为什么你的 LLM 返回了 200 OK 但答案却是错的

import time
import uuid
import requests

class TraceNode:
    def __init__(self, name, parent_id=None):
        self.name = name
        self.node_id = str(uuid.uuid4())
        self.parent_id = parent_id
        self.start_time = None
        self.end_time = None
        self.inputs = None
        self.outputs = None
        self.metadata = {}

    def start(self, inputs):
        self.start_time = time.time()
        self.inputs = inputs

    def end(self, outputs, metadata=None):
        self.end_time = time.time()
        self.outputs = outputs
        if metadata:
            self.metadata.update(metadata)

    def to_dict(self):
        return {
            "node_id": self.node_id,
            "parent_id": self.parent_id,
            "name": self.name,
            "duration_ms": (self.end_time - self.start_time) * 1000 if self.end_time else 0,
            "inputs": self.inputs,
            "outputs": self.outputs,
            "metadata": self.metadata
        }

class TracedRAGPipeline:
    def __init__(self, api_key):
        self.api_key = api_key
        # n1n.ai 提供统一的 LLM API 接入端点
        self.api_url = "https://api.n1n.ai/v1/chat/completions"

    def retrieve_documents(self, query, trace_parent_id):
        node = TraceNode("Vector_Retrieval", parent_id=trace_parent_id)
        node.start({"query": query})
        
        # 模拟向量数据库检索延迟
        time.sleep(0.15)
        retrieved_docs = [
            {"id": 101, "content": "DeepSeek-V3 是一款先进的混合专家（MoE）语言模型。", "score": 0.89},
            {"id": 102, "content": "该模型拥有 6710 亿总参数，每个 Token 激活 370 亿参数。", "score": 0.82}
        ]
        
        node.end(outputs=retrieved_docs, metadata={"vector_db": "Qdrant", "top_k": 2})
        return retrieved_docs, node

    def call_llm(self, prompt, trace_parent_id):
        node = TraceNode("LLM_Generation", parent_id=trace_parent_id)
        node.start({"prompt": prompt})

        headers = {
            "Authorization": f"Bearer {self.api_key}",
            "Content-Type": "application/json"
        }
        
        # 通过 n1n.ai 统一 API 接口调用 DeepSeek-V3
        data = {
            "model": "deepseek-v3",
            "messages": [
                {"role": "system", "content": "你是一个严谨的技术助手。"},
                {"role": "user", "content": prompt}
            ],
            "temperature": 0.1
        }

        response = requests.post(self.api_url, headers=headers, json=data)
        response_json = response.json()
        
        generation = response_json["choices"][0]["message"]["content"]
        usage = response_json.get("usage", {})
        
        node.end(outputs=generation, metadata={
            "model": "deepseek-v3",
            "tokens_used": usage.get("total_tokens", 0),
            "prompt_tokens": usage.get("prompt_tokens", 0),
            "completion_tokens": usage.get("completion_tokens", 0)
        })
        return generation, node

    def execute(self, query):
        pipeline_trace = []
        root_id = str(uuid.uuid4())
        
        # 步骤 1：文档检索
        docs, retrieval_node = self.retrieve_documents(query, root_id)
        pipeline_trace.append(retrieval_node.to_dict())
        
        # 步骤 2：构建带有上下文的 Prompt
        context_str = "\n".join([d["content"] for d in docs])
        formatted_prompt = f"上下文信息：\n{context_str}\n\n问题：{query}\n回答："
        
        # 步骤 3：调用 LLM 生成回答
        answer, llm_node = self.call_llm(formatted_prompt, root_id)
        pipeline_trace.append(llm_node.to_dict())
        
        return {
            "answer": answer,
            "trace": pipeline_trace
        }

# 运行示例
if __name__ == "__main__":
    # 请替换为你在 n1n.ai 申请的真实 API Key
    N1N_API_KEY = "your-n1n-api-key"
    pipeline = TracedRAGPipeline(api_key=N1N_API_KEY)
    
    result = pipeline.execute("DeepSeek-V3 每个 Token 激活多少参数？")
    print("回答:", result["answer"])
    
    # 打印链路追踪数据
    print("\n--- 执行链路追踪 ---")
    for span in result["trace"]:
        print(f"步骤: {span['name']} | 耗时: {span['duration_ms']:.2f}ms")
        print(f"元数据: {span['metadata']}\n")