The Response Healing plugin automatically validates and repairs malformed JSON responses from AI models. When models return imperfect formatting – missing brackets, trailing commas, markdown wrappers, or mixed text – this plugin attempts to repair the response so you receive valid, parseable JSON.
Overview
Response Healing provides:
- Automatic JSON repair: Fixes missing brackets, commas, quotes, and other syntax errors
- Markdown extraction: Extracts JSON from markdown code blocks
How It Works
The plugin activates for non-streaming requests when you use response_format with either type: "json_schema" or type: "json_object", and include the response-healing plugin in your plugins array. See the Complete Example below for a full implementation.
What Gets Fixed
The Response Healing plugin handles common issues in LLM responses:
JSON Syntax Errors
Input: Missing closing bracket
{"name": "Alice", "age": 30
{"name": "Alice", "age": 30}
Markdown Code Blocks
Input: Wrapped in markdown
```json
{"name": "Bob"}
```
Mixed Text and JSON
Input: Text before JSON
Here's the data you requested:
{"name": "Charlie", "age": 25}
{"name": "Charlie", "age": 25}
Trailing Commas
Input: Invalid trailing comma
{"name": "David", "age": 35,}
{"name": "David", "age": 35}
Unquoted Keys
Input: JavaScript-style
Output: Fixed
{"name": "Eve", "age": 40}
Complete Example
Limitations
Non-Streaming Requests OnlyResponse Healing only applies to non-streaming requests.
Will not repair all JSONSome malformed JSON responses may still be unrepairable. In particular, if the response is truncated by max_tokens, the plugin will not be able to repair it.
