Custom Metric

Custom Metrics: Multi-Step Evaluation

This functionality allows you to create multi-step evaluation pipelines, enabling complex workflows that uses Large Language Models (LLMs) and/or Python-based scripting. Each metric can be composed of multiple steps, where each step’s output is available to subsequent steps. By chaining steps together, you can design flexible and robust evaluation metrics for your unique use cases.

You can find a dedicated Custom Metrics tab in the main navigation menu of the RagaAI Catalyst platform. Here, you can:

• View all custom metrics you have created (or have access to).

• Create new metrics from scratch.

• Edit existing metrics—add, remove, or modify steps.

• Manage them through version control and deployments.

Creating a New Metric

Initiation:

1. Click on Create New Metric.

2. Provide a Metric Name (required, must be unique).

3. Add a Description (optional, up to 30 characters).

Adding Steps

After naming and describing the metric:

1. Click Add Step.

2. Choose the Step Type:

   1. Custom Metric (LLM Call)

   2. Python

You can add as many steps as you need. Each step can reference the outputs of prior steps, giving you virtually unlimited flexibility in how you design your metric.

Custom Metric (LLM Call)

Custom Metric (LLM Call)

When you add a Custom Metric (LLM Call) step, you will see the following configuration sections:

1. Prompt Editor

   • Insert variables using the syntax {{variable_name}}.

   • You can configure the System Role and User Role instructions for the LLM.

2. Model Configuration

   • Select from available LLMs (e.g., GPT-4).

   • Set parameters like Max Tokens, Temperature, etc.

   • Any output from previous steps can be referenced using {{step_name.response}}.

3. Reasoning

   • Reasoning Checkbox: If you want the LLM’s reasoning for how it arrived at its score or response, check the Reasoning checkbox. This ensures the LLM call also returns a reasoning string in addition to the primary output.

   • Reasoning Dropdown: In the main evaluation configuration (outside the step editor), you can select which step’s reasoning will be displayed when running Evals. This gives you the flexibility to see the most relevant step’s reasoning during your final evaluation review.

Note: Enabling reasoning will store an additional field from the same LLM call. You can use or display this reasoning later for review, debugging, or transparency into how the LLM arrived at its final answer.

Python

• Function Definition:

  • The function name must match the step name and must be uniquely defined once.

  • Only one top-level function definition is allowed. If more than one are found, you will see an error: Error: Only one top-level function definition is allowed. Found {count}. Nested functions are allowed.

• Inputs can include outputs from previous steps using {{step_name.response}}.

• Code is securely run in a sandboxed environment.

• The output can be used in subsequent steps.

Multi-Step Workflow

Execution Order

Steps run in the sequence they appear in the Steps Panel (top to bottom). Each step’s output is stored for reference and possible use by subsequent steps.

Inter-Step Variable Referencing

To reference a step’s output, use:

{{step_name.response}}

Ensure step_name is correct. A missing or misspelled name will lead to an error.

Example Workflow

Metric: Hallucination Check

Step 1: Claims Generation (LLM Call)

Step 2: Claims Verification (LLM Call)

Step 3: Scoring (Python)

Detailed Flow:

Step 1: Prompt the LLM to extract factual claims from a given {{response}}.

Step 2: Prompt the LLM again to verify each claim against a provided {{context}}.

Step 3: Use Python to score the final verification. If any claim is “refuted,” return a score of 1 (indicating a hallucination).

Grading Criteria

After all steps are executed, RagaAI Catalyst looks at the final step’s output to determine whether to produce a:

• Score (0 – 1)

• Boolean (0/1)

During metric setup, you can select which grading criteria is appropriate. Only the last step’s output (and not intermediate steps) is considered for this grading assessment.

Testing and Verification

Test Eval Button

• Runs the entire metric workflow from the first step to the last.

• Provides a Results Viewer, showing output for each step.

Grading Criteria Verification

• After the test run, click Verify Grading Criteria to validate that the final output meets the chosen criteria.

• On success/failure, you’ll see a status alert, and the last step’s output is revealed for review.

Version Control and Deployment

Commit Versions

• Like the Playground, each commit creates a new version of your metric.

• Keep track of iterations as you refine your steps.

Deploy to Eval

• Click Deploy to Eval to open a modal.

• You’ll see the Current Deployed Version and can select another version to deploy. (If none is deployed, it shows None.)

• Commit Version: You can commit your current configuration as a new version and then deploy it.

• Unlike the Playground, there’s no “default” version. Instead, you explicitly choose which version to deploy.

Using a Custom Metric in Your Evaluations

Once a metric is committed:

• It becomes visible in a new tab called Custom Metric when configuring evaluations on your dataset.

• Map Variables to your dataset columns. For example:

• • prompt → Response

  • context → Context

  • response → User Response Column

Upon running, the system automatically executes the steps in your custom metric and produces a final evaluation.