llm-json-guard

Deterministic JSON repair and schema validation for LLM outputs.

Large Language Models frequently return malformed JSON containing:

• Missing quotes
• Trailing commas
• Invalid tokens
• Broken object structures

llm-json-guard repairs malformed JSON and optionally validates it against a JSON Schema — locally, instantly, and without network dependencies.

Installation

npm install llm-json-guard

Requirements

• Node.js 18+
• ESM environment ("type": "module" in package.json)

No API keys.
No external services.
Runs fully local inside your application.

Why Use This?

LLMs do not guarantee valid JSON. Even a single trailing comma can crash production systems.

This package provides:

• Deterministic JSON repair (no extra model calls)
• Confidence scoring based on repair intensity
• Optional JSON Schema validation (AJV)
• Structured error responses
• Production-safe output handling

It acts as a reliability layer between your LLM and your business logic.

Quick Example

import { LLMJsonGuard } from "llm-json-guard";

const guard = new LLMJsonGuard();

const rawOutput = `
{
  name: "Harsh",
  age: 21,
}
`;

const schema = {
  type: "object",
  properties: {
    name: { type: "string" },
    age: { type: "number" }
  },
  required: ["name", "age"]
};

// Repair only
const sanitized = guard.sanitize(rawOutput);
console.log("Sanitized:", sanitized);

// Repair + Validate
const validated = guard.guard(rawOutput, schema);
console.log("Validated:", validated);

Run Local Example (From Repository)

git clone https://github.com/harshxframe/llm-json-guard.git
cd llm-json-guard
npm install
node examples/basic.js

API Methods

sanitize(rawOutput)

Repairs malformed JSON and safely parses it.

Returns:

• success
• stage
• meta (repair status + confidence)
• data
• errors

guard(rawOutput, schema)

Repairs malformed JSON and validates it against a JSON Schema.

Returns:

• validated stage if schema passes
• validation_failed if schema check fails
• Structured validation errors

Response Structure

Example successful response:

{
  "success": true,
  "stage": "validated",
  "meta": {
    "repaired": true,
    "confidence": 0.95
  },
  "data": {
    "name": "Harsh",
    "age": 21
  },
  "errors": []
}

Failure Stages

The package clearly indicates failure states:

• parse_failed — JSON could not be repaired
• repair_suspicious — Repair heavily modified input
• validation_failed — Schema validation failed

This ensures your application can safely branch logic based on reliability.

When To Use

• AI agents generating structured output
• RAG pipelines
• Backend systems consuming LLM JSON
• Automation workflows
• Webhook normalization
• Contract enforcement

If your system depends on structured AI output, this acts as a guardrail between the LLM and your production logic.

License

MIT