Structured Extraction

Structured extraction — text in, a typed object you can trust out

The deliverable is a typed object that conforms to a schema you defined — not prose, not "roughly JSON." The whole skill rests on one distinction the rest of the file keeps returning to:

Native structured outputs make the JSON valid and typed. They never make the values correct.

Constrained decoding guarantees the model cannot emit a token that breaks your schema, so JSON.parse errors, missing keys, wrong types, and stray markdown fences disappear at the source. It does nothing to stop the model from putting a plausible-but-wrong email in a string field, snapping a fuzzy category to the wrong enum, or coercing "$1,200" into 1200.0 when the currency mattered. Owning both halves — the shape (decoding) and the values (validation) — is this skill. If you only do the first half you ship a database full of well-typed lies.

Boundary test (bytes vs. schema). If the input is a PDF, scan, DOCX, or HTML and the deliverable is the raw text/Markdown/cells of that document, that is upstream: document-processing produces the text, this skill turns that text into typed fields. If you're holding text and want it shaped, you're in the right place.

Current as of 2026-06-02: OpenAI Structured Outputs (strict: true json_schema), Anthropic Structured Outputs (GA since the 2025-11-14 public beta; output_config.format), and Instructor (built on Pydantic, ~3M downloads/month). Exact request/response shapes and the per-provider limit tables live in references/providers.md so this file stays lean.

Default: native constrained decoding when the provider has it

If the model and provider support native structured outputs, use them. This is not a tuning knob — it is the difference between ~100% schema conformance and hoping a regex catches the fence.

Bad — prompt-and-pray, then parse raw text:

resp = client.chat.completions.create(
    model="gpt-5.1",
    messages=[{"role": "user", "content": f"Return JSON with name and email:\n{text}"}],
)
data = json.loads(resp.choices[0].message.content)  # markdown fence / preamble / missing key -> crash

Good — OpenAI strict json_schema (Chat Completions):

resp = client.chat.completions.create(
    model="gpt-5.1",
    messages=[{"role": "user", "content": text}],
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "contact",
            "strict": True,
            "schema": {
                "type": "object",
                "additionalProperties": False,
                "required": ["name", "email"],
                "properties": {
                    "name": {"type": "string", "description": "Full name as written."},
                    "email": {"type": ["string", "null"],
                              "description": "Email exactly as written, or null if none is stated."},
                },
            },
        },
    },
)
data = json.loads(resp.choices[0].message.content)  # now guaranteed valid + typed

On the OpenAI Responses API the same block moves under text.format instead of response_format. On Anthropic, the equivalent is output_config={"format": {"type": "json_schema", "schema": {...}}} on Claude Opus 4.5–4.8 / Sonnet 4.5+ / Haiku 4.5; Anthropic compiles your schema into a grammar and caches it for 24h, and the SDKs ship helpers (client.messages.parse(...) in Python, zodOutputFormat(schema) in TS). The older deprecated output_format param and the deprecated structured-outputs-2025-11-13 beta header still work in a transition window — do not write new code against them. Full shapes in references/providers.md.

The non-negotiable strict-schema rule (OpenAI and Anthropic both):

• every object sets "additionalProperties": false;
• every property is listed in required;
• "optional" is expressed as a union with null ("type": ["string", "null"]), never by leaving the field out of required. Omitting it is the single most common strict-mode error.

Schema design rules (each prevents a specific failure)

Keep schemas flat and shallow. If you find yourself nesting four levels deep or describing a tree, that is two extractions, not one heroic schema.

The reliability ladder — escalate only on failure, and cap it

Climb from the cheapest mechanism upward. Each rung catches what the rung below cannot; you stop at the first rung that holds for your data.

Rung 1 is mandatory when available. Rung 2 is mandatory the moment any field has a value rule (a range, a format, a "must match the order date") — because rung 1 structurally cannot enforce values. Rungs 3 and 4 are opt-in. Never make rung 3 unbounded: a retry loop with no cap turns one bad document into an unbounded bill.

Value validation the schema can't enforce

This is the half native decoding leaves on the table. Validate values after you have a typed object.

Pydantic — value rules + normalization the schema can't carry:

from pydantic import BaseModel, field_validator

class Order(BaseModel):
    amount_cents: int
    discount_pct: float | None  # nullable: may be absent
    order_date: str             # we'll normalize to ISO

    @field_validator("discount_pct")
    @classmethod
    def pct_in_range(cls, v):
        if v is not None and not (0 <= v <= 100):
            raise ValueError("discount_pct must be between 0 and 100")
        return v

    @field_validator("amount_cents", mode="before")
    @classmethod
    def strip_currency(cls, v):
        if isinstance(v, str):  # "$1,200.00" -> 120000
            return int(round(float(v.replace("$", "").replace(",", "")) * 100))
        return v

The Zod equivalent uses .refine() for cross-field and range checks and .transform() for normalization. Three normalizations bite constantly: currency ("$1,200" vs 1200 vs 120000 cents — pick one and enforce it), dates (free text → ISO 8601, and decide what a missing year means), and enum snapping (the model rounds "kinda urgent" to urgent; validate that the snap was legitimate, or widen the enum).

Null vs. absent. A nullable field with a clear instruction is the entire fix for "the model invents an email." "email": {"type": ["string","null"], "description": "...or null if the text states no email"} plus a one-line system instruction ("use null for any field not present in the source; never guess"). If you make the field non-nullable, you have told the model to produce a value — it will.

Bounded reask with Instructor — failed validation is fed back to the model as an error message:

import instructor

client = instructor.from_provider("openai/gpt-5.1")
order = client.chat.completions.create(
    response_model=Order,            # your Pydantic model, validators and all
    max_retries=2,                   # BOUND it; each retry is another paid call
    messages=[{"role": "user", "content": text}],
)

On a validation failure Instructor reasks with the ValueError text, so @field_validator rules the model never saw in the schema still get enforced through the loop. Log every reask (count + reason): a quietly climbing reask rate is your early signal that a field's instruction or schema is wrong.