Structured outputs from AI: Pydantic, Instructor and JSON Schema in production

Structured Outputs with JSON Schema or the Instructor library guarantee that the model returns data in exactly the format your code expects — validated either at the API level or in client code. This eliminates the entire string-parsing nightmare: no markdown fences, no comments inside JSON, no fields in the wrong type. If you are building a data extraction pipeline or an LLM integration with any downstream system, this is the only approach suitable for production.

How to stop parsing strings from GPT and start getting data ready to use in code — JSON Schema, Pydantic and the Instructor library.

Tuesday morning, production deployment. The model returns: `{"name": "Jan Kowalski", "age": "thirty-two", "tags": "python, django"}`. Your code expected `age` as int, `tags` as list — and it throws an exception. The model "tried its best", but it couldn't know that the list is `["python", "django"]`, not a string. This isn't an edge case — it's the daily reality when an LLM and code communicate through a string.

Three Approaches — and Why the First Two Fail

Most teams go through the same phases. Phase 1 — "tell GPT to return JSON" — works for a week, then the model adds a markdown fence or a comment and `json.loads` blows up. Phase 2 — JSON mode (`response_format={"type": "json_object"}`) — stable JSON, but without a schema the model decides the field shapes itself. Phase 3 — Structured Outputs with JSON Schema or Instructor — you get exactly what you described, validated either at the API level or in code.

JSON Schema and strict mode — API-side validation

OpenAI Structured Outputs (from GPT-4o) enforce the schema at the tokenisation level — the model only generates tokens matching the defined structure. `strict: true` + `response_format` with `json_schema` guarantees the response always parses without error. Requirements: every object needs `additionalProperties: false` and all fields in `required` — you implement optionality via `anyOf` with `{"type": "null"}`.

from openai import OpenAIimport jsonclient = OpenAI()SCHEMA = {    "name": "order_extraction",    "strict": True,    "schema": {        "type": "object",        "properties": {            "customer_name": {"type": "string"},            "order_id": {"type": "string"},            "items": {                "type": "array",                "items": {                    "type": "object",                    "properties": {                        "product": {"type": "string"},                        "quantity": {"type": "integer"},                        "price_pln": {"type": "number"}                    },                    "required": ["product", "quantity", "price_pln"],                    "additionalProperties": False                }            },            "total_pln": {"type": "number"}        },        "required": ["customer_name", "order_id", "items", "total_pln"],        "additionalProperties": False    }}resp = client.chat.completions.create(    model="gpt-4o",    messages=[{"role": "user", "content": "Extract: Jan Kowalski, ORD-001234, 3x coffee 12.99 PLN, 1x tea 8.50 PLN"}],    response_format={"type": "json_schema", "json_schema": SCHEMA})order = json.loads(resp.choices[0].message.content)print(order["total_pln"])

The result is always valid JSON matching the schema — zero exceptions from `json.loads`. The downside is verbosity: for complex objects, JSON Schema quickly becomes unreadable and hard to maintain.

Pydantic as the schema description layer

Instead of writing JSON Schema by hand, describe the structure as a Pydantic class. `Model.model_json_schema()` generates the schema automatically from type hints and validators. The key: `Field(description=...)` — the LLM reads field descriptions and fills data far more accurately when it knows what you expect. `field_validator` lets you add business rules that JSON Schema can't express — sum validation, ID format, conditional rules.

from pydantic import BaseModel, Field, field_validatorfrom typing import Optionalimport reclass OrderItem(BaseModel):    product: str = Field(description="Product name exactly as written in the text")    quantity: int = Field(ge=1, description="Number of units, min 1")    price_pln: float = Field(gt=0, description="Unit price in PLN")class Order(BaseModel):    customer_name: str = Field(description="Customer's first and last name")    order_id: str = Field(description="Order ID in format ORD-XXXXXX")    items: list[OrderItem] = Field(description="List of all order items")    total_pln: float = Field(description="Sum of all items in PLN")    notes: Optional[str] = Field(default=None, description="Notes if provided, otherwise null")    @field_validator("order_id")    @classmethod    def validate_order_id(cls, v: str) -> str:        if not re.match(r"ORD-d{6}$", v):            raise ValueError(f"order_id must be ORD-XXXXXX, got: {v}")        return v    @field_validator("total_pln")    @classmethod    def validate_total(cls, v: float, info) -> float:        if "items" in info.data:            expected = sum(i.price_pln * i.quantity for i in info.data["items"])            if abs(v - expected) > 0.01:                raise ValueError(f"total_pln {v} != sum of items {expected:.2f}")        return v

`field_validator` lets you define business rules — sum validation, ID format, date ranges — that JSON Schema can't handle. A validation error gives you a concrete message you can pass back to the model in the next retry.

Instructor — 3 lines of code instead of your own parser

Instructor wraps the OpenAI client (and 10+ other providers) and turns the response directly into a validated Pydantic object. You don't need `json.loads`, `model.model_validate` or manual retry — the library handles it for you with 3 retries by default, sending the validation error message back to the model as context.

import instructorfrom openai import OpenAIfrom pydantic import BaseModel, Fieldfrom typing import Literalclient = instructor.from_openai(OpenAI())class ProductReview(BaseModel):    sentiment: Literal["positive", "negative", "neutral"]    score: int = Field(ge=1, le=5, description="Rating 1–5")    key_issues: list[str] = Field(description="List of main problems or strengths, max 5 points")    would_recommend: bool    summary: str = Field(max_length=200, description="One-sentence summary")review = client.chat.completions.create(    model="gpt-4o",    response_model=ProductReview,    messages=[        {"role": "user", "content": "Analyse: 'Product arrived damaged, support didn't pick up for 3 days, eventually got a refund but wasted my time. Never again.'"}    ])print(review.sentiment)print(review.score)print(review.key_issues)

`response_model=ProductReview` is all you need — Instructor generates the JSON Schema from the class, calls the API, parses the response, validates with Pydantic, and on failure automatically retries with the error appended to the conversation context.

Patterns: extraction, classification, normalisation

Three main use cases differ in their approach to the schema. Extraction (pulling data from text) — use `Optional` for fields that may not appear; never force fields the model can't fill. Classification — use `Literal` or `Enum` instead of `str`, the model will only choose from allowed values. Normalisation — describe the exact output format with an example in `description` and use `field_validator` to verify.

from enum import Enumfrom typing import Optional, Literalfrom pydantic import BaseModel, Field, field_validatorfrom datetime import datetimeimport instructorfrom openai import OpenAIclass Priority(str, Enum):    LOW = "low"    MEDIUM = "medium"    HIGH = "high"    CRITICAL = "critical"class TicketExtraction(BaseModel):    title: str = Field(max_length=100, description="Short ticket title")    priority: Priority = Field(description="Priority based on urgency and business impact")    category: Literal["bug", "feature", "question", "billing"]    affected_users: Optional[int] = Field(default=None, ge=1, description="Number of affected users if stated, otherwise null")    reported_at: Optional[str] = Field(default=None, description="Date in ISO 8601 format e.g. 2026-06-05T10:30:00Z, null if unknown")    is_regression: bool = Field(description="True if it worked before")    @field_validator("reported_at")    @classmethod    def validate_date(cls, v: Optional[str]) -> Optional[str]:        if v is None:            return v        try:            datetime.fromisoformat(v.replace("Z", "+00:00"))        except ValueError:            raise ValueError(f"reported_at must be ISO 8601, got: {v}")        return vclient = instructor.from_openai(OpenAI())ticket = client.chat.completions.create(    model="gpt-4o",    response_model=TicketExtraction,    messages=[{"role": "user", "content": "URGENT: login stopped working at 10:30, around 500 users can't log in, it worked before"}])print(ticket.priority)print(ticket.affected_users)

Instructor works with multiple providers — `instructor.from_anthropic()`, `instructor.from_gemini()`, `instructor.from_mistral()` — same Pydantic code, different client.

When structured output fails — 4 scenarios

Even with Instructor you hit walls. Here are the four main ones and how to get out.

1. The model can't fill a required field. Symptom: retry loop, model hallucinates a value just to put "something". Fix: change the field to `Optional` and add `description="null if unknown"` — let the model admit missing information.

2. The schema is too complex. Symptom: the model fills a field with a random value instead of null. Fix: simplify to a flat structure. If you need complexity, split into two calls — first extracts flat data, second classifies or normalises.

3. Business validation fails after 3 retries. Symptom: `InstructorRetryException`. Fix: catch the exception and log the model's last attempt — often the rule is too restrictive or the prompt doesn't contain the information the validator expects. Loosen the validator or enrich the prompt with an example of a correct response.

4. The list has too few elements. Symptom: `items` has 2 instead of 5 entries. Fix: add `"Extract ALL items — don't skip any"` to the `description`. Instructor also supports `Iterable[Model]` as `response_model` — the model streams objects incrementally.

---

I build data extraction and classification systems for companies — from simple pipelines to complex multi-step architectures with business validation and monitoring. Get in touch — I start with an analysis of your input data and schema design.