JSON Normalizer

JSON Normalizer ensures that LLM JSON outputs are clean, parseable, and optionally schema-conformant before they reach your application. It handles the common real-world issues with LLM JSON output: markdown code fences, surrounding prose, inconsistent whitespace, and structural deviations.

All extraction behaviors are always active — the module always strips fences, extracts JSON from prose, and parses the result. Schema validation is optional.

Stage

What It Does

The module applies the following steps to every LLM response, in order:

1. Fence removal: Strips ```json / ``` Markdown code fences if present.
2. JSON extraction: Finds and extracts the first valid JSON object {} or array [] from the surrounding text, even if the LLM added explanatory prose before or after.
3. Parsing: Parses the extracted string to confirm it is valid JSON.
4. Schema validation (optional): If a JSON Schema is configured, the parsed object is validated against it. If validation fails, the pipeline raises an exception immediately.

Example

Raw LLM response:

Sure! Here is the result:

```json
{
  "status": "success",
  "items": ["apple", "banana"],
  "count": 2
}

Let me know if you need anything else!

**After normalization:**
```json
{"status": "success", "items": ["apple", "banana"], "count": 2}

Configuration

Expected Output Schema

A JSON Schema used to validate the extracted JSON output. You can build the schema visually in the Infralo UI by pasting an example JSON response — field types and required constraints are inferred automatically and can be adjusted.

Leave empty to skip schema validation and use this module purely for extraction and parsing.

Allow Additional Properties

Controls whether the LLM response may include fields not declared in the schema.

Failure Behavior

If the extracted JSON does not match the configured schema and Allow Additional Properties is disabled, the module raises a pipeline exception immediately. It does not:

• Retry the LLM call
• Attempt to repair the JSON
• Fall back to the raw response

This is intentional — it ensures that your downstream application never receives malformed or non-conformant data silently. Handle this at the application level by catching gateway errors and retrying if needed.

Example Use Cases

• Structured output APIs: Ensure every LLM call that is expected to return JSON actually does so — gate on it at the infrastructure level.
• Schema enforcement: Validate that required fields are present and correctly typed before the response reaches your database or downstream service.
• Fence stripping: Many LLMs wrap JSON in markdown code fences by default; this module removes them reliably without any prompt engineering.
• Prose cleanup: When using few-shot prompting that sometimes generates explanatory text around JSON, this module extracts only the JSON automatically.