Application of large model network search in to B intelligent agents

In my previous article "Understanding AI Agents in One Article", I shared with you the components of an agent, which consists of four parts. Among them, the large model (LLM) mainly plays the key role of the "brain". To build a complete agent, the large model needs to select the appropriate tools according to the specific situation, so as to achieve the established goals and tasks more efficiently.

This article will share with you how to integrate the capability of “ online search ” into the exploration and practice of to B intelligent agents.

2. What is the online search capability of large models?

First of all, the big model itself does not have the attribute of online search. Its online search capability refers to obtaining the latest information through real-time Internet search, thereby generating more accurate and reasonable answers .

This capability mainly solves the problem of real-time knowledge lag caused by the timeliness of the pre-trained knowledge base of the large model. Online search is essentially an auxiliary tool that helps the large model obtain real-time information and use it as input with the question, and then combines this information to generate more accurate responses.

The following is the business process of large model network search:

• User questions : Users ask questions to the big model, and the model first determines whether the question requires an online search to obtain the latest information.

• Calling the search API : If an online search is required, the model will call the search engine's API and send the keywords in the question as search parameters to the search engine.

• Parsing search results : After the search engine returns the search results, the model will parse these results, extract the text content most relevant to the question, and sort them by similarity.

• Generate the final answer: Use the rearranged result content as the context of the question and answer, and interact with the large model again in combination with the question. The model generates and aggregates the content, and finally generates and returns an accurate and natural answer.

In daily life, scenarios such as weather inquiries, news information inquiries, and commodity price inquiries are greatly affected by real-time factors, and it is difficult to accurately answer them with a large model knowledge base alone. At this time, it is necessary to use external tools such as online search.

3. Application scenarios of online search capabilities in toB software

In the daily customer management and follow-up scenarios of salesmen, it is crucial to understand the latest dynamics of corporate customers. These dynamics include changes in industrial and commercial information, the direction of news and public opinion, etc., which directly affect the adjustment of sales strategies and the formulation of follow-up plans.

Last year, my team launched an intelligent agent called Business Opportunity Assistant , which mainly helps users obtain business information of corporate customers more efficiently.

Users can ask questions to the Business Opportunity Assistant through chatting. The intelligent body will use the RAG technology of the local knowledge base to recall business information with high similarity to the questions, and pass these fragments into the big model for integration and generation, ultimately providing users with accurate responses .

During the implementation process, we discovered a problem: since the local knowledge base usually stores data in a structured form , the content recalled through RAG also exhibits structured characteristics . This makes the agent's answers appear very stiff, and the overall performance is more like a simple knowledge retrieval tool rather than a real intelligent assistant. For example:

When a user queries " what is a certain company's product ", the content of RAG recall may only be the specific name of the product. However, users need more than just the product name, but also expect a brief introduction and the latest updates related to the product. In this case, the intelligent agent's answer cannot meet the user's needs for information richness and depth, and the user experience is greatly reduced.

Secondly, in news and public opinion query scenarios that require higher knowledge timeliness, large models cannot provide real-time answers. For example:

When asking the big model " Which is the latest model of Xiaomi car? ", due to the timeliness lag of the pre-trained knowledge base of the big model, it cannot provide the latest and accurate information, and thus cannot meet the user's demand for real-time knowledge query.

For the above two scenarios, in order to provide users with a better product experience, it is necessary to use the ability of online search.

4. Implementation ideas

You may wonder, users can use a general large model with online search capabilities to query corporate business information and news and public opinion, so what is the value of giving the [Business Opportunity Assistant] intelligent agent online search?

This is because enterprise information query is not the ultimate demand of users in this scenario. The essence of users querying enterprise information is to dynamically formulate targeted follow-up strategies based on the latest information . Therefore, after the intelligent agent replies to the content, it can automatically associate the specific customer information card in the CRM system, and support functions such as connection (making calls, sending text messages), follow-up (setting to-do tasks, labeling), etc., which is more conducive to ensuring the integrity and coherence of the entire enterprise information query business.

So, is the ability to search on the Internet self-developed?

The answer is no. Currently, many model cloud vendors have opened up a series of rich component capabilities, such as AI real-time search, image content understanding, and text-based graphs. As an AI business application party, we can make the intelligent body more powerful by integrating these tools.

The final optimization plan for [Business Opportunity Assistant] is as follows:

• Company name extraction: When a user asks a question, the company name is first extracted from the question using a large model.

• Company name matching : Since the company names mentioned by users in their questions are usually abbreviations, in order to avoid query errors caused by incomplete or inaccurate names, business search (keyword + similarity search) is required to match the extracted abbreviations with the company information stored in the system to accurately locate the specific company referred to in the user's question.

• Online search capability judgment: The big model determines whether the online search capability needs to be called based on the content of the user's question. If necessary, the latest enterprise-related information (such as changes in business information, news and public opinion, etc.) is obtained through online search based on the user's question to ensure the timeliness and accuracy of the answer.

• Information association and business operations: Combine the content returned by the big model and automatically associate it with the customer data card in the CRM. On this basis, a series of targeted business operations are provided to users, such as:

• When the user is not in the database, operations such as "add customer to the database" and "contact customer" are provided .
• When the user follows up for himself, operations such as "view customer details", "set to-do tasks", and "tag" are provided .
• When the user is not yourself, a "Request to Share" action is provided.

Through the above solution, users can not only quickly obtain the required information when using the [Business Opportunity Assistant] intelligent body, but also follow up with customers directly in the business process, achieving seamless connection between enterprise information query and sales follow-up, and improving work efficiency.

In the above scenarios, the method of actively querying the intelligent agent still has certain limitations. After all, users usually cannot know in real time when corporate information will change.

In order to better meet user needs, we can introduce a "push" approach . For example, the system proactively monitors changes in key corporate information (such as changes in legal persons, adjustments to business scope, or major news reports, negative public opinion, etc.) and pushes them to users, allowing users to grasp the latest developments of corporate customers at the first time, thereby better formulating targeted follow-up strategies.

VI. Conclusion

"Online search" is just one of the many tools of the big model. In the process of building a to B intelligent body, there are still many practical tools and scenarios worth exploring. In subsequent articles, the author will continue to share some interesting ideas and practices.

The above is the author’s recent personal learning and thinking, I hope it can help you