Repair structured output from an LLM

Medium~20 min

Models asked for JSON often wrap it in prose or code fences, or add trailing commas. Extract the first JSON object from the text:

If the text contains a fenced block (json … or …), use the content of the first one.
Otherwise use the substring from the first { to the matching closing } (track nesting and ignore braces inside strings).
Remove trailing commas before } or ] (outside strings) and parse.

Return the parsed dictionary, or None if no valid object can be extracted.

Function

Implement extract_json. Input: text: str. Returns: dict | None.

>>> extract_json('Sure! Here it is:\n```json\n{"name": "web", "ports": [80, 443,],}\n```\nAnything else?')

{'name': 'web', 'ports': [80, 443]}

>>> extract_json('The answer is {"msg": "use {braces}, carefully", "ok": true} as requested.')

{'ok': True, 'msg': 'use {braces}, carefully'}

Level: API integration against fakes, data processing, error handling, decorators and asyncio. Several cooperating parts.
Roles: Generative AI Engineer, AI Engineer, Forward Deployed Engineer
Topics: LLM integration, JSON, Parsing text
Tests: 2 public (run with Run), 5 hidden (graded on Submit)
Grading: Exact match of the return value

Models asked for JSON often wrap it in prose or code fences, or add trailing commas. Extract the first JSON object from the text:

If the text contains a fenced block (json … or …), use the content of the first one.
Otherwise use the substring from the first { to the matching closing } (track nesting and ignore braces inside strings).
Remove trailing commas before } or ] (outside strings) and parse.

Return the parsed dictionary, or None if no valid object can be extracted.

Implement extract_json. Input: text: str. Returns: dict | None.

>>> extract_json('Sure! Here it is:\n```json\n{"name": "web", "ports": [80, 443,],}\n```\nAnything else?')

{'name': 'web', 'ports': [80, 443]}

>>> extract_json('The answer is {"msg": "use {braces}, carefully", "ok": true} as requested.')

{'ok': True, 'msg': 'use {braces}, carefully'}

Level: API integration against fakes, data processing, error handling, decorators and asyncio. Several cooperating parts.
Roles: Generative AI Engineer, AI Engineer, Forward Deployed Engineer
Topics: LLM integration, JSON, Parsing text
Tests: 2 public (run with Run), 5 hidden (graded on Submit)
Grading: Exact match of the return value

Python 3.13

import json

def extract_json(text: str):
    try:
        return json.loads(text)
    except ValueError:
        return None

Ctrl/⌘ + Enter runs · Ctrl/⌘ + Shift + Enter submits

Run executes the public tests. Submit grades against public and hidden tests.

Repair structured output from an LLM · Python practice