An automatic complaint management system and method based on intelligent agent workflow

Abstract

The present invention relates to the field of intelligent management systems, and specifically to an automatic complaint management system and method based on an agent workflow, wherein the system includes five agents: complaint acceptance, complaint classification, complaint dispatch, complaint processing, and user feedback. The method includes a complaint acceptance agent preprocessing information, and a complaint classification agent performing hierarchical classification; a complaint dispatching agent classifies complaints into automatic processing procedures or offline work order procedures according to classification; a complaint processing agent retrieves processing solutions from a system knowledge base and responds; and a user feedback agent updates data on the system knowledge base, etc., according to user evaluations. The present invention realizes automated management by using an agent workflow, and combines sentiment word embedding technology to hierarchically process complaints. Problems that require escalation of complaints are handled manually in conjunction with a work order system; problems that do not require escalation of complaints are automatically processed by the agent workflow, thereby improving processing efficiency and user experience.

Description

Technical Field

[0001] The present invention relates to the field of intelligent management systems, and in particular to an automatic complaint management system and method based on intelligent agent workflow. Background Art

[0002] At present, common complaint management systems include customer service management systems and work order management systems.

[0003] The customer service management system can effectively respond to user questions through preset QA and manual customer service intervention. The original QA system mainly relied on pre-set rules and templates to answer common questions, lacking personalization and flexibility. With the development of technology, personalized recommendations and context awareness have been introduced into the QA system, enabling the system to provide more intelligent and personalized answers and suggestions based on the user's historical information and current context.

[0004] The work order management system is more rigorous and professional in form, and has better support capabilities for long-term businesses. Currently, artificial intelligence and machine learning technologies have been applied to the work order management system to achieve intelligent work order processing and services. The intelligent system can automatically identify and classify work orders, intelligently recommend solutions, and automatically respond to and handle common problems, thereby improving response speed and user satisfaction.

[0005] In the current existing technology, the work order management system and the customer service management system are independent of each other and do not integrate with each other, resulting in the following defects in actual applications:

[0006] (1) In terms of the system, the data on user complaints and work order processing cannot be interconnected. When a customer initiates a complaint, customer service staff often need to query the work order based on the complaint content, or manually create a work order based on customer requirements. The staff need to take into account both the work order system and the customer service system at the same time. The construction of related systems has instead brought more burdens to the staff.

[0007] (2) In terms of content, templated automatic replies often cannot be customized according to the individual needs of specific users. Each user's questions may be slightly different, but automatic replies cannot effectively provide completely personalized answers for each case. At the same time, automatic replies can usually only respond to pre-set standard questions or situations. For complex or non-standard questions, robots may not be able to provide effective solutions, resulting in a poor user experience.

[0008] (3) In terms of AI capability application, in the current complaint management system, AI capabilities are mostly used as auxiliary means in the system and cannot play a role in driving business flow, resulting in the inability to truly optimize efficiency. It is impossible to effectively utilize historical records and interface query service capabilities. Intelligent robots can only process specified content. For content that requires certain business permissions, customer service must intervene and process it. The effective information that can be provided by manual customer service during non-working hours is relatively limited.

[0009] Therefore, there is an urgent need for an automatic complaint management system and method based on intelligent agent workflow. Summary of the invention

[0010] In order to overcome the problems existing in the prior art, the purpose of the present invention is to provide an automatic complaint management system and method based on intelligent agent workflow, so that AI natural language processing becomes the driving force for business operations in the complaint management system, connects the customer service system, work order management system and other data modules of the business platform, and realizes real-time updating of the business knowledge base, automation of government work orders and other applications.

[0011] To achieve the above object, the present invention provides the following technical solution: an automatic complaint management method based on intelligent agent workflow, comprising the following steps:

[0012] S1: User complaint consultation: Users initiate complaints on the business platform;

[0013] S2: Data is input into the complaint acceptance agent for preliminary text preprocessing, including sentence segmentation, content sorting by part of speech, extracting key factors such as personnel, time, place, and event, sorting complaint content using the LLM prompt template, and transmitting it to the complaint classification agent;

[0014] S3: After receiving the input content, the complaint classification agent performs word embedding processing on the data, and combines it with the system knowledge base to extract information such as the complaint business type, whether the user has made repeated complaints, and the user's emotional state, and generates complaint information through the LLMprompt template and transmits it to the complaint dispatch agent.

[0015] S4: The complaint dispatching agent uses the internal logic identification module to identify and process the complaint information, and then automatically processes the information and sends it to the complaint processing agent; or it processes the information in an offline work order process and is manually processed.

[0016] For general complaints, they are automatically handled by the complaint handling agent. For more complex demands, business problems that have never been solved before, and complaints from users with more emotional emotions, the agent API tool is used to upgrade the complaint content to a difficult work order and hand it over to the relevant offline staff for processing and resolution.

[0017] S5: The complaint handling agent determines whether the assigned complaint is the first complaint from the customer and responds based on the system knowledge base;

[0018] S6: Input the user's reply to the complaint handling agent and the evaluation of the work order process into the user feedback agent, analyze and organize the user's feedback information, and update the system knowledge base based on the analysis results.

[0019] It should be noted that agent is a concept proposed in the natural language processing framework technology in recent years. It usually refers to an artificial intelligence entity that can complete one or more specific tasks. Agent mainly consists of core module, memory module, work plan module and tool set.

[0020] The core module is a decision-making and execution module, which is used to define the overall work goals of the agent, operate the tools in the tool set, decide which plan content to execute through logical judgment, query and retrieve historical information, configure the agent's personalized information, etc. The memory module provides the agent with the ability to store and manage historical information. By combining various types of memory information including long-term and short-term, it forms a business context to ensure that the agent can have a global perspective when making judgments and decisions on the business. The work plan module can define logical processing flows for more complex businesses. These tasks will be broken down into more fine-grained sub-tasks and executed separately. At the same time, this module will also optimize the processing and decision-making logic by using critical reflection techniques such as ReAct and Reflexion. The tool set mainly includes workflows that implement various specific business contents. In most scenarios, these tools will exist in the form of interface APIs, and the agent can use these interface APIs to achieve various business needs.

[0021] The present invention is further configured as follows: the system knowledge base includes a public knowledge base, a desensitized knowledge base and a private knowledge base.

[0022] The present invention is further configured as follows: Step S2 specifically comprises the following steps:

[0023] S21: Use a pre-trained word segmentation model to split the user complaint content into multiple words and phrases. In the present invention, the Hidden Markov Model (HMM) based on statistical probability is mainly used. The HMM processes the word segmentation problem as a sequence labeling problem, and uses the Hidden Markov Model to estimate the probability of each word, so as to determine the optimal word segmentation method;

[0024] S22: Remove meaningless connecting word content in the words and phrases; such as "de", "shi", "en", etc.;

[0025] S23: Use a pre-trained word embedding model to convert the words and phrases into a word vector representation, and then use a bidirectional Transformer encoder to encode the part-of-speech of the text to generate context-related embeddings for each word;

[0026] S24: Judge the part-of-speech of the words and phrases according to the encoding, count the occurrence frequency of the encoding combinations, and identify the business entity information and complaint event content in the words and phrases;

[0027] There is a relatively complete solution in this part. During actual processing, an open-source large language model can be directly used for identification. The business entities include personnel, time, location, etc.; the complaint events include identifying the content of preset complaint keywords, as well as identifying the key words and phrases of historical complaints;

[0028] S25: Map the extracted key content vectors into words or vocabulary, and perform formatted content output through the LLM prompt template.

[0029] The present invention is further configured as: the step S3 specifically includes:

[0030] S31: For the content input in step S2, use word embedding technology to convert the input content into a business word vector with business significance;

[0031] And use a pre-trained emotion word embedding model to convert the input content into an emotion expression, perform emotion type discrimination, and use an emotion classifier to classify the emotion into positive, negative, and neutral;

[0032] S32: Input the business word vector into the private knowledge base and the desensitized knowledge base for data retrieval, and obtain the personal complaint record information of this user and the complaint handling situation information of similar events;

[0033] S33: Use a type discrimination model to judge the complaint event type corresponding to the business word vector, and combine step S32 to judge whether it is a personal repeated complaint or an others' repeated complaint;

[0034] S34: The user's personal complaint record information and the information on the handling of complaints of similar events are mapped and converted into text form, combined with the user's emotion type identification result, formatted content output is performed through the LLM prompt template, and output to the complaint dispatching agent.

[0035] The present invention is further configured as follows: Step S4 specifically comprises: extracting information from the output content of step S3 to determine whether to upgrade the complaint problem; for upgrading the complaint problem, first extracting request parameter information according to the input content, using the platform API interface, creating an event work order, and executing the offline work order process;

[0036] For complaints that do not need to be escalated, use the LLM prompt template to generate a dispatch method, input it into the complaint handling agent, and execute the automatic processing process.

[0037] The present invention is further configured as follows: the request parameter information includes user ID, complaint type, complaint time, event location, event content, historical complaints and processing records.

[0038] The present invention is further configured to: determine whether to upgrade the complaint problem specifically includes the following logic judgment module:

[0039] Logical judgment module A determines whether it is a new type of event complaint;

[0040] Search the system knowledge base. If there has never been a complaint about this incident and the system knowledge base cannot provide an effective reference for handling it, it will be handled as an upgraded complaint issue.

[0041] If there is complaint information about the event, enter the logic judgment module B;

[0042] Logical judgment module B determines whether the user has complained about the event multiple times;

[0043] If the complaint is initiated by the current user for the first time, it will be handled as a complaint that does not need to be upgraded;

[0044] If the user has made multiple complaints about the incident, the user's emotions are judged and the logic judgment module C is entered;

[0045] Logical judgment module C, judging the user's emotion type;

[0046] If the user sentiment obtained in step S3 is negative, it is treated as an upgraded complaint issue;

[0047] If the user emotion obtained in step S3 is positive or neutral, then the complaint will be handled without escalation in combination with the way other people handle similar incidents.

[0048] The present invention is further configured as follows: Step S5 specifically comprises:

[0049] S51: Determine whether the user has complained about the event multiple times; if the user has complained about the event multiple times, proceed to step S52; if the user has not complained about the event, proceed directly to step S53;

[0050] S52: Retrieving the historical complaint records and historical work order records of the user in the private knowledge base for a processing solution;

[0051] S53: searching the public knowledge base and the desensitized knowledge base for solutions to similar user complaints;

[0052] S54: Summarize the complaint handling solutions obtained by searching the private knowledge base, the public knowledge base, and the desensitized knowledge base, and map and convert them into text forms respectively;

[0053] S55: Generate the response content of the complaint handling plan through the LLM prompt template.

[0054] The present invention is further configured as follows: Step S6 specifically comprises: the user evaluates the results of the automatic processing flow and the offline work order processing flow; the evaluation content is word-embedded to form a user's private word vector, and the private knowledge base is updated; and after deleting sensitive information from the evaluation content, word embedding processing is performed to form a desensitized word, and the desensitized knowledge base is updated.

[0055] The present invention is further configured as follows: the sensitive information includes the user's personal name, address and contact information.

[0056] The present invention also provides an automatic complaint management system based on intelligent agent workflow, comprising:

[0057] Complaint acceptance agent: used to receive user complaint information received by the business platform and perform preliminary pre-processing;

[0058] Complaint classification agent: used to extract and identify complaint types, user sentiment types, and whether the complaint is repeated;

[0059] Complaint dispatch agent: used to identify and process complaint information according to the logic identification module, and dispatch it to the automatic processing process or offline work order process;

[0060] Complaint handling agent: used to retrieve and output the handling solutions in the system knowledge base;

[0061] User feedback agent: used to analyze user feedback results and update the system knowledge base;

[0062] System knowledge base: used to store information in public knowledge bases, private databases, and desensitized databases;

[0063] Public knowledge base includes community public information and Internet information;

[0064] The private database includes the user's historical complaint records, historical work order records, and personal data of the user;

[0065] The desensitized database includes business data and data after the sensitive information in the private database is blurred. For example, the private information such as the name and address of the complainant in the historical complaint record is blurred or deleted.

[0066] In summary, the beneficial effects of the above technical solution of the present invention are as follows:

[0067] 1. The present invention realizes the automated management of complaint handling by using agentic workflows, and uses an internal dialogue mechanism to drive the agent data flow. It uses sentiment word embedding technology to optimize the sentiment analysis capability of user complaints, quickly locate and identify difficult and upgraded complaints, and classify complaints. Problems that need to be upgraded are handled manually in conjunction with the work order system and handled in a timely manner according to user needs to avoid reducing user experience. Problems that do not need to be upgraded are automatically handled by the agent workflow to improve processing efficiency and user experience.

[0068] 2. By dividing the knowledge base into public, private and desensitized knowledge base, user complaints and inquiries can be processed and answered in different dimensions. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0070] Figure 1 This is the flow chart of the complaint handling agent.

[0071] Figure 2 This is the flow chart of complaint classification agent.

[0072] Figure 3 Flowchart for assigning agents to complaints.

[0073] Figure 4 This is the complaint handling agent flow chart.

[0074] Figure 5 Flowchart of the user feedback agent.

[0075] Figure 6 Schematic diagram of the automatic complaint management system with intelligent agent workflow. Detailed implementation manners

[0076] In order to enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the present invention. Based on the embodiments in the present invention, other similar embodiments obtained by those of ordinary skill in the art without creative efforts shall all fall within the protection scope of the present invention.

[0077] The present invention will be further described below in conjunction with the accompanying drawings and preferred embodiments.

[0078] Embodiment 1:

[0079] As Figure 1-Figure 5 shown, it is a preferred embodiment of the present invention. An automatic complaint management method based on an agent workflow includes the following steps:

[0080] S1: User complaint consultation: The user initiates a complaint on the business platform;

[0081] S2: Input the data into the complaint acceptance agent for preliminary text preprocessing, including sentence word segmentation processing, sorting the content according to parts of speech, extracting key factors such as personnel, time, place, event, etc., organizing the complaint content in the LLM prompt template manner, and transmitting it to the complaint classification agent; As Figure 1 shown.

[0082] S21: Use a pre-trained word segmentation model to split the user's complaint content into multiple words. In the present invention, the mainly used is the Hidden Markov Model (HMM) based on statistical probability. HMM processes the word segmentation problem as a sequence labeling problem, and uses the Hidden Markov Model to estimate the probability of each word, so as to determine the optimal word segmentation method;

[0083] S22:剔除字词中的无意义连接词内容;例如“的”、“是”、“嗯”等;

[0084] S23: Use a pre-trained word embedding model (such as BERT) to represent the words in vector form. The bidirectional Transformer encoder encodes the parts of speech of the text to generate context-related embeddings for each word, including the following steps:

[0085] Input: First, the BERT model will represent the input data in the way of word embedding, position embedding, and segment embedding, that is, for any input sequence X = x1, x2,..., x n , its input representation is:

[0086] Input(x i ) =

[0087] TokenEmbedding(x i )+PositionEmbedding(i)+SegmentEmbedding(s);

[0088] Self-Attention: The model uses a multi-layer Transformer encoder and processes the self-attention mechanism and feed-forward neural network in each layer:

[0089]

[0090] Where Q represents the query value, K represents the key value, V represents the information corresponding to the key value, T represents the matrix transpose, and d k Represents the dimension of the word vector.

[0091] Output: For an input sequence (X), BERT generates a corresponding context-dependent embedding H, where each h i Representation word x i Contextual embeddings:

[0092] H=h1,h2,…,h n .

[0093] S24: judging the part of speech of the word according to the code, counting the frequency of occurrence of the code combination, and identifying the business entity information and complaint event content in the word;

[0094] S25: Map the extracted key content vectors into text or words, and format the content output through the LLM prompt template.

[0095] S3: After receiving the input content, the complaint classification agent performs word embedding processing on the data, and combines it with the system knowledge base to extract information such as the complaint business type, whether the user has made repeated complaints, and the user's emotional state, and generates complaint information through the LLMprompt template and transmits it to the complaint dispatching agent. Figure 2 shown.

[0096] S31: for the content input in step S2, convert the input content into a business word vector with business significance through word embedding technology;

[0097] The pre-trained sentiment word embedding model is used to convert the input content into sentiment expression, and the sentiment type is identified. The sentiment is divided into positive, negative and neutral using a sentiment classifier. The joint loss of the word embedding loss and the sentiment loss is calculated, and the joint loss data is input into a multi-classifier to generate the sentiment identification result.

[0098] S32: Input the business word vector into the private knowledge base and the desensitized knowledge base for data retrieval to obtain the user's personal complaint record information and the complaint handling information of similar events;

[0099] S33: Use the type discrimination model to determine the type of complaint event corresponding to the business word vector, and combine it with step S32 to determine whether it is a repeated complaint by an individual or repeated complaint by others;

[0100] S34: The user's personal complaint record information and the information on the handling of complaints of similar events are mapped and converted into text form, combined with the user's emotion type identification result, formatted content output is performed through the LLM prompt template, and output to the complaint dispatching agent.

[0101] S4: The complaint dispatching agent uses the internal logic judgment module to judge and process the complaint information, and then automatically processes it and sends it to the complaint handling agent; or it conducts an offline work order process for manual processing.

[0102] like Figure 3 As shown in the figure, general complaints are automatically handled by the complaint handling agent. For more complex demands, business problems that have never been solved before, and complaints from users with more emotional emotions, the agent API tool is used to upgrade the complaint content to a difficult work order and hand it over to relevant offline staff for processing and resolution.

[0103] The discrimination process is to execute the following logic discrimination modules in sequence:

[0104] Logical judgment module A determines whether it is a new type of event complaint;

[0105] Search the system knowledge base. If there has never been a complaint about this incident and the system knowledge base cannot provide an effective reference for handling it, it will be handled as an upgraded complaint issue.

[0106] If there is complaint information about the event, enter the logic judgment module B;

[0107] Logical judgment module B determines whether the user has complained about the event multiple times;

[0108] If the complaint is initiated by the current user for the first time, it will be handled as a complaint that does not need to be upgraded;

[0109] If the user has made multiple complaints about the incident, the user's emotions are judged and the logic judgment module C is entered;

[0110] Logical judgment module C, judging the user's emotion type;

[0111] If the user sentiment obtained in step S3 is negative, it is treated as an upgraded complaint issue;

[0112] If the user emotion obtained in step S3 is positive or neutral, then the complaint will be handled without escalation in combination with the way other people handle similar incidents.

[0113] S5: The complaint handling agent determines whether the assigned complaint content is the first complaint from the customer and responds based on the system knowledge base; Figure 4 shown.

[0114] S51: Determine whether the user has complained about the event multiple times; if the user has complained about the event multiple times, proceed to step S52; if the user has not complained about the event, proceed directly to step S53;

[0115] S52: Retrieving the historical complaint records and historical work order records of the user in the private knowledge base for a processing solution;

[0116] S53: searching the public knowledge base and the desensitized knowledge base for solutions to similar user complaints;

[0117] S54: Summarize the complaint handling solutions obtained by searching the private knowledge base, the public knowledge base, and the desensitized knowledge base, and map and convert them into text forms respectively;

[0118] S55: Generate the response content of the complaint handling plan through the LLM prompt template.

[0119] S6: Input the user's reply to the complaint handling agent and the evaluation of the work order process into the user feedback agent, analyze and organize the user's feedback information, and update the system knowledge base based on the analysis results. Figure 5 As shown, the user evaluates the results of the automatic processing flow and the offline work order processing flow; the evaluation content is word-embedded to form the user's private word vector, and the private knowledge base is updated; the evaluation content is then deleted from the user's personal name, address, and contact information, and then word-embedded to form desensitized words, and the desensitized knowledge base is updated.

[0120] Embodiment 2:

[0121] like Figure 6 As shown, an automatic complaint management system based on an agent workflow is used in conjunction with the above-mentioned automatic complaint management method based on an agent workflow, including:

[0122] Complaint acceptance agent: used to receive user complaint information received by the business platform and perform preliminary pre-processing;

[0123] Complaint classification agent: used to extract and identify complaint types, user sentiment types, and whether the complaint is repeated;

[0124] Complaint dispatch agent: used to identify and process complaint information according to the logic identification module, and dispatch it to the automatic processing process or offline work order process;

[0125] Complaint handling agent: used to retrieve and output the handling solutions in the system knowledge base;

[0126] User feedback agent: used to analyze user feedback results and update the system knowledge base;

[0127] System knowledge base: used to store information in public knowledge bases, private databases, and desensitized databases;

[0128] Public knowledge base includes community public information and Internet information;

[0129] The private database includes the user's historical complaint records, historical work order records, and personal data of the user;

[0130] The desensitized database includes business data and data after the sensitive information in the private database is blurred. For example, the private information such as the name and address of the complainant in the historical complaint record is blurred or deleted.

[0131] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, rather than to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions of the technical solution of the present invention by ordinary technicians in this field do not deviate from the essence and scope of the technical solution of the present invention.

Claims

1. An automatic complaint management method based on intelligent agent workflow, characterized in that: The following steps are involved: S1: The user initiates a complaint on the business platform; S2: Input the user's complaint information data into the complaint acceptance agent, perform preliminary text preprocessing, organize the complaint content using the LLMprompt template, and transmit it to the complaint classification agent; S3: After receiving the input content, the complaint classification agent performs word embedding processing on the data, and extracts the complaint business type, whether the user has made repeated complaints, and the user's emotional state information in combination with the system knowledge base, and generates complaint information through the LLM prompt template and transmits it to the complaint dispatching agent; The system knowledge base includes a public knowledge base, a desensitized knowledge base and a private knowledge base; S4: The complaint dispatching agent uses the internal logic identification module to identify and process the complaint information, and then automatically processes the information and sends it to the complaint processing agent; or it processes the information in an offline work order process and is manually processed. Extract information from the output content of step S3 to determine whether to escalate the complaint; For upgraded complaint issues, first extract the request parameter information based on the input content, use the platform API interface to create an event ticket, and execute the offline ticket process; For complaints that do not need to be escalated, use the LLM prompt template to generate a dispatch method, input it into the complaint handling agent, and execute the automatic processing process; Determining whether to escalate a complaint specifically includes: determining whether it is a new type of complaint, determining whether the user has complained about the incident multiple times, and determining the user's emotional type; S5: The complaint handling agent determines whether the assigned complaint is the first complaint from the customer, retrieves the solution from the system knowledge base and responds; S6: Input the user's reply to the complaint handling agent and the evaluation of the work order process processing results into the user feedback agent to update the system knowledge base.

2. According to claim 1, the method for automatic complaint management based on agent workflow is characterized in that: Step S2 specifically includes the following steps: S21: Use the pre-trained word segmentation model to split the user complaint content into multiple words; S22: Eliminate meaningless conjunctions in words; S23: Use the pre-trained word embedding model to convert words into word vector representation, and then use the bidirectional Transformer encoder to encode the text part of speech to generate context-dependent embeddings for each word; S24: judging the part of speech of the word according to the code, counting the frequency of occurrence of the code combination, and identifying the business entity information and complaint event content in the word; S25: Map the extracted key content vectors into text or words, and output the formatted content through the LLM prompt template.

3. The method for automatic complaint management based on agent workflow according to claim 2 is characterized in that: Step S3 specifically includes: S31: for the content input in step S2, convert the input content into a business word vector with business significance through word embedding technology; The pre-trained sentiment word embedding model is used to convert the input content into sentiment expression, identify the sentiment type, and use the sentiment classifier to classify the sentiment into positive, negative, and neutral. S32: Input the business word vector into the private knowledge base and the desensitized knowledge base for data retrieval to obtain the user's personal complaint record information and the complaint handling information of similar events; S33: Use the type discrimination model to determine the type of complaint event corresponding to the business word vector, and combine it with step S32 to determine whether it is a repeated complaint by an individual or repeated complaint by others; S34: The user's personal complaint record information and the information on the handling of complaints of similar events are mapped and converted into text form, combined with the user's emotion type identification result, formatted content output is performed through the LLM prompt template, and output to the complaint dispatching agent.

4. The method for automatic complaint management based on agent workflow according to claim 3 is characterized in that: The determination of whether to escalate a complaint issue specifically includes the following logic judgment modules: Logical judgment module A determines whether it is a new type of event complaint; Search the system knowledge base. If there has never been a complaint about this incident and the system knowledge base cannot provide an effective reference for handling it, it will be handled as an upgraded complaint issue. If there is complaint information about the event, enter the logic judgment module B; Logical judgment module B determines whether the user has complained about the event multiple times; If the complaint is initiated by the current user for the first time, it will be handled as a complaint that does not need to be upgraded; If the user has made multiple complaints about the incident, the user's emotions are judged and the logic judgment module C is entered; Logical judgment module C, judging the user's emotion type; If the user sentiment obtained in step S3 is negative, it is treated as an upgraded complaint issue; If the user emotion obtained in step S3 is positive or neutral, then the complaint will be handled without escalation in combination with the way other people handle similar incidents.

5. The method for automatic complaint management based on agent workflow according to claim 4 is characterized in that: Step S5 is specifically as follows: S51: Determine whether the user has complained about the event multiple times; if the user has complained about the event multiple times, proceed to step S52; if the user has not complained about the event, proceed directly to step S53; S52: Retrieving the historical complaint records and historical work order records of the user in the private knowledge base for a processing solution; S53: searching the public knowledge base and the desensitized knowledge base for solutions to similar user complaints; S54: Summarize the complaint handling solutions obtained by searching the private knowledge base, the public knowledge base, and the desensitized knowledge base, and map and convert them into text forms respectively; S55: Generate the response content of the complaint handling plan through the LLM prompt template.

6. The method for automatic complaint management based on agent workflow according to claim 1, characterized in that: Specifically, step S6 includes the user evaluating the results of the automatic processing flow and the offline work order processing flow; embedding the evaluation content into words to form a user's private word vector, and updating the private knowledge base; and then deleting sensitive information from the evaluation content, embedding the words into words to form desensitized words, and updating the desensitized knowledge base.

7. The method for automatic complaint management based on agent workflow according to claim 6, characterized in that: The sensitive information includes the user's personal name, address and contact information.

8. An automatic complaint management system based on intelligent workflow, in conjunction with an automatic complaint management method based on intelligent workflow as claimed in any one of claims 1 to 7, characterized in that: include: Complaint acceptance agent: used to receive user complaint information received by the business platform and perform preliminary pre-processing; Complaint classification agent: used to extract and identify complaint types, user sentiment types, and whether the complaint is repeated; Complaint dispatch agent: used to identify and process complaint information according to the logic identification module, and dispatch it to the automatic processing process or offline work order process; Complaint handling agent: used to retrieve and output the handling solutions in the system knowledge base; User feedback agent: used to analyze user feedback results and update the system knowledge base; System knowledge base: used to store information in public knowledge bases, private databases, and desensitized databases; Public knowledge base includes community public information and Internet information; The private database includes the user's historical complaint records, historical work order records, and personal data of the user; The desensitized database includes business data and data that has been obfuscated from sensitive information in private databases.

Images