Intelligent customer opinion processing system based on multiple agents

Abstract

The invention discloses an intelligent customer opinion processing system based on multiple agents. A receiving module receives complaint information of a user and standardizes the complaint information. And the task distribution module realizes splitting of customer complaint tasks and intelligent agent adaptation for standardized customer complaint data through a multi-intelligent agent iterative learning framework. And the subtask execution module executes the received customer complaint task. And the feedback optimization module collects full-link indexes in real time, detects abnormal events by adopting a sliding time window, and optimizes the construction of a decision tree in the sub-task execution module. And the content filtering module is used for carrying out post-processing on output texts of the consultation agent and the risk identification agent by using a text filtering and optimizing framework based on an off-line large model. And the result storage module adopts a multi-level storage strategy to record a full-link operation log. According to the method, the problems of low efficiency, insufficient accuracy, weak dynamic adaptive capacity and the like of a traditional customer complaint processing method can be effectively solved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent information processing technology, specifically relating to an intelligent customer feedback processing system based on multi-agent technology, which is based on multi-agent service design and system construction. Background Technology

[0002] With the digitalization of enterprise services and the increasing awareness of consumer rights, customer complaint handling plays an increasingly important role in service quality management and user experience optimization. However, traditional complaint handling methods have many problems, mainly reflected in slow response speed, low efficiency of cross-departmental collaboration, insufficient standardization of solutions, and limited ability to identify potential risks. Especially in high-concurrency service scenarios such as e-commerce, finance, and telecommunications, traditional manual handling methods are difficult to respond quickly and accurately, leading to decreased customer satisfaction and even potentially triggering group complaints.

[0003] Currently, most customer complaint systems still rely primarily on manual processing or single models, resulting in the following key problems: 1. Low efficiency of manual processing. Slow response speed: Manual work order assignment requires hierarchical review, easily leading to backlogs in high-concurrency scenarios. Weak collaboration capabilities: Cross-departmental collaboration relies on manual communication, easily resulting in misjudgments or duplicate processing due to information asymmetry. Insufficient standardization: Human decision-making is influenced by subjective experience, making it difficult to ensure consistency in processing logic. 2. Limitations of single-model capabilities. Limited coverage: Single models (such as classification or retrieval models) can only handle specific tasks (such as work order assignment) and cannot cope with multiple types of complex complaints (such as risk identification + consultation). Poor dynamic adaptability: Fixed model parameters cannot dynamically optimize strategies based on real-time load, departmental efficiency, or newly added knowledge. Weak risk control: Lack of multi-dimensional analysis capabilities (such as entity recognition + intent detection) makes it difficult to identify potential fraud or compliance risks. 3. Challenges in high-concurrency scenarios: Traditional methods are prone to problems such as work order mismatch and response delay when traffic surges (such as e-commerce promotions), while single models may lead to a significant increase in error rate due to insufficient computing power or generalization ability.

[0004] To address these issues, Multi-Agent Systems (MAS) have gradually become an important technology in intelligent and automated task processing in recent years. Compared with traditional manual or single-model methods, MAS has significant advantages. However, when existing MAS technologies are directly applied to customer complaint scenarios, significant shortcomings remain: 1. Insufficient domain adaptability. The state space (such as task priority and resource availability) used by general MAS systems does not include key indicators for customer complaint scenarios. Furthermore, relevant domain knowledge is lacking, making it impossible to flexibly adapt to changes in the scenario. 2. Weak toolchain integration. Existing systems often use isolated functional modules (such as independent model classifiers and rule engines), lacking cross-module knowledge fusion mechanisms. 3. Rigid feedback mechanisms. Traditional optimization only adjusts the assignment strategy by comparing time consumption, lacking in-depth tracing of failure cases and system-level knowledge updates. Summary of the Invention

[0005] To address the aforementioned issues, this invention constructs an intelligent customer complaint system based on multi-agent technology. Through a three-tiered architecture of task planning, tool execution, and feedback optimization, it achieves efficient, dynamic, and accurate customer complaint handling. The system employs a configurable task decomposition engine, automatically breaking down customer complaints into sub-tasks such as work order assignment, product consultation, dispute mediation, and risk warning. Based on real-time constructed multi-dimensional feature parameters (including complaint type, urgency, departmental suitability, and risk level), it generates a dynamic priority queue, combined with a load-aware intelligent scheduling algorithm to achieve precise cross-departmental allocation, thus improving the efficiency of high-concurrency customer complaint handling. For different industry scenarios, the system supports customizable knowledge bases and toolchains. Knowledge base customization: Users can upload industry-specific product manuals, service terms, historical cases, etc., to form a structured knowledge graph. Real-time semantic matching is achieved through RAG (Retrieval-Augmented Generation) technology, improving response accuracy. Toolchain Expansion: Supports flexible integration with natural language processing models (such as complaint classification and sentiment recognition), external services (such as logistics tracking APIs and payment platform query interfaces), and custom rule engines (such as compensation calculation and work order priority strategies) to meet the personalized needs of enterprises. A dual-loop optimization mechanism is introduced: 1. Strategy Optimization: Feedback agents monitor task execution efficiency and customer satisfaction in real time, using case backtracking learning to dynamically adjust assignment rules and departmental adaptation weights, reducing the error routing rate. 2. Knowledge Optimization: Automatically identifies knowledge base vulnerabilities based on failed cases, triggering incremental learning and policy updates to ensure continuous system evolution. All processing results are stored in a vector database supporting multi-dimensional analysis after cross-validation by multiple agents, enabling in-depth analysis functions such as complaint type clustering, risk trend prediction, and service effectiveness tracing. Ultimately, a closed-loop system from intelligent response to strategy iteration is constructed, breaking through the technical bottlenecks of traditional customer complaint systems in dynamic adaptability and cross-departmental collaboration.

[0006] A multi-agent-based intelligent customer feedback processing system includes the following modules:

[0007] Receiver module:

[0008] Users submit complaints through the intelligent customer service system. Complaints can include, but are not limited to, product quality issues, service complaints, logistics anomalies, and after-sales inquiries. The receiving module receives user complaint information through multiple terminal interfaces (such as web pages, mobile apps, and online customer service) and performs standardized preprocessing on the submitted information: 1. Text standardization: removing special symbols, correcting typos, and standardizing date formats. 2. Key information extraction: automatically identifying and structuring fields such as order number, product model, and time of problem occurrence. 3. Multimodal data processing: transcribing voice complaints and generating text descriptions for image / video complaints. Standardized customer service data is generated after preprocessing. Then, the receiving module pushes the standardized work order to the task allocation module in real time via a message queue, and simultaneously sends a work order receipt confirmation notification to the user (including the work order number and estimated processing time).

[0009] Task assignment module:

[0010] The task allocation module is based on a multi-agent iterative learning framework. For standardized customer complaint data, it achieves accurate task splitting and agent adaptation through dynamic linkage of experience exploration mechanism and Elo-based utility learning mechanism.

[0011] First, after receiving standardized customer complaint data, the module initializes a root decision node as the starting point for task allocation. This node contains basic characteristics of the customer complaint, such as complaint type, involved products, and urgency level, and generates an initial batch of candidate allocation paths based on these characteristics. Each path corresponds to a possible sub-task allocation sequence (e.g., "work order assignment → department A processing" or "consultation and answering → department B processing"). Then, the module enters an experience exploration phase, using a Monte Carlo tree search to expand the decision tree. For each existing decision node, the exploration range is controlled based on dynamic characteristics such as the sub-task type, department real-time load, and historical processing success rate, combined with a temperature parameter (T): when the temperature is high, new allocation paths that have not been tried before (e.g., cross-departmental collaborative processing) are explored first; when the temperature is low, paths with good historical performance are focused on. Simultaneously, to avoid getting trapped in local optima, the module introduces a "reject decision" mechanism—if none of the existing paths are applicable (e.g., the target department is overloaded), a completely new allocation sequence is automatically generated (e.g., temporarily scheduling a backup team) and incorporated into the decision tree as a new exploration node.

[0012] Next, the module initiates a utility learning phase, using the Elo scoring system to quantitatively evaluate the utility of each allocation path. Specifically, for the newly generated allocation sequence (τ... n ), randomly select an explored sequence (τ) i The large language model performs pairwise comparisons of the processing performance (such as assignment accuracy and estimated processing time) of the two sequences, and outputs the comparison results (τ). n Better than τ i A score is recorded as 1 for a positive result, 0 for a negative result, and 0.5 for a tie. Based on the comparison results, the scores of the two sequences are updated using the Elo scoring formula: v n ′=v n +K·(R ni -E ni ), v i ′=v i +K·(R ni -E in Among them, v n and v i Corresponding to the allocation sequence τ n and the allocation sequence τ i The rating, R represents the expected win rate (r is the Elo coefficient). ni For the actual comparison results (1 represents τ) n Better, 0 represents τ i Better (0.5 represents a tie), K is the update step size. For intermediate nodes in the decision tree (not the final allocation step), their scores are obtained by weighted aggregation of the scores of their child nodes, with the weights being the selection probabilities of the child nodes (calculated based on temperature parameters). Simultaneously, the module updates the number of times (M) based on the score of each allocation path. a The temperature parameter is dynamically adjusted using the following formula: T0 represents the initial temperature. The more updates, the more stable the score, the lower the temperature, and the more focused the exploration becomes on high-scoring paths, ensuring that efficient allocation strategies are prioritized in high-concurrency scenarios.

[0013] T0 represents the initial temperature. The more updates, the more stable the score, the lower the temperature, and the more focused the exploration becomes on high-scoring paths, ensuring that efficient allocation strategies are prioritized in high-concurrency scenarios.

[0014] Finally, when the exploration reaches the preset number of iterations or the upper limit of computing resources, the module selects the sequence with the highest Elo score of the final decision node from all generated allocation sequences as the optimal allocation strategy, and accurately dispatches the subtasks to the corresponding departments or agents to complete the task allocation.

[0015] Subtask execution module:

[0016] Once a task is assigned to the corresponding agent, the agent begins executing the customer complaint task. Different types of agents will process different customer complaint tasks according to their preset functions and algorithms. Meanwhile, the agents in the subtask execution module support user customization; users can configure different agents according to their needs and usage scenarios, and deploy appropriate development tools and knowledge bases to facilitate agent scheduling and retrieval. The system will initialize three types of agents: a work order dispatch agent, a consultation agent, and a risk identification agent.

[0017] The work order dispatching agent primarily employs two mechanisms: dynamic state construction and decision reflection. Upon receiving a customer complaint, the agent first uses various tools (including pre-trained models and API interfaces) to dynamically construct a state set. This state set includes a self-state and an environment state. The self-state mainly consists of the work order description, model results, and risk perception. The environment state mainly consists of historical feedback, departmental responsibilities, and processing rules. After obtaining the state set, the agent dispatches the letter to the designated department based on its current decision-making logic. Subsequently, by receiving feedback and opinions from the department, the agent performs retrospective analysis of the dispatch results, summarizes lessons learned from failed cases, and optimizes the dispatch logic. Simultaneously, it performs state correction, checking whether the self-state and environment state are correct and whether specific environmental information has been overlooked.

[0018] Consultation Agent: In response to a user's question or inquiry, the consultation agent first extracts key fields from the question to ensure an accurate understanding of the user's needs. Next, it searches for relevant information online by calling a Web API, or uses RAG (Retrieval-Augmented Generation) to retrieve relevant documents from a knowledge base. Afterward, the agent integrates the search results with the retrieved information to generate and output a detailed analysis report.

[0019] Risk Identification Agent: For high-risk work orders involving sensitive information or potentially negative impacts, the risk identification agent conducts in-depth analysis. The agent extracts the risk subjects from the case by calling a pre-trained Named Entity Recognition (NER) model and combines it with a pre-trained NLP model (such as RoBERTa) to classify the case content into risks. After classification, the agent combines the risk subjects, risk categories, and specific case content to generate a detailed risk analysis report, including a comprehensive explanation of the causes of potential risks, their possible impacts, and early warning recommendations.

[0020] Feedback optimization module:

[0021] The feedback agent collects end-to-end metrics (such as delivery accuracy, F1 score of consultation response, and risk recall) in real time through distributed monitoring probes, and uses a sliding time window (with dynamically adjusted window size) to detect abnormal events (such as three consecutive failed delivery orders from the same department). Feedback is used to optimize the construction of the GBDT decision tree, outputting: departmental adaptation weight adjustments for the work order dispatch agent, timeliness completion instructions for the consultation knowledge base, and false positive case feature vectors for the risk model. Optimization instructions are pushed to each agent in real time via a gRPC bidirectional streaming channel.

[0022] Content filtering module:

[0023] For the output text of the consulting agent and the risk identification agent, a text filtering and optimization framework based on an offline large model is used to post-process the generated text. This method dynamically adjusts the generation path, removes sensitive or inappropriate content, and optimizes the text expression to make it suitable for public release.

[0024] Result storage module:

[0025] A multi-level storage strategy is adopted to achieve efficient management of structured and unstructured data. The storage module records full-link operation logs through data lineage tracing technology to meet enterprise-level audit and compliance requirements.

[0026] The beneficial effects of this invention are as follows: The method of this invention effectively solves the problems of low efficiency, insufficient accuracy, and weak dynamic adaptability in traditional customer complaint handling methods. The system, through task decomposition and parallel processing mechanisms, breaks down customer complaint tasks into multiple sub-tasks (such as work order dispatch, product consultation, risk warning, etc.), which are then executed independently by different intelligent agents, effectively shortening processing time and improving overall response efficiency. The intelligent customer complaint system of this invention has the following core advantages: 1. Highly efficient and accurate processing: The tool's intelligent agent integrates advanced Natural Language Processing (NLP) technology, enabling it to accurately extract key information from customer complaint work orders (such as product model, complaint type, customer sentiment, etc.), and combines this with the enterprise knowledge base for intelligent classification, automatic work order dispatch, and risk assessment, significantly reducing misjudgments and omissions. 2. Dynamic adaptive capability: The system dynamically adjusts task allocation strategies and optimizes resource scheduling based on task priority, agent load, and real-time feedback, ensuring stable and efficient response capabilities even in high-concurrency customer complaint scenarios. 3. Automation and Human Collaboration: By automating standardized customer complaint tasks (such as answering frequently asked questions and dispatching work orders), the need for human intervention is reduced, allowing customer service personnel to focus on handling complex, high-value customer complaints, significantly improving overall service quality and operational efficiency. 4. Customizable Expansion: Enterprises can configure their own knowledge bases, departmental mapping rules, and risk warning thresholds according to their business needs, flexibly adapting to the customer complaint management needs of different industries (such as e-commerce, finance, and manufacturing). Attached Figure Description

[0027] Figure 1 The flowchart of the entire system;

[0028] Figure 2 Dynamic state construction and decision reflection of intelligent agents;

[0029] Figure 3 Text filtering and optimization based on offline large models. Detailed Implementation

[0030] The present invention will be further described below with reference to the accompanying drawings:

[0031] The entire system's operation flow is shown in the attached diagram. Figure 1 As shown.

[0032] First, the receiving module receives user input and preprocesses the data to generate standard customer complaint data, which is then transmitted to the task allocation module. Next, the task allocation module analyzes the task, breaking it down into multiple sub-tasks, such as work order assignment, consultation, and risk identification, and dynamically allocates them to appropriate tool agents for execution based on task priority, complexity, and the real-time load of the tool agents.

[0033] In the subtask execution module, each tool agent executes its corresponding function upon receiving a task: the work order dispatch agent first calls various tools to dynamically construct a state set, then dispatches the work order to the designated department based on the current decision logic. Afterwards, by receiving feedback and suggestions from the department, the agent performs retrospective analysis of the dispatch results, summarizes lessons learned from failed cases, and optimizes the dispatch logic; the consultation agent extracts key information fields from user questions and generates targeted answers through RAG knowledge base retrieval or Web search API; the risk identification agent calls the NER model to extract the risk subjects in the case, uses the LLM model to determine the case risk type, and generates a detailed risk analysis report based on the case content.

[0034] While the subtask module is executing its tasks, the feedback optimization module monitors in real time, dynamically building optimization models and outputting optimization instructions by detecting abnormal events and states. After the subtask module has completed all its tasks, all outputs undergo text optimization and post-processing by the content filtering module, making the agent's outputs more suitable for the public and society.

[0035] In the final stage, the results storage module stores all results, employing a multi-level storage strategy to store both structured and unstructured data, and managing their lifecycle according to relevant regulations. The module also supports OLAP intelligent analysis.

[0036] The dynamic state construction and decision reflection process of the agent in the subtask execution module is shown in the appendix. Figure 2 As shown.

[0037] The dynamic state construction and decision reflection process of the agent in the subtask execution module achieves intelligent upgrades through a three-order mechanism of "dynamic perception - collaborative reasoning - closed-loop evolution," breaking through the limitations of static decision-making in traditional single models. Regarding the dynamic construction of the state set, unlike the fixed parameter state definition of traditional systems, the agent adopts a continuously iterative dynamic perception framework to achieve real-time updates of the state set. Among them, the self-state performs multimodal fusion and confidence evolution, which not only includes customer complaint work order information and model output results, but also introduces a real-time confidence decay mechanism. The confidence of the model results will be dynamically adjusted with changes in the environment. When the risk identification model determines that the confidence of the model results changes, the risk perception parameters will be adjusted upward in real time and the self-state set will be updated. The construction of the environmental state is mainly achieved by dynamically capturing the functional rules and historical cases of each department, breaking through the static dimension of "departmental functions + historical data". It can capture departmental processing rule updates, cross-departmental resource conflicts, etc. in real time. For example, when a department temporarily adjusts the processing scope, the intelligent agent can quickly capture and correct the dispatch logic. The tool call also realizes adaptive scheduling, dynamically selecting tool combinations based on task complexity. Complex work orders will automatically trigger multi-tool collaboration and even call the historical failure case library to avoid errors.

[0038] In terms of decision-making reflection, it is no longer limited to summarizing experiences from failed cases, but rather achieves cognitive deepening through a dual-drive approach of causal reasoning and knowledge accumulation. For failed cases, the intelligent agent constructs a decision causal graph to conduct multi-dimensional causal attribution analysis, tracing the root cause. For example, when a work order is assigned to the wrong department, it will analyze whether it is due to model misjudgment, omission of environmental perception, or conflict in decision-making logic. Through a knowledge-sharing bus, cross-agent collaborative reflection is achieved. The failed experiences of the work order assignment agent are synchronized to other agents, promoting an exponential improvement in the overall decision-making capability of the system. The reflection results also trigger dual iterations of rules and models, optimizing the assignment logic rules and updating model parameters, taking into account both data-driven and logically traceable aspects. This process upgrades the intelligent agent from a "passive execution tool" to a "cognitive subject with environmental adaptability and self-evolution capabilities," significantly improving the robustness of decision-making in complex customer complaint scenarios.

[0039] The text filtering and optimization process based on an offline large model in the content filtering module is shown in the attached figure. Figure 3 As shown.

[0040] This method dynamically adjusts the generation path, removes sensitive or inappropriate content, and optimizes the text expression to make it suitable for public release.

[0041] 1. Forward generation: Given input text X = (x1, x2, ..., x...) n For each sentence unit x iThe model generates multiple candidate sequences Y through autoregression. i:j :

[0042] Y i:j =(y i ,y i+1 ,…,y j ),j≥i

[0043] The probability distribution of the generated candidate sequences is as follows:

[0044]

[0045] Among them, y k Y represents the element at position k in the candidate sequence. i:k-1 X represents the partially generated candidate sequence. 1:i-1 This represents the segment of the input text from the beginning to position i-1. p(y k |Y i:k-1 ,X 1:i-1 ) indicates that in the already generated partial sequence Y i:k-1 and input text X 1:i-1 Under the condition of generating the current element y k The probability of the sequence is determined. In the initial stage, the sequence with the highest probability is selected as the candidate path for the current node.

[0046] 2. Self-evaluation: For the generated candidate sequence Y i:j The offline large model performs a self-assessment of its quality and generates a score s(Y). i:j The scores are generated through large-scale model analysis. The scores are based on the following two aspects:

[0047] harmless s harmless : Check whether there are sensitive or inappropriate expressions in the content;

[0048] Applicability applicability : Assess whether the text's expression conforms to public communication norms in the field of intelligent customer feedback processing.

[0049] The final scoring formula is:

[0050] s(Y i:j )=α·s harmless (Y i:j )+β·s applicability (Y i:j )

[0051] Here, α and β are weight parameters, satisfying α + β = 1.

[0052] 3. Backtracking and Regeneration: When s(Y) i:j When τ < τ, the model backtracks to the previous node.

[0053] The entire path is generated and a new candidate sequence is generated. Sensitive words w are replaced with words w′: y′ that are more suitable for public contexts. k =replace(y k (w, w′). Logical optimization and recombination are performed on incoherent or semantically conflicting sentences. The adjusted sequence is resampled and generated, and finally, candidate paths Y that meet the threshold are selected. i:j .

[0054] The results storage module employs a multi-level storage strategy to achieve efficient management of structured and unstructured data. All processing results first pass through an intelligent classification engine. Structured data (such as work order numbers, processing times, and responsible departments) is stored in a MySQL relational database, supporting complex queries. Unstructured data (such as original customer complaint texts, voice recordings, and processing logs) is archived through the MinIO object storage engine, using a distributed storage architecture to ensure high availability. The system has a built-in OLAP analysis engine that supports real-time generation of multi-dimensional business reports (such as department processing efficiency heatmaps, complaint type trend analysis, and business opportunity conversion rate statistics), providing historical references for subsequent intelligent decision-making. The storage module records end-to-end operation logs through data lineage tracing technology, meeting enterprise-level auditing and compliance requirements.

[0055] Example:

[0056] Experimental Setup: A self-built customer complaint dataset was used, covering customer complaint data from multiple sectors including e-commerce and catering. To prevent privacy leaks, the data was preprocessed before use. The experiment evaluated two main aspects: first, the accuracy of the task planning agent in different task assignments; and second, the response quality of the work order dispatch agent, consultation agent, and risk identification agent. In evaluating task assignment accuracy, a batch of test datasets was first manually labeled, and then input into the agent for testing, ultimately achieving an accuracy rate of 92.5%.

[0057] A comparative experiment was used to evaluate the quality of the agent's responses. For each agent, a large model was set up with the same initialization as the agent. The responses of the large model and the agents were compared using the same dataset. The comparative experiment included automated metric evaluation. The large model used GPT-4 and glm-4.

[0058] Table 1 shows the results of the automated evaluation:

[0059] Table 1

[0060]

[0061] The experiment compared three types of tasks: work order assignment, consultation, and risk identification. Each task was generated by an agent, glm, and gpt. The generated content was then compared using different metrics. A total of five metrics were used. BERTScore is a metric used to evaluate the quality of text generation models (such as machine translation and summarization), utilizing the contextual embeddings of a pre-trained language model (e.g., BERT) to compare the semantic similarity between predicted and reference text. BLEURT is an automated evaluation metric for assessing text quality in natural language generation tasks. It primarily measures the semantic similarity between machine-generated text and reference text. Perplexity is a commonly used evaluation metric in Natural Language Processing (NLP), primarily used to measure the quality of text generated by a language model and its fit to the data distribution. BARTScore is an automated evaluation metric based on the pre-trained language model BART (Bidirectional and Auto-Regressive Transformer), used to evaluate text quality in natural language generation tasks. NPMI (Normalized Pointwise Mutual Information) is a statistical indicator used to measure the degree of association between two events or words.

Claims

1. A multi-agent-based intelligent customer feedback processing system, characterized in that, Includes the following modules: Receiving module: Receives user complaint information through multi-terminal interfaces, performs standardized preprocessing on the submitted information, generates standardized customer complaint data, and sends a work order receipt confirmation notification to the user; Task allocation module: Used to accurately split customer complaint tasks and adapt agents for standardized customer complaint data through a multi-agent iterative learning framework; Subtask execution module: used to execute received customer complaint tasks, including work order dispatch agent, consultation agent and risk identification agent; Feedback optimization module: Collects full-link metrics in real time through distributed monitoring probes, detects abnormal events using a sliding time window, and optimizes the construction of the decision tree in the subtask execution module; Content filtering module: For the output text of the consulting agent and the risk identification agent, a text filtering and optimization framework based on an offline large model is used to post-process the generated text; Results storage module: Employs a multi-level storage strategy to record end-to-end operation logs.

2. The intelligent customer feedback processing system based on multi-agent technology according to claim 1, characterized in that, The standardization preprocessing in the receiving module includes: Text standardization: removing special characters, correcting typos, and standardizing date formats; Key information extraction: Automatically identify and structure order number, product model, and time of problem occurrence; Multimodal data processing: Transcribe voice complaints and generate text descriptions for image / video complaints.

3. The intelligent customer feedback processing system based on multi-agent technology according to claim 2, characterized in that, The task allocation module is implemented as follows: First, standardized customer complaint data is collected to initialize a root decision node as the starting point for task allocation. This node contains the basic characteristics of the customer complaints, and the first batch of candidate allocation paths are generated based on these basic characteristics. Each path corresponds to a sub-task allocation sequence. Then, the empirical exploration phase begins, using Monte Carlo tree search to expand the decision tree. For each existing decision node, the exploration range is controlled based on the dynamic characteristics of the current node and the temperature parameter T. At the same time, a rejection decision mechanism is introduced. If none of the existing paths are applicable, a completely new allocation sequence is automatically generated and incorporated into the decision tree as a new exploration node. Next, the task allocation module initiates the utility learning phase, using Elo scoring to quantitatively evaluate the utility of each allocation path: for the newly generated allocation sequence τ n A randomly selected, explored assignment sequence τi is used. The large language model performs a pairwise comparison of the processing effects of the two sequences, outputting the comparison results. Based on the comparison results, the scores of the two sequences are updated using the Elo scoring formula: v′ n =v n +K·(R ni -E ni ), v′ i =v i +K·(1-R ni -E in ), where v n and v i Corresponding to the allocation sequence τ n And the score of the assigned sequence τi; R is the expected win rate; r is the Elo coefficient; ni For the actual comparison results: 1 represents τ n "Better" is defined as follows: 0 indicates τi is better, and 0.5 indicates a draw; K is the update step size; for intermediate nodes in the decision tree, their scores are obtained by weighted aggregation of the scores of their child nodes, with the weights being the selection probabilities of the child nodes calculated based on the temperature parameter; simultaneously, the task allocation module updates the number of times M is updated based on the score of each allocation path. a Dynamically adjust temperature parameters; Finally, when the exploration reaches the preset number of iterations or the upper limit of computing resources, the sequence with the highest Elo score of the final decision node is selected from all generated allocation sequences as the optimal allocation strategy, and the sub-tasks are accurately assigned to the corresponding departments or agents to complete the task allocation.

4. The intelligent customer feedback processing system based on multi-agent technology according to claim 3, characterized in that, The basic characteristics include the type of complaint, the product involved, and the urgency level. The dynamic features include subtask type, department real-time load, and historical processing success rate.

5. The intelligent customer feedback processing system based on multi-agent technology according to claim 4, characterized in that, The work order dispatching intelligent agent is specifically implemented as follows: the work order dispatching intelligent agent includes two mechanisms: dynamic state construction and decision reflection; after receiving a customer complaint task, the work order dispatching intelligent agent first dynamically constructs a state set, which includes self-state and environmental state. The self-state consists of work order description, model results, and risk perception; the environmental state consists of historical feedback, departmental responsibilities, and processing rules; after obtaining the state set, the intelligent agent dispatches the letter to the designated department according to the current decision logic; then, by receiving the results and opinions from the department, the intelligent agent performs retrospective analysis on the dispatch results, summarizes experience from failed cases, and optimizes the dispatch logic; The consultation agent is specifically implemented as follows: In response to a user's question or consultation, the consultation agent first extracts the key fields in the question; then it uses RAG to retrieve relevant files in the knowledge base, integrates the search results with the retrieval information, generates a detailed analysis report, and outputs it. The risk identification agent is specifically implemented as follows: the agent extracts the risk subjects in the case by calling a pre-trained Named Entity Recognition (NER) model, and classifies the case content for risk by combining it with a pre-trained NLP model. After classification, the intelligent agent combines the risk subject, risk category, and specific case details to generate a risk analysis report, including a comprehensive explanation of the causes, impacts, and early warning recommendations of potential risks.

6. The intelligent customer feedback processing system based on multi-agent technology according to claim 5, characterized in that, The feedback optimization module is specifically implemented as follows: the feedback agent collects full-link indicators in real time through distributed monitoring probes, detects abnormal events using a sliding time window, and optimizes the construction of the decision tree GBDT through feedback; The content filtering module is specifically implemented by using a text filtering and optimization framework based on an offline large model to post-process the text generated by the consulting agent and the risk identification agent. By dynamically adjusting the generation path, sensitive or inappropriate content is deleted, while the text expression is optimized.

Images