Prompt Engineering实测：从7.2%到91.4%准确率的金融意图识别调优全记录

import asyncio
from openai import AsyncAzureOpenAI

client = AsyncAzureOpenAI(
    api_key="sk-xxx",
    api_version="2024-02-15-preview",
    azure_endpoint="https://your-resource.openai.azure.com/"
)

async def classify_p0(text: str) -> str:
    """P0: 零样本，纯文本输出"""
    response = await client.chat.completions.create(
        model="gpt-4o-mini-2024-07-18",
        temperature=0.2,
        messages=[
            {"role": "system", "content": "你是一个金融客服意图分类器。"
                                          "请判断以下用户输入的意图类别，"
                                          "类别包括：账户查询、转账汇款、贷款咨询、"
                                          "投诉建议、信用卡服务、投资理财、外汇兑换、"
                                          "密码重置、账户冻结、利率咨询、网点查询、"
                                          "其他问题。"},
            {"role": "user", "content": f"用户输入：{text}"}
        ]
    )
    return response.choices[0].message.content

# 测试第一条
print(asyncio.run(classify_p0("我想查一下余额")))
# 输出：这是一个关于账户信息查询的请求，用户希望了解当前账户的余额情况。
# 解析后：无法匹配到13个类别中的任何一个，只能算“其他问题”，但实际是“账户查询”。

async def classify_p1(text: str) -> dict:
    """P1: 零样本 + JSON Schema输出"""
    response = await client.chat.completions.create(
        model="gpt-4o-mini-2024-07-18",
        temperature=0.2,
        messages=[
            {"role": "system", "content": "你是金融意图分类器。"
                                          "输出必须严格符合JSON格式，"
                                          "包含两个字段：intent和confidence。"},
            {"role": "user", "content": f"请对以下输入分类：{text}"}
        ],
        tools=[
            {
                "type": "function",
                "function": {
                    "name": "classify_intent",
                    "description": "分类用户意图",
                    "parameters": {
                        "type": "object",
                        "properties": {
                            "intent": {
                                "type": "string",
                                "enum": ["账户查询", "转账汇款", "贷款咨询",
                                         "投诉建议", "信用卡服务", "投资理财",
                                         "外汇兑换", "密码重置", "账户冻结",
                                         "利率咨询", "网点查询", "其他问题"]
                            },
                            "confidence": {"type": "number", "minimum": 0, "maximum": 1}
                        },
                        "required": ["intent", "confidence"]
                    }
                }
            }
        ],
        tool_choice={"type": "function", "function": {"name": "classify_intent"}}
    )
    # 解析工具调用参数
    tool_call = response.choices[0].message.tool_calls[0]
    import json
    return json.loads(tool_call.function.arguments)

# 测试
print(asyncio.run(classify_p1("我想查一下余额")))
# 输出：{'intent': '账户查询', 'confidence': 0.95}

async def classify_p3(text: str, similar_examples: list[dict]) -> dict:
    """P3: 少样本+CoT+动态示例"""
    messages = [
        {"role": "system", "content": 
         "你是金融意图分类器。请先分析用户输入的关键信息，"
         "然后在JSON中输出reasoning（分析过程）和intent（最终类别）。"
         "严格按JSON输出，不要额外解释。"}
    ]
    # 动态示例
    for ex in similar_examples:
        messages.append({"role": "user", "content": f"输入：{ex['text']}"})
        messages.append({"role": "assistant", "content": 
                         f"{{'reasoning': '{ex['reasoning']}', "
                         f"'intent': '{ex['intent']}'}}"})
    # 当前测试样本
    messages.append({"role": "user", "content": f"输入：{text}"})

    response = await client.chat.completions.create(
        model="gpt-4o-mini-2024-07-18",
        temperature=0.2,
        messages=messages,
        response_format={"type": "json_object"}
    )
    return json.loads(response.choices[0].message.content)