Your personal operating system for career transition.

Parse and Extract Structured Data from LLM JSON Responses

LLMs asked to return JSON do so imperfectly. They wrap output in markdown code fences, add explanatory prose before the JSON, omit required fields, or return strings where you expect numbers. A single fragile json.loads(response) call is not a production parsing strategy.

What to build

Implement three functions and a Pydantic model:

1. strip_fences(text: str) -> str — remove markdown code fences (```json, ```) from the start and end of a string.

2. extract_json_object(text: str) -> str | None — find the first complete {...} or [...] block in the text using a regex. Return None if none is found.

3. parse_into(text: str, model_class: type[T]) -> T | None — combine the above: strip fences, extract JSON, parse with json.loads, validate with Pydantic. Return None on any failure, never raise.

4. EntityExtraction — a Pydantic model with:

• entities: list[str] — required, at least 1 item
• sentiment: Literal["positive", "neutral", "negative"]
• confidence: float — between 0.0 and 1.0

Test cases to pass

Input: '{"entities": ["Paris", "Eiffel Tower"], "sentiment": "positive", "confidence": 0.9}'
Expected: EntityExtraction parsed successfully

Input: '```json\n{"entities": ["London"], "sentiment": "neutral", "confidence": 0.8}\n```'
Expected: fences stripped, parsed successfully

Input: 'Here is the result: {"entities": ["NYC"], "sentiment": "positive", "confidence": 0.95}'
Expected: JSON extracted from prose, parsed successfully

Input: '{"sentiment": "positive", "confidence": 0.7}'  (missing entities)
Expected: returns None (validation failure)

Input: 'No JSON here at all'
Expected: returns None

Why this matters

This three-layer pattern — fence stripping, JSON extraction, schema validation — handles 95% of real LLM structured output failures. Build it once and reuse it across every prompt task that returns structured data.