> ## Documentation Index
> Fetch the complete documentation index at: https://docs.shim.so/llms.txt
> Use this file to discover all available pages before exploring further.

# Shim JSON Repair API Documentation - Fix Broken LLM Outputs

> Complete guide to Shim JSON repair API. Fix malformed JSON from GPT-4, Claude, and AI agents in 8-50ms. Faster than OutputFixingParser, cheaper than LLM retries. Streaming and batch support.

# Why Shim Exists

LLMs hallucinate. They truncate JSON mid-stream. They wrap valid objects in markdown fences. They forget closing brackets.

**Your options:**

1. **Retry** → Expensive ($0.002/retry × 1000 failures = $2/day)
2. **OutputFixingParser** → Slow (waits for full output, 2-5s delay)
3. **Regex hacks** → Brittle (breaks on nested objects)

**Shim:** Receive. Repair. Return. Sub-millisecond repair, `<10ms` API response.

***

## What Shim Does

Shim is a **streaming reliability layer** for LLM outputs. Route all your LLM traffic through Shim to get:

* **Real-time repair** (token-by-token, no buffering)
* **Schema validation** (type coercion, missing field detection)
* **Confidence scoring** (high/medium/low, never silent corruption)
* **Zero data storage** (privacy-first, GDPR compliant)

***

## How It Works

```typescript theme={null}
// Before: Broken JSON from LLM
const raw = '{"name": "John", "age": "30"'; // Missing closing brace

// After: Shim repairs it
const response = await fetch('https://api.shim.so/v1/repair', {
  method: 'POST',
  headers: { 'Authorization': 'Bearer sk_xxx' },
  body: JSON.stringify({ raw_output: raw })
});

const { repaired, metadata } = await response.json();
// repaired: { name: "John", age: 30 }
// metadata.confidence: "high"
// metadata.repairs_applied: ["closed_object_bracket", "type_coercion"]
```

***

## Key Features

### 1. Streaming-First Architecture

Batch repair is a tool. **Streaming repair is infrastructure.** Shim repairs JSON token-by-token without buffering delays.

### 2. Schema-Aware Repair

Provide a JSON Schema and Shim will:

* Coerce types (`"30"` → `30`)
* Detect missing required fields
* Remove extra fields (strict mode)
* Validate nested objects

### 3. Never Fails Silently

Every response includes:

* Confidence score (high/medium/low)
* Exact repairs applied
* Warnings for risky changes
* Errors for unrecoverable issues

### 4. Zero Data Persistence

Shim **never stores**:

* Raw LLM outputs
* Repaired JSON
* Field names or values

We only log metadata: repair types, confidence, latency.

### 5. Performance

Sub-millisecond repair, edge-deployed globally:

| Metric            | Performance                               |
| ----------------- | ----------------------------------------- |
| Repair engine     | `<0.1ms` (P99: 0.03ms)                    |
| Total API latency | 5-15ms (including auth, logging, network) |
| KV lookup         | `<1ms`                                    |
| Edge deployment   | 330+ locations globally                   |

**200-500x faster than OutputFixingParser** (which waits 2-5 seconds for full output).

***

## Who Uses Shim

**Indie Developers:** Building AI apps with LangChain/LlamaIndex\
**Startups:** Shipping AI features to production\
**Enterprises:** Deploying AI at scale with compliance requirements

***

## Next Steps

<CardGroup cols={2}>
  <Card title="Quick Start" icon="rocket" href="/quick-start">
    Get your first repair working in 5 minutes
  </Card>

  <Card title="API Reference" icon="code" href="/reference/batch-endpoint">
    Complete API documentation
  </Card>

  <Card title="Demo" icon="flask" href="/demo">
    Try Shim without signing up
  </Card>

  <Card title="TypeScript SDK" icon="npm" href="/sdks/typescript">
    Install the official SDK
  </Card>
</CardGroup>