Perplexity Pro Search深度搜索

import refrom typing import List, Dict, Tuple, Optionalfrom dataclasses import dataclass, field  @dataclassclass SearchResult:    """单条搜索结果"""    url: str    title: str    snippet: str    content: str = ""    source_quality: float = 0.5  # 来源质量评分  @dataclassclass SubQuestion:    """子问题"""    question: str    query: str  # 转化后的搜索查询    order: int  # 搜索顺序    status: str = "pending"  # pending / searching / done    results: List[SearchResult] = field(default_factory=list)  @dataclassclass InfoFact:    """提取的信息事实"""    fact: str    source_url: str    confidence: float = 0.8    verified: bool = False  # 是否多源验证  class SimpleSearchEngine:    """简化的搜索引擎模拟        实际场景中应该调用真实的搜索API    """        def __init__(self):        self.knowledge_base = {            "gpt4 code": "GPT-4在HumanEval上达到67%的pass@1，在代码生成任务上表现优异。支持多种编程语言，包括Python、JavaScript、C++等。",            "claude code": "Claude 3 Opus在代码生成任务上表现出色，HumanEval pass@1约为84.9%，特别擅长长代码理解和修改。",            "gpt4 benchmark": "GPT-4在MMLU上达到86.4%，在GSM8K数学题上达到92%，是多模态能力最强的模型之一。",            "claude benchmark": "Claude 3 Opus在MMLU上达到86.8%，在GSM8K上达到95%，推理能力突出。",            "gpt4 context": "GPT-4支持128K token上下文窗口，约合96000个单词，可以处理长文档。",            "claude context": "Claude 3支持200K token上下文窗口，约合150000个单词，长文档处理能力更强。",        }        def search(self, query: str, num_results: int = 5) -> List[SearchResult]:        """模拟搜索"""        results = []        query_lower = query.lower()                for key, content in self.knowledge_base.items():            # 简单的关键词匹配            query_words = set(re.findall(r'\w+', query_lower))            key_words = set(key.split())            overlap = len(query_words & key_words)                        if overlap > 0:                results.append(SearchResult(                    url=f"https://example.com/{key.replace(' ', '-')}",                    title=f"{key} - 技术资料",                    snippet=content[:100] + "...",                    content=content,                    source_quality=0.7 + 0.1 * overlap                ))                return results[:num_results]  class ProSearchSimulator:    """Pro Search 简化模拟器        模拟问题拆解、多轮搜索、信息提取、答案合成的核心流程    """        def __init__(self, search_engine: SimpleSearchEngine):        self.search_engine = search_engine        self.sub_questions: List[SubQuestion] = []        self.facts: List[InfoFact] = []        def decompose_question(self, user_question: str) -> List[SubQuestion]:        """问题拆解（简化版：基于关键词匹配）                实际中由LLM完成，这里用规则模拟        """        sub_questions = []                # 简化：如果是比较类问题，拆成三个子问题        if "对比" in user_question or "比较" in user_question or "差异" in user_question or "vs" in user_question.lower():            sub_questions.append(SubQuestion(                question="GPT-4的代码能力如何？",                query="GPT-4 code generation capability benchmark",                order=1            ))            sub_questions.append(SubQuestion(                question="Claude的代码能力如何？",                query="Claude code generation evaluation",                order=2            ))            sub_questions.append(SubQuestion(                question="GPT-4和Claude的综合对比？",                query="GPT-4 vs Claude benchmark comparison",                order=3            ))        else:            # 普通问题直接搜索            sub_questions.append(SubQuestion(                question=user_question,                query=user_question,                order=1            ))                self.sub_questions = sub_questions        return sub_questions        def execute_search_round(self, sub_q: SubQuestion) -> List[SearchResult]:        """执行一轮搜索"""        sub_q.status = "searching"        results = self.search_engine.search(sub_q.query)        sub_q.results = results        sub_q.status = "done"        return results        def extract_facts(self, results: List[SearchResult]) -> List[InfoFact]:        """从搜索结果中提取信息事实（简化版）"""        facts = []        for result in results:            if result.content:                # 简化：直接将内容作为事实提取                facts.append(InfoFact(                    fact=result.content,                    source_url=result.url,                    confidence=result.source_quality                ))                # 交叉验证（简化：内容相似即视为验证通过）        for i, fact1 in enumerate(facts):            for j, fact2 in enumerate(facts):                if i != j and fact1.fact[:50] == fact2.fact[:50]:                    fact1.verified = True                    fact2.verified = True                self.facts.extend(facts)        return facts        def synthesize_answer(self, user_question: str) -> Dict:        """合成答案（简化版）"""        # 按主题归类信息        verified_facts = [f for f in self.facts if f.verified]        unverified_facts = [f for f in self.facts if not f.verified]                # 生成直接回答        direct_answer = f"关于'{user_question}'，经过多轮搜索和信息整合，以下是分析结果："                # 生成详细分析        analysis_parts = []        for i, fact in enumerate(self.facts[:5]):            source_marker = f"[{i+1}]"            analysis_parts.append(f"{source_marker} {fact.fact}")                detailed_analysis = "\n".join(analysis_parts)                # 生成来源列表        sources = list(set(f.source_url for f in self.facts))                return {            "direct_answer": direct_answer,            "detailed_analysis": detailed_analysis,            "sources": sources,            "verified_count": len(verified_facts),            "total_facts": len(self.facts),            "search_rounds": len(self.sub_questions)        }        def run(self, user_question: str) -> Dict:        """运行完整的Pro Search流程"""        print(f"用户问题: {user_question}")        print("-" * 50)                # 1. 问题拆解        print("[步骤1] 问题拆解...")        sub_questions = self.decompose_question(user_question)        print(f"  拆解为 {len(sub_questions)} 个子问题:")        for sq in sub_questions:            print(f"    {sq.order}. {sq.question}")                # 2. 多轮搜索 + 信息提取        print("\n[步骤2] 多轮搜索与信息提取...")        all_facts = []        for sq in sub_questions:            print(f"  搜索子问题 {sq.order}: {sq.query}")            results = self.execute_search_round(sq)            print(f"    找到 {len(results)} 条结果")            facts = self.extract_facts(results)            all_facts.extend(facts)            print(f"    提取 {len(facts)} 条信息事实")                # 3. 答案合成        print("\n[步骤3] 答案合成...")        answer = self.synthesize_answer(user_question)        print(f"  已合成答案，包含 {answer['total_facts']} 条事实，{answer['verified_count']} 条已验证")                return answer  if __name__ == "__main__":    print("=" * 60)    print("Pro Search 简化模拟器演示")    print("=" * 60)        search_engine = SimpleSearchEngine()    pro_search = ProSearchSimulator(search_engine)        # 测试比较类问题    test_question = "GPT-4和Claude在代码生成上的差异对比"    print(f"\n测试问题: {test_question}")    print("=" * 60)        result = pro_search.run(test_question)        print(f"\n{'=' * 60}")    print("最终答案:")    print(f"{'=' * 60}")    print(f"\n【直接回答】")    print(result['direct_answer'])    print(f"\n【详细分析】")    print(result['detailed_analysis'])    print(f"\n【来源列表】")    for i, src in enumerate(result['sources']):        print(f"  [{i+1}] {src}")    print(f"\n【统计】")    print(f"  搜索轮数: {result['search_rounds']}")    print(f"  信息事实数: {result['total_facts']}")    print(f"  已验证事实: {result['verified_count']}")    print(f"{'=' * 60}")