📄 中文摘要
在实际客户支持聊天机器人中，RAG 方法在回答“专业计划的价格是多少？”时仅需 800 毫秒，而函数调用则需要 3.1 秒。尽管函数调用在理论上应该更快，因为它是确定性的，涉及解析架构、调用端点并返回结构化数据，但实际测试显示 RAG 的性能更优。函数调用需要进行多次 API 调用，包括生成函数调用和执行响应的往返，而 RAG 则能一次性完成所有操作。这一发现揭示了在特定场景下，RAG 方法的效率优势。
函数调用与 RAG：生产中的 2.3 秒延迟差距
出处: Function Calling vs RAG: 2.3s Latency Gap in Production
发布: 2026年3月9日
📄 English Summary
Function Calling vs RAG: 2.3s Latency Gap in Production
In a real customer support chatbot scenario, the RAG method answered the question 'What's the price of the Pro plan?' in just 800 milliseconds, while function calling took 3.1 seconds. Although function calling is theoretically faster due to its deterministic nature, which involves parsing schemas, calling endpoints, and returning structured data, the actual tests showed that RAG consistently outperformed function calling. The reason lies in the sequential API calls required by function calling, which involves multiple round trips to generate the function call and execute the response, whereas RAG accomplishes everything in one go. This finding highlights the efficiency advantage of the RAG approach in specific contexts.