Group intelligent question and answer method based on user behavior recognition

Abstract

The invention discloses a group intelligent question-answering method based on user behavior recognition, which is characterized by comprising the following steps: adding an intelligent question-answering assistant in a group of instant messaging software, monitoring a group message in real time by the intelligent question-answering assistant and forwarding the group message to a service application so as to recognize and answer questions of group members; messages are classified into instruction messages or session messages, the instruction messages trigger built-in actions, the session messages are combined with historical messages of senders, whether a question intention exists or not is judged through a large language model, the messages with the question intention are rewritten into questions through the large language model, and reply content is generated by means of a knowledge vector library. And the effects of monitoring the messages with the questioning intention in the group in real time and automatically replying, ignoring the messages without the questioning intention in the group, merging the question fragments dispersed in the plurality of messages into the complete question and replying are achieved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent question answering of computer information processing technology, and in particular to a method for group intelligent question answering based on user behavior recognition. Background Art

[0002] With the rapid development of artificial intelligence technology, artificial intelligence has made great technological breakthroughs in large language models, which has affected people's lives in all aspects. Especially in the field of interactive question and answer, many technology companies have launched intelligent question and answer assistants. For example, OpenAI's ChatGPT, Baidu's Wenxin Yiyan, Alibaba's Tongyi Qianwen, and Kimi from the Dark Side of the Moon. These intelligent question and answer assistants process user messages through large language models and provide users with intelligent question and answer functions. However, the current interactive form of intelligent question and answer assistants is basically a one-question-one-answer form, which cannot meet the needs of intelligent question and answer assistants in group scenarios to automatically identify the questions of group members and actively answer them.

[0003] Therefore, a group intelligent question-answering method is needed, which can automatically identify messages of group members in the group, analyze potential problems and generate reply messages, and does not generate messages when group members are chatting. Summary of the invention

[0004] The technical problem to be solved by the present invention is to overcome the deficiencies of the prior art and provide a method for group intelligent question answering based on user behavior recognition.

[0005] In order to solve the above technical problems, the present invention provides a method for group intelligent question answering based on user behavior recognition, which is characterized by comprising the following steps:

[0006] Step 1: Add an intelligent question-and-answer assistant to the group of the instant messaging software; the intelligent question-and-answer assistant is used to call the object of the intelligent question-and-answer service;

[0007] Step 2: The intelligent question-and-answer assistant monitors the current message in the group in real time, and forwards the monitored current message to the intelligent question-and-answer service; the intelligent question-and-answer service is a service application that provides intelligent question-and-answer capabilities for the intelligent question-and-answer assistant; the forwarded current message at least includes the message text, the message sender, and the message recipient; the message recipient is designated as the group where the current message is located in the group conversation scenario; the purpose of real-time monitoring of messages in the group is to promptly identify potential problems in group member messages and actively answer them, thereby reducing the reply workload of the group administrator;

[0008] Step 3: Determine the type of the current message; if the current message contains a preset specific instruction, determine that the type of the current message is an instruction message; otherwise, determine that the type of the current message is a conversation message; the instruction is an action built into the intelligent question-and-answer service, including but not limited to group management, conversation management, and knowledge vector library management; the knowledge vector library is used to provide knowledge vectors for the intelligent question-and-answer service to generate reply content; the knowledge vector includes word elements that can be recognized by the AI ​​algorithm converted from natural language, and the word element is the smallest information unit that can be recognized by the AI ​​algorithm;

[0009] Step 4: If the type of the current message is a command message, a corresponding command function is matched according to the command name in the current message to process the command;

[0010] Step 5: If the type of the current message is a conversation message, then first obtain the historical messages of the sender of the current message; the historical message range does not set a fixed value, and the historical message range is adjusted according to the speech interval of the group members. The historical message range includes but is not limited to: the last ten minutes, the last minute, and the last ten historical messages;

[0011] Step 6: Pass the current message and historical messages to the large language model service to determine whether the current message has the intention to ask a question; if the large language model service determines that there is no intention to ask a question, the current message is marked as processed and the processing flow is forced to end; if the large language model service determines that there is an intention to ask a question, the large language model service rewrites the complete question based on the current message and historical messages; if the large language model service fails to rewrite the complete question, the current message is marked as processed and the processing flow is forced to end; if the large language model service successfully extracts the complete question, the complete question is vectorized to obtain a question vector;

[0012] Step 7: pass the question vector to the knowledge vector library and recall the knowledge vectors with the highest similarity ranking; pass the current message, historical messages and recalled knowledge vectors to the large language model service to generate reply content; if the large language model service fails to successfully generate a reply, mark the current message as processed and force the processing flow to end; if the large language model service successfully generates a reply, push the generated reply to the group where the current message is located through the push channel provided by the instant messaging software.

[0013] In step 1, the instant messaging software includes but is not limited to WeChat, QQ, WeChat for Enterprise, DingTalk, and Feishu; in different instant messaging software, the intelligent question-and-answer assistant has different forms of expression, for example, it is an ordinary WeChat account in WeChat, and it is a custom application in WeChat for Enterprise.

[0014] In step 3, the knowledge vector library includes but is not limited to Milvus and Pgvector.

[0015] In step 4, the instruction function is a method for processing user instructions, including but not limited to operating a database, calling a file system, and calling internal and external interfaces.

[0016] In step 5, the method for adjusting the scope of historical messages is specifically: if the sending time of the messages that constitute a complete question in the historical messages differs by no more than one minute, the historical message scope is set to the most recent minute; similarly, if there is a sending time difference of more than one minute but less than ten minutes, the historical message scope is set to the most recent ten minutes; if there is a sending time difference of more than ten minutes, it is set to the most recent ten historical messages.

[0017] In step 6, the method of judging whether the current message has the intention to ask a question is to provide the large language model with prompt words for judging whether the intention to ask a question is based on the current message and historical messages, and the large language model generates a judgment result according to the instructions of the prompt words; the method of rewriting the complete question is to provide the large model with prompt words for rewriting the question according to the current message and historical messages, and the large language model generates the complete question according to the instructions of the prompt words.

[0018] In step 7, the top-ranked knowledge vectors do not have a fixed value in the ranking range, and need to be dynamically adjusted according to the richness and quality of the knowledge, including but not limited to the top 3 knowledge vectors, the top 5 knowledge vectors, the top 10 knowledge vectors, etc.

[0019] The beneficial effects achieved by the present invention are:

[0020] (1) The present invention implements a group intelligent question-and-answer assistant that actively identifies and answers questions of group members in group conversation scenarios, thereby preventing group members from actively asking questions to the intelligent question-and-answer assistant, improving the utilization rate of the intelligent question-and-answer assistant, and saving the cost of answering questions in the group;

[0021] (2) The present invention realizes that in a group conversation scenario, a group member may split a question into multiple messages, and the intelligent question-answering assistant can reply in a unified manner after all the messages have been sent, thereby preventing the intelligent question-answering assistant from disturbing group members by replying to a question multiple times, and preventing the intelligent question-answering assistant from replying inaccurately before the question is fully asked.

[0022] (3) The present invention realizes the ability of the intelligent question-and-answer assistant in the group conversation scenario to actively distinguish whether group members are chatting or asking questions, so as to achieve the effect of not replying during chatting and actively replying during asking questions. It integrates the automatic reply capability into the group in an imperceptible manner without interfering with the daily communication in the group. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 The method flow diagram of the exemplary embodiment of the present invention is shown in FIG.

[0024] Figure 2 A simplified flow chart of actively identifying group questions and responding to them according to an exemplary embodiment of the present invention;

[0025] Figure 3 A simplified diagram of a process flow in which a problem is broken down into multiple messages for an exemplary embodiment of the present invention;

[0026] Figure 4 A simplified flow chart of proactively identifying group questions but failing to respond according to an exemplary embodiment of the present invention;

[0027] Figure 5 A simplified flowchart of a process in which a group member does not reply during small talk according to an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0028] The present invention will be further described below in conjunction with the accompanying drawings and exemplary embodiments:

[0029] like Figure 1 As shown in the method flow chart of an exemplary embodiment of the present invention, there is a method for group intelligent question answering based on user behavior recognition, comprising the following steps:

[0030] Step 101: Add a smart question-and-answer assistant to a group of an instant messaging software; the smart question-and-answer assistant is used to call an object of a smart question-and-answer service; the instant messaging software includes but is not limited to WeChat, QQ, WeChat for Enterprise, DingTalk, and Feishu; in different instant messaging software, the smart question-and-answer assistant has different forms of expression, for example, in WeChat it is an ordinary WeChat account, in WeChat for Enterprise it is a custom application, etc.

[0031] Step 102: The intelligent question-and-answer assistant monitors the current message in the group in real time, and forwards the monitored current message to the intelligent question-and-answer service; the intelligent question-and-answer service is a service application that provides intelligent question-and-answer capabilities for the intelligent question-and-answer assistant; the forwarded current message at least includes the message text, the message sender, and the message recipient; the message recipient is designated as the group where the current message is located in the group conversation scenario; the purpose of real-time monitoring of messages in the group is to promptly identify potential problems in group member messages and actively answer them, thereby reducing the reply workload of the group administrator;

[0032] Step 103: Determine the type of the current message; if the current message contains a preset specific instruction, determine that the type of the current message is an instruction message; otherwise, determine that the type of the current message is a conversation message; the instruction is a built-in action of the intelligent question and answer service, including but not limited to group management, conversation management, and knowledge vector library management; the knowledge vector library is used to provide knowledge vectors for the intelligent question and answer service to generate reply content, and the knowledge vector includes word elements that can be recognized by the AI ​​algorithm converted from natural language, and the word element is the smallest information unit that can be recognized by the AI ​​algorithm; the knowledge vector library includes but is not limited to Milvus and Pgvector.

[0033] Step 104: If the type of the current message is an instruction message, a corresponding instruction function is matched according to the instruction name in the current message to process the instruction; the instruction function is a method for processing user instructions, including but not limited to operating a database, calling a file system, and calling internal and external interfaces.

[0034] Step 105: If the type of the current message is a conversation message, then the historical messages of the sender of the current message are first obtained; the historical message range does not set a fixed value, and the historical message range is adjusted according to the speaking interval of the group members. The historical message range includes but is not limited to: the last ten minutes, the last minute and the last ten historical messages; the method for adjusting the historical message range is specifically: if the sending time of the messages that constitute a complete question in the historical messages does not differ by more than one minute, then the historical message range is set to the last minute. Similarly, if there is a sending time difference of more than one minute but less than ten minutes, then the historical message range is set to the last ten minutes. If there is a sending time difference of more than ten minutes, then it is set to the last ten historical messages.

[0035] Step 106: pass the current message and the historical message to the large language model service to determine whether the current message has the intention to ask a question; if the judgment result of the large language model service is that there is no intention to ask a question, mark the current message as processed and force the processing flow to end; if the judgment result of the large language model service is that there is an intention to ask a question, the large language model service rewrites the complete question according to the current message and the historical message; if the large language model service fails to rewrite the complete question, mark the current message as processed and force the processing flow to end; if the large language model service successfully extracts the complete question, vectorize the complete question to obtain a question vector; the method of determining whether the current message has the intention to ask a question is to provide the large language model with a prompt word for determining whether there is the intention to ask a question according to the current message and the historical message, and the large language model generates a determination result according to the indication of the prompt word; the method of rewriting the complete question is to provide the large model with a prompt word for rewriting the question according to the current message and the historical message, and the large language model generates a complete question according to the indication of the prompt word; the large language model service includes but is not limited to the online large language model service provided by OpenAI, Cluade, Kimi, and Tongyi Qianwen, and the model service built by itself based on the pre-training scheme such as LLaMA.

[0036] Step 107: pass the question vector to the knowledge vector library, and recall the knowledge vectors with the highest similarity ranking; pass the current message, historical messages and recalled knowledge vectors to the large language model service to generate reply content; if the large language model service fails to successfully generate a reply, mark the current message as processed and force the processing flow to end; if the large language model service successfully generates a reply, push the generated reply to the group where the current message is located through the push channel provided by the instant messaging software. The top-ranked knowledge vectors do not have a fixed value in the ranking range, and need to be dynamically adjusted according to the richness and quality of the knowledge, including but not limited to the top 3 knowledge vectors, the top 5 knowledge vectors, the top 10 knowledge vectors, etc.

[0037] like Figure 2 The flowchart of the exemplary embodiment of the present invention for actively identifying group questions and responding to them includes the following steps:

[0038] Step 201: Create an intelligent question-and-answer assistant application in WeChat for Enterprise, and add the intelligent question-and-answer assistant to WeChat for Enterprise Group A.

[0039] Step 202: The intelligent question-and-answer assistant monitors the messages of the enterprise WeChat group A in real time and forwards them to the intelligent question-and-answer service. The current message content is "Who is on duty for orders today?"

[0040] Step 203: The intelligent question-answering service determines the type of the current message. The current message does not contain a specific instruction and is determined to be a conversation message.

[0041] Step 204: Get the historical messages of the current message sender in the last 10 minutes. The current message sender has only sent one message in the last 10 minutes, and the message content is "Hello".

[0042] Step 205: pass the current message "Who is on duty for today's order?" and the message context "Hello" to the large language model service.

[0043] Step 206: The large language model service determines that the user sending the current message in the group is asking a question in the group.

[0044] Step 207: The large language model service re-refines the complete question based on the current message "Who is on duty for orders today?" and the message context "Hello." After re-refining, the generated complete question is "Hello, who is on duty for orders today?".

[0045] Step 208: Vectorize the complete question “Hello, who is on duty for orders today?” to generate a question vector.

[0046] Step 209: The question vector is passed to the vector database, and the top five pieces of knowledge with the highest similarity are recalled. The top five pieces of knowledge with the highest similarity are "order duty personnel list", "order duty arrangement", "order duty specification", "order duty record" and "order duty situation".

[0047] Step 210: The current message "Who is on duty for orders today?", the message context "Hello" and the related knowledge "List of order duty personnel", "Order duty schedule", "Order duty specifications", "Order duty records" and "Order duty status" are passed to the large language model to generate a reply.

[0048] Step 211: The large language model refers to the content in the knowledge "List of Order Duty Personnel" and generates a reply content of "Hello, Zhang San is on duty for the order today.".

[0049] Step 212: Push the generated reply "Hello, Zhang San is on duty for the order today." to the enterprise WeChat group A through the push interface of the enterprise WeChat application.

[0050] like Figure 3 The illustrated embodiment of the present invention is a simplified flowchart of a problem being split into multiple messages, and the specific steps are:

[0051] Step 301: Create an intelligent question-and-answer assistant application in WeChat for Enterprise, and add the intelligent question-and-answer assistant to WeChat for Enterprise Group A.

[0052] Step 302: The intelligent question-and-answer assistant monitors the messages of the enterprise WeChat group A in real time and forwards them to the intelligent question-and-answer service. The current message content is "Please ask today".

[0053] Step 303: The intelligent question-answering service determines the type of the current message. The current message does not contain a specific instruction and is determined to be a conversation message.

[0054] Step 304: Get the historical messages of the current message sender in the last 10 minutes. The current message sender has only sent one message in the last 10 minutes, and the message content is "Hello".

[0055] Step 305: pass the current message "Excuse me today" and the message context "Hello" to the large language model service.

[0056] Step 306: The large language model service determines that the user sending the current message in the group is asking a question in the group.

[0057] Step 307: The large language model service re-refines the complete question based on the current message "May I ask today" and the message context "Hello". After re-refining, the large model fails to refine the complete question. The current message is set as processed and the subsequent processing flow is terminated.

[0058] Step 308: The intelligent question-and-answer assistant monitors the messages of the enterprise WeChat group A in real time and forwards them to the intelligent question-and-answer service. The current message content is "Who is on duty for the order?"

[0059] Step 309: The intelligent question-answering service determines the type of the current message. The current message does not contain a specific instruction and is determined to be a conversation message.

[0060] Step 310: Get the historical messages of the current message sender in the last 10 minutes. The current message sender has sent two messages in the last 10 minutes, and the message contents are "Hello" and "May I ask today".

[0061] Step 311: pass the current message "Who is on duty for the order" and the message context "Hello" and "May I ask today" to the large language model service.

[0062] Step 312: The large language model service determines that the user sending the current message in the group is asking a question in the group.

[0063] Step 313: The large language model service re-refines the complete question based on the current message "Who is on duty for the order?" and the message context "Hello," "May I ask who is on duty today?" After re-refining, the generated complete question is "Hello, who is on duty for the order today?".

[0064] Step 314: vectorize the complete question “Hello, who is on duty for orders today?” to generate a question vector.

[0065] Step 315: The question vector is passed to the vector database, and the top five pieces of knowledge with the highest similarity are recalled. The top five pieces of knowledge with the highest similarity are "order duty personnel list", "order duty arrangement", "order duty specification", "order duty record" and "order duty situation".

[0066] Step 316: The current message "Who is on duty for orders today?", the message context "Hello" and the related knowledge "List of order duty personnel", "Order duty schedule", "Order duty specifications", "Order duty records" and "Order duty status" are passed to the large language model to generate a reply.

[0067] Step 317: The large language model refers to the content in the knowledge "List of Order Duty Personnel" and generates a reply content of "Hello, Zhang San is on duty for the order today.".

[0068] Step 318: Push the generated reply "Hello, Zhang San is on duty for the order today." to the enterprise WeChat group A through the push interface of the enterprise WeChat application.

[0069] like Figure 4 The flowchart of actively identifying group questions but failing to respond in the embodiment of the present invention shown in the figure has the following specific steps:

[0070] Step 401: Create an intelligent question and answer assistant application in WeChat for Enterprise, and add the intelligent question and answer assistant to WeChat for Enterprise Group A.

[0071] Step 402: The intelligent question-and-answer assistant monitors the messages of the enterprise WeChat group A in real time and forwards them to the intelligent question-and-answer service. The current message content is "Who is on duty for orders today?"

[0072] Step 403: The intelligent question-answering service determines the type of the current message. The current message does not contain a specific instruction and is determined to be a conversation message.

[0073] Step 404: Get the historical messages of the current message sender in the last 10 minutes. The current message sender has only sent one message in the last 10 minutes, and the message content is "Hello".

[0074] Step 405: pass the current message "Who is on duty for today's order?" and the message context "Hello" to the large language model service.

[0075] Step 406: The large language model service determines that the user sending the current message in the group is asking a question in the group.

[0076] Step 407: The large language model service re-refines the complete question based on the current message "Who is on duty for orders today?" and the message context "Hello." After re-refining, the generated complete question is "Hello, who is on duty for orders today?".

[0077] Step 408: vectorize the complete question “Hello, who is on duty for orders today?” to generate a question vector.

[0078] Step 409: The question vector is passed to the vector database, and the top five pieces of knowledge in terms of similarity are recalled. The top five pieces of knowledge in terms of similarity are respectively "What to do if an order cannot be submitted", "Duty adjustment and holiday arrangement", "Order duty specification", "Order duty record" and "Order duty status".

[0079] Step 410: The current message "Who is on duty for orders today?", the message context "Hello" and related knowledge "What should I do if the order cannot be submitted?", "Duty adjustment schedule", "Order duty specifications", "Order duty records" and "Order duty status" are passed to the large language model to generate a reply.

[0080] Step 411: The large language model does not detect any information that clearly describes the order duty scheduling in the relevant knowledge, and fails to successfully generate a reply content.

[0081] Step 412: Set the current message as processed and terminate the subsequent processing flow.

[0082] like Figure 5 This is a simplified flow chart of a group member not replying during a chat in an embodiment of the present invention, and the specific steps are:

[0083] Step 501: Create an intelligent question and answer assistant application in WeChat for Enterprise, and add the intelligent question and answer assistant to WeChat for Enterprise Group A.

[0084] Step 502: The intelligent question-and-answer assistant monitors the messages of the enterprise WeChat group A in real time and forwards them to the intelligent question-and-answer service. The current message content is "I opened a big order today" and the sender of the message is Zhang San.

[0085] Step 503: The intelligent question-answering service determines the type of the current message. The current message does not contain a specific instruction and is determined to be a conversation message.

[0086] Step 504: Get Zhang San's historical messages in the last 10 minutes. Zhang San has only sent one message in the last 10 minutes, and the message content is "It's great".

[0087] Step 505: The current message "I opened a big order today" and the message context "It's great" are passed to the large language model service.

[0088] Step 506: The large language model service determines that the user is not asking a question in the group when sending the current message in the group.

[0089] Step 507: Set the current message as processed and terminate the subsequent processing flow.

[0090] Step 508: The intelligent question-and-answer assistant monitors the messages of the enterprise WeChat group A in real time and forwards them to the intelligent question-and-answer service. The current message content is "I'm so envious" and the sender of the message is Li Si.

[0091] Step 509: The intelligent question-answering service determines the type of the current message. The current message does not contain a specific instruction and is determined to be a conversation message.

[0092] Step 510: Obtain the historical messages of Li Si in the last 10 minutes. Li Si has not sent any message in the last 10 minutes.

[0093] Step 511: pass the current message "I'm so envious" to the large language model service.

[0094] Step 512: The large language model service determines that the user is not asking a question in the group when sending the current message in the group.

[0095] Step 513: Set the current message as processed and terminate the subsequent processing flow.

[0096] The present invention is mainly used to provide a method for group intelligent question answering based on user behavior recognition, and its beneficial effects are:

[0097] (1) The present invention implements a group intelligent question-and-answer assistant that actively identifies and answers questions of group members in group conversation scenarios, thereby preventing group members from actively asking questions to the intelligent question-and-answer assistant, improving the utilization rate of the intelligent question-and-answer assistant, and saving the cost of answering questions in the group;

[0098] (2) The present invention realizes that in a group conversation scenario, a group member may split a question into multiple messages, and the intelligent question-answering assistant can reply in a unified manner after all the messages have been sent, thereby preventing the intelligent question-answering assistant from disturbing group members by replying to a question multiple times, and preventing the intelligent question-answering assistant from replying inaccurately before the question is fully asked.

[0099] (3) The present invention realizes the ability of the intelligent question-and-answer assistant in the group conversation scenario to actively distinguish whether group members are chatting or asking questions, so as to achieve the effect of not replying during chatting and actively replying during asking questions. It integrates the automatic reply capability into the group in an imperceptible manner without interfering with the daily communication in the group.

[0100] The present invention may also have many other implementation modes. The above embodiments do not limit the present invention in any way. Without departing from the spirit and essence of the present invention, technicians familiar with the field may make various corresponding changes and modifications based on the present invention. All other improvements and applications made to the above embodiments in an equivalent transformation manner should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A method for group intelligent question answering based on user behavior recognition, characterized in that: The steps include: Step 1: Add an intelligent question-and-answer assistant to the group of the instant messaging software; the intelligent question-and-answer assistant is used to call the object of the intelligent question-and-answer service; Step 2: The intelligent question-and-answer assistant monitors the current message in the group in real time, and forwards the monitored current message to the intelligent question-and-answer service; the intelligent question-and-answer service is a service application that provides intelligent question-and-answer capabilities for the intelligent question-and-answer assistant; the forwarded current message at least includes the message text, the message sender, and the message recipient; the message recipient is specified as the group where the current message is located in the group conversation scenario; Step 3: Determine the type of the current message; if the current message contains a preset specific instruction, determine that the type of the current message is an instruction message; Otherwise, the type of the current message is determined to be a conversation message; the instruction is a built-in action of the intelligent question-and-answer service, including but not limited to group management, conversation management, and knowledge vector library management; the knowledge vector library is used to provide knowledge vectors for the intelligent question-and-answer service to generate reply content; the knowledge vector includes word elements that can be recognized by the AI ​​algorithm converted from natural language, and the word element is the smallest information unit that can be recognized by the AI ​​algorithm; Step 4: If the type of the current message is a command message, a corresponding command function is matched according to the command name in the current message to process the command; Step 5: If the type of the current message is a conversation message, then first obtain the historical messages of the sender of the current message; the historical message range does not set a fixed value, and the historical message range is adjusted according to the speech interval of the group members. The historical message range includes but is not limited to: the last ten minutes, the last minute, and the last ten historical messages; Step 6: Pass the current message and historical messages to the large language model service to determine whether the current message contains a question intention; If the large language model service determines that there is no question intention, the current message is marked as processed and the processing flow is forced to end; If the large language model service determines that there is an intention to ask a question, the large language model service rewrites the complete question based on the current message and historical messages; If the large language model service fails to rewrite the complete question, the current message is marked as processed and the processing flow is forced to end; If the large language model service successfully extracts the complete question, it vectorizes the complete question to obtain a question vector; Step 7: passing the question vector to the knowledge vector library, and recalling the knowledge vectors with the highest similarity ranking; passing the current message, historical messages, and the recalled knowledge vectors to the large language model service to generate the reply content; If the large language model service fails to successfully generate a reply, the current message is marked as processed and the processing flow is forced to end; If the large language model service successfully generates a reply, the generated reply is pushed to the group where the current message is located through the push channel provided by the instant messaging software.

2. The method for group intelligent question answering based on user behavior recognition according to claim 1, characterized in that: In step 1, the instant messaging software includes but is not limited to WeChat, QQ, Enterprise WeChat, DingTalk, and Feishu.

3. The method for group intelligent question answering based on user behavior recognition according to claim 2, characterized in that: In step 3, the knowledge vector library includes but is not limited to Milvus and Pgvector.

4. The method for group intelligent question answering based on user behavior recognition according to claim 3, characterized in that: In step 4, the instruction function includes but is not limited to operating a database, calling a file system, and calling internal and external interfaces.

5. The method for group intelligent question answering based on user behavior recognition according to claim 4, characterized in that: In step 5, the method for adjusting the scope of historical messages is specifically: if the sending time of the messages that constitute a complete question in the historical messages differs by no more than one minute, the historical message scope is set to the most recent minute; similarly, if there is a sending time difference of more than one minute but less than ten minutes, the historical message scope is set to the most recent ten minutes; if there is a sending time difference of more than ten minutes, it is set to the most recent ten historical messages.

6. A method for group intelligent question answering based on user behavior recognition as claimed in claim 5, characterized in that: In step 6, the method of determining whether the current message has a question intention is to provide the large language model with a prompt word for determining whether the question intention exists based on the current message and historical messages, and the large language model generates a determination result according to the prompt word; The method for rewriting the complete question is to provide the large model with prompt words for rewriting the question according to the current message and the historical message, and the large language model generates the complete question according to the instructions of the prompt words.

7. The method for group intelligent question answering based on user behavior recognition according to claim 6, characterized in that: In step 7, the top-ranked knowledge vectors do not have a fixed value in the ranking range, and need to be dynamically adjusted according to the richness and quality of the knowledge, including but not limited to the top 3 knowledge vectors, the top 5 knowledge vectors, the top 10 knowledge vectors, etc.

Images