clean-llm-output

Clean and sanitize AI model output text and JSON.

Features

• Remove backticks, code fences (including ```json), leading/trailing quotes
• Strip hallucinated preambles like "Here is the JSON:" lines
• Fix invalid trailing commas in JSON-ish output
• Options: stripMarkdown, stripQuotes, trimLines, collapseWhitespace
• Utilities: cleanText() and cleanJson()
• TypeScript, ESM + CJS builds, no dependencies

Install

npm i clean-llm-output

Usage

import { cleanText, cleanJson } from 'clean-llm-output';

const messy = `Here is the JSON:\n\n\`\`\`json\n{\n  "a": 1,\n  "b": [2,3,],\n}\n\`\`\``;

const t = cleanText(messy);
// Before:
// ```json\n{\n  "a": 1,\n  "b": [2,3,],\n}\n```
// After:
// {\n"a": 1\n"b": [2,3]\n}

const obj = cleanJson(messy);
// { a: 1, b: [2,3] }

const plain = cleanText('  “Hello”   world  ', { collapseWhitespace: true });
// Hello world

API

type CleanOptions = {
  stripMarkdown?: boolean;
  stripQuotes?: boolean;
  trimLines?: boolean;
  collapseWhitespace?: boolean;
}

cleanText(input: string, options?: CleanOptions): string
cleanJson<T = unknown>(input: string, options?: CleanOptions): T

Notes

• cleanJson extracts the first JSON object/array it finds, removes comments and trailing commas, then parses.
• cleanText focuses on presentation cleanup and will also remove trailing commas when the text looks like JSON.