llmatch-js

A JavaScript library for invoking LLMs and matching their responses with patterns. This is a JavaScript port of the Python llmatch package.

Installation

npm install llmatch-js

Features

• Pattern-matching on LLM responses using regular expressions
• Automatic retries with exponential backoff
• Support for any LLM client with a compatible interface
• Detailed verbose logging option

Usage

Basic Example

const { llmatch } = require('llmatch-js');
import { ChatLLM7 } from "langchain-llm7";

// Initialize your LLM client
const chat = new ChatLLM7()

// Create a wrapper for the LLM that matches the expected interface
const llm = {
  invoke: async (messages, options = {}) => {
    const response = await chat.invoke({
      messages: messages,
      ...options
    });
    
    return {
      content: response.content,
      raw: response
    };
  }
};

// Use llmatch to extract structured data
async function extractJson() {
  const result = await llmatch({
    llm,
    query: "Generate a JSON object with information about a random book.",
    pattern: /```json\n([\s\S]*?)```/,
    verbose: true
  });
  
  if (result.success) {
    // Parse the extracted JSON
    const bookData = JSON.parse(result.extractedData[0]);
    console.log("Book data:", bookData);
  } else {
    console.error("Failed to extract JSON:", result.errorMessage);
  }
}

extractJson().catch(console.error);

API Reference

llmatch(options)

Main function that invokes an LLM and matches the response against a pattern.

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | llm | Object | null | An already initialized instance of the LLM client | | query | string | "Extract relevant data." | The main query or instruction for the LLM | | pattern | string | RegExp | /(.*?)/s | A regex string or compiled RegExp object | | context | string | null | Optional additional text to provide context to the LLM | | promptTemplate | string | "{query}\n\n{context}\nWrite the answer in the next format: {format}" | Template string | | maxRetries | number | 15 | Maximum number of attempts to make | | initialDelay | number | 1.0 | Seconds to wait before the first retry | | backoffFactor | number | 1.5 | Multiplier for the delay between retries | | verbose | boolean | false | If true, prints detailed logs of the process | | passThrough | Object | {} | Additional parameters to pass to the LLM invocation |

Return Value

The function returns a Promise that resolves to an object with the following properties:

| Property | Type | Description | |----------|------|-------------| | success | boolean | True if a valid response was found | | extractedData | Array | null | Array of strings matching the pattern (capturing groups), or null if no pattern was provided or matched | | finalContent | string | null | The content of the last successful or final failed LLM response | | retriesAttempted | number | Number of retries made (0 means success on first try) | | errorMessage | string | null | Description of the error if success is false | | rawResponse | any | The raw response object from the last LLM call |

LLM Interface Requirements

Your LLM client should implement the following interface:

interface LLMClient {
  invoke(messages: Array<{role: string, content: string}>, options?: any): Promise<{
    content: string;
    [key: string]: any;
  }>;
}

License

MIT