On November 1, 2024, Beijing time, OpenAI ChatGPT officially launched its AI search feature.
Usually, we use search engines like Baidu and Google to find information, but obtaining useful answers this way can require a lot of effort: we often need to conduct multiple searches and sift through numerous links to find high-quality content sources, then read the content on the web pages to find and organize the information we expect.
AI search is designed to solve this problem, allowing us to obtain better answers more easily through dialogue: we can ask questions in a more natural, conversational manner, and AI search will provide high-quality answers based on the information it retrieves from the web, allowing for deeper follow-up questions to extract more valuable information.
AI search is an important supplement to the issues inherent in large language models. First, AI search compensates for the shortcomings of large language models in terms of data timeliness and accuracy; large models often have data lags and cannot cover the latest information. By utilizing real-time web searches, AI can obtain the latest content, ensuring the timeliness of answers. Second, AI models sometimes generate fabricated information (known as “hallucinations”), leading to inaccurate responses. By leveraging web searches, AI can incorporate authoritative and real information sources, significantly improving the reliability and credibility of answers and reducing misleading responses.
In fact, the working principle of AI search is not complicated and can be simply divided into three steps (of course, in real systems, this process is more complex, and the implementation details vary among companies).
-
According to the user’s question, it calls third-party search engines to find relevant web content. In this step, the user’s question may be rewritten or expanded to find higher-quality web information.
-
From the numerous web contents, it filters out the most relevant authoritative information related to the user’s question.
-
Based on the filtered web information and the user’s question, the large language model organizes the corresponding answer and returns it to the user.
AI search is suitable for scenarios where one wants to get answers directly and quickly, reducing the cost of searching, analyzing, and organizing information. Some typical examples include:
-
What are the best Sichuan restaurants within 1 kilometer of me?
-
What will the weather be like tomorrow?
-
What movies has Leonardo DiCaprio recently starred in?
-
I plan to visit Jiuzhaigou next month; can you recommend a detailed itinerary and guide?
Such scenarios cover over 90% of the questions we use Baidu for in our daily lives, so the emergence of AI search poses a significant challenge to search engines like Google and Baidu.
Before integrating search functionality, OpenAI ChatGPT had already attracted widespread attention for AI search, with one of the most notable being Perplexity.ai, which specializes in providing AI search services. In China, there are similar services like Metaso (metaso.cn), which focuses on this service model. Additionally, dialogue models like Kimi and Tongyi Qianwen have also integrated search functions, and many users have become reliant on AI search features.