Multi-agent collaborative product closed-loop optimization system and method based on user feedback data

Abstract

The present application relates to a multi-agent collaborative product closed-loop optimization system and method based on user feedback data, which is used to solve the technical problem that the existing technology cannot efficiently utilize user feedback data to obtain optimization improvement requirements, resulting in poor optimization effect. It includes: a data collection module that captures user feedback data in real time from public and private data sources; a data analysis module that calls a multi-dimensional analysis agent to perform AI multi-dimensional analysis and mining on the user feedback data after data normalization, generating feedback analysis results; an image construction module that constructs and calls expert role prompt information templates, fills in the context by inputting feedback analysis results into the templates, obtains expert role prompt information, and calls an intelligent portrait generation agent to construct a structured user intelligent portrait; a task generation module that calls a business design expert agent to automatically generate product optimization requirements and task collaboration based on feedback analysis results and structured user intelligent portraits.

Description

Technical Field

[0001] This invention relates to the field of computer information processing, and in particular to a multi-agent collaborative product closed-loop optimization system, method, electronic device, computer storage medium, and computer program product based on user feedback data. Background Technology

[0002] With the development of information technology, existing business products often include functional systems that can be used in multiple application scenarios. Therefore, the user feedback data obtained is diverse in form and large in volume. Collecting user feedback data allows for the systematic collection, analysis, and understanding of user needs, expectations, preferences, and feedback regarding products, services, or experiences. This enables the capture of users' true needs and facilitates efficient business optimization and improvement based on user experience or feedback.

[0003] Current technologies rely on a limited range of methods for collecting user feedback data. These methods often involve internal customer service calls, official website forms, or app store reviews, using manual or semi-automatic approaches to gather feedback, which is then analyzed using basic statistical tools. However, this approach suffers from a singular source of feedback, leading to a disconnect from actual user experience and the potential for overlooking genuine needs. Furthermore, the manual nature of the collection and analysis process hinders the accurate acquisition of multi-dimensional information on user needs. Additionally, the low level of collaboration in optimizing user feedback results in slow iteration and an inability to meet rapidly changing user demands. Therefore, existing technologies struggle to accurately and efficiently utilize user feedback data, often leading to technical problems such as ineffective product improvements and optimizations. Summary of the Invention

[0004] The main objective of this invention is to solve the technical problem in the prior art that the inability to efficiently utilize user feedback data to obtain optimization and improvement needs leads to poor product optimization and improvement results.

[0005] The first aspect of this invention provides a multi-agent collaborative product closed-loop optimization system based on user feedback data, comprising:

[0006] The data collection module is used to capture user feedback data in real time from public and private data sources, and to perform data normalization on the user feedback data.

[0007] The data analysis module is used to call a multi-dimensional analysis agent to perform AI multi-dimensional analysis and mining on the user feedback data after data normalization, and generate feedback analysis results. The feedback analysis results include at least: text dimension analysis results, sentiment dimension analysis results, topic dimension analysis results, and trend analysis results.

[0008] The profile building module is used to build and retrieve the expert role prompt information template, input the feedback analysis results into the expert role prompt information template to fill in the context content, and obtain the expert role prompt information; and call the intelligent profile generation agent to build a structured user intelligent profile based on the expert role prompt information.

[0009] The task generation module is used to invoke the business design expert intelligent agent to automatically generate product optimization requirements and task collaboration based on the feedback analysis results and the structured user intelligent profile.

[0010] Optionally, in a first implementation of the first aspect of the present invention, the user feedback data includes public domain feedback data and private domain feedback data;

[0011] The data collection module is specifically used for:

[0012] By using data interfaces, we can capture private domain feedback data in real time from multiple private domain data source systems for the reception optimization product.

[0013] By invoking robot operation processes or distributed crawlers, relevant information about the product to be optimized can be monitored from multiple public domain data sources, and public domain feedback data can be captured in real time.

[0014] Optionally, in a second implementation of the first aspect of the present invention, the data collection module is further configured to:

[0015] The user feedback data is preprocessed, including deduplication, desensitization, and noise filtering;

[0016] A natural language processing model is invoked to perform language detection and format standardization on the preprocessed user feedback data, transforming the user feedback data into a preset standard format.

[0017] Optionally, in a third implementation of the first aspect of the present invention, the data analysis module is specifically used for:

[0018] The multi-dimensional analysis agent is invoked to analyze the user feedback data from the text dimension, extracting factual information reflecting customer usage behavior and problem descriptions, and obtaining text dimension analysis results;

[0019] The multi-dimensional analysis agent is invoked to analyze the user feedback data from the emotional dimension, identify the emotional category and intensity of the user's emotions, and obtain the emotional dimension analysis results;

[0020] The multi-dimensional analysis agent is invoked to analyze the user feedback data from the topic dimension, identify multiple topic tags related to the user feedback data, and obtain the topic dimension analysis results;

[0021] Historical user feedback data corresponding to the user feedback data is obtained, and the frequency and time change information of similar feedback data are statistically analyzed based on the historical user feedback data. A multi-dimensional analysis agent is invoked to perform trend analysis on the frequency and time change information to obtain the trend analysis results of the user feedback data. The similar feedback data refers to historical user feedback data whose similarity to the user feedback data in the sentiment dimension and the topic dimension is higher than the preset similarity threshold of the corresponding dimension.

[0022] Optionally, in a fourth implementation of the first aspect of the present invention, the expert role prompt information includes:

[0023] Role constraint information used to define the role of the intelligent agent and the task objectives to be performed; reasoning constraint information used to define the order of user profile generation or reasoning path; and output structure constraint information used to define the output fields of the user profile and the dependencies between fields.

[0024] Optionally, in a fifth implementation of the first aspect of the present invention, the portrait construction module is specifically used for:

[0025] Construct a profile input feature set based on the feedback analysis results;

[0026] The intelligent profile generation agent receives the profile input feature set and generates a structured user intelligent profile according to the preset profile generation format based on the expert role prompts.

[0027] The structured user intelligent profile includes profile type information, user background information, user behavior patterns, user core pain points, user potential needs, and profile emotional attitude.

[0028] Optionally, in a sixth implementation of the first aspect of the present invention, the task generation module is specifically used for:

[0029] At least one task decision feature is extracted from the feedback analysis results and the structured user intelligent profile, wherein the task decision feature includes at least one of user type, risk level, problem type and emotional attitude;

[0030] Based on the task decision characteristics, task parameters are determined, wherein the task parameters include actual optimization requirements, task priority, task type, and the corresponding processing department.

[0031] Based on the task parameters, product optimization requirement tasks associated with the user feedback data are generated.

[0032] Optionally, in a seventh implementation of the first aspect of the present invention, the data analysis module is further used to evaluate whether the feedback analysis results of each dimension meet the analysis requirement threshold.

[0033] If the analysis requirement threshold is not met, the task generation module is also used to generate data collection point requirements and generate a data collection point requirement task based on the data collection point requirements.

[0034] A second aspect of this invention provides a closed-loop optimization method for multi-agent collaborative products based on user feedback data, comprising:

[0035] User feedback data is captured in real time from public and private data sources, and the user feedback data is then organized.

[0036] The multi-dimensional analysis agent is invoked to perform AI multi-dimensional analysis and mining on the user feedback data after data normalization, and to generate feedback analysis results, wherein the feedback analysis results include at least: text dimension analysis results, sentiment dimension analysis results, topic dimension analysis results and trend analysis results;

[0037] Construct and retrieve the expert role prompt information template, input the feedback analysis results into the expert role prompt information template to fill in the context content, and obtain the expert role prompt information;

[0038] The intelligent profile is invoked to generate an intelligent agent that constructs a structured user intelligent profile based on the expert role prompts.

[0039] The business design expert intelligent agent is invoked to automatically generate product optimization requirements and task collaboration based on the feedback analysis results and the structured user intelligent profile.

[0040] Optionally, in a first implementation of the second aspect of the present invention, the user feedback data includes public domain feedback data and private domain feedback data;

[0041] The real-time capture of user feedback data from public and private data sources includes:

[0042] By using data interfaces, we can capture private domain feedback data in real time from multiple private domain data source systems for the reception optimization product.

[0043] By invoking robot operation processes or distributed crawlers, relevant information about the product to be optimized can be monitored from multiple public domain data sources, and public domain feedback data can be captured in real time.

[0044] Optionally, in a second implementation of the second aspect of the present invention, the data normalization of the user feedback data includes:

[0045] The user feedback data is preprocessed, including deduplication, desensitization, and noise filtering;

[0046] A natural language processing model is invoked to perform language detection and format standardization on the preprocessed user feedback data, transforming the user feedback data into a preset standard format.

[0047] Optionally, in a third implementation of the second aspect of the present invention, the step of invoking a multi-dimensional analysis agent to perform AI multi-dimensional analysis and mining on user feedback data and generating feedback analysis results includes:

[0048] The multi-dimensional analysis agent is invoked to analyze the user feedback data from the text dimension, extracting factual information reflecting customer usage behavior and problem descriptions, and obtaining text dimension analysis results;

[0049] The multi-dimensional analysis agent is invoked to analyze the user feedback data from the emotional dimension, identify the emotional category and intensity of the user's emotions, and obtain the emotional dimension analysis results;

[0050] The multi-dimensional analysis agent is invoked to analyze the user feedback data from the topic dimension, identify multiple topic tags related to the user feedback data, and obtain the topic dimension analysis results;

[0051] Historical user feedback data corresponding to the user feedback data is obtained, and the frequency and time change information of similar feedback data are statistically analyzed based on the historical user feedback data. A multi-dimensional analysis agent is invoked to perform trend analysis on the frequency and time change information to obtain the trend analysis results of the user feedback data. The similar feedback data refers to historical user feedback data whose similarity to the user feedback data in the sentiment dimension and the topic dimension is higher than the preset similarity threshold of the corresponding dimension.

[0052] Optionally, in a fourth implementation of the second aspect of the present invention, the expert role prompt information includes:

[0053] Role constraint information used to define the role of the intelligent agent and the task objectives to be performed; reasoning constraint information used to define the order of user profile generation or reasoning path; and output structure constraint information used to define the output fields of the user profile and the dependencies between fields.

[0054] Optionally, in a fifth implementation of the second aspect of the present invention, the step of invoking the intelligent profile generation agent to construct a structured user intelligent profile based on the expert role prompt information includes:

[0055] Construct a profile input feature set based on the feedback analysis results;

[0056] The intelligent profile generation agent receives the profile input feature set and generates a structured user intelligent profile according to the preset profile generation format based on the expert role prompts.

[0057] The structured user intelligent profile includes profile type information, user background information, user behavior patterns, user core pain points, user potential needs, and profile emotional attitude.

[0058] Optionally, in a sixth implementation of the second aspect of the present invention, the automatic generation of product optimization requirements and task collaboration based on the feedback analysis results and the structured user intelligent profile includes:

[0059] At least one task decision feature is extracted from the feedback analysis results and the structured user intelligent profile, wherein the task decision feature includes at least one of user type, risk level, problem type and emotional attitude;

[0060] Based on the task decision characteristics, task parameters are determined, wherein the task parameters include actual optimization requirements, task priority, task type, and the corresponding processing department.

[0061] Based on the task parameters, product optimization requirement tasks associated with the user feedback data are generated.

[0062] Optionally, in the seventh implementation of the second aspect of the present invention, after calling the multi-dimensional analysis agent to perform AI multi-dimensional analysis and mining on the data-normalized user feedback data and generating feedback analysis results, the method further includes:

[0063] Evaluate whether the feedback analysis results for each dimension meet the analysis requirement thresholds;

[0064] If the analysis requirement threshold is not met, a data collection point requirement is generated, and a point collection point requirement task is generated based on the data collection point requirement.

[0065] A third aspect of the present invention provides a multi-agent collaborative product closed-loop optimization device based on user feedback data, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the multi-agent collaborative product closed-loop optimization device based on user feedback data to perform the steps of the above-described multi-agent collaborative product closed-loop optimization method based on user feedback data.

[0066] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the steps of the above-described multi-agent collaborative product closed-loop optimization method based on user feedback data.

[0067] A fifth aspect of the present invention provides a computer program product, including a computer program / instruction that, when executed by a processor, implements the steps of the multi-agent collaborative product closed-loop optimization method based on user feedback data as described above.

[0068] The technical solution provided by this invention includes a multi-agent collaborative product closed-loop optimization system based on user feedback data, comprising a data collection module, a data analysis module, a profile construction module, and a task generation module. The data collection module captures user feedback data in real time from public and private data sources and performs data normalization on the user feedback data. The data analysis module calls a multi-dimensional analysis agent to perform AI multi-dimensional analysis and mining on the normalized user feedback data, generating feedback analysis results, which include at least: text dimension analysis results, sentiment dimension analysis results, topic dimension analysis results, and trend analysis results. The profile construction module constructs and retrieves an expert role prompt information template, inputs the feedback analysis results into the expert role prompt information template to fill in contextual content, and obtains expert role prompt information. It then calls an intelligent profile generation agent to construct a structured user intelligent profile based on the expert role prompt information. The task generation module calls a business design expert agent to automatically generate product optimization requirements and task collaboration based on the feedback analysis results and the structured user intelligent profile. Based on this, this invention can efficiently utilize user feedback data, combined with the constructed structured user intelligent profile, to extract comprehensive optimization and improvement requirements from multiple perspectives, thereby improving the optimization and improvement effect of the product. Attached Figure Description

[0069] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0070] Figure 1 This is a flowchart illustrating the first embodiment of the multi-intelligent collaborative product closed-loop optimization method based on user feedback data in this invention.

[0071] Figure 2 This is a schematic diagram of the modules of the first embodiment of the multi-intelligent collaborative product closed-loop optimization system based on user feedback data in this invention.

[0072] Figure 3 This is a flowchart illustrating a second embodiment of the multi-intelligent collaborative product closed-loop optimization method based on user feedback data in this invention.

[0073] Figure 4 This is another flowchart illustrating a second embodiment of the multi-intelligent collaborative product closed-loop optimization method based on user feedback data in this invention.

[0074] Figure 5 This is a schematic diagram of the architecture of the second embodiment of the multi-intelligent collaborative product closed-loop optimization system based on user feedback data in this invention.

[0075] Figure 6 This is a flowchart illustrating the third embodiment of the multi-agent collaborative product closed-loop optimization method based on user feedback data in this invention.

[0076] Figure 7 This is a schematic diagram of the architecture of the third embodiment of the multi-intelligent collaborative product closed-loop optimization method based on user feedback data in this invention. Detailed Implementation

[0077] Exemplary embodiments of the invention will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limiting the invention to the embodiments set forth herein. Rather, these exemplary embodiments are provided to make the invention more comprehensive and complete, and to facilitate a full communication of the inventive concept to those skilled in the art. The same reference numerals in the drawings denote the same or similar elements, components, or parts, and therefore repeated descriptions of them will be omitted.

[0078] Subject to the inventive concept, the features, structures, characteristics or other details described in a particular embodiment may be combined in one or more other embodiments in a suitable manner.

[0079] In the description of specific embodiments, the features, structures, characteristics, or other details described in this invention are intended to enable those skilled in the art to fully understand the embodiments. However, it is not excluded that those skilled in the art can practice the technical solutions of this invention without one or more of the specific features, structures, characteristics, or other details.

[0080] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0081] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0082] The terms “and / or” or “and / or” include all combinations of any one or more of the listed items.

[0083] See Figure 1 The specific details of the multi-agent collaborative product closed-loop optimization method based on user feedback data in the first embodiment of the present invention are as follows:

[0084] S101. Capture user feedback data in real time from public domain data sources and / or private domain data sources, and perform data normalization on user feedback data based on semantic purification intelligent agents;

[0085] It is understood that the executing entity of this invention can be a multi-agent collaborative product closed-loop optimization system based on user feedback data, or it can be a terminal or a server; the specific implementation is not limited here. This embodiment of the invention will be described using a server as an example.

[0086] In this embodiment, when collecting user feedback data, different feedback information, experience information, and public opinion data are mainly collected from both public and / or private data sources. Specifically, private feedback data is captured in real time from multiple private data source systems of the product to be optimized through data interfaces. The operation process of the intelligent extraction robot is invoked to listen to the relevant information of the product to be optimized in multiple public data sources and capture public feedback data in real time.

[0087] Private domain data sources include, but are not limited to, brand apps, official websites, customer service phone numbers, official social media accounts, and SMS messages; public domain data sources include, but are not limited to, social media, offline surveys, forum discussions, blog posts, news reports, and app store reviews.

[0088] After obtaining user feedback data, the process also includes preprocessing the user feedback data, including deduplication, desensitization, and noise filtering; calling a natural language processing model to perform language detection and format standardization on the preprocessed user feedback data, and converting the user feedback data into a preset standard format.

[0089] S102. Call the multi-dimensional analysis agent to combine the dynamic knowledge graph and the multi-hop topological reasoning capability of the GraphRAG engine to perform AI multi-dimensional analysis on the user feedback data after data normalization and generate feedback analysis results.

[0090] The multi-agent collaborative product closed-loop optimization method based on user feedback data in this embodiment is implemented based on a multi-agent collaboration model, which can be deployed using a pre-built AI (Artificial Intelligence) analysis engine. This multi-agent collaboration model includes multiple agents built on a Large Language Model (LLM), such as a multi-dimensional analysis agent, an intelligent profile generation agent, and a business design expert agent. Through the cooperation of these agents, appropriate agents are selected to execute specific steps at different times, achieving intelligent closed-loop optimization of the product based on the collaboration of these agents. In this step, the multi-dimensional analysis agent in the multi-agent collaboration model is invoked to perform AI multi-dimensional analysis and mining on the user feedback data after data normalization.

[0091] Specifically, the feedback analysis results in this embodiment include at least text dimension analysis results, sentiment dimension analysis results, topic dimension analysis results, and trend analysis results. When obtaining the analysis results, a multi-dimensional analysis agent is invoked to analyze the user feedback data from different dimensions, including: analyzing the user feedback data from a text dimension perspective to extract factual information reflecting customer usage behavior and problem descriptions, thus obtaining text dimension analysis results; analyzing the user feedback data from a sentiment dimension perspective to identify the emotion category and intensity of user emotions, thus obtaining sentiment dimension analysis results; and analyzing the user feedback data from a topic dimension perspective to identify multiple topic tags related to the user feedback data, thus obtaining topic dimension analysis results.

[0092] The multi-dimensional analysis agent first receives a request for AI multi-dimensional analysis of standardized semantic data. Then, the orchestration agent breaks down the request and coordinates a group of domain expert agents, including specialist agents for sentiment analysis, topic analysis, and trend analysis, to conduct collaborative analysis. Each specialist agent works with a dynamic knowledge graph and the GraphRAG engine to first transform the standardized semantic data into "entity-relationship" triples and mount them onto the graph to complete semantic mapping. Then, GraphRAG uses a hybrid retrieval method of locally matching entity neighbors to obtain micro-details and globally matching topic communities to capture macro-trends. Combined with graph algorithms, multi-hop topological reasoning is used to mine deep causal relationships. After completing the analysis of the corresponding dimensions, the results are integrated to form a feedback analysis result that includes at least sentiment, topic, and trend analysis results. Finally, AI multi-dimensional analysis of user feedback data is achieved.

[0093] In a preferred embodiment, the method further includes obtaining trend analysis results. Specific steps include: obtaining historical user feedback data corresponding to the user feedback data; statistically analyzing the frequency and time variation information of similar feedback data based on the historical user feedback data; and calling a multi-dimensional analysis agent to perform trend analysis on the frequency and time variation information to obtain trend analysis results for the user feedback data. The similar feedback data refers to historical user feedback data whose similarity to the user feedback data in the text, sentiment, and topic dimensions is higher than a preset similarity threshold for the corresponding dimension.

[0094] S103. During the analysis process, the uncertainty of the calculated information is triggered. When the uncertainty value exceeds the dynamic judgment threshold, active data completion is triggered.

[0095] This embodiment also includes an active learning data completion algorithm based on information entropy, which introduces uncertainty measurement to reduce data bias caused by forced reasoning when users report missing data.

[0096] Specifically, based on the feedback analysis results, the information entropy of the output probability distribution is calculated, and the prediction variance is calculated through multiple forward propagations using the Monte Carlo Dropout (MC Dropout) algorithm to obtain the uncertainty value. A preset dynamic judgment threshold is obtained. When the uncertainty value exceeds the dynamic judgment threshold, the current information is considered insufficient to support decision-making, triggering an exploration task. Inference is paused, and a completion query task is issued to the exploration agent. The exploration agent attributes missing data based on the context of the feedback analysis results, generates a completion query requirement, obtains the completion data, and updates the feedback analysis results based on the completion data. The feedback analysis results include at least: sentiment dimension analysis results, topic dimension analysis results, and trend analysis results.

[0097] S104. Construct and retrieve the expert role prompt information template, input the feedback analysis results into the expert role prompt information template to fill in the context content, and obtain the expert role prompt information;

[0098] In this embodiment, the step of generating a user intelligent profile is implemented based on an intelligent profile generation agent, which is also built based on a large language model. Therefore, before executing the specific structured user intelligent profile generation steps, a standardized profile generation process needs to be constructed to achieve the computer-recognizable and interpretable nature of the generated user intelligent profile.

[0099] In this step, to standardize the profile generation process, several expert role prompt information templates for different scenarios are pre-constructed. These templates include role constraint information templates for defining the expert role and task objectives of the intelligent agent, reasoning constraint information templates for defining the user profile generation order or reasoning path, and output structure constraint information templates for defining the user profile output fields and their dependencies. However, they do not include contextual content regarding specific user feedback information. When constructing expert role prompt information, the expert role prompt information template is retrieved, and the contextual content of the feedback analysis results is input into the template to fill in the contextual content, thus obtaining the expert role prompt information corresponding to the feedback analysis results collected in this embodiment.

[0100] The contextual content of the user feedback information may include basic context, such as user identity information, feedback timestamp, geographical location information, and feedback device identifier; it may also include behavioral context, such as the user's operation path information before providing feedback, interaction log information, etc.

[0101] S105. Call the intelligent profile to generate an intelligent agent and construct a structured user intelligent profile based on the expert role prompts.

[0102] In this step, based on the expert role prompts obtained in the preceding steps, the intelligent profile generation agent is invoked to construct a profile input feature set based on the feedback analysis results; the profile input feature set is received and a structured user intelligent profile is generated, wherein the structured user intelligent profile includes structured profile type information, user background information, user behavior patterns, user core pain points, user potential needs, and profile emotional attitudes, etc.

[0103] S106. Invoke the business design expert intelligent agent to generate a visual report based on feedback analysis results and structured user intelligent profiles, including full-process closed-loop monitoring feedback data, product optimization requirements, and task collaboration.

[0104] At least one task decision feature is extracted from the feedback analysis results and the structured user intelligent profile, wherein the task decision feature includes at least one of risk level, problem type, and emotional attitude; a business design expert intelligent agent is invoked to determine task parameters based on the task decision feature, wherein the task parameters include user type, task priority, task type, and corresponding processing department; subsequently, the business design expert intelligent agent generates product optimization requirement tasks associated with the user feedback data based on the task parameters, and submits a task collaboration request to the corresponding processing department.

[0105] Furthermore, in this step, when generating product optimization requirements and task collaboration requests, full-process closed-loop monitoring feedback data will be generated based on the feedback analysis results and the structured user intelligent profile. This full-process closed-loop monitoring feedback data is used at least to guide the capture of user feedback data and to feed back into the dynamic knowledge graph and the real-time updates of the structured user intelligent profile.

[0106] Based on the feedback data from the closed-loop monitoring of the entire process, product optimization needs, and task collaboration information, a visual report is generated for archiving or presentation to managers.

[0107] The method provided in this embodiment of the invention can obtain user feedback data from public and / or private data sources, enabling automated integration of multi-source data. It constructs user profiles based on feedback analysis results and generates product optimization requirements based on these profiles. This method efficiently utilizes user feedback data to extract comprehensive optimization and improvement requirements from multiple perspectives, thereby enhancing the effectiveness and efficiency of product optimization.

[0108] See Figure 2 The following is a description of the module composition of the multi-agent collaborative product closed-loop optimization system based on user feedback data in the first embodiment of the present invention:

[0109] Data collection module 201 is used to capture user feedback data in real time from public domain data sources and / or private domain data sources; and to perform data normalization on the user feedback data based on semantic purification intelligent agent;

[0110] The data analysis module 202 is used to call the multi-dimensional analysis agent to combine the multi-hop topological reasoning capabilities of the dynamic knowledge graph and the GraphRAG engine to perform AI multi-dimensional analysis on the user feedback data after data normalization, and generate feedback analysis results; during the analysis process, the uncertainty of information is calculated, and when the uncertainty value exceeds the dynamic judgment threshold, active data completion is triggered; wherein, the feedback analysis results include at least: sentiment dimension analysis results, topic dimension analysis results, and trend analysis results;

[0111] The profile construction module 203 is used to construct and retrieve the expert role prompt information template, input the feedback analysis results into the expert role prompt information template to fill in the context content, and obtain the expert role prompt information; and call the intelligent profile generation agent to construct a structured user intelligent profile based on the expert role prompt information.

[0112] The task generation module 204 is used to invoke the business design expert intelligent agent to generate a visual report based on the feedback analysis results and the structured user intelligent profile, including full-process closed-loop monitoring feedback data, product optimization requirements, and task collaboration; wherein, the full-process closed-loop monitoring feedback data is used at least to guide the capture of the user feedback data and to feed back into the dynamic knowledge graph and the real-time update of the structured user intelligent profile.

[0113] In another optional embodiment of this system, the user feedback data includes public domain feedback data and private domain feedback data;

[0114] The data collection module 201 is specifically used for:

[0115] Through data interfaces and / or the client side, one or more private domain data source systems of the reception optimization product are connected; wherein, when connecting through the client side, structured private domain feedback data is extracted through business system probes, and offline sentiment computing models are called to obtain sentiment vectors and intent tags;

[0116] The system invokes an intelligent extraction robot and / or a browser-based intelligent agent to monitor relevant information about the product to be optimized from one or more public domain data sources. When the browser-based intelligent agent is invoked, it combines a visual language model to perform real-time screenshot analysis and DOM tree semantic parsing of the page, and invokes a multimodal parser to extract facial expressions, body language, and keyframe text from the video stream data to obtain real-time captured public domain feedback data.

[0117] In another optional embodiment of this system, the data collection module 201 is further used to perform semantic equivalence determination, context-aware desensitization, and information entropy gain filtering on user feedback data;

[0118] By identifying and merging user feedback data with semantic repetition or semantic similarity within a preset threshold range through semantic equivalence determination, intent-level deduplication is achieved.

[0119] Based on context-aware desensitization technology, personal privacy information is accurately identified and desensitized by combining the context of user feedback, avoiding over-desensitization or incomplete desensitization.

[0120] The information entropy value of each feedback data is calculated by filtering the information entropy gain, and redundant data with too low entropy value and / or no effective information gain are removed, and finally standardized semantic data is output.

[0121] In another optional embodiment of this system, the multi-dimensional analysis agent is composed of a group of domain expert agents, and the AI ​​multi-dimensional analysis is completed through orchestration agents. The data analysis module 202 is specifically used for:

[0122] The orchestration agent is invoked to break down the requirements corresponding to the multi-dimensional AI analysis and coordinate the domain expert agent group to carry out collaborative analysis; wherein, the domain expert agent group includes at least specialized expert agents for sentiment dimension analysis, topic dimension analysis and trend analysis; more specifically, the domain expert agent group includes at least specialized expert agents for sentiment dimension analysis, topic dimension analysis and trend analysis.

[0123] Each specialized expert agent combines dynamic knowledge graphs with the multi-hop topological reasoning capabilities of the GraphRAG engine to collaboratively conduct multi-dimensional AI analysis.

[0124] In another optional embodiment of this system, the data analysis module 202 is further configured to:

[0125] The orchestration agent is invoked to calculate the information entropy of the output probability distribution, and the prediction variance is calculated through multiple forward propagations using the Dropout Monte Carlo sampling algorithm to obtain the uncertainty value.

[0126] A preset dynamic judgment threshold is obtained. When the uncertainty value exceeds the dynamic judgment threshold, it is considered that the current information is insufficient to support the decision and an exploration task is triggered.

[0127] The inference process is paused, and the orchestration agent issues a completion query task to the exploration agent. The exploration agent attributes the missing data based on the context of the feedback analysis result, generates a completion query requirement, obtains the completion data, and updates the feedback analysis result based on the completion data.

[0128] In another optional embodiment of this system, the task generation module 204 is further configured to:

[0129] The orchestration agent is invoked to coordinate the business design expert agent and the domain expert agent group to carry out collaborative review and optimization. The generated product optimization requirements and the task collaboration are isolated and tested in a simulation sandbox, and the execution trajectory is recorded and the business impact is predicted.

[0130] Once the execution trajectory and business impact have passed security testing and meet the audit requirements of the compliance risk control intelligent agent, the business design expert intelligent agent integrates monitoring information from all stages of the entire process to generate closed-loop monitoring feedback data. Then, the closed-loop monitoring feedback data, the product optimization requirements, and the tasks are collaboratively integrated to generate a visual report.

[0131] In one optional embodiment of this system, the data analysis module 202 is specifically used for:

[0132] The user feedback data is analyzed from an emotional dimension to identify the emotion category and intensity of the user's emotions, thus obtaining the emotional dimension analysis results;

[0133] The user feedback data is analyzed from a topic dimension to identify multiple topic tags related to the user feedback data, and the topic dimension analysis results are obtained.

[0134] Historical user feedback data corresponding to the user feedback data is obtained, and the frequency and time change information of similar feedback data are statistically analyzed based on the historical user feedback data. The frequency and time change information are then analyzed to obtain the trend analysis results of the user feedback data. The similar feedback data refers to historical user feedback data whose similarity to the user feedback data in the text dimension, sentiment dimension, and topic dimension is higher than the preset similarity threshold of the corresponding dimension.

[0135] In another optional implementation of this system embodiment, the expert role prompt information includes:

[0136] Role constraint information used to define the role of the intelligent agent and the task objectives to be performed; reasoning constraint information used to define the order of user profile generation or reasoning path; and output structure constraint information used to define the output fields of the user profile and the dependencies between fields.

[0137] See Figure 3 as well as Figure 4 The second embodiment of the multi-agent collaborative product closed-loop optimization method based on user feedback data in this invention is as follows:

[0138] It is understood that the executing entity of this invention can be a product intelligent closed-loop optimization system based on user feedback data, or it can be a terminal or a server; the specific implementation is not limited here. This embodiment of the invention will be illustrated using a product intelligent closed-loop optimization system as the executing entity.

[0139] In this embodiment, the user feedback data can be the Voice of the Customer (VoC), also known as customer voice, which actually represents customer needs or customer feedback information. In recent years, it has been widely used in product quality management, user experience optimization, and product development.

[0140] S301. Capture user feedback data in real time from public domain data sources and / or private domain data sources, and perform data normalization on the user feedback data;

[0141] The multi-agent collaborative product closed-loop optimization method based on user feedback data described in this embodiment can be implemented based on a multi-agent collaboration model, which can be deployed relying on a pre-built AI (Artificial Intelligence) analysis engine. This multi-agent collaboration model includes multiple agents built on a Large Language Model (LLM), such as a multi-dimensional analysis agent, an intelligent profile generation agent, and a business design expert agent. Through the cooperation of these agents, suitable agents are selected to execute specific steps at different times, achieving intelligent closed-loop optimization of the product based on the collaboration of these agents. In this step, the multi-dimensional analysis agent in the multi-agent collaboration model is invoked to perform AI multi-dimensional analysis and mining on the user feedback data after data normalization. The Large Language Model refers to a deep learning-based AI model designed to understand and generate human language, possessing a large parameter scale and rich application scenarios.

[0142] The product intelligent closed-loop optimization system provided in this embodiment can be composed of multiple servers, including a data acquisition server, an analysis server, and a task management server. The client (such as a user terminal or management backend) is only used to communicate and connect with each server, report and receive data, and view results.

[0143] The method provided in this embodiment is mainly applied to customer feedback management and product optimization scenarios of C-end (Customer end, referring to the client used by the end user) products, such as consumer apps (Application), e-commerce platforms, social media applications, etc.

[0144] Based on this application scenario, this embodiment collects user feedback data primarily from two sources: public domain data sources and / or private domain data sources. These sources collect various feedback information, experience information, and public opinion data, resulting in public domain feedback data and private domain feedback data. Private domain data sources include, but are not limited to, information from brand apps, official websites, customer service hotlines, official social media accounts, and SMS messages. Public domain data sources include, but are not limited to, information from social media, offline surveys, forum discussions, blog posts, news reports, and app store reviews.

[0145] Preferably, the process of collecting public domain feedback data and private domain feedback data includes: capturing private domain feedback data in real time from multiple private domain data source systems for receiving and optimizing products through data interfaces; and calling the operation process of the intelligent extraction robot to listen to relevant information of the products to be optimized in multiple public domain data sources and capture public domain feedback data in real time.

[0146] In one specific implementation, this step can be performed by a data acquisition module deployed on a data acquisition server. Specifically, in conjunction with... Figure 4 The data collection processes for Private Domain Data (401), Public Domain Data (402), and Voice of the Customer (403) can be as follows: When collecting public opinion data from private domain channels, the data can be proactively integrated with internal business systems, such as CRM (Customer Relationship Management) systems, customer service systems, and task systems, through business interfaces of various private domain channels, such as Application Programming Interfaces (APIs). This allows for the acquisition of structured data (such as rating scores) and unstructured data (such as call recordings converted to text and feedback forms). When collecting data from public domain channels, RPA (Robotic Process Automation) technology can be used to build an automated process for intelligent extraction robots to simulate browser operations and obtain real-time Voice of the Customer (VoC) as user feedback data. This includes information from social media, e-commerce platforms, and industry forums for public opinion monitoring, resulting in real-time unstructured text, images, and video metadata. Furthermore, when acquiring input data, relevant metadata about the Voice of the Customer, such as collection time, source platform, user identifier, and product or feature identifier, will also be obtained.

[0147] After collecting customer feedback data, it is stored in a central database. This stored data undergoes data standardization, specifically including preprocessing and format standardization. Preprocessing steps include deduplication, anonymization, and noise filtering. Format standardization involves using a Natural Language Processing (NLP) model to perform language detection and format standardization on the preprocessed user feedback data, thereby converting it into a preset standard format. See also... Figure 4 As shown in the relevant process of the AI ​​data analysis 404 section, after obtaining the customer's voice 403 in the standard format, a multi-dimensional analysis agent can be called to perform AI data analysis.

[0148] In this step, by coordinating different data acquisition processes, data from multiple data sources is obtained, not limited to internal customer service data or a single social channel. This achieves the fusion of multi-source heterogeneous data between private domain business systems and public domain public opinion information, solving the problem of incomplete data acquisition. This enables enterprises to capture real user feedback from a global perspective and avoids analytical biases caused by data fragmentation.

[0149] S302. Call the multi-dimensional analysis agent to analyze user feedback data from the text dimension, extract factual information reflecting customer usage behavior and problem description, and obtain text dimension analysis results;

[0150] In one specific implementation, steps S302-S305 of this embodiment are executed by a data analysis module deployed on a data acquisition server. The data analysis module calls a multi-dimensional analysis agent to analyze the collected user feedback data and obtain analysis results from different dimensions. The multi-dimensional analysis agent is built based on a Large Language Model (LLM).

[0151] Combination Figure 4 In the Chinese text analysis section 405, step S302 of this embodiment calls a multi-dimensional analysis agent to extract factual information and obtain text-dimensional analysis results. Specifically, the multi-dimensional analysis agent can call Natural Language Processing (NLP) algorithms to parse the user feedback data stored in the central database from the text dimension, extract keywords and entity names, obtain information such as the functional points mentioned by the user, the core descriptive phrases of the problems that occurred, and keywords of usage scenarios, and can save and output them in the form of a list of factual fragments to obtain the text-dimensional analysis results.

[0152] S303. Call the multi-dimensional analysis agent to analyze the user feedback data from the emotional dimension, identify the emotional category and intensity of the user's emotions, and obtain the emotional dimension analysis results;

[0153] Combination Figure 4 In the sentiment analysis section 406, in step S303 of this embodiment, a large language model is called to perform sentiment analysis on the parsing results based on the text content, and the user's emotions are identified to obtain the sentiment analysis results.

[0154] For example, user emotions can be categorized into different types such as positive, neutral, and negative, and the sentiment analysis results can be numerically expressed based on the intensity of the emotion, such as using an emotion intensity score to represent the category and intensity of the emotion; finally, a vectorized representation of the sentiment features is output, and the vectorized sentiment features are used as the result of the sentiment dimension analysis.

[0155] S304. Call the multi-dimensional analysis agent to analyze the user feedback data from the topic dimension, identify multiple topic tags related to the user feedback data, and obtain the topic dimension analysis results;

[0156] Combination Figure 4In the topic analysis section 407, in step S304 of this embodiment, topic modeling and clustering are performed on different content based on the text content to automatically identify the main hot topics that users are recently concerned about; for example, certain product defect issues, service issues, experience issues, etc. Corresponding tags are constructed based on the obtained main topics to obtain a set of topic tags, which serves as the result of topic dimension analysis.

[0157] S305. Obtain historical user feedback data corresponding to user feedback data, and statistically analyze the frequency and time change information of similar feedback data based on historical user feedback data. Call the multi-dimensional analysis agent to perform trend analysis on the frequency and time change information to obtain the trend analysis results of user feedback data.

[0158] In a preferred embodiment, the feedback analysis results also include trend characteristics. Figure 4 In the predictive analysis section 408, this embodiment, when calling a large language model to perform AI multi-dimensional analysis and mining on the user feedback data and generate feedback analysis results, further includes: obtaining historical user feedback data corresponding to the user feedback data to generate historical time-series data; statistically analyzing the frequency and time change information of similar feedback data based on the historical time-series data; and generating trend characteristics of the user feedback data based on the frequency and time change information to achieve predictive analysis. Specifically, the similar feedback data corresponding to the user feedback data refers to historical user feedback data whose similarity to the user feedback data in the text, sentiment, and topic dimensions is higher than the corresponding preset similarity thresholds. The corresponding preset similarity thresholds include preset similarity thresholds for the text, sentiment, and topic dimensions.

[0159] In this embodiment, a multi-dimensional analysis agent can be invoked to analyze historical time-series data, obtain the frequency and time trends of similar problems, and thus obtain trend characteristics to predict and analyze future public opinion trends, such as determining whether there will be potential risks.

[0160] S306. Generate feedback analysis results based on text dimension analysis results, sentiment dimension analysis results, topic dimension analysis results, and trend analysis results;

[0161] In this step, the obtained text dimension analysis results, sentiment dimension analysis results, topic dimension analysis results, and trend analysis results are concatenated to obtain the final feedback analysis results. Specifically, the various feedback analysis results are structured and transformed, including a list of factual fragments, sentiment feature vectors, a set of topic tags, and trend features. These analysis results will serve as the raw feature input for subsequent user profile generation.

[0162] In this embodiment, steps S302-S306 integrate information from multiple aspects, including NLP text analysis, multi-granularity sentiment quantification, topic content clustering, and time series prediction models. This enables automated and high-precision identification of user sentiment polarity and trending topics. In particular, with the help of the predictive analysis module, the system can identify potential quality defects or brand risks before the outbreak of public opinion, transforming "passive handling" into "proactive early warning." Compared with simple screening and keyword matching, it can achieve sentiment and trend early warning based on deep learning, improving the timeliness of decision-making.

[0163] In a preferred embodiment, this embodiment further includes evaluating whether the feedback analysis results of each dimension meet the analysis requirement threshold; if the analysis requirement threshold is not met (e.g., due to missing original data fields preventing in-depth analysis), a feedback loop is triggered to generate data collection and tracking requirements; the system automatically sends a "add business tracking" instruction to the data collection module or the front-end APP development team to optimize the subsequent private domain data collection dimensions and generate tracking requirement tasks based on the collection and tracking requirements. This closed-loop mechanism solves the long-standing pain point of data collection being disconnected from business needs, enabling the method in this embodiment to have the ability to continuously evolve and ensuring that the quality of the data source is dynamically optimized as business needs change.

[0164] S307. Construct and retrieve the expert role prompt information template, input the feedback analysis results into the expert role prompt information template to fill in the context content, and obtain the expert role prompt information;

[0165] In this embodiment, while obtaining user feedback information, basic user information is also obtained, including user ID (which can be anonymized or desensitized if sensitive content needs to be protected), user device information, user account information, or historical feedback records. In this step, the analysis results obtained in the above steps are first organized into a profile input feature set, which includes frequently occurring question types, typical usage scenarios, emotional changes, and feedback frequency.

[0166] Specifically, step S307 is executed by the profile building module deployed on the analytics server.

[0167] To standardize the user profile generation process, expert role prompt information templates for various scenarios are pre-constructed. These templates include role constraint information templates to define the expert role and task objectives of the intelligent agent, reasoning constraint information templates to define the user profile generation order or reasoning path, and output structure constraint information templates to define the user profile output fields and their dependencies. However, they do not include contextual content related to specific user feedback information. When constructing expert role prompt information, the templates are retrieved, and the contextual content of the feedback analysis results is input to fill in the contextual content, thus obtaining expert role prompt information corresponding to the feedback analysis results collected in this embodiment. The expert role prompt information in this step can be implemented using the Prompt (prompt word) of a large language model.

[0168] For example, when the actual need is to build an expert role of "C-end product - user profile expert" using a large language model, the corresponding expert role prompt information will be generated.

[0169] S308. Call the intelligent profile to generate an intelligent agent and construct a structured user intelligent profile based on expert role prompts;

[0170] In this embodiment, S308 is executed by the profile building module deployed on the analysis server.

[0171] After constructing the expert role prompts, the intelligent profile generation agent can be invoked to receive the profile input feature set based on the analysis results, and generate a structured user intelligent profile according to the preset profile generation format based on the expert role prompts. Each profile field corresponds to one type of analysis result mentioned above; the structured user intelligent profile includes user behavior patterns, core user needs, main pain points, and emotional attitudes; the preset profile generation format can be Markdown (a lightweight markup language format).

[0172] In one specific implementation, the structured user intelligent profile includes the user's basic background information and behavioral characteristics, such as age, gender, city, occupation, etc.; user stories, such as job descriptions, lifestyle habits, hobbies, etc.; user goals and needs; user behavior patterns, such as usage scenarios, frequency, platform preferences, etc.; as well as the user's core pain points, potential user needs, and the profile's emotional attitude, etc.

[0173] In this embodiment, the generated structured user intelligent profile is a pyramid structure, with the bottom layer associated with the original implementation and data, the middle layer associated with supporting arguments, and the top layer associated with the core profile conclusions and viewpoints.

[0174] This embodiment innovatively utilizes a large language model combined with specific expert role templates through the specific implementation steps of S307 and S308 to transform the analysis results into a standard Markdown-formatted user profile. This ensures a high degree of standardization and logical consistency in the user profile across dimensions such as work background, pain points, and behavioral patterns, thereby improving the interpretability and guidance value of the subsequently generated content.

[0175] S309. Invoke the business design expert intelligent agent to automatically generate product optimization requirements and task collaboration based on feedback analysis results and structured user intelligent profiles.

[0176] In this embodiment, step S309 is executed by the task generation module deployed on the task management server. (See also...) Figure 4 The process includes the analysis results report 409, product optimization requirement task 410, task requirement document design 411, and requirement proposal 412. In this step, the feedback analysis results obtained in the previous steps and the information contained in the structured user intelligent profile are obtained. Under the condition of task generation, the analysis results report is automatically generated, and the product optimization requirement task is generated according to the specific requirement content of the analysis results report.

[0177] When generating product optimization requirement tasks, at least one task decision feature is extracted from the feedback analysis results and the structured user intelligent profile. This task decision feature includes at least one of user type, risk level, problem type, and emotional attitude. Based on the task decision feature, task parameters are determined, including the actual optimization requirement, task priority, task type, and the corresponding processing department. Based on the task parameters, a product optimization requirement task associated with the user feedback data is generated. After obtaining the product optimization requirement task, task collaboration requirements are generated based on the improvement steps involved in the task, and tasks are automatically assigned to the corresponding departments. Taking a specific product optimization requirement task as an example, information flow between departments can be performed in the following order to achieve optimization and adjustment: the quality department proposes an improvement plan, the R&D department iterates the product, the sales department adjusts its strategy, the customer service department optimizes the process, and the marketing department adjusts its marketing strategy.

[0178] In one specific implementation, the system automatically calculates task priorities based on the degree of negative sentiment and audience reach, achieving automatic task classification. Furthermore, based on the problem domain (quality, R&D, sales, service, etc.), the system pushes tasks to the corresponding collaborative systems via data interfaces (such as APIs) to achieve precise task distribution. Figure 4As shown in steps 410 (Product Optimization Requirement Task), 411 (Processing Task Requirement Document Design), and 412 (Requirement Proposal), the business design expert agent can also propose optimization requirements to the corresponding processing department based on the preliminary design of the online Product Requirements Document (PRD).

[0179] Specifically, the product optimization task also includes a tracking ID. The system can monitor the task's status in real time during the R&D, testing, and deployment stages based on the tracking ID, and assess the processing progress. For example, time information can be added, and timeout reminders can be issued based on the time information. Ultimately, a long-term improvement report is generated to ensure closed-loop intelligent optimization of the product.

[0180] The following are some specific examples to illustrate the importance of structured user intelligent profiles for generating product optimization requirements in this embodiment:

[0181] (1) In one scenario, two complaints about "App crashes" were captured during the data collection step. Through the profile building step, it was identified that complainant A in the first complaint was a "high-frequency deep user / KOL profile," and complainant B was a "negative profit user / low activity profile." At this time, based on the specific user feedback content (i.e., complaint information) and the user profile information of the two different complainants, the priority information of the product optimization requirement task was calculated. The task priority of complainant A was raised to the emergency handling level, and the "usage path" feature contained in the user profile was directly pushed to the developer for positioning. It can be seen that in this scenario, the structured user intelligent profile can help determine the credibility and importance of user feedback data.

[0182] (2) In one scenario, the feedback from complainant C is "Cannot find the settings button". Through the profile building process, complainant C's profile is "over 50 years old / visually impaired / non-technical". According to the Agent model, it is inferred that his actual optimization need is to design "elderly mode" or "voice guidance", rather than a simple "enlarge button". It can be seen that in this scenario, the structured user intelligent profile can help determine the user's actual needs that are not clearly expressed, and increase the priority of the relevant tasks of designing "elderly mode" or "voice guidance".

[0183] In a preferred embodiment, a set of mapping rules or strategy models can be predefined. If the emotional attitude of a user profile is of high negative risk, the priority parameter of the task will be increased. If the profile features include pain points such as functional defects, they will be mapped to the R&D department in the task. If they include pain points in the service process, they will be mapped to the customer service department in the task. If the profile features include high-frequency group problems, the relevant information will also be mapped to the product improvement task in order to increase the priority parameter.

[0184] The method provided in this embodiment can acquire user feedback data from public and / or private data sources, enabling automated integration of multi-source data and avoiding information omissions during the acquisition process. Furthermore, the method enhances the real-time performance and accuracy of data analysis through multi-agent collaboration, identifying patterns, sentiments, and trends. It utilizes a structured user intelligent profile generation process to construct user profiles based on feedback analysis results, improving the interpretability and usability of these profiles and enabling efficient utilization of user feedback data. After generating product optimization requirements based on the feedback analysis results and structured user intelligent profiles, the method automates task allocation and tracking, achieving closed-loop cross-departmental collaboration, resolving inter-departmental collaboration gaps, promoting product improvement, reducing product risks, and facilitating optimization. This approach achieves the technical effect of comprehensively extracting multi-faceted optimization and improvement needs, enhancing the effectiveness and efficiency of product optimization.

[0185] See Figure 5 In conjunction with the foregoing first and second method embodiments, the present invention also provides a multi-agent collaborative product closed-loop optimization system based on user feedback data. The second embodiment of the multi-agent collaborative product closed-loop optimization system based on user feedback data is as follows:

[0186] The architecture of the intelligent closed-loop optimization system described in this embodiment is divided into three layers: a first perception layer 501, a first middleware layer 502, and an application layer 503. These are described in detail below:

[0187] First Perception Layer 501:

[0188] The first perception layer 501 is configured with a multi-source data acquisition server.

[0189] The multi-source data acquisition server is equipped with a public domain feedback data acquisition process. It can use RPA (Robotic Process Automation) technology to build an automated process for intelligent extraction robots or use a web crawler engine to crawl content from social media, forums, competitor apps, and news information in public domain data sources to obtain public domain feedback data.

[0190] The multi-source data acquisition server also deploys a private domain feedback data acquisition process, which uses business system interfaces or APIs (Application Programming Interfaces) to capture content from the product's APP, official website, customer service, and SMS messages in the private domain data source to obtain private domain feedback data.

[0191] After obtaining public and private domain feedback data, it can be saved to the central database of the first middle platform layer 502.

[0192] First Middle Platform Floor 502:

[0193] In this embodiment, the first middleware layer 502 is mainly configured with a Voice of the Customer (VoC) center and an analysis server.

[0194] The Customer Voice Center stores public and private feedback data in a central database. In one specific implementation, before the data is stored in the database, the public and private feedback data need to be preprocessed, including cleaning, deduplication, and desensitization.

[0195] The analysis server is equipped with an AI analysis engine, which is based on a multi-agent (intelligent agent) collaborative model. This model includes multi-dimensional analytical agents for analyzing the normalized user feedback data. Specifically, these multi-dimensional analytical agents include:

[0196] (1) NLP text analysis agent, used to obtain text dimension analysis results;

[0197] (2) A sentiment analysis agent, used to obtain the results of sentiment dimension analysis;

[0198] (3) Topic modeling or clustering agent, used to obtain the topic dimension analysis results;

[0199] (4) Trend prediction and analysis agent, used to obtain trend analysis results.

[0200] In addition, the AI ​​analysis engine also includes an intelligent profile generation agent for generating structured user intelligent profiles.

[0201] After obtaining the text dimension analysis results, sentiment dimension analysis results, topic dimension analysis results, trend analysis results, and structured user intelligent profiles, this information will be sent to the application layer 503 to execute specific task decisions.

[0202] Application layer 503:

[0203] In the application layer 503, the sufficiency of the data from the first middle platform layer 502 is first checked. When the data dimension or data information is insufficient, the tracking point requirement is generated based on the insufficient part, and the tracking point requirement task is constructed. Based on the tracking point requirement task, the tracking point information in the public domain data source and private domain data source in the first perception layer 501 is iterated in reverse to add the corresponding data tracking points.

[0204] In the application layer 503, data from the first middleware layer 502 can trigger the task decision server within the intelligent task management system to generate corresponding intelligent optimization tasks. These tasks enable collaborative projects across departments such as R&D iteration, marketing strategies, and customer service optimization, while simultaneously acquiring processing results. With sufficient data, a visual report is generated based on the processing results and user feedback data, and this report is fed back to the intelligent task management system to achieve comprehensive control over the intelligent closed-loop optimization of the product.

[0205] The multi-agent collaborative product closed-loop optimization method based on user feedback data described in the foregoing method embodiments can be implemented based on the multi-agent collaborative product closed-loop optimization system based on user feedback data described in this embodiment. For specific implementation details, please refer to the content in the foregoing method embodiments, which will not be repeated here.

[0206] The system provided in this embodiment of the invention can efficiently acquire and utilize valuable information contained in user feedback data, and combine it with the constructed structured user intelligent profile to extract comprehensive optimization and improvement needs from multiple perspectives, thereby improving the optimization and improvement effect of the product and increasing the optimization efficiency of the product.

[0207] Furthermore, the hierarchical system in this embodiment can be built in a distributed manner on the server, which decouples the core business links. When a certain link (such as a surge in public domain data) faces load pressure, the data collection server can be horizontally scaled separately without affecting the stability of analysis and task processing, which significantly improves the reliability of the system when processing massive concurrent data.

[0208] In addition to the aforementioned problems with existing technologies, related technologies also have other shortcomings in the entire processing chain, such as: 1. Passive data collection and data bias: Existing systems only passively receive data and cannot identify missing data (especially non-random missing data, MNAR), resulting in "survivorship bias," overlooking silent user experiences. Furthermore, data collection relies on hard-coded rules, making them poorly adaptable to dynamic web pages and complex interactive scenarios (such as multi-step click pop-ups), and easily detected by anti-scraping / risk control measures. 2. Limitations in data processing technology: Deduplication relies on character similarity matching, which cannot identify semantically equivalent but drastically different feedback, leading to data redundancy; anonymization relies on predefined regular expressions / entity dictionaries, which are insufficient for handling obscure privacy information and emerging expressions, easily resulting in omissions or over-anonymization; noise filtering relies on stop word lists and length filtering, which are highly indiscriminate and easily delete "hidden signal" data containing core emotions. 3. Lack of proactive completion capability: When data dimensions are missing, the system cannot "perceive" uncertainty and only forces inference, leading to biased analysis results. 4. Insufficient security and robustness: Automated instructions lack prior simulation verification, posing a high operational risk when executed directly in the production environment. 5. Purely cloud-based centralized processing requires uploading raw sensitive data, violating compliance requirements, and the massive data transmission leads to high bandwidth costs and latency.

[0209] See Figure 3 , Figure 6 as well as Figure 7 The third embodiment of the multi-agent collaborative product closed-loop optimization method based on user feedback data in this invention is as follows:

[0210] First, the executing entity in this embodiment can be a product intelligent closed-loop optimization system based on user feedback data, or it can be a terminal or a server; the specific implementation is not limited here. The method provided by this invention can acquire Voice of the Customer (VoC) or other public opinion data as user feedback data to achieve closed-loop optimization of the product. This executing entity can achieve the following during operation: Figure 6 The process shown, and Figure 6 The process shown can rely on Figure 7 The system logical architecture shown is implemented. Based on... Figure 6 Phases one through four are designed as a closed-loop OODA (Observe-Orient-Decide-Act) cycle. The third embodiment of the multi-agent collaborative product closed-loop optimization method based on user feedback data in this example is described in stages below.

[0211] Phase 1: Global Awareness and Privacy Computing (Observe)

[0212] In this embodiment, stage one can rely on... Figure 7The second perception layer 701 shown is implemented to perform federated collection of multimodal heterogeneous data. This second perception layer 701 can collect data from public and / or private data sources (i.e., it can collect data from one type of data source or multiple data sources simultaneously), specifically including edge SDK, business system probe, public / private domain intelligent extraction RPA (Robotic Process Automation) tools, browser use agent, and semantic refinement agent (SRA). The edge SDK can perform sentiment computing based on the LiteRT (TensorFlow Lite Runtime) framework; the business system probe can be used to connect to Customer Relationship Management (CRM), Enterprise Resource Planning (ERP), and customer service ticketing systems to extract structured business data in real time; the public domain intelligent extraction RPA tool can be used to intelligently capture public opinion in public and / or private domains; the browser operation agent is used to mimic human understanding to achieve content perception and understanding of complex pages; and the semantic purification agent is used to perform high-dimensional information distillation on multimodal data through a Large Language Model (LLM).

[0213] Combination Figure 6 In a preferred embodiment, stage one specifically includes:

[0214] S611. Public domain public opinion RPA technology;

[0215] Referring to S301 of the second method embodiment of the present invention, this step mainly involves capturing user feedback data from public domain data sources. This step uses a public domain intelligent extraction RPA (Robotic Process Automation) robot to capture user feedback data and public opinion information from public domain data sources, including monitoring relevant feedback information or other public opinion information related to the product to be optimized from multiple public domain data sources.

[0216] Specifically, the first step is to receive input information, which includes relevant information about the product to be optimized (such as product name, brand, model, etc.); it also includes the specific types of multiple public domain data sources to be monitored (such as social media, forums, e-commerce platforms, etc.), or specific information about the public domain data sources (such as which social media software or forums, etc.); based on this, the specific collection scope can be obtained.

[0217] After determining the collection scope, the RPA collection task is initiated, invoking an intelligent extraction robot to perform targeted crawling tasks on the specified public domain data source using pre-set intelligent crawling tools. Specifically, this may also include utilizing a browser user agent to simulate real user behavior, enabling deep interaction with complex dynamic pages through actions such as clicking, scrolling, and loading. Furthermore, it involves listening for and capturing user feedback data and public opinion information related to the product to be optimized on the target page; including but not limited to comments, discussions, ratings, and complaints.

[0218] Furthermore, the intelligent crawler tool in this step can also determine whether new public opinion data is continuously being received. If so, it continues to listen; if not, it remains in standby mode until the next round of triggering.

[0219] This embodiment can achieve targeted crawling through a preset intelligent crawler tool. When crawling information, it uses a browser use agent to imitate human behavior and perform in-depth interaction on complex pages. This changes the inefficiency of traditional crawlers storing all data and turns it into instruction-driven collection based on upper-layer requirements.

[0220] S612. Public Domain Autonomous Perception (Browser-Operated Intelligent Agent);

[0221] Referring to S301 of the second method embodiment of the present invention, this step also involves capturing user feedback data from a public domain data source. This step uses a browser use agent to obtain user feedback data and public opinion information from the public domain data source. The browser use agent includes a browser control engine based on a large language model, a visual-language model (VLM), and a multimodal parser.

[0222] In one specific implementation, the browser control agent, driven by a browser control engine based on a large language model, combines a Vision-Language Model (VLM) to perform real-time screenshot analysis and DOM (Document Object Model) tree semantic parsing on the page, enabling cross-platform autonomous form filling, multi-level linked clicks, and semantic bypassing of complex CAPTCHAs. The multimodal parser utilizes the MediaPipe framework to extract facial expressions, body language, and keyframe text from video stream data. The browser control engine can be implemented using browser automation testing tools such as Playwright or Selenium as the underlying driver.

[0223] To illustrate with a concrete example, when receiving an optimization instruction (such as optimizing competitor A's after-sales strategy in a specific region), the Browser Use Agent is activated. It autonomously plans the search path within the browser, dynamically handles page scrolling, asynchronous loading, and anti-scraping queries, and extracts deep, unstructured data. Through the work of the Browser Use Agent, even if the target website undergoes a large-scale redesign, the system can continue to operate through semantic understanding (e.g., finding elements that "look like search boxes"), significantly reducing maintenance costs.

[0224] S613. Edge-side data acquisition / inference (Mobile Edge SDK / LiteRT);

[0225] Referring to S301 of the second method embodiment of the present invention, this step mainly involves capturing user feedback data from private domain data sources. This step includes using an SDK (Software Development Kit) embedded in the client-side APP (user-side application, which can serve as a private domain data source) to run a lightweight model based on the TensorFlow Lite Runtime (LiteRT) framework to directly analyze input text or speech intonation on the user-side, obtaining sentiment vectors and intent labels. When connecting via the client-side, business system probe nodes can be used to connect to Customer Relationship Management (CRM), Enterprise Resource Planning (ERP), or customer service ticket systems to extract structured business data in real time.

[0226] In one specific implementation, this step first receives the identification information of the product to be optimized (such as product name, model, etc.), or receives a list of multiple private domain data source systems to be connected, thereby obtaining the scope of the private domain data sources. The list of private domain data source systems includes: Customer Relationship Management System (CRM), Enterprise Resource Planning System (ERP), and Customer Service Ticket System.

[0227] When the private domain data source is a user-side application (APP), an SDK (Software Development Kit) is embedded in the user-side application (APP). The SDK is a mobile edge SDK. The mobile edge SDK enables a lightweight model based on the TensorFlow Lite Runtime (LiteRT) framework; this model has offline running capabilities and can perform local computations without a network connection, thereby completing the preparation for collecting private domain data.

[0228] Once preparation is complete, user actions within the client software can be monitored, including text input (such as the content of a feedback box) and voice input (such as voice calls and voice messages); furthermore, user action logs (such as click paths and dwell time) can also be included.

[0229] After data collection is complete, the offline sentiment analysis model built into the client-side SDK is invoked to perform real-time analysis of the collected text or speech intonation, outputting sentiment vectors, intent labels, and sentiment scores; specifically including:

[0230] (1) Emotional vector: For example, a multidimensional numerical value representing the degree of positivity / negativity;

[0231] (2) Intent tags: such as category tags like "returns", "complaints", and "inquiries";

[0232] (3) Emotional rating: For example, a normalized rating between 0 and 1, used to quantify the intensity of emotion.

[0233] To illustrate with a concrete example, when a user inputs feedback or makes a voice call in the client software, the SDK integrated into the client software can call the local LiteRT model to extract sentiment scores (such as normalized scores of 0-1) and key intentions in real time, such as the intention to return the product or to file a complaint.

[0234] Specifically, the emotion vector and intent label are explained in detail below:

[0235] a. Sentiment vectors are the core data support for sentiment dimension analysis. There is no need to perform secondary sentiment calculations on the original feedback text. The quantitative statistics of sentiment tendencies (such as the proportion of positive / negative / neutral) can be completed directly based on the vectors. They are also used to help distinguish the differences in sentiment tendencies of consensus map feedback, avoid deleting high-value sentiment-differentiated feedback due to simple deduplication, and provide a reference for information entropy gain filtering. That is to say, feedback with high sentiment discrimination and high intent concentration will be judged as high information gain data and retained.

[0236] b. Intent tags assist in the classification modeling of topic dimension analysis, improving the efficiency and accuracy of topic clustering, while providing a foundation for trend analysis. That is to say, by tracking the changes in the amount of feedback with the same intent tag, the evolution trend of user needs can be judged; furthermore, intent tags can be directly used for semantic equivalence determination, quickly clustering user feedback with the same core needs, and achieving precise deduplication at the intent level;

[0237] c. In the process of building a structured user intelligent profile, the intelligent profile generation agent associates intent tags with user identifiers to depict the user's core needs and preferences (such as the function optimization intent that a certain type of user frequently reports); the emotion vector supplements the user's product experience attitude characteristics (such as the intensity of negative emotions towards a certain function). The combination of the two upgrades the user profile from "behavioral description" to a three-dimensional profile of "needs + attitudes", enhancing the guidance value of the profile.

[0238] This step also includes connecting the business system probe node to the customer relationship management system (CRM), enterprise resource planning system (ERP), and customer service ticket system; and extracting structured private domain feedback data related to the product to be optimized from each system in real time, including but not limited to user service records, ticket processing status, and order change information.

[0239] S614. Differential privacy desensitization;

[0240] Referring to S301 of the second method embodiment of the present invention, this step mainly involves the standardization of user feedback data. The privacy alignment and desensitization processing described in this step is mainly achieved through differential privacy (DP). All raw data acquired in this stage needs to be processed by differential privacy algorithm. The edge does not upload the raw data (such as text, audio, etc.), but performs differential privacy processing on the inferred information (such as "emotion feature vector" and "intent label"), and uploads the encrypted content to the central server or sends it to other intelligent agents for processing, ensuring that the personally identifiable information (PII) involved does not leave the edge.

[0241] Specifically, the system first receives the inference results data from the edge, including inference results generated by the local model such as sentiment vectors, intent labels, and normalized sentiment scores. Then, a privacy protection mechanism is initiated, employing a differential privacy (DP) algorithm to anonymize the aforementioned inference results; this differential privacy processing is applied to all raw inference output data acquired in this stage.

[0242] When performing specific differential privacy noise addition and encryption, this includes adding noise (such as Laplace noise or Gaussian noise) that meets the differential privacy requirements to the inference results at the end to mask individual contributions; and encrypting the processed data to ensure security during transmission.

[0243] After processing, the content to be differentially privacy processed and encrypted is uploaded to the central server for storage for later use, or the data is sent directly to other intelligent agents for collaborative processing.

[0244] S615. Semantic purification agent;

[0245] Referring to S301 in the second method embodiment of the present invention, this step mainly involves the standardization of user feedback data. The Semantic Refinement Agent (SRA) in this step directly adopts a text semantic understanding architecture based on a large language model. Through a pre-set inference-based cleaning protocol, it performs high-dimensional information distillation on the extracted data by intelligently understanding the semantics of the text through the large language model.

[0246] This step first receives the user feedback data processed with differential privacy in step S614, and then calls the semantic purification agent module to enter the purification process. Specifically, this semantic purification agent utilizes the logical reasoning capabilities of the large language model to perform a three-in-one processing of the semantic equivalence consensus process, the contextual PII de-identification process, and the entropy-gain based noise filtering process.

[0247] In one specific implementation, it includes:

[0248] (1) Semantic Equivalence Consensus:

[0249] First, a preprocessing algorithm is invoked to project the anonymized user feedback data from step S614 onto the high-dimensional vector space of the large language model. Then, an intent alignment algorithm is used to analyze the semantic similarity between the current feedback and other anonymized feedback data in the historical cache, in order to identify and merge semantically repetitive or highly similar user feedback data, and to determine whether the two texts logically point to the same product defect, thus achieving intent-level deduplication.

[0250] In one specific implementation, when determining semantic equivalence, a semantic summary of the current feedback can be generated autonomously. This summary is then compared semantically with historical summaries in the cache. It is determined whether semantic equivalence is detected; if so, semantic duplication is considered, semantic aggregation is performed, and the data is merged into a representative statement, achieving generative deduplication based on intent alignment. Furthermore, when analyzing the semantic similarity between the current feedback and other desensitized feedback data in the historical cache using the intent alignment algorithm, the semantic similarity is also calculated. When the semantic similarity is within a preset threshold range, it is considered highly similar.

[0251] (2) Contextual PII De-identification:

[0252] Leveraging the zero-shot reasoning capabilities of large language models, this approach understands the contextual logic structure of the entire feedback sequence, proactively identifying unstructured descriptions with privacy attributes. Based on corporate compliance policies, it automatically generates masking labels for Personally Identifiable Information (PII), ensuring that the anonymized text retains semantic coherence and contextual plausibility. By combining the context of user feedback, it accurately identifies and anonymizes personal privacy information, avoiding over-anonymization or incomplete anonymization. This information is then used for subsequent Retrieval-Augmented Generation (RAG) analysis, thus achieving a dynamic privacy masking technology based on contextual logic understanding.

[0253] To illustrate with a concrete example, when an unstructured description with privacy attributes is identified as "I live across from that famous XX Plaza, on the third floor, the one to the right," it is replaced with "the residential area near a certain plaza" according to the company's compliance policy.

[0254] (3) Entropy-Gain Based Noise Pruning:

[0255] First, the semantic purification agent initiates the Chain-of-Thought (CoT) evaluation mechanism to pre-scan the raw data and assess the information gain of each data point for knowledge graph construction. Information entropy gain is used to filter and calculate the information entropy value of each feedback data point, eliminating redundant data with excessively low entropy values ​​or no effective information gain, ultimately outputting standardized semantic data.

[0256] Specifically, based on the sentiment scores and intent tags obtained in the aforementioned steps, the type of user feedback data is initially determined. If content with "strong emotions but unclear intent tags" is identified, it is judged as "high-entropy disordered noise" that is emotional venting and has no business entity; at this time, the semantic purification agent automatically performs information compression or noise blocking to filter out such content.

[0257] In addition, this section also includes a low-entropy, high-value content enhancement process. Specifically, for seemingly brief but key logical transitions in feedback (such as "I wanted to give a good review, but it's still not fixed after the third repair"), semantic augmentation is performed using a large language model to enhance its feature expression.

[0258] After content filtering, the output is clean semantic data processed by a three-in-one approach, including: a deduplicated set of high-value feedback, contextually coherent text that has been dynamically anonymized, key cases enhanced with semantic augmentation, and structured semantic input that can be directly used for knowledge graph construction or RAG analysis.

[0259] This step utilizes the workflow of the semantic purification agent to transform noisy data into knowledge assets. Based on a large language model, high-dimensional information distillation is performed on multimodal data, preparing it for subsequent... Figure 7 The second middleware layer 702 shown provides clean semantic input.

[0260] S616. Standardized semantics;

[0261] In a preferred embodiment, referring to the content described in S301 of the second method embodiment of the present invention, the part that regulates user feedback data may further include the standardized semantic process described in this step.

[0262] Specifically, in this step, the structured data from the edge and the unstructured data from the public domain, which have been purified by the semantic purification agent, are normalized. The structured and unstructured data are converted into standard input formats suitable for knowledge graph construction, such as triple form (subject-predicate-object), semantic web compatible formats such as JSON-LD or RDF, and structured statement representations containing semantic role labeling (SRL).

[0263] Synonyms in unstructured data are normalized; for example, "black screen," "screen not lit," and "no display" are unified as "display malfunction." Intent tags are mapped to predefined categories in a standard semantic ontology library. Fields such as time, region, and product model are formatted according to unified standards (e.g., standardized time formats like ISO, administrative division codes, etc.). The final output is a normalized and semantically standardized dataset. All data is stored in a standard knowledge graph input format and can be used for subsequent knowledge graph construction or semantic reasoning tasks.

[0264] Phase Two: Cognitive Restructuring and Graph Fusion (Orient)

[0265] In this embodiment, stage two can rely on... Figure 7The second middle platform layer 702 shown in the diagram works in conjunction with the system to achieve long-term memory and reasoning, thereby analyzing the user feedback data after data normalization. This second middle platform layer 702 includes three parts: a dynamic knowledge graph, a dynamic profiling engine, and a GraphRAG inference engine. Specifically, the dynamic knowledge graph can dynamically maintain the transformation of refined semantic units into triple information; the dynamic profiling engine can index dynamic profiling based on the event subject (user / organization); and the GraphRAG (Graph-based Retrieval-Augmented Generation) inference engine can perform multi-hop reasoning to find underlying connections between seemingly unrelated events.

[0266] In the specific implementation process, in Phase Two, a multi-dimensional analysis agent is invoked to combine the dynamic knowledge graph and the multi-hop topological reasoning capability of the GraphRAG engine to perform AI multi-dimensional analysis on the user feedback data after data normalization, generating feedback analysis results. During the analysis process, information uncertainty is calculated, and when the uncertainty value exceeds the dynamic judgment threshold, active data completion is triggered. The feedback analysis results include at least: sentiment dimension analysis results, topic dimension analysis results, and trend analysis results.

[0267] Furthermore, combined Figure 6 In a preferred embodiment, stage two specifically includes:

[0268] S621. Dynamic graph construction (entity linking / triples);

[0269] Referring to S301-S306 of the second method embodiment of the present invention, this step mainly involves analyzing and understanding user feedback data and constructing a knowledge graph. First, it is necessary to obtain the normalized user feedback data and acquire multi-dimensional analysis results, and then construct a dynamic knowledge graph based on these results. The multi-dimensional analysis results include text-level analysis results, sentiment-level analysis results, topic-level analysis results, and trend analysis results. For specific acquisition methods, please refer to the second method embodiment; they will not be repeated here.

[0270] In a preferred embodiment, this step first receives input information, including: receiving user feedback data that has been purified and standardized; wherein the data sources include: edge-side structured data (such as sentiment vectors, intent tags); and public-domain unstructured data (such as comment text, speech-to-text content). Furthermore, all of the above data has been converted to a standard graph input format (such as triples, JSON-LD, etc.).

[0271] Next, the multi-dimensional analysis agent module is invoked to prepare for a data analysis task based on GraphRAG and multi-hop topological reasoning algorithms. Specifically, the following three operations are performed: entity relation extraction, ontology mapping, and community detection clustering, and a dynamic knowledge graph is constructed based on the above results.

[0272] In one specific implementation, the Dynamic KG Builder can be invoked to perform map construction and indexing. The Dynamic KG Builder executes the following sub-steps:

[0273] a. Text unit segmentation:

[0274] Long documents or large blocks of feedback content are segmented into semantically complete text chunks to facilitate subsequent fine-grained processing.

[0275] b. Entity extraction:

[0276] The large language model is used to extract entities E and relations R from each text unit. In one specific implementation: zero-shot entity-relation extraction is performed using the large language model to extract triples (entity-relation-entity, or entity-attribute-attribute value) from unstructured text.

[0277] c. Hybrid index generation:

[0278] Generate a summary for each community in the graph. And quantize to obtain the summary vector. .

[0279] After extracting entity mapping information, the extracted entities are mapped to the enterprise's standard data model through ontology mapping (e.g., based on the user's specific device information, mapping "that screen" described in the feedback information to the specific model code in the BOM table).

[0280] In a preferred embodiment, the method also includes using the Leiden algorithm to implement a community detection function, including clustering graph nodes to automatically discover potential topic communities (such as "heat dissipation problem community", "logistics delay community", etc.).

[0281] Furthermore, this step also includes an update function for the dynamic graph. The graph builder can materialize new data and insert it into the existing enterprise knowledge graph. It determines whether there is a strong association between the new entity and the existing entity. If a strong association is found between the new entity and the existing entity (such as a newly appearing fault word co-occurring with an old module), a new edge will be established.

[0282] Furthermore, this step also includes an anomaly detection function, which can identify potential public opinion outbreak points or product defects based on changes in the topology of the dynamic graph (such as a sudden spike in the degree centrality of a node).

[0283] Furthermore, the construction of the dynamic graph also includes an active learning data completion algorithm based on information entropy, introducing uncertainty measurement to reduce data bias caused by forced reasoning when user feedback data is missing. This includes: calculating uncertainty values ​​based on feedback analysis results, and attributing and completing the data analysis results when the uncertainty value exceeds a dynamic judgment threshold. In one specific implementation, it includes:

[0284] The information entropy of the output probability distribution is calculated based on the feedback analysis results, and the prediction variance is calculated through multiple forward propagations using the Monte Carlo Dropout (MC Dropout) algorithm to obtain the uncertainty value. A preset dynamic judgment threshold is obtained. When the uncertainty value exceeds the dynamic judgment threshold, it is considered that the current information is insufficient to support the decision, and an exploration task is triggered. Reasoning is paused and a completion query task is issued to the exploration agent. The exploration agent performs missing data attribution based on the context of the feedback analysis results, generates a completion query requirement, obtains the completion data, and updates the feedback analysis results based on the completion data.

[0285] Let's take a specific example to illustrate:

[0286] (1) Uncertainty calculation:

[0287] For any classification or reasoning task T, the probability distribution of the model output is P(y|x); the system calculates the Shannon entropy of this distribution as the information entropy:

[0288] ;

[0289] Simultaneously, by combining the Dropout Monte Carlo sampling (MC Dropout) algorithm, the prediction variance is calculated through multiple forward propagations, serving as a second uncertainty indicator;

[0290] (2) Active trigger threshold determination:

[0291] Set dynamic threshold τ ,when At this time, the system determines that the current information is insufficient to support the decision. For example, the user reports the fault phenomenon, but the device model and operating system version are missing, resulting in a very high degree of uncertainty in the root cause analysis.

[0292] (3) Scout Mission Generation:

[0293] at this time, Figure 7 The Orchestrator Agent in the decision orchestration layer 703 shown in the diagram pauses the current inference and issues a task to the Scout Agent. The Scout Agent generates a "complete query" based on the context, for example:

[0294] (3-1) Strategy 1 (Private Domain Reverse Inquiry): If the missing context is an individual, generate an App push or customer service message for that user: "Dear user, we have detected that you are experiencing a problem. What is your system version?" (This is an automatic generation of reverse data tracking requirements).

[0295] (3-2) Strategy 2 (Public Domain Proactive Search): If the missing information is common knowledge (such as "Does a certain system version have Bluetooth bugs?"), the Scout Agent calls a search engine or crawler to retrieve relevant information in technical forums.

[0296] In a preferred embodiment, based on the above strategy and generating completion query requirements, after obtaining completion data, the system further includes incremental learning and graph update steps. The newly obtained data is cleaned and updated to the knowledge graph in real time. The system then re-executes inference until the entropy value is reached. If the value drops below the threshold, update the feedback analysis results.

[0297] In this step, by introducing an active learning mechanism based on information entropy, the system in this embodiment acquires "metacognitive" capabilities, that is, the ability to "know what it doesn't know." When faced with fuzzy data, the system proactively schedules web crawlers or triggers reverse data tracking to obtain key missing information, which fundamentally solves the decision-making bias problem caused by non-random missing data (MNAR).

[0298] S622. Dynamic profiling and cognitive entropy calculation;

[0299] In this step, a deep user profile can be created without infringing on privacy through distributed dynamic profiling via LLM, using a cloud-edge collaborative approach; for example, the dynamic profile (persona) can be indexed based on the event subject (user / organization).

[0300] In this step, when obtaining the dynamic profile, you can refer to the content described in S307 and S308 of the second method embodiment of the present invention to construct and retrieve the expert role prompt information template, input the feedback analysis results into the expert role prompt information template to fill in the context content, and obtain the expert role prompt information; and construct a structured user intelligent profile based on the expert prompt information. The specific method will not be described in detail here.

[0301] In this specific implementation, user feedback data from the preceding steps is received, the dynamic profile extraction process is initiated, the dynamic profile engine is invoked, and user attribute inference based on semantic understanding is performed. Specifically, this includes: performing deep semantic analysis on the input text using a large language model to extract and infer the user's multi-dimensional profile attributes, thereby performing attribute mapping. For example, from the text "This is the third time this month I've encountered the login loop," the large language model infers the user's profile attributes as: "Loyalty: High," "Tolerance: Extremely Low," and "Technical Proficiency: Medium."

[0302] When generating a dynamic profile, the first step is to output a structured preliminary dynamic profile, which includes the following fields: loyalty level, tolerance level, technical proficiency, and derived indicators such as operation frequency and number of historical issues.

[0303] Furthermore, this step includes calculating cognitive entropy to assess the subject's current state uncertainty. Based on the user's recent interaction sequence, the cognitive entropy of their current state is calculated to quantify cognitive uncertainty; specific calculation criteria include: the continuity of failed operations, changes in emotional intensity, and the frequency of repeated intentions. The user is then tagged according to the value of the cognitive entropy and associated with their real-time conversation or event stream for subsequent personalized responses. For example, three consecutive failed operations result in an increased entropy value, indicating extreme anxiety for the user.

[0304] In a preferred embodiment, stage two further includes a group profile clustering and anomaly detection step: the dynamic profile engine in S622 uses community detection algorithms (e.g., Leiden Algorithm) to search for user clusters with "profile similarity" in the graph. Clustering dimensions may include: region, age group, gender, usage habits, and current cognitive entropy level. In the anomaly detection step, by monitoring the changes in cognitive entropy values ​​of each cluster in real time, anomalies in the state of specific groups are detected, and it is determined whether there is a sudden surge in the cognitive entropy value of a specific profile group. If so, an anomaly may have occurred. If an anomaly occurs in a group, it is automatically identified as a "high-priority group event," triggering the product closed-loop optimization cycle described in this embodiment, and generating product optimization requirements and task collaboration for this group event as soon as possible.

[0305] S623.GraphRAG and multi-hop topological reasoning;

[0306] In this step, a multi-dimensional analysis agent is invoked based on GraphRAG and multi-hop topological reasoning algorithms to perform vector retrieval and graph traversal to mine hidden connection paths and generate feedback analysis results. The feedback analysis results include at least: text dimension analysis results, sentiment dimension analysis results, topic dimension analysis results, and trend analysis results.

[0307] Unlike traditional vector retrieval, GraphRAG, as described in this step, combines vector search with graph traversal algorithms. When faced with complex queries, it not only retrieves similar text but also performs multi-hop queries on the graph, traversing multiple logical hops to find implicit relationships. This enhances contextual reasoning based on a dynamic knowledge graph, and the reasoning results are then sent to... Figure 7 Subsequent processing is carried out in the decision orchestration layer 703 shown.

[0308] For example, an implicit connection can be derived through reasoning: "Supplier B changed the electrolyte formula in a certain batch, causing voltage instability at low temperatures, which in turn led to user complaints." This connects seemingly unrelated complaints in region A with changes in supplier B's raw materials. Alternatively, when a surge in complaints is observed in a certain region, a topological relationship can be established between these complaints and negative videos posted by a social media KOL.

[0309] In a preferred embodiment, the GraphRAG inference engine, when combining vector retrieval and graph traversal algorithms for retrieval and attribution, specifically includes a retrieval process and the generation of attribution paths, wherein:

[0310] In the retrieval process, upon receiving a user query Q, the query Q is first represented as a vector, and then local and global searches are performed. Local search uses the vector of Q to match entity nodes, traversing its neighboring nodes to obtain micro-level details; global search uses the vector of Q to match community summary vectors. To obtain macro trends, such as "the main pain points are mainly concentrated in the battery life community".

[0311] In generating attribution paths, graph algorithms (such as Shortest Path or Steiner Tree) are used to find the optimal path from the "fault phenomenon node" to the "system component node" or "organizational structure node" in the graph. For example, an optimal path might be: User Feedback("Unable to charge"), Concept("Charging failed"), Component("BMS module"), Vendor("Supplier A"), Batch("2024-Q1"). This path directly reveals potential supply chain issues, rather than simply focusing on "user complaints about charging."

[0312] In this embodiment, the combined operation of the dynamic knowledge graph, user profile graph, and GraphRAG inference engine built in Phase 2 can quickly grasp information such as sudden public opinion events or serious defects and bugs, automatically triggering the perception, inference, and order / early warning processes, thereby improving the efficiency of the product optimization process.

[0313] Phase Three: Multi-Agent Cooperative Decision Making (Decide)

[0314] See Figure 7 In this embodiment, stage three is... Figure 7 The decision orchestration layer 703 shown in the diagram works in conjunction with the orchestration agent within it to drive a Mixture-of-Agents (MOA) architecture. This MOA architecture, based on debate and review mechanisms, simulates the working mode of a human expert team to automatically generate product optimization requirements and construct task collaboration. Specifically, the orchestration agent is invoked to decompose the requirements corresponding to the AI's multi-dimensional analysis, coordinating the domain expert agent group to conduct collaborative analysis. This domain expert agent group includes at least specialized expert agents for sentiment analysis, topic analysis, and trend analysis. Each specialized expert agent collaborates on analysis using a dynamic knowledge graph and the multi-hop topological reasoning capabilities of the GraphRAG engine.

[0315] After the hybrid expert architecture outputs instructions, pre-execution judgments can also be performed based on the isolated execution sandbox.

[0316] Refer to steps S302-S306 in the second method embodiment of the present invention. The Mixture-of-Agents (MOA) architecture described in this step can realize some functions of multi-dimensional analysis agents, and realize the analysis of various contents contained in user feedback data from multiple links and perspectives.

[0317] S631. Task breakdown and arrangement;

[0318] In this step, the Orchestrator Agent is invoked, acting as the central control brain of the decision orchestration layer 703, to prepare for task scheduling and decomposition. The system checks whether trigger conditions are met: if an abnormal signal is detected (such as a critical defect alarm, a public opinion crisis, or the identification of a supply chain risk path), task generation is triggered; or if the scheduled closed-loop optimization time point is reached according to a preset cycle, a routine analysis task is triggered.

[0319] In this step, the Orchestrator Agent is invoked to receive the inference results provided in Phase Two, breaking down the overall task into multiple expert tasks, including but not limited to the following:

[0320] Root Cause Analysis: Delving into the root causes at the technical or process level;

[0321] Script Generation: Generates appropriate communication text for customer service or user outreach scenarios;

[0322] Content Topic Analysis: Identify the core issues and functional modules that are frequently discussed in user feedback (such as "login failure" and "payment lag"), and categorize them into preset product dimensions to form a traceable topic distribution view;

[0323] Sentiment Analysis: Determines the emotional tendency (e.g., positive, negative, neutral) in user feedback, quantifies the intensity of emotions, and identifies extreme emotional expressions (e.g., anger, disappointment) in conjunction with context for prioritization.

[0324] Text Analysis: Performing linguistic processing on unstructured text, such as word frequency statistics, keyword extraction, and sentence structure analysis, to uncover high-frequency expression patterns and potential semantic clues;

[0325] Predictive Analysis: Based on historical trends and current signals, it extrapolates the likelihood of a problem spreading (such as the growth curve of complaint volume and the scope of affected users), and predicts the risk level or optimization benefits in the future.

[0326] Risk Control Assessment: Assess the compliance, security, or brand risks that this issue may pose.

[0327] S632. Hybrid Expert Reasoning (MoA);

[0328] The Mixture of Agents (MoA) used in this step is mainly implemented through a swarm of domain expert agents. This swarm consists of multiple specialized expert agents that can work through a hierarchical or iterative collaborative mechanism. Each expert is called upon to delve into their respective domains, and a multi-round debate mechanism is used to combat the hallucination of the large model, ensuring the rigor of the output strategy and thus jointly completing the complex task.

[0329] In one specific implementation, step S632 includes the following: First, an expert task instruction is received from the Orchestrator Agent. This task includes contextual data, such as: inference paths generated by GraphRAG, user feedback text, dynamic knowledge graph fragments, sentiment vectors, intent tags, and intermediate analysis results such as cognitive entropy values. Next, the hybrid expert inference architecture is loaded, and the collaborative work of expert agents from various domains is activated.

[0330] In addition to subject-dimensional analysis experts and trend analysis experts, the hybrid expert reasoning unit also includes at least the following four types of core expert agents:

[0331] (1) Sentiment Agent: Focuses on fine-grained emotion mining, performs high-precision emotion recognition, and distinguishes complex emotion types such as anger, disappointment, and anxiety.

[0332] (2) Root Cause Diagnosis Expert Agent (RCA Agent): Focuses on combining knowledge graphs for fault tree analysis, supporting drill-down from phenomena to components, batches, and suppliers.

[0333] (3) Scout Agent: Responsible for executing an active learning process when information is insufficient. It can access external systems or user touchpoints to obtain supplementary data.

[0334] (4) Compliance Agent: reviews the compliance of input and output content in real time and performs red-line scanning for privacy protection requirements such as GDPR and PII.

[0335] In the specific collaborative decision-making process, the Root Cause Diagnosis (RCA) Agent is invoked to generate a fault tree by querying a dynamic knowledge graph to construct a causal chain from user feedback to potential technical causes; and outputs a structured fault tree, providing a technical root cause judgment. The Scout Agent performs an information integrity assessment based on information entropy to determine if there are missing contents in the current feedback; if the information is deemed insufficient, it triggers an active completion function, initiating a query request to external data sources (such as retrieving logs or sending questionnaires). The Sentiment Agent performs sentiment intensity analysis, combining user profile entropy values ​​and historical behavior sequences to analyze the sentiment intensity level of the current feedback (e.g., "extremely negative," "mild dissatisfaction"). The Compliance Agent performs a compliance scan, conducting real-time review of all input text and upcoming conclusions; and performs a compliance red line scan to ensure that no sensitive words, un-anonymized PII, or content violating regional regulations are involved.

[0336] Each expert agent outputs its analysis results, introducing a multi-round debate mechanism. Based on a pre-set multi-round debate protocol, this mechanism promotes interactive verification among experts. Multiple domain expert agents engage in multiple rounds of questioning and response. For example, a root cause diagnosis expert might question the sentiment analysis expert's conclusion: "This problem is a clear hardware defect; why is it judged as high emotional fluctuation?" After multiple rounds of interaction, each expert adjusts their judgment based on rebuttal evidence. The system employs weighted voting or confidence fusion strategies to generate a final, consistent comprehensive conclusion. Converging to a more accurate conclusion through multiple rounds of interaction effectively suppresses hallucinations in large models.

[0337] S633. Sandbox environment;

[0338] In one specific implementation, after obtaining the strategy based on the above-mentioned Orchestrator Agent, Specialist Agents, and Debate & Review mechanism, it is also necessary to perform verification.

[0339] In this step, a virtualized execution environment isolated from the production system is obtained by constructing an isolated execution sandbox. This mirrors a portion of the business system's logic and data, allowing for trial runs before the browser-based intelligent agent officially releases commands, recording execution trajectories, and predicting business impact. Specific implementation steps include:

[0340] Before the instructions are officially sent to the API gateway, the policy instructions output by hybrid expert inference are first sent to the sandbox environment. In the sandbox environment, the Browser Use Agent is invoked to simulate real business operations (such as simulating the creation of error correction work orders in the backend and simulating customer service responses), capturing DOM (Document Object Model, or DOM tree) feedback of system anomalies. Simultaneously, tracing is performed to record the operation fingerprints of each step in the sandbox environment, assessing whether these operations will lead to system crashes or logical conflicts. If a system crash or logical conflict is caused, the process returns to step S631 to re-decompose and orchestrate the tasks.

[0341] In one specific implementation, the method also includes predicting the business impact of the simulated operation and speculating on the potential consequences of each strategy, such as calculating the impact of each strategy on the Net Promoter Score (NPS).

[0342] In this step, once the execution trajectory and business impact have passed security tests and meet the audit requirements of the compliant risk control intelligent agent, the product optimization requirements and tasks will be collaboratively executed on the blockchain to ensure the security of the decision.

[0343] Phase Four: Closed-Loop Execution and Feedback (Act):

[0344] See Figure 7 In this embodiment, stage four is... Figure 7 The closed-loop execution layer 704 shown in the figure is implemented in cooperation with the automation toolchain encapsulated in the closed-loop execution layer 704, which includes an action executor, an API (Application Programming Interface) gateway, a visualization report and an impact analysis unit.

[0345] Referring to the second method embodiment of the present invention, stage four is used to implement the invocation of the business design expert intelligent agent as described in S309, and to automatically generate product optimization requirements and task collaboration schemes based on feedback analysis results and structured user intelligent profiles.

[0346] Continue reading Figure 6 The specific execution process is as follows:

[0347] S641. API Gateway / Action Executor;

[0348] After successful sandbox verification, the action executor transforms the operation fingerprints pre-executed in the sandbox into specific API call sequences. The transformation rules are based on a preset mapping table, mapping browser-level behaviors (such as "clicking the submit button") to backend interface requests (such as POST / tickets / create).

[0349] The API gateway encapsulates operation interfaces for multiple external systems, including Jira (order creation), Salesforce (lead updates), Slack (alerts), and Email.

[0350] S642.Jira / Salesforce / Slack / end side;

[0351] It receives specific API call instructions generated in phase S641, and invokes R&D management tools such as Jira to create issues, such as initiating an issue creation request in the Jira system, automatically filling in fields such as issue title, description, priority, and responsible person. Subsequently, it calls CRM systems such as Salesforce to update leads, updating customer status or tagging (e.g., "Service remediation triggered," "Entered care process") in Salesforce, while simultaneously providing contextual feedback for account managers to follow up. It uses collaboration platforms such as Slack for alerts and notifications, sending structured messages to designated Slack channels or groups, including event summaries, scope of impact, and suggested response actions. It also includes feedback channels at the contact origination level, displaying service response progress on the user side via push notifications, app pop-ups, or SMS.

[0352] S643. Closed-loop monitoring of effectiveness (NPS / complaint rate);

[0353] The action executor initiates a continuous monitoring mechanism, enabling real-time monitoring of business metrics after strategy implementation. It continuously observes the trends in the following key performance indicators (KPIs): changes in complaint volume (e.g., the number of new complaints per unit time), fluctuations in NPS scores (whether the Net Promoter Score has rebounded), and auxiliary indicators such as user activity and session interruption rate, to determine whether the core indicators have improved.

[0354] This step also includes generating a visualization report, which includes a list of specific operations performed, the time points and system feedback for each operation, KPI change curves (such as complaint trend line charts, NPS trends, etc.), attribution analysis summaries (such as "the decrease in complaints may be related to the rapid response after Slack alerts"), full-process closed-loop monitoring feedback data, product optimization requirements, and task collaboration.

[0355] In a preferred embodiment, the method further includes evaluating whether the feedback analysis results of each dimension meet the analysis requirement threshold. If the analysis requirement threshold is not met (e.g., due to missing original data fields preventing in-depth analysis), a feedback loop is triggered to generate a data collection and tracking requirement. The system automatically sends a "add business tracking" instruction to the data collection module or the front-end APP development team to optimize the subsequent private domain data collection dimensions and generate tracking requirement tasks based on the collection and tracking requirements. This closed-loop mechanism solves the long-standing pain point of data collection being disconnected from business needs, enabling the method in this embodiment to continuously evolve and ensuring that the quality of the data source is dynamically optimized as business needs change.

[0356] S644. Self-evolving cyclical reward feedback;

[0357] In one specific implementation, a self-evolving reward feedback loop is also included. Specifically, this includes: converting successful execution paths into new knowledge edges and storing them in the dynamic knowledge graph constructed in step S621 of phase two; and adjusting the cognitive tags and emotional thresholds corresponding to the customer user profile in S622 based on the user's response to the strategy.

[0358] Furthermore, it also includes feeding back the reward value of this decision to the Orchestrator Agent described in S631, so as to optimize future task orchestration logic and further improve the optimization effect.

[0359] In one specific implementation, Phase Four also generates end-to-end closed-loop monitoring feedback data. Based on the product optimization requirements and specific task collaboration solutions generated in Phase Three, and the end-to-end closed-loop monitoring feedback data, a visual report is generated. This report can be sent to management personnel for display. More importantly, the execution results (such as NPS changes and fault repair rates) are fed back to the second perception layer 701 and the dynamic profiling engine, triggering the next round of the OODA loop and realizing the system's self-evolution. The end-to-end closed-loop monitoring feedback data is used at least to guide the capture of user feedback data and to feed back into the dynamic knowledge graph and the real-time updates of the structured user intelligent profile.

[0360] Furthermore, the end-to-end closed-loop monitoring feedback data is a collection of various high-value information generated throughout the entire process, including: standardized semantic data output by the semantically purified intelligent agent, causal relationship conclusions generated by the GraphRAG engine through multi-hop reasoning, information entropy data, cognitive entropy data, feedback data generated by AI multi-dimensional analysis, new user feedback information obtained through proactive data completion, product optimization requirements and sandbox simulation results, compliance audit conclusions, changes in business indicators after implementation, and task collaborative execution trajectory data, etc. Feedback to the dynamic graph specifically includes: supplementing / correcting entities and relationships: transforming standardized semantic data into entity-relationship triples, mounting them to the existing graph, and correcting erroneous relationship links; refreshing entity attributes: synchronously updating changes in business indicators and analysis results; iterating knowledge timeliness: eliminating outdated knowledge and adding emerging requirements or competitor dynamics. The feedback to structured user intelligent profiling specifically includes: iterating demand preferences: associating user identifiers with feedback intent tags to update core needs; correcting emotional attitudes: adjusting users' emotional inclinations towards product functions based on emotional vectors and cognitive entropy data; and dynamically adjusting stratification: updating user stratification tags by combining feedback behavior and execution results, and supplementing non-sensitive features through edge-cloud collaboration to ensure privacy compliance.

[0361] During the execution of steps S643-S644, at least one task decision feature is extracted from the feedback analysis results and the structured user intelligent profile. The task decision feature includes at least one of user type, risk level, problem type, and emotional attitude.

[0362] Subsequently, task parameters are determined based on the extracted task decision features. These parameters at least cover the actual optimization needs, task priority, task type, and corresponding processing department. Based on these task parameters, product optimization requirement tasks directly related to user feedback data are generated, and then integrated to form preliminary product optimization requirements and task collaboration schemes. Subsequently, the orchestration agent coordinates a group of domain expert agents and a business design expert agent to conduct collaborative review and optimization. The optimized scheme is then verified in a simulation environment, and its business impact is evaluated.

[0363] After the solution passes compliance and security verification, the business design expert intelligent agent integrates monitoring information from all stages of the process to generate closed-loop monitoring feedback data. This closed-loop monitoring feedback data, the core content of the solution (including product optimization requirements generated based on task decision characteristics), and verification conclusions are then integrated to generate a visual report. Simultaneously, the verified solution is pushed to the execution stage for implementation. The closed-loop monitoring feedback data covers key monitoring information from at least the user feedback capture, data organization, multi-dimensional analysis, solution generation, and verification stages. This data guides the accurate capture of subsequent user feedback data and feeds back into the real-time updates of the dynamic knowledge graph and structured user intelligent profile.

[0364] The method provided in this embodiment can acquire multi-faceted user feedback data through holistic perception, and combined with dynamic knowledge graphs and dynamic profiles, it utilizes path tracing capabilities to achieve a shift from "fuzzy qualitative" to "precise attribution," thereby improving the interpretability of user feedback data analysis and reducing inference overhead. By introducing an active learning mechanism based on information entropy, when faced with fuzzy data, the system no longer guesses but actively schedules crawlers or triggers reverse tracking requirements to obtain key missing information, fundamentally solving the decision-making bias problem caused by non-random data omissions. Utilizing an OODA closed-loop architecture and a multi-agent architecture, the system can automatically trigger closed-loop optimization, greatly improving decision-making efficiency, while continuously correcting its own errors during use. Based on the introduced Browser Use Agent, it can work continuously through semantic understanding technology, maintaining stable operation even if the target website is redesigned. Sandbox technology is used for stress testing and effect evaluation to ensure the safety of decision-making. Thus, it can comprehensively extract multi-faceted optimization and improvement needs, improving the technical effects of product optimization and improvement, and enhancing optimization efficiency. Furthermore, the solution in this embodiment combines edge-side inference with federated learning, enabling the training of high-precision industry models without touching user privacy data. This addresses data compliance pain points in industries such as finance and healthcare, and offers extremely high compliance security. The use of sandbox technology for stress testing and performance evaluation further ensures the security of decision-making.

[0365] This embodiment achieves a leapfrog upgrade from "passive processing" to "proactive intelligence" through multi-dimensional technological innovation, enabling a closed-loop optimization method and system for multi-agent collaborative products based on user feedback data. Its core value lies in both technological breakthroughs and business implementation: at the data collection end, it constructs a proactive perception mechanism and semantic autonomous interaction capabilities. Through information entropy-driven proactive learning, it identifies data uncertainty, triggering targeted public domain collection or private domain reverse inquiry to eliminate survivorship bias. Simultaneously, it relies on Browser Use... Agent's visual-semantic understanding breaks through the barriers of collecting data from dynamic web pages and complex interactive scenarios, significantly improving the system's environmental adaptability. The preprocessing stage upgrades from "text cleaning" to "knowledge purification," using large-model generative reasoning to achieve intent-level deduplication, context-aware desensitization, and information entropy gain filtering, outputting standardized semantic data and completely overcoming the limitations of traditional pattern matching. At the analysis and reasoning level, it integrates GraphRAG and dynamic knowledge graphs, accurately locating the root cause of problems through multi-hop causal reasoning. Combined with a hybrid expert agent (MoA) collaborative debate mechanism, it effectively reduces the model illusion rate, achieving a deeper understanding from "phenomenon description" to "root cause diagnosis." Decision execution builds Agentic. The OODA closed-loop architecture, coupled with an isolated sandbox simulation verification mechanism, simulates before execution. This achieves minute-level automation of work order distribution and policy implementation through standardized APIs, while mitigating business operation risks and reducing maintenance costs. For privacy protection, it adopts a federated edge computing architecture with end-to-cloud collaboration. Sentiment computing and intent recognition are completed locally on the edge, transmitting only encrypted feature vectors or model gradients, ensuring that raw sensitive data does not leave the domain. This perfectly balances compliance requirements with the need for collaborative learning across the entire data domain. Simultaneously, the system possesses self-evolution capabilities, feeding back execution results to the knowledge graph and decision-making model through closed-loop feedback. Combined with dynamic user profiles, it enables differentiated and precise operations, significantly improving attribution diagnosis accuracy and business response efficiency. This provides an interpretable, highly secure, and highly adaptable intelligent solution for closed-loop optimization of products based on user feedback data.

[0366] Based on the same inventive concept, embodiments of this specification also provide an electronic device for multi-agent collaborative product closed-loop optimization based on user feedback data. The multi-agent collaborative product closed-loop optimization device based on user feedback data includes: a memory and at least one processor. The memory stores instructions. The at least one processor calls the instructions in the memory to cause the multi-agent collaborative product closed-loop optimization device based on user feedback data to perform the steps of the multi-agent collaborative product closed-loop optimization method based on user feedback data as described in the above method embodiments.

[0367] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described in this invention can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this invention can be embodied in the form of a software product, which can be stored in a computer-readable storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, or network device, etc.) to execute the method described above according to this invention. When the computer program is executed by a data processing device, it enables the computer-readable medium to implement the method described above, i.e.: as... Figure 1 , Figure 3 , Figure 4 or Figure 6 The method shown in any of the figures.

[0368] Based on this, the embodiments described herein also provide a computer-readable medium for implementing... Figure 1 , Figure 3 , Figure 4 or Figure 6 A computer program for any of the methods shown in any of the figures can be stored on one or more computer-readable media. A computer-readable medium can be a readable signal medium or a readable storage medium. A readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0369] The computer-readable storage medium may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The readable storage medium may also be any readable medium other than a readable storage medium, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0370] In addition, the present invention also provides a computer program product, including a computer program / instruction that, when executed by a processor, implements the multi-agent collaborative product closed-loop optimization method based on user feedback data as described in any of the above embodiments.

[0371] In summary, the present invention can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that in practice, general-purpose data processing devices such as microprocessors or digital signal processors (DSPs) can be used to implement some or all of the functions of some or all of the components according to the embodiments of the present invention. The present invention can also be implemented as a device or apparatus program (e.g., a computer program and computer program product) for performing part or all of the methods described herein. Such programs implementing the present invention can be stored on a computer-readable medium or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.

[0372] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0373] If the technical solution of this application involves personal information, the product using this technical solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product using this technical solution has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of "express consent". For example, at personal information collection devices such as cameras, clear and prominent signs are set up to inform users that they have entered the scope of personal information collection and that personal information will be collected. If an individual voluntarily enters the collection scope, it is deemed that they have agreed to the collection of their personal information; or on the personal information processing device, with clear signs / information informing users of the personal information processing rules, authorization is obtained from the individual through pop-up information or by asking the individual to upload their personal information; wherein, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

[0374] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A multi-agent collaborative product closed-loop optimization system based on user feedback data, characterized in that, include: The data collection module is used to capture user feedback data in real time from public and / or private data sources; Based on the semantic purification intelligent agent, the user feedback data is normalized and standardized semantic data is output. The data analysis module is used to call the orchestration agent for overall scheduling, decompose the AI ​​multi-dimensional analysis requirements of the standardized semantic data received by the multi-dimensional analysis agent, coordinate the domain expert agent group included in the multi-dimensional analysis agent, transform the standardized semantic data into triples and attach them to the dynamic knowledge graph to complete semantic mapping, obtain micro-details by matching entity neighbors locally in GraphRAG and capture macro-trends by matching topic communities globally, and carry out multi-hop topological reasoning to mine deep causal relationships by combining graph algorithms, and complete the analysis of the corresponding dimensions to realize AI multi-dimensional analysis and generate feedback analysis results. During the analysis process, the uncertainty of the information is calculated. When the uncertainty value exceeds the dynamic judgment threshold, active data completion is triggered. The domain expert intelligent agent group includes at least specialized expert intelligent agents for sentiment dimension analysis, topic dimension analysis, and trend analysis. The feedback analysis results include at least the sentiment dimension analysis results, topic dimension analysis results, and trend analysis results. The profile building module is used to construct and retrieve expert role prompt information templates, input the feedback analysis results into the expert role prompt information templates to fill in context content, and obtain expert role prompt information; call the intelligent profile generation agent to construct a profile input feature set based on the feedback analysis results; receive the profile input feature set, and construct a structured user intelligent profile according to the expert role prompt information in a preset Markdown structured format; the structured user intelligent profile is a pyramid structure with the bottom layer associated with raw data, the middle layer associated with supporting arguments, and the top layer associated with core profile conclusions and viewpoints; The task generation module is used to extract at least one task decision feature from the feedback analysis results and the structured user intelligent profile; invoke a business design expert intelligent agent to determine task parameters based on the task decision feature, and generate product optimization requirement tasks and task collaboration schemes associated with the user feedback data based on the task parameters; and then generate a visual report including full-process closed-loop monitoring feedback data, product optimization requirement tasks, and task collaboration schemes based on the feedback analysis results and the structured user intelligent profile; wherein, the full-process closed-loop monitoring feedback data is used at least to guide the capture of the user feedback data and to feed back into the dynamic knowledge graph and the real-time update of the structured user intelligent profile.

2. The multi-agent collaborative product closed-loop optimization system based on user feedback data according to claim 1, characterized in that, The user feedback data includes public domain feedback data and / or private domain feedback data; The data collection module is specifically used for: Through data interfaces and / or the client side, one or more private domain data source systems of the reception optimization product are connected; wherein, when connecting through the client side, structured private domain feedback data is extracted through business system probes, and offline sentiment computing models are called to obtain sentiment vectors and intent tags; The system invokes an intelligent extraction robot and / or a browser-based intelligent agent to monitor relevant information about the product to be optimized from one or more public domain data sources. When the browser-based intelligent agent is invoked, it combines a visual language model to perform real-time screenshot analysis and DOM tree semantic parsing of the page, and invokes a multimodal parser to extract facial expressions, body language, and keyframe text from the video stream data to obtain real-time captured public domain feedback data.

3. The multi-agent collaborative product closed-loop optimization system based on user feedback data according to claim 1, characterized in that, The data collection module is also specifically used to perform semantic equivalence determination, context-aware desensitization, and information entropy gain filtering on the user feedback data. By identifying and merging user feedback data with semantic repetition or semantic similarity within a preset threshold range through semantic equivalence determination, intent-level deduplication is achieved. Based on context-aware desensitization technology, personal privacy information is accurately identified and desensitized by combining the context of user feedback, avoiding over-desensitization or incomplete desensitization. The information entropy value of each feedback data is calculated by filtering the information entropy gain, and redundant data with too low entropy value and / or no effective information gain are removed, and finally standardized semantic data is output.

4. The multi-agent collaborative product closed-loop optimization system based on user feedback data according to claim 1, characterized in that, The data analysis module is also specifically used for: The orchestration agent is invoked to calculate the information entropy of the output probability distribution, and the prediction variance is calculated through multiple forward propagations using the Dropout Monte Carlo sampling algorithm to obtain the uncertainty value. A preset dynamic judgment threshold is obtained. When the uncertainty value exceeds the dynamic judgment threshold, it is considered that the current information is insufficient to support the decision and an exploration task is triggered. The inference process is paused, and the orchestration agent issues a completion query task to the exploration agent. The exploration agent attributes the missing data based on the context of the feedback analysis result, generates a completion query requirement, obtains the completion data, and updates the feedback analysis result based on the completion data.

5. The multi-agent collaborative product closed-loop optimization system based on user feedback data according to claim 1, characterized in that, The task generation module is also used for: The orchestration agent is invoked to coordinate the business design expert agent and the domain expert agent group to carry out collaborative review and optimization. The generated product optimization requirements task and the task collaboration scheme are isolated and tested in a simulation sandbox to record the execution trajectory and predict the business impact. Once the execution trajectory and the business impact have passed security testing and meet the audit requirements of the compliance risk control intelligent agent, the business design expert intelligent agent integrates the monitoring information of each link in the entire process to generate full-process closed-loop monitoring feedback data. Then, the full-process closed-loop monitoring feedback data, the product optimization requirement tasks, and the task collaboration scheme are integrated to generate a visual report.

6. The multi-agent collaborative product closed-loop optimization system based on user feedback data according to claim 1, characterized in that, The data analysis module is specifically used for: The user feedback data is analyzed from an emotional dimension to identify the emotion category and intensity of the user's emotions, thus obtaining the emotional dimension analysis results; The user feedback data is analyzed from a topic dimension to identify multiple topic tags related to the user feedback data, and the topic dimension analysis results are obtained. Historical user feedback data corresponding to the user feedback data is obtained, and the frequency and time change information of similar feedback data are statistically analyzed based on the historical user feedback data. The frequency and time change information are then analyzed to obtain the trend analysis results of the user feedback data. The similar feedback data refers to historical user feedback data whose similarity to the user feedback data in the emotional dimension and the topic dimension is higher than the preset similarity threshold of the corresponding dimension.

7. The multi-agent collaborative product closed-loop optimization system based on user feedback data according to claim 1, characterized in that, The expert role prompts include: Role constraint information used to define the role of the intelligent agent and the task objectives to be performed; reasoning constraint information used to define the order of user profile generation or reasoning path; and output structure constraint information used to define the output fields of the user profile and the dependencies between fields.

8. A closed-loop optimization method for multi-agent collaborative products based on user feedback data, characterized in that, include: Capture user feedback data in real time from public and / or private data sources; Based on the semantic purification intelligent agent, the user feedback data is normalized and standardized semantic data is output. The system invokes an orchestration agent for overall scheduling, decomposes the AI ​​multi-dimensional analysis requirements of the standardized semantic data received by the multi-dimensional analysis agent, coordinates the domain expert agent group included in the multi-dimensional analysis agent, transforms the standardized semantic data into triples and attaches them to a dynamic knowledge graph to complete semantic mapping, uses GraphRAG's hybrid retrieval of local matching entity neighbors to obtain micro-details and global matching topic communities to capture macro-trends, and combines graph algorithms to carry out multi-hop topological reasoning to mine deep causal relationships, completes the analysis of the corresponding dimensions to achieve AI multi-dimensional analysis, and generates feedback analysis results; During the analysis process, the uncertainty of the information is calculated. When the uncertainty value exceeds the dynamic judgment threshold, active data completion is triggered. The domain expert intelligent agent group includes at least specialized expert intelligent agents for sentiment dimension analysis, topic dimension analysis, and trend analysis. The feedback analysis results include at least the sentiment dimension analysis results, topic dimension analysis results, and trend analysis results. Construct and retrieve the expert role prompt information template, input the feedback analysis results into the expert role prompt information template to fill in the context content, and obtain the expert role prompt information; The intelligent profile generation agent is invoked to construct a profile input feature set based on the feedback analysis results; the profile input feature set is received, and a structured user intelligent profile is constructed according to the expert role prompt information in a preset Markdown structured format; the structured user intelligent profile is a pyramid structure with the bottom layer associated with the original data, the middle layer associated with supporting arguments, and the top layer associated with the core profile conclusions and viewpoints; At least one task decision feature is extracted from the feedback analysis results and the structured user intelligent profile; a business design expert intelligent agent is invoked to determine task parameters based on the task decision feature, and product optimization requirement tasks and task collaboration schemes associated with the user feedback data are generated based on the task parameters; then, a visualization report including full-process closed-loop monitoring feedback data, product optimization requirement tasks, and task collaboration schemes is generated based on the feedback analysis results and the structured user intelligent profile; wherein, the full-process closed-loop monitoring feedback data is used at least to guide the capture of the user feedback data and to feed back into the dynamic knowledge graph and the real-time update of the structured user intelligent profile.

9. The closed-loop optimization method for multi-agent collaborative products based on user feedback data according to claim 8, characterized in that, The user feedback data includes public domain feedback data and / or private domain feedback data; The real-time capture of user feedback data from public and / or private data sources includes: Through data interfaces and / or the client side, one or more private domain data source systems of the reception optimization product are connected; wherein, when connecting through the client side, structured private domain feedback data is extracted through business system probes, and offline sentiment computing models are called to obtain sentiment vectors and intent tags; The system invokes an intelligent extraction robot and / or a browser-based intelligent agent to monitor relevant information about the product to be optimized from one or more public domain data sources. When the browser-based intelligent agent is invoked, it combines a visual language model to perform real-time screenshot analysis and DOM tree semantic parsing of the page, and invokes a multimodal parser to extract facial expressions, body language, and keyframe text from the video stream data to obtain real-time captured public domain feedback data.

10. The multi-agent collaborative product closed-loop optimization method based on user feedback data according to claim 8, characterized in that, The process of the semantic purification agent performing data normalization on the user feedback data and outputting standardized semantic data specifically includes: The user feedback data is subjected to semantic equivalence determination, context-aware desensitization, and information entropy gain filtering. By identifying and merging user feedback data with semantic repetition or semantic similarity within a preset threshold range through semantic equivalence determination, intent-level deduplication is achieved. Based on context-aware desensitization technology, personal privacy information is accurately identified and desensitized by combining the context of user feedback, avoiding over-desensitization or incomplete desensitization. The information entropy value of each feedback data is calculated by filtering the information entropy gain, and redundant data with too low entropy value and / or no effective information gain are removed, and finally standardized semantic data is output.

11. The multi-agent collaborative product closed-loop optimization method based on user feedback data according to claim 8, characterized in that, The uncertainty of the calculated information during the analysis process specifically includes: The orchestration agent is invoked to calculate the information entropy of the output probability distribution, and the prediction variance is calculated through multiple forward propagations using the Dropout Monte Carlo sampling algorithm to obtain the uncertainty value. A preset dynamic judgment threshold is obtained. When the uncertainty value exceeds the dynamic judgment threshold, it is considered that the current information is insufficient to support the decision and an exploration task is triggered. The inference process is paused, and the orchestration agent issues a completion query task to the exploration agent. The exploration agent attributes the missing data based on the context of the feedback analysis result, generates a completion query requirement, obtains the completion data, and updates the feedback analysis result based on the completion data.

12. The multi-agent collaborative product closed-loop optimization method based on user feedback data according to claim 8, characterized in that, Before generating the visualized report that includes full-process closed-loop monitoring feedback data, product optimization requirements, and task collaboration solutions, the following steps are also included: The orchestration agent is invoked to coordinate the business design expert agent and the domain expert agent group to carry out collaborative review and optimization. The generated product optimization requirements task and the task collaboration scheme are isolated and tested in a simulation sandbox to record the execution trajectory and predict the business impact. Once the execution trajectory and the business impact have passed security testing and meet the audit requirements of the compliance risk control intelligent agent, the business design expert intelligent agent integrates the monitoring information of each link in the entire process to generate full-process closed-loop monitoring feedback data. Then, the full-process closed-loop monitoring feedback data, the product optimization requirement tasks, and the task collaboration scheme are integrated to generate a visual report.

13. The multi-agent collaborative product closed-loop optimization method based on user feedback data according to claim 8, characterized in that, The generated feedback analysis results include at least the following: The user feedback data is analyzed from an emotional dimension to identify the emotion category and intensity of the user's emotions, thus obtaining the emotional dimension analysis results; The user feedback data is analyzed from a topic dimension to identify multiple topic tags related to the user feedback data, and the topic dimension analysis results are obtained. Historical user feedback data corresponding to the user feedback data is obtained, and the frequency and time change information of similar feedback data are statistically analyzed based on the historical user feedback data. The frequency and time change information are then analyzed to obtain the trend analysis results of the user feedback data. The similar feedback data refers to historical user feedback data whose similarity to the user feedback data in the emotional dimension and the topic dimension is higher than the preset similarity threshold of the corresponding dimension.

14. The closed-loop optimization method for multi-agent collaborative products based on user feedback data according to claim 8, characterized in that, The expert role prompts include: Role constraint information used to define the role of the intelligent agent and the task objectives to be performed; reasoning constraint information used to define the order of user profile generation or reasoning path; and output structure constraint information used to define the output fields of the user profile and the dependencies between fields.

15. A closed-loop optimization device for multi-agent collaborative products based on user feedback data, characterized in that, The multi-intelligent collaborative product closed-loop optimization device based on user feedback data includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor invokes the instructions in the memory to cause the multi-agent collaborative product closed-loop optimization device based on user feedback data to perform the steps of the multi-agent collaborative product closed-loop optimization method based on user feedback data as described in any one of claims 8-14.

16. A computer-readable storage medium storing a computer program / instructions thereon, characterized in that, When the program / instruction is executed by the processor, it implements the steps of the multi-agent collaborative product closed-loop optimization method based on user feedback data as described in any one of claims 8-14.

17. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the multi-agent collaborative product closed-loop optimization method based on user feedback data as described in any one of claims 8-14.

Images