提示词优化

# 初始提示词模板
baseline_prompt = """
Task: {task_description}
Input: {input_data}
Output:
"""

# 性能基准测试
baseline_metrics = {
    "accuracy": 0.72,
    "relevance": 0.68,
    "consistency": 0.65
}

{
  "system_role": "专业数据分析师",
  "task_context": {
    "objective": "分析用户行为数据",
    "constraints": ["准确性优先", "包含置信度"],
    "output_format": "structured_json"
  },
  "examples": [
    {"input": "...", "output": "...", "reasoning": "..."}
  ],
  "chain_of_thought": true,
  "self_consistency": 3
}

# OPRO优化循环
for iteration in range(max_iterations):
    # 生成候选提示词
    candidates = generate_prompt_variants(current_best)
    
    # 评估性能
    scores = evaluate_prompts(candidates, test_set)
    
    # 选择最优
    current_best = select_best(candidates, scores)
    
    # 早停检查
    if converged(scores):
        break

# 优化前（准确率72%）
prompt_v1 = "Classify this text: {text}"

# 优化后（准确率90%）
prompt_v2 = """
You are a text classification expert. Classify the following text into one of these categories:
- Technical: Programming, software, technology
- Business: Finance, marketing, management  
- Academic: Research, education, science

Text: {text}

Step 1: Identify key terms and context
Step 2: Match with category definitions
Step 3: Provide classification with confidence score

Output format:
Category: [category]
Confidence: [0.0-1.0]
Reasoning: [brief explanation]
"""

# 第1轮：基础提示词
"Analyze this data and provide insights"

# 第2轮：添加结构
"Analyze the data following these steps:
1. Data overview
2. Key patterns
3. Anomalies
4. Recommendations"

# 第3轮：添加约束和示例
"As a data scientist, analyze...
Constraints: Focus on actionable insights
Format: Executive summary (100 words) + Detailed analysis
Example: [provided]"

# 第4轮：元提示词+自动优化
# 使用OPRO迭代优化，性能提升47%

class PromptOptimizer:
    def __init__(self):
        self.variants = []
        self.metrics = {}
    
    def add_variant(self, prompt_template):
        self.variants.append(prompt_template)
    
    def run_test(self, test_data, sample_size=1000):
        for variant in self.variants:
            results = evaluate(variant, test_data[:sample_size])
            self.metrics[variant.id] = calculate_metrics(results)
        
        return self.select_winner()