Rules

Rules allow you to define specific conditions for your evaluation metrics that can trigger alerts and notifications when met. They serve as the foundation for the alerting system and help you monitor your AI system's performance against predetermined thresholds.

Overview

A rule consists of one or more conditions, each tied to a specific metric, that is supported by our Scorer (like Faithfulness or AnswerRelevancy). When evaluations are performed, the rules engine checks if the measured scores satisfy the conditions set in your rules. Based on the rule's configuration, alerts can be triggered and notifications sent through various channels.

Creating Rules

Rules can be created using the Rule class from the judgeval.rules (Python) or judgeval (Typescript) module. Each rule requires:

• A name
• A list of conditions
• A combine type (how conditions should be evaluated together)

Optional parameters include:

• A description
• Notification configuration

Basic Rule Structure

from judgeval.rules import Rule, Condition
from judgeval.scorers import FaithfulnessScorer, AnswerRelevancyScorer

# Create a rule
rule = Rule(
    name="Quality Check",
    description="Check if quality metrics meet thresholds",
    conditions=[
        Condition(metric=FaithfulnessScorer(threshold=0.7)),
        Condition(metric=AnswerRelevancyScorer(threshold=0.8))
    ],
    combine_type="all"  # "all" = AND, "any" = OR
)

Conditions

Conditions are the building blocks of rules. Each condition specifies a metric (must be a built-in API Scorer like FaithfulnessScorer or AnswerRelevancyScorer). The condition is met when the score for that metric is greater than or equal to the threshold specified in the scorer.

Creating Conditions

from judgeval.rules import Condition
from judgeval.scorers import FaithfulnessScorer

# Create a condition that passes when faithfulness score is greater than or equal to 0.7
condition = Condition(
    metric=FaithfulnessScorer(threshold=0.7)
)

How Conditions are Evaluated

When a condition is evaluated, it uses the scorer's threshold and internal evaluation logic:

1. By default, a condition passes when the actual score is greater than or equal to the threshold
2. If the scorer has a custom success_check() method, that method will be used instead
3. The threshold is retrieved from the scorer's threshold attribute

Combine Types

Rules support two combine types that determine how multiple conditions are evaluated:

• "all": The rule triggers when all conditions fail (logical AND)
• "any": The rule triggers when any condition fails (logical OR)

This design is meant for setting up alerts that trigger when your metrics indicate a problem with your AI system's performance.

Using Rules with the Tracer

Rules are most commonly used with the Tracer to monitor your AI system's performance:

from judgeval.common.tracer import Tracer
from judgeval.rules import Rule, Condition
from judgeval.scorers import FaithfulnessScorer, AnswerRelevancyScorer

# Create rules
rules = [
    Rule(
        name="Quality Check",
        description="Check if quality metrics meet thresholds",
        conditions=[
            Condition(metric=FaithfulnessScorer(threshold=0.7)),
            Condition(metric=AnswerRelevancyScorer(threshold=0.8))
        ],
        combine_type="all"  # Trigger when all conditions fail
    )
]

# Initialize tracer with rules
judgment = Tracer(
    project_name="your_project", 
    rules=rules
)

For more information on configuring notifications with rules, see the Notifications documentation.