Preface

In the previous series of articles, we mentioned some limitations of AI big models, including insufficient data timeliness and reliance on the timeliness of model training. There are several solutions to this problem. The previous article "Let the model understand your business better! RAG Principles and Solutions Analysis" mentioned the RAG plug-in knowledge base method. Another method is to use online search.

The popularity of DeepSeek some time ago officially brought the ability of large-model online search into the public eye, and it became the talk of the town.

Its emergence also makes up for the problem of timeliness of AI large model data, as well as the defect that traditional search engines mainly rely on keyword inverted index matching and cannot understand semantics and context.

Have you ever wondered what big model online search is all about? Can it completely replace search engines?

This article is going to explore the implementation principle of large model network search based on the previous series of articles. If you need background knowledge, you can review "AI Agent: The "Last Mile" of Large Model Land Application" and "Practical Tips! How to Use Java to Build an Enterprise-level AI Agent Service Framework?".

The basic principles and processes of large model network search

Let me give you the conclusion first: AI online search is an application of agent capabilities, and it is not something that comes with the model itself.

A simple diagram of the AI ​​network search process is as follows:

When users use AI online search to ask questions, the big model first semantically parses the user's question and extracts the keywords that need to be retrieved through the Internet. For example, when a user asks "Today in History", the big model has no timeliness information, it needs to retrieve relevant information through the Internet. Based on the user's question, the search keyword it may extract is, "Today in History March 15th".

Then the big model uses the Agent capability to call the browser to retrieve relevant keywords. The Agent capability here is usually a search engine API. After the search engine returns the search web page results, the big model continues to call the Agent capability to parse the returned web page and extract the text content from it.

The extracted search result content is then used as the context of the question and answer, combined with the prompt to interact with the big model again. The big model generates and aggregates content and returns the final result.

Does this look familiar? It seems to be quite similar to the process described in "Let the model understand your business better! Analysis of RAG principles and solutions". Indeed! The following section will compare the similarities and differences between RAG and AI network search.

Analysis of core technical points

1. Query Generation

   At this stage, it may be necessary to extract effective search query keywords from the natural language questions entered by the user, and of course it may also be necessary to determine whether a search is necessary. If the questions entered by the user are not related to timeliness, there may be no need to involve online search at all.

   This is naturally the strength of the big model. It can correctly understand the user's vague and ambiguous content through question rewriting, intent recognition, multi-round dialogue, etc., and extract appropriate keywords for subsequent retrieval.

2. Content Extraction

   How to filter out the web page results that are most relevant to the question from the results returned by the search engine? This may be a difficult question, involving two aspects:

   One is how to choose which web pages to use from the thousands of results returned by the search engine;

   The second is how to obtain content from the returned web pages and obtain the web page fragments that are most relevant to the question;

   Here I can only talk about my own speculations, and I may conduct some experiments to verify the ideas later.

   First, the search engine's relevance score may be used, and the results with high scores are returned first;

   However, we all know that search engines do not always return the most relevant results first. The bidding ranking mechanism has a great influence on the results, so it is possible to further judge based on the summary combined with semantic similarity;

   There may be some other mechanisms, such as website weight, etc. The large model gives priority to obtaining content from some professional and authoritative websites;

   After obtaining the web link returned by the search engine, AI online search also needs to access the web page, parse the text content, and find the fragment most relevant to the original question. This involves web crawling and parsing tools, which will not be elaborated here.

3. Information Integration

   At this stage, the big model may encounter information from different sources and with inconsistent data accuracy. Outdated, redundant, and conflicting information will bring challenges to the processing of online searches. Similar to the illusion of the big model, this will also cause AI online searches to give incorrect information.

   When using reasoning models such as DeepSeek R1, we sometimes see contradictory results output during the model's reasoning process. The model will use fact-checking and its own pre-trained knowledge to fuse opinions, but there is no guarantee that it can always distinguish the true from the false.

4. Answer Generation

   At this stage, the big model has obtained enough contextual content. Generating content based on user questions is the strength of the big model. As long as the data is correct, the big model can basically generate fluent, accurate and natural answers.

   The main problem at this stage is the hallucination problem based on the big model itself, which is also the most troublesome problem. It is precisely because the content it generates is fluent and natural enough, the expression is confident, and even logically clear and seems to be well-founded that it is difficult for people to quickly identify the errors in the AI-generated results without sufficient background knowledge.

Comparison of RAG and AI network search

Through the above principle analysis, we can see that the process of AI network search is very similar to that of RAG:

RAG is AI+knowledge base, while online search is AI+Web. RAG relies on vector retrieval, while online retrieval relies on search engines.

Below is a detailed comparison of the similarities and differences between the two:

Similarities:

Differences:

Summarize

Therefore, big model online search is not a simple big model capability. It is essentially an AI Agent application: it combines external tools such as search engines and web content parsing to empower search engines with the intelligence of big models and improve the content retrieval experience.

Back to the question at the beginning, will AI online search completely replace search engines?

From this perspective, I'm afraid not. Search engines are an agent capability of AI network search products. If there is no search engine, there will naturally be no AI network search. However, search engines may be upgraded to AI network search, combining AI to improve the intelligence of search scenarios, avoiding becoming an API form, and the entrance is replaced by AI network search products.

Finally, we all know that large models are subject to contextual limitations. If AI online retrieval obtains results by calling search engines, how does it break through the contextual limitations of the large model itself?