Add pydantic model support for extraction schemas

[jordanalexmeyer]

Summary

Adds typed extraction schemas via pydantic, mirroring the TypeScript SDK's Zod support. Pass a pydantic.BaseModel subclass as config["schema"] and the SDK:

1. Converts the model to Extend's JSON Schema subset for the request (primitives forced nullable, all properties required, enums get a null option, additionalProperties: false)
2. Validates the extraction output back into model instances, returned as a TypedExtractRun[Model] — a step beyond the TS SDK, which only casts at compile time

from typing import List, Optional
from pydantic import BaseModel, Field
from extend_ai import Extend, ExtendCurrency, ExtendDate

class Invoice(BaseModel):
    invoice_number: Optional[str] = Field(None, description="The invoice number")
    invoice_date: ExtendDate = Field(None, description="The invoice date")
    total: Optional[ExtendCurrency] = Field(None, description="Total amount due")

client = Extend(token="...")
result = client.extract(file={"url": "https://example.com/invoice.pdf"}, config={"schema": Invoice})

result.output.value.invoice_number   # str | None — a validated Invoice instance
result.output.value.invoice_date     # datetime.date | None
result.output.value.total.amount     # float | None

Integration points (full TS parity)

• client.extract() (sync + async)
• client.extract_runs.create_and_poll() (sync + async), including extractor["override_config"]["schema"]
• client.extractors.create() / update() (new wrapper clients)
• client.extractor_versions.create() (new wrapper client)

@overload signatures give full static inference — mypy resolves result.output.value to the schema model, while untyped (dict) configs keep returning plain ExtractRun.

Details

• New wrapper/schema/ module: ExtendDate / ExtendCurrency / ExtendSignature field types (plain datetime.date annotations also map to the date type), pydantic_to_extend_schema() conversion, SchemaConversionError carrying the field path, typed run wrappers
• Supported field types: str, float, int, bool, datetime.date, Literal[...] / string enums, nested models, and lists of these; unsupported constructs (unions, dicts, datetime.datetime) raise SchemaConversionError
• Works with both pydantic v1 and v2 (via the existing IS_PYDANTIC_V2 compat pattern)
• All code lives under the .fernignore-protected wrapper/ directory
• README gains a "Typed extraction with Pydantic" section

Out of scope (matching TS): extend:name, per-enum-value extend:descriptions, typed schemas on plain extract_runs.create().

Testing

• Ported the TS schema.test.ts conversion suite (~45 cases) plus typed-extraction integration tests for all four endpoints
• pytest tests/wrapper tests/custom: 212 passed (1 pre-existing skip)
• Schema tests also verified against pydantic 1.10.13 in a separate environment
• mypy .: clean (1,134 files); ruff clean on all touched files

[jordanalexmeyer]

Superseded by #78, which stacks these pydantic commits on top of a fresh 1.17.0 Fern regeneration from documentation main.