Rules

Learn about defining rules for runtime protection

Rules are conditions that you never want your application to break. A rule is composed of three parts:

• A metric
• An operator
• A target value

These rules will take the value from an evaluated metric, then use the operator to compare with the target value. If this comparison returns true, then the rule has been broken and is triggered. You can then use this in your code to change the response from your application.

Create rules

To create a rule, create an instance of the Rule, setting the metric and operator. Set the target values as needed (for example, for the empty and not_empty operators, there is no need to set a target value).

The metric uses the GalileoMetrics enum, and the operator uses the RuleOperator enum.

Python

from galileo import GalileoMetrics
from galileo_core.schemas.protect.rule import Rule, RuleOperator

rule = Rule(
    metric=GalileoMetrics.input_toxicity,
    operator=RuleOperator.gt,
    target_value=0.10
)

Valid values for the operator are:

Value	Type	Description	Notes
RuleOperator.gt	Numerical	Greater than
RuleOperator.lt	Numerical	Less than
RuleOperator.gte	Numerical	Greater than or equal
RuleOperator.lte	Numerical	Less than or equal
RuleOperator.eq	Numerical or Categorical	Equals
RuleOperator.neq	Numerical or Categorical	Not equals
RuleOperator.contains	Categorical	Contains	Does the list of categories contain the single target values
RuleOperator.all	Categorical	All	Does the list of categories contain all the target values
RuleOperator.any	Categorical	Any	Does the list of categories contain any of the target values
RuleOperator.empty	Categorical	Empty	Is the list of categories empty
RuleOperator.not_empty	Categorical	Not empty	Is the list of categories not empty

Metrics and Operators supported

Rules support the following Luna-2 metrics:

• Action Advancement
• Action Completion
• Completeness
• Context Adherence
• PII
• Prompt Injection
• Sexism
• Tone
• Tool Errors
• Tool Selection Quality
• Toxicity

You can also use custom code-based metrics

Metrics can have different output values (e.g. numerical, categorical), therefore the available operators and target values differ depending on the metric.

Action advancement

This rule measures if your agent accomplishes, or is making progress towards a goal.

Read more about action advancement.

Python

Setting	Value
Metric	GalileoMetrics.action_advancement
Required Payload Fields	Both input and output must be included
Values	0.0 - 1.0
Operators supported	Greater than (RuleOperator.gt) Less than (RuleOperator.lt) Greater than or equal (RuleOperator.gte) Less than or equal (RuleOperator.lte)

Python

from galileo import GalileoMetrics
from galileo_core.schemas.protect.rule import Rule, RuleOperator

rule = Rule(
    metric=GalileoMetrics.action_advancement,
    operator=RuleOperator.lt,
    target_value=0.90
)

Action completion

This rule measures if your agent successfully accomplished all of the user’s goals.

Read more about action completion.

Python

Setting	Value
Metric	GalileoMetrics.action_completion
Required Payload Fields	Both input and output must be included
Values	0.0 - 1.0
Operators supported	Greater than (RuleOperator.gt) Less than (RuleOperator.lt) Greater than or equal (RuleOperator.gte) Less than or equal (RuleOperator.lte)

Python

from galileo import GalileoMetrics
from galileo_core.schemas.protect.rule import Rule, RuleOperator

rule = Rule(
    metric=GalileoMetrics.action_completion,
    operator=RuleOperator.lt,
    target_value=0.90
)

Completeness

This rule measures how thoroughly your model’s response covered the relevant information available in the context provided.

Read more about completeness.

Python

Setting	Value
Metric	GalileoMetrics.completeness
Required Payload Fields	Both input and output must be included
Values	0.0 - 1.0
Operators supported	Greater than (RuleOperator.gt) Less than (RuleOperator.lt) Greater than or equal (RuleOperator.gte) Less than or equal (RuleOperator.lte)

Python

from galileo import GalileoMetrics
from galileo_core.schemas.protect.rule import Rule, RuleOperator

rule = Rule(
    metric=GalileoMetrics.completeness,
    operator=RuleOperator.lt,
    target_value=0.90
)

Context Adherence

This rule measures whether your model’s response was purely based on the context provided. It can be used to stop hallucinations from reaching your end users.

Read more about context adherence.

Python

Setting	Value
Metric	GalileoMetrics.context_adherence
Required Payload Fields	Both input and output must be included
Values	0.0 - 1.0
Operators supported	Greater than (RuleOperator.gt) Less than (RuleOperator.lt) Greater than or equal (RuleOperator.gte) Less than or equal (RuleOperator.lte)

Generally, 0.1 is a good threshold below which the response is not adhering to the context. Creating a rule for less than 0.1 will trigger the rule when the response does not adhere to the provided context.

Python

from galileo import GalileoMetrics
from galileo_core.schemas.protect.rule import Rule, RuleOperator

rule = Rule(
    metric=GalileoMetrics.context_adherence,
    operator=RuleOperator.lt,
    target_value=0.10
)

PII (Personal Identifiable Information)

This rule is used to detect and stop Personal Identifiable Information (PII). When applied on the input, it can be used to stop the user or company PII from being included in API calls to external services. When applied on the output, it can be used to prevent data leakage or PII being shown back to the user.

Read more about PII categories and their definitions.

Python

Setting	Value
Metric	GalileoMetrics.input_pii (for detecting PII in the input) GalileoMetrics.output_pii (for detecting PII in the output)
Required Payload Fields	input for input PII, output for output PII
Values	Depending on the operator, one or more of: account_info address credit_card_info date_of_birth email name network_info password phone_number ssn username
Operators supported	Any (RuleOperator.any) - A list of categories All (RuleOperator.all) - A list of categories Contains (RuleOperator.contains) - A single category Equal (RuleOperator.eq) - A single category Not equal (RuleOperator.neq) - A single category Empty (RuleOperator.empty) Not empty (RuleOperator.not_empty)

Python

from galileo import GalileoMetrics
from galileo_core.schemas.protect.rule import Rule, RuleOperator

rule = Rule(
    metric=GalileoMetrics.output_pii,
    operator=RuleOperator.any,
    target_value=["ssn", "address"]
)

Prompt Injection

This rule is used to detect and stop prompt injections in the input.

Prompt Injection Protect rules now compare a float score instead of categorical attack labels. Existing rules that used operators like any, contains, or eq against label values should be migrated to numeric thresholds such as gte 0.5.

Read more about Prompt Injection.

Python

Setting	Value
Metric	GalileoMetrics.prompt_injection
Required Payload Fields	input
Values	0.0 - 1.0. Higher values indicate a higher probability of prompt injection.
Operators supported	Greater than (RuleOperator.gt) Less than (RuleOperator.lt) Greater than or equal (RuleOperator.gte) Less than or equal (RuleOperator.lte)

Generally, 0.5 is a good threshold above which the input should be treated as likely prompt injection. Creating a rule for greater than or equal to 0.5 will trigger the rule when prompt injection risk is elevated.

Python

from galileo import GalileoMetrics
from galileo_core.schemas.protect.rule import Rule, RuleOperator

rule = Rule(
    metric=GalileoMetrics.prompt_injection,
    operator=RuleOperator.gte,
    target_value=0.5
)

Sexism

This rule is used to detect sexist or biased language. When applied on the input, it can be used to detect sexist remarks in user queries. When applied on the output, it can be used to prevent your application from using an making biased or sexist comments in its responses.

Read more about sexism.

Python

Setting	Value
Metric	GalileoMetrics.input_sexism (for detecting sexism in the input) GalileoMetrics.output_sexism (for detecting sexism in the output)
Required Payload Fields	input for input sexism, output for output sexism
Values	0.0 - 1.0. Higher values indicate higher sexism.
Operators supported	Greater than (RuleOperator.gt) Less than (RuleOperator.lt) Greater than or equal (RuleOperator.gte) Less than or equal (RuleOperator.lte)

Python

from galileo import GalileoMetrics
from galileo_core.schemas.protect.rule import Rule, RuleOperator

rule = Rule(
    metric=GalileoMetrics.input_sexism,
    operator=RuleOperator.gt,
    target_value=0.95
)

Tone

This rule is used to detect the primary tone from the text. When applied on the input, it can be used to detect negative tones in user queries. When applied on the output, it can be used to prevent your application from using an undesired tone in its responses.

Read more about tone.

Python

Setting	Value
Metric	GalileoMetrics.input_tone (for detecting tone in the input) GalileoMetrics.output_tone (for detecting tone in the output)
Required Payload Fields	input for input tone, output for output tone
Values	One of: anger annoyance confusion fear joy love sadness surprise neutral
Operators supported	Equal (RuleOperator.eq) - A single category Not equal (RuleOperator.neq) - A single category

Python

from galileo import GalileoMetrics
from galileo_core.schemas.protect.rule import Rule, RuleOperator

rule = Rule(
    metric=GalileoMetrics.input_tone,
    operator=RuleOperator.neq,
    target_value="neutral"
)

Tool errors

This rule measures any errors when executing tools.

Read more about tool errors.

Python

Setting	Value
Metric	GalileoMetrics.tool_error_rate
Required Payload Fields	Both input and output must be included
Values	0.0 - 1.0
Operators supported	Greater than (RuleOperator.gt) Less than (RuleOperator.lt) Greater than or equal (RuleOperator.gte) Less than or equal (RuleOperator.lte)

Python

from galileo import GalileoMetrics
from galileo_core.schemas.protect.rule import Rule, RuleOperator

rule = Rule(
    metric=GalileoMetrics.tool_error_rate,
    operator=RuleOperator.lt,
    target_value=0.90
)

Tool selection quality

This rule measures whether the agent selected the correct tool, and for each tool passed the correct arguments.

Read more about tool selection quality.

Python

Setting	Value
Metric	GalileoMetrics.tool_selection_quality
Required Payload Fields	Both input and output must be included
Values	0.0 - 1.0
Operators supported	Greater than (RuleOperator.gt) Less than (RuleOperator.lt) Greater than or equal (RuleOperator.gte) Less than or equal (RuleOperator.lte)

Python

from galileo import GalileoMetrics
from galileo_core.schemas.protect.rule import Rule, RuleOperator

rule = Rule(
    metric=GalileoMetrics.tool_selection_quality,
    operator=RuleOperator.lt,
    target_value=0.90
)

Toxicity

This rule is used to detect and stop toxic or foul language in the input (user query) or output (response shown to the user).

Read more about toxicity.

Python

Setting	Value
Metric	GalileoMetrics.input_toxicity (for detecting toxicity in the input) GalileoMetrics.output_toxicity (for detecting toxicity in the output)
Required Payload Fields	input for input toxicity, output for output toxicity
Values	0.0 - 1.0. Higher values indicate higher toxicity.
Operators supported	Greater than (RuleOperator.gt) Less than (RuleOperator.lt) Greater than or equal (RuleOperator.gte) Less than or equal (RuleOperator.lte)

Python

from galileo import GalileoMetrics
from galileo_core.schemas.protect.rule import Rule, RuleOperator

rule = Rule(
    metric=GalileoMetrics.input_toxicity,
    operator=RuleOperator.gt,
    target_value=0.10
)

Custom code-based metrics

You can use custom code-based metrics in your runtime protection rulesets.

Python

from galileo_core.schemas.protect.rule import Rule, RuleOperator

rule = Rule(
    metric="your-metric-name",
    operator=RuleOperator.gt,
    target_value=0.95
)

The operators and target values here should match the type of data that the registered scorer is expected to produce.

Related resources

Runtime protection basics

Learn the basics of running runtime protection.

Rulesets

Learn about defining rulesets for runtime protection.

Stages

Learn about defining stages for runtime protection to be used during different stages in your application workflow.

Invoke runtime protection

Learn how to invoke runtime protection in code using the Galileo SDK.