeholders for dynamic data |
| **Evaluation Parameters** | The data parameters the judge needs for evaluation (automatically selected based on the specification in spec mode, or from the metric description in completion mode) |
| **Tags** | Categorization tags |
| **Evaluator Model** | The LLM model used for evaluation (null for Human Evaluation metrics) |
| **Test Type** | Inherited from the source specification (null in completion mode when no specification is provided) |
## Specification Linking
When you save a metric generated from a specification, it is automatically linked to that specification. This creates a traceable connection between your requirements and your evaluation criteria.
Metrics created via **Complete with AI** are not automatically linked to a specification — you can link them manually from the [Specification Hub](/concepts/product/specification) if needed.
You can view linked specifications directly from a metric's detail page, and manage metric-specification links from the [Specification Hub](/concepts/product/specification).