我用3轮实验把Prompt从78%优化到96%：提示词工程Token消耗与输出质量对比

Python: 3.11.6
openai: 1.35.0
模型: gpt-4o-mini (snapshot: 2024-07-18)
temperature: 0（固定，减少随机性）
max_tokens: 128
测试集: 200条人工标注的电商评论
评测指标: 准确率 / 平均Token消耗 / 格式合规率 / 平均延迟

from openai import OpenAI
import json, time

client = OpenAI(api_key="sk-xxx")

def classify_v1(text: str) -> dict:
    prompt = f"""判断这条电商评论的情感倾向，正面、负面还是中性：
{text}"""

    start = time.time()
    resp = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": prompt}],
        temperature=0,
        max_tokens=128,
    )
    latency = time.time() - start
    content = resp.choices[0].message.content.strip()

    # 暴力解析，先看能不能用
    label = "unknown"
    for k in ["正面", "负面", "中性"]:
        if k in content:
            label = k
            break

    return {
        "label": label,
        "raw": content,
        "tokens": resp.usage.total_tokens,
        "latency": round(latency, 3),
    }

准确率: 78.5%
平均Token: 142
格式合规率: 61%  (能稳定解析出标签的比例)
平均延迟: 0.82s

SYSTEM_PROMPT = """你是一个电商评论情感分析引擎，专门处理中文用户评论。
你的输出必须严格符合以下JSON schema，不要输出任何其他文字：
{"label": "正面|负面|中性", "intensity": 1-5, "reason": "20字以内的判断依据"}

分类标准：
- 正面：明确表达满意、推荐、复购意愿
- 负面：明确表达不满、投诉、退货意愿  
- 中性：仅陈述事实、无明确情绪倾向，或正负抵消

intensity评分标准：
1 = 极其负面（要求退款/差评威胁）
2 = 轻微负面（小瑕疵但能接受）
3 = 完全中性（纯事实描述）
4 = 轻微正面（基本满意）
5 = 极其正面（强烈推荐/复购）"""

FEW_SHOT = [
    {"role": "user", "content": "这衣服质量还行，就是物流有点慢"},
    {"role": "assistant", "content": '{"label":"中性","intensity":3,"reason":"褒贬抵消"}'},
    {"role": "user", "content": "客服态度太差了，申请退款！"},
    {"role": "assistant", "content": '{"label":"负面","intensity":1,"reason":"投诉退款意图"}'},
    {"role": "user", "content": "用了一周，真心好用，已经推荐给朋友了"},
    {"role": "assistant", "content": '{"label":"正面","intensity":5,"reason":"推荐复购意愿"}'},
]

def classify_v3(text: str) -> dict:
    messages = [{"role": "system", "content": SYSTEM_PROMPT}]
    messages.extend(FEW_SHOT)
    messages.append({"role": "user", "content": text})

    start = time.time()
    resp = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=messages,
        temperature=0,
        max_tokens=128,
        response_format={"type": "json_object"},  # 关键：强制JSON
    )
    latency = time.time() - start
    content = resp.choices[0].message.content

    try:
        parsed = json.loads(content)
        label = parsed.get("label", "unknown")
    except json.JSONDecodeError:
        parsed, label = {}, "unknown"

    return {
        "label": label,
        "parsed": parsed,
        "tokens": resp.usage.total_tokens,
        "latency": round(latency, 3),
    }