AI 垂直 Agent 全景解析：从通用到专用的范式转变

from transformers import AutoModelForCausalLM, AutoTokenizer, Trainer, TrainingArguments
from datasets import load_dataset

# 1. 加载基础模型和领域数据
base_model = "openai/gpt-5"
tokenizer = AutoTokenizer.from_pretrained(base_model)

# 领域专用数据集（生命科学）
domain_data = load_dataset("bio-instructions", split="train")

def format_example(example):
    """将领域数据格式化为指令微调格式"""
    return f"""<|system|>你是一个生命科学研究助手。请基于科学证据回答问题，
并在不确定时明确标注置信度。

<|user|>{example['instruction']}

<|assistant|>{example['output']}<|confidence|>{example['confidence_score']}"""

tokenized_data = domain_data.map(
    lambda x: {"text": format_example(x)},
    remove_columns=domain_data.column_names
)

# 2. 领域继续预训练（Domain-Adaptive Pretraining）
training_args = TrainingArguments(
    output_dir="./gpt-rosalind-domain",
    num_train_epochs=3,
    per_device_train_batch_size=8,
    learning_rate=1e-5,  # 低学习率避免破坏已有知识
    fp16=True,
    gradient_accumulation_steps=4,
)

trainer = Trainer(
    model=AutoModelForCausalLM.from_pretrained(base_model),
    args=training_args,
    train_dataset=tokenized_data,
    tokenizer=tokenizer,
)

trainer.train()
trainer.save_model("./gpt-rosalind-domain")

# 3. 使用微调后的模型进行领域推理
from openai import OpenAI

client = OpenAI()
response = client.chat.completions.create(
    model="gpt-rosalind-domain",
    messages=[
        {"role": "user", "content": "分析以下 DNA 序列的基因变异：ATCGTACG..."},
    ],
    response_format={"type": "json_object"},  # 结构化输出
)

result = response.choices[0].message.content
print(f"变异分析结果: {result}")
# 输出包含置信度: {"variant": "T→A at position 5", "confidence": 0.94}

from agents import Agent, function_tool, Runner
from typing import List
import json

# 1. 定义领域专用工具
@function_tool
def generate_layout(description: str, style: str = "modern") -> dict:
    """生成 UI 布局方案"""
    # 调用设计引擎 API
    layout = {
        "grid": "12-column",
        "components": [
            {"type": "header", "position": "top", "height": "64px"},
            {"type": "hero", "position": "below-header", "height": "400px"},
            {"type": "content", "position": "main", "cols": "8"},
            {"type": "sidebar", "position": "right", "cols": "4"},
        ],
        "style_guide": style,
    }
    return layout

@function_tool
def generate_color_palette(base_color: str, mood: str = "professional") -> dict:
    """生成配色方案"""
    palettes = {
        "professional": ["#1a365d", "#2d3748", "#4a5568", "#edf2f7"],
        "creative": ["#6b46c1", "#805ad5", "#d6bcfa", "#faf5ff"],
        "warm": ["#c05621", "#dd6b20", "#ed8936", "#fefcbf"],
    }
    return {"colors": palettes.get(mood, palettes["professional"]), "base": base_color}

@function_tool
def export_to_code(layout: dict, palette: dict, format: str = "react") -> str:
    """将设计导出为代码"""
    if format == "react":
        return f"""import React from 'react';

export default function DesignComponent() {{
  return (
    <div style={{{{
      display: 'grid',
      gridTemplateColumns: 'repeat(12, 1fr)',
      gap: '16px',
      padding: '24px',
    }}}}>
      {/* Header */}
      <header style={{{{ gridColumn: '1 / -1', height: '64px',
        background: '{palette['colors'][0]}', color: '#fff' }}}}>
        Header
      </header>
      {/* Main Content */}
      <main style={{{{ gridColumn: '1 / 9' }}}}>Content</main>
      <aside style={{{{ gridColumn: '9 / -1' }}}}>Sidebar</aside>
    </div>
  );
}}"""
    return ""

# 2. 构建垂直 Agent
design_agent = Agent(
    name="设计专家 Agent",
    instructions="""你是一个专业 UI 设计师。
    根据用户需求生成设计方案，包括布局、配色和代码。
    始终考虑用户体验和设计规范。""",
    tools=[generate_layout, generate_color_palette, export_to_code],
)

# 3. 运行 Agent
async def main():
    result = await Runner.run(
        design_agent,
        "为一家金融科技公司设计一个落地页，风格专业、可信",
    )
    print(result.final_output)