langchain-aws-utils

Utilities and helpers for LangChain and LangGraph with AWS services

Install

npm install langchain-aws-utils

or

yarn add langchain-aws-utils

Usage

Guardrails

import { isGuardrailStopped } from 'langchain-aws-utils/guardrails';

// Check if an LLM result was stopped by guardrails
const llOutput = await model.invoke(messages);
const isStopped = isGuardrailStopped(llOutput);

if (isStopped) {
  console.log('LLM result was stopped by guardrails');
}

Memory Management

import { cleanMemoryFromBlockedMessages } from 'langchain-aws-utils/memory';
// Clean memory checkpoint from blocked messages
const { cleanedData, shouldOverride } = cleanMemoryFromBlockedMessages(memoryCheckpoint);

if (shouldOverride) {
  // Update your memory with the cleaned data
  await updateMemory(cleanedData);
}

API

Guardrails

isGuardrailStopped(llOutput)

Check if an LLM result was stopped by guardrails. This function examines the LLM result object to determine if guardrails intervened during execution.

Parameters:

• llOutput (LLMResult): The LLM result object to check, usually from a LangGraph node callback

Returns: boolean - True if the LLM result was stopped by guardrails

Example:

// In a LangGraph node callback
async handleLLMEnd(llOutput: LLMResult): Promise<any> {
  const isStopped = isGuardrailStopped(llOutput);
  if (isStopped) {
    console.log('LLM result was stopped by guardrails');
  }
}

Memory Management

cleanMemoryFromBlockedMessages(memoryCheckpoint)

Cleans memory checkpoint traces by removing messages that were blocked by guardrails. This function filters out blocked assistant messages and their causative human messages to maintain clean conversation memory.

Parameters:

• memoryCheckpoint (unknown): The memory checkpoint trace object to clean

Returns: { cleanedData: CheckpointTuple; shouldOverride: boolean }

• cleanedData: The cleaned checkpoint data with blocked messages removed
• shouldOverride: Whether the memory should be updated with the cleaned data

Example:

// Clean memory from blocked messages
const { cleanedData, shouldOverride } = cleanMemoryFromBlockedMessages(memoryCheckpoint);

if (shouldOverride) {
  // Update your LangGraph memory with the cleaned data
  await graph.updateState(cleanedData);
}

What it removes:

• Assistant messages that were blocked by guardrails (detected by stop reasons like "blocked" or "guardrail_intervened")
• The causative human messages that led to the blocked assistant responses
• Maintains conversation flow by removing the problematic interaction pairs

Use cases:

• Clean conversation memory after guardrail interventions
• Remove blocked content from chat history
• Maintain clean conversation state in LangGraph applications