Fix invalid JSON returned by an AI assistant

An application expects a JSON object but receives an explanation, a Markdown fence or a trailing comma. The text may be readable to a person while still being unusable by the parser.

Why this happens

Natural-language instructions alone do not create a strict output contract. Even supported structured-output features need application handling for refusals, incomplete responses and schema constraints.

Step-by-step solution

1. Define the smallest schema your application needs, including required fields and allowed values. Use the provider’s supported structured-output mechanism when available.

2. Parse the response and validate its types before any database write. Handle a refused or incomplete result separately from a valid business answer.

3. On failure, log a redacted diagnostic and return a controlled error. Do not evaluate model text as executable code or strip arbitrary characters until it happens to parse.

Worked example

import json
value = json.loads('{"category":"Technology","approved":false}')
assert isinstance(value, dict)
assert value['category'] in {'Technology','Culture'}
assert type(value['approved']) is bool

Verify the fix

The example passes. Test a missing field, a string instead of a boolean and plain explanatory text; none should reach the publishing action as a valid approval.

Prevent a repeat

Version the schema and save representative malformed-response fixtures. A syntactically valid object can still contain an incorrect answer, so validate meaning separately.

Reference

Official documentation or standards guidance

An original Gistmax troubleshooting guide. Scenarios and sample values are illustrative; configuration depends on your software version and environment. Cover diagram: Gistmax Tutorials.