Instructor / Outlines dla gwarantowanej struktury

Dwie główne biblioteki

Instructor Basics

# pip install instructor openai
import instructor
from openai import OpenAI
from pydantic import BaseModel

client = instructor.from_openai(OpenAI())

class User(BaseModel):
    name: str
    age: int

user = client.chat.completions.create(
    model='gpt-4o-mini',
    response_model=User,
    messages=[{'role': 'user', 'content': 'Alice, 30'}]
)
print(user.name, user.age)

Magia Instructor

Instructor with Validation

from pydantic import field_validator

class User(BaseModel):
    name: str
    age: int

    @field_validator('age')
    @classmethod
    def positive(cls, v):
        if v <= 0:
            raise ValueError('Age must be positive')
        return v

# Instructor catches ValidationError and retries automatically

Streaming Instructor

for partial in client.chat.completions.create_partial(
    model='gpt-4o-mini',
    response_model=User,
    messages=[{'role': 'user', 'content': 'Alice, 30'}]
):
    print(partial)
# Streams partial Pydantic instances as fields fill in.

Instructor with Anthropic

from anthropic import Anthropic
client = instructor.from_anthropic(Anthropic())

user = client.messages.create(
    model='claude-sonnet-4-5',
    max_tokens=1024,
    response_model=User,
    messages=[{'role': 'user', 'content': 'Alice, 30'}]
)

Outlines dla modeli OSS

# pip install outlines
import outlines

model = outlines.models.transformers('mistralai/Mistral-7B-Instruct-v0.2')

generator = outlines.generate.json(model, User)
user = generator('Alice, 30')

Dlaczego dekodowanie ograniczone?

Wyrażenia regularne i wybór w Outlines

import outlines.text.generate as g
yes_no = g.choice(model, ['yes', 'no'])
result = yes_no('Are bananas fruits?')   # 'yes'

Łączenie obu podejść

Porównanie kosztów

Kiedy używać którego rozwiązania

Podejście Outlines

Podsumowanie