Skip to content

prxtenses/repair-json-stream

Repository files navigation

Alt text

repair-json-stream

Blazing-fast JSON repair for LLM streaming

npm CI License Zero Dependencies Bundle Size


When streaming responses from LLMs like OpenAI or Anthropic, JSON often arrives incomplete:

{"message": "Hello, I'm currently generating your resp

JSON.parse() crashes. repair-json-stream fixes it instantly.

import { repairJson } from 'repair-json-stream'

repairJson('{"message": "Hello, I\'m currently generating your resp')
// β†’ '{"message": "Hello, I\'m currently generating your resp"}'

⚑ Performance

Built for real-time streaming. 1.9x faster than alternatives.

β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚ Benchmark          β”‚ repair-json-stream β”‚ jsonrepair β”‚ Speedup β”‚
β”œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”Όβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”Όβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”Όβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€
β”‚ Small (15 KB)      β”‚ 0.85 ms            β”‚ 1.53 ms    β”‚ 1.8x    β”‚
β”‚ Large (3.9 MB)     β”‚ 295 ms             β”‚ 392 ms     β”‚ 1.3x    β”‚
β”‚ Streaming (1K ops) β”‚ 426 ms             β”‚ 583 ms     β”‚ 1.4x    β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”΄β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”΄β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”΄β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜

πŸ“¦ Installation

npm install repair-json-stream
pnpm add repair-json-stream
yarn add repair-json-stream

πŸ”§ What It Fixes

Issue Example Result
Truncated strings {"text": "Hello {"text": "Hello"}
Missing brackets {"a": [1, 2 {"a": [1, 2]}
Unquoted keys {name: "John"} {"name": "John"}
Single quotes {'key': 'val'} {"key": "val"}
Python constants {"x": None, "y": True} {"x": null, "y": true}
Trailing commas [1, 2, 3,] [1, 2, 3]
Comments {"a": 1} // comment {"a": 1}
Fenced code blocks ```json {"a":1} ``` {"a":1}
JSONP wrappers callback({"a": 1}) {"a": 1}
MongoDB types NumberLong(123) 123
String concatenation "hello" + "world" "helloworld"
NDJSON {"a":1}\n{"b":2} [{"a":1},{"b":2}]

πŸš€ Usage

Basic

import { repairJson } from 'repair-json-stream'

const broken = '{"users": [{"name": "Alice'
const fixed = repairJson(broken)
// β†’ '{"users": [{"name": "Alice"}]}'

JSON.parse(fixed) // Works!

Streaming (Node.js)

import { createReadStream, createWriteStream } from 'fs'
import { pipeline } from 'stream'
import { jsonrepairTransform } from 'repair-json-stream/stream'

pipeline(
  createReadStream('broken.json'),
  jsonrepairTransform(),
  createWriteStream('fixed.json'),
  (err) => console.log(err || 'Done!')
)

CLI

# Install globally
npm install -g repair-json-stream

# Repair a file
repair-json-stream broken.json > fixed.json

# Pipe from stdin
cat broken.json | repair-json-stream

# Overwrite in place
repair-json-stream broken.json --overwrite

# Verbose output (show repair actions)
repair-json-stream broken.json --verbose

Browser

<script src="https://unpkg.com/repair-json-stream/dist/repair-json-stream.browser.global.js"></script>
<script>
  const fixed = RepairJsonStream.repairJson('{"broken": true')
  console.log(fixed) // '{"broken": true}'
</script>

πŸ“– API Reference

repairJson(json: string): string

Repairs broken JSON and returns valid JSON. Never throws.

repairJson('{"a": tru')     // '{"a": true}'
repairJson("{'b': None}")   // '{"b": null}'
repairJson('[1, 2,')        // '[1, 2]'

repairJson(input, onRepair?)

Track what was fixed using the callback:

repairJson('{"a": 1,', (action, idx, context) => {
  console.log(`Action: ${action} at index ${idx} (${context})`)
})
// Output: Action: closed_object at index 7 (Closing missing })

preprocessJson(input: string): string

Strips wrappers (fenced blocks, JSONP, escaped strings) before repair.

preprocessJson('```json\n{"a": 1}\n```')  // '{"a": 1}'
preprocessJson('callback({"a": 1})')      // '{"a": 1}'

jsonrepairTransform(options?): Transform

Node.js Transform stream for piping.

import { jsonrepairTransform } from 'repair-json-stream/stream'

Web Streams API

Universal TransformStream for Deno, Bun, Cloudflare Workers, and modern browsers.

import { jsonRepairStream } from 'repair-json-stream/web-stream'

// Use with fetch
const response = await fetch('/api/llm')
const repairedStream = response.body
  .pipeThrough(new TextDecoderStream())
  .pipeThrough(jsonRepairStream())

// Read repaired JSON
const reader = repairedStream.getReader()
const { value } = await reader.read()
JSON.parse(value) // Always valid!

Incremental Stateful Repair

True streaming repair with immediate partial output. Perfect for live UI updates.

import { IncrementalJsonRepair } from 'repair-json-stream/incremental'

const repairer = new IncrementalJsonRepair()

// As LLM streams chunks...
let output = ''
for await (const chunk of llmStream) {
  output += repairer.push(chunk)
  updateUI(output) // Live update!
}
output += repairer.end()

JSON.parse(output) // Always valid!

Methods:

  • .push(chunk) - Process chunk, returns repaired output
  • .end() - Finalize and close open structures
  • .snapshot() - Get valid JSON at any point (non-destructive)
  • .reset() - Reset state for reuse

LLM Garbage Filtering

Extract JSON from messy LLM outputs containing prose, thinking blocks, or markdown.

import { extractJson, extractAllJson, stripLlmWrapper } from 'repair-json-stream/extract'

// Extract JSON from prose
extractJson('Sure! Here is the data: {"name": "John"} Hope this helps!')
// β†’ '{"name": "John"}'

// Handle thinking blocks (DeepSeek, Claude, etc.)
extractJson('<thought>Let me think...</thought>\n{"result": true}')
// β†’ '{"result": true}'

// Extract multiple JSON blocks
extractAllJson('First: {"a": 1} Second: {"b": 2}')
// β†’ ['{"a": 1}', '{"b": 2}']

// Full cleanup (markdown, prose, thinking blocks)
stripLlmWrapper(`
<thinking>reasoning here</thinking>
\`\`\`json
{"data": [1, 2, 3]}
\`\`\`
Let me know if you need anything else!
`)
// β†’ '{"data": [1, 2, 3]}'

πŸ—οΈ Architecture

  • O(n) single-pass - No multiple iterations
  • Stack-based state machine - No regex parsing
  • Zero dependencies - Minimal attack surface
  • ReDoS-safe - No exponential backtracking
  • 5.5 KB minified - Lightweight for browsers

πŸ“„ License

MIT Β© sonka