结构化输出和限制式解码

// BAD
{"answer": "yes", "reasoning": "because ..."}

// GOOD
{"reasoning": "... therefore ...", "answer": "yes"}

def mask_logits(logits, valid_token_ids):
    mask = [float("-inf")] * len(logits)
    for tid in valid_token_ids:
        mask[tid] = logits[tid]
    return mask


def generate_constrained(model, tokenizer, prompt, fsm):
    ids = tokenizer.encode(prompt)
    state = fsm.initial_state
    while not fsm.is_accept(state):
        logits = model.next_token_logits(ids)
        valid = fsm.valid_tokens(state, tokenizer)
        logits = mask_logits(logits, valid)
        tok = sample(logits)
        ids.append(tok)
        state = fsm.transition(state, tok)
    return tokenizer.decode(ids)

from pydantic import BaseModel
from typing import Literal
import outlines


class Review(BaseModel):
    sentiment: Literal["positive", "negative", "neutral"]
    confidence: float
    evidence_span: str


model = outlines.models.transformers("meta-llama/Llama-3.2-3B-Instruct")
generator = outlines.generate.json(model, Review)

result = generator("Classify: 'The wait staff was attentive and the food arrived hot.'")
print(result)
# Review(sentiment='positive', confidence=0.93, evidence_span='attentive ... hot')

import instructor
from anthropic import Anthropic
from pydantic import BaseModel, Field


class Invoice(BaseModel):
    vendor: str
    total_usd: float = Field(ge=0)
    line_items: list[str]


client = instructor.from_anthropic(Anthropic())
invoice = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=1024,
    response_model=Invoice,
    messages=[{"role": "user", "content": "Extract from: 'Acme Corp $420. Widget, Gizmo.'"}],
)

from openai import OpenAI

client = OpenAI()
response = client.responses.create(
    model="gpt-5",
    input=[{"role": "user", "content": "Classify: 'The food was cold.'"}],
    text={"format": {"type": "json_schema", "name": "sentiment",
          "schema": {"type": "object", "required": ["sentiment"],
                     "properties": {"sentiment": {"type": "string",
                                                  "enum": ["positive", "negative", "neutral"]}}}}},
)
print(response.output_parsed)

---
name: structured-output-picker
description: Choose a structured output approach, schema design, and validation plan.
version: 1.0.0
phase: 5
lesson: 20
tags: [nlp, llm, structured-output]
---

Given a use case (provider, latency budget, schema complexity, failure tolerance), output:

1. Mechanism. Native vendor structured output, Instructor retries, Outlines FSM, or XGrammar CFG. One-sentence reason.
2. Schema design. Field order (reasoning first, answer last), nullable fields for "unknown", enum vs regex, required fields.
3. Failure strategy. Max retries, fallback model, graceful `null` handling, out-of-distribution refusal.
4. Validation plan. Schema compliance rate (target 100%), semantic validity (LLM-judge), field-coverage rate, latency p50/p99.

Refuse any design that puts `answer` or `decision` before reasoning fields. Refuse to use bare JSON mode without a schema. Flag recursive schemas behind an FSM-only library.