Class DecisionModelOutputGuardrail

java.lang.Object

dev.langchain4j.guardrails.DecisionModelOutputGuardrail

All Implemented Interfaces:

Guardrail<OutputGuardrailRequest, OutputGuardrailResult>, OutputGuardrail

@Experimental
public class DecisionModelOutputGuardrail
extends Object
implements OutputGuardrail

An OutputGuardrail that checks the responses of the model with a DecisionModel.

Each check is a yes/no question where "yes" means the response must be rejected. All checks are answered in a single call. The decision model receives the conversation that led to the response, ending with the response: the messages sent to the model to produce the response (OutputGuardrailRequest.chatRequest()), including the system message, tool calls and tool results, so that checks can compare the response with the instructions of the assistant and with the data its tools returned (see DecisionModelOutputGuardrail.Builder.maxMessages(Integer), DecisionModelOutputGuardrail.Builder.maxMessageLength(Integer) and DecisionModelOutputGuardrail.Builder.messageFilter(Predicate)). The user message comes with the content retrieved for it (RAG), and after a reprompt, the reprompt and its tool calls are included. When the request is not known, the messages are taken from the chat memory, or, without one, the user message is sent with the response. Each question is preceded by a fixed English sentence saying that the response is the last assistant message of the conversation. For example:

OutputGuardrail guardrail = DecisionModelOutputGuardrail.builder()
        .decisionModel(decisionModel)
        .check("personalData", "Does the response reveal personal data, such as contact details or account numbers?")
        .check("unanswered", "Does the response fail to address the user message?")
        .reprompt("Answer the user message without revealing personal data.")
        .build();

A rejected response fails with a fatal result or, if DecisionModelOutputGuardrail.Builder.reprompt(String) is set, the model is asked again with that instruction. Responses without text (for example, only tool calls) are not checked. If the decision model fails, the exception is propagated, so the request fails. A check that the decision model refuses to answer fails, so the response is rejected.

Since:

1.21.0

Nested Class Summary

Nested Classes

Modifier and Type	Class	Description
static class	DecisionModelOutputGuardrail.Builder

Constructor Summary

Constructors

Modifier	Constructor	Description
protected	DecisionModelOutputGuardrail(DecisionModelOutputGuardrail.Builder builder)
	DecisionModelOutputGuardrail(DecisionModel decisionModel, Map<String,String> checks)	Creates a guardrail with the given checks and the default threshold.

Method Summary

Modifier and Type	Method	Description
static DecisionModelOutputGuardrail.Builder	builder()
protected String	failureMessage(List<String> failedChecks)	The message of the failure when checks fail.
OutputGuardrailResult	validate(OutputGuardrailRequest request)	Validates the response from the LLM.

Methods inherited from class Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Methods inherited from interface Guardrail

name, validateAsync

Methods inherited from interface OutputGuardrail

failure, failure, failureWithMessageRemoval, failureWithMessageRemoval, fatal, fatal, fatalWithMessageRemoval, fatalWithMessageRemoval, reprompt, reprompt, retry, retry, success, successWith, successWith, successWith, successWith, successWithResult, validate

Constructor Details

DecisionModelOutputGuardrail

public DecisionModelOutputGuardrail(DecisionModel decisionModel,
 Map<String,String> checks)

Creates a guardrail with the given checks and the default threshold.

Parameters:

decisionModel - the decision model that answers the checks.

checks - the checks, as yes/no questions keyed by check name, where "yes" means the response must be rejected.

DecisionModelOutputGuardrail

protected DecisionModelOutputGuardrail(DecisionModelOutputGuardrail.Builder builder)

Method Details

validate

public OutputGuardrailResult validate(OutputGuardrailRequest request)

Description copied from interface: OutputGuardrail

Validates the response from the LLM.

Unlike OutputGuardrail.validate(AiMessage), this method allows you to access the memory and the augmentation result (in the case of a RAG).

Implementation must not attempt to write to the memory or the augmentation result.

Specified by:

validate in interface Guardrail<OutputGuardrailRequest, OutputGuardrailResult>

Specified by:

validate in interface OutputGuardrail

Parameters:

request - the parameters, including the response from the LLM, the memory, and the augmentation result.

Returns:

The result of the validation

failureMessage

protected String failureMessage(List<String> failedChecks)

The message of the failure when checks fail. Override it, for example, to hide which checks failed from users who can see the message.

Parameters:

failedChecks - the names of the failed checks.

builder

public static DecisionModelOutputGuardrail.Builder builder()