A customer service interaction method and system based on multi-agent collaborative decision-making

Abstract

The application provides a customer service interaction method and system based on multi-agent collaborative decision, belonging to the technical field of agent, specifically comprising: determining the construction strategy of the front-end agent combination based on the sorting result of the analysis duration of the front-end agent and the analysis duration deviation between the front-end agent and other front-end agents; performing the construction processing of the front-end agent combination based on the construction strategy; determining the verification processing method of the front-end agent combination based on the composition data of the front-end agent of the front-end agent combination and the deviation of the composition data of the front-end agent combination and other front-end agent combinations; and determining the customer service interaction method based on the front-end agent combination based on the verification processing result and the analysis duration data, thereby improving the efficiency and reliability of the customer service interaction processing.

Description

Technical Field

[0001] This invention belongs to the field of intelligent agent technology, and in particular relates to a customer service interaction method and system based on multi-agent collaborative decision-making. Background Technology

[0002] To reduce the labor costs of customer service, the invention patent application CN202411931431.X, "An Integrated Customer Service Method and System Based on Multi-Agents," implements integrated customer service using multiple agents. This involves the collaborative division of labor among multiple agents, each focusing on its specific business processes to provide better service to customers. Furthermore, it can proactively contact customers by analyzing customer data, track customer issues, or conduct proactive marketing, effectively improving customer satisfaction. However, the above technical solution has the following technical problems: By processing different types of semantic recognition results through front-end intelligent agents and transmitting the semantic recognition results to the central intelligent agent for semantic understanding, the final intelligent agent decision-making output can be achieved. While this improves the accuracy of semantic recognition results, the parsing time of different front-end intelligent agents may vary. Therefore, if the parsing time of one front-end intelligent agent is too slow, it will inevitably lead to a longer overall parsing time. This makes it crucial to solve the technical problem of how to construct the combination of front-end intelligent agents and verify the reliability of the recognition results. In order to improve the efficiency of customer service interaction processing by using a combination of front-end intelligent agents with higher reliability in the parsing time of some front-end intelligent agents when the parsing time of some front-end intelligent agents is too long, the solution is to use a combination of front-end intelligent agents with higher reliability in the recognition results.

[0003] Therefore, there is an urgent need for a customer service interaction method and system based on multi-agent collaborative decision-making. Summary of the Invention

[0004] To achieve the objectives of this invention, the following technical solution is adopted: Specifically, this application provides a customer service interaction method based on multi-agent collaborative decision-making, which includes: S1 uses the parsing data of the intelligent agent as a basis to determine the parsing time deviation of the front-end intelligent agent in the process of recognizing user questions. Based on the parsing time deviation, when it is determined that the intelligent agent combination needs to be optimized, the construction strategy of the front-end intelligent agent combination is determined by the ranking result of the parsing time of the front-end intelligent agent among all front-end intelligent agents and the parsing time deviation between it and other front-end intelligent agents. S2 constructs a front-end intelligent agent combination based on the construction strategy. Based on the composition data of the front-end intelligent agents in the combination and the deviation of the composition data from other front-end intelligent agent combinations, the verification processing method of the front-end intelligent agent combination is determined. Based on the verification processing results and parsing time data, the customer service interaction method based on the front-end intelligent agent combination is determined.

[0005] The beneficial effects of this invention are as follows: Based on the deviation in parsing time of the front-end agents in the process of recognizing user questions, it can be determined whether the agent combination needs to be optimized. In this way, when the parsing time of some front-end agents is significantly longer than that of others, it can be determined in a timely and effective manner that only some front-end agents should be used. That is, the combination of front-end agents can accurately perform semantic recognition processing, thereby improving the efficiency and reliability of semantic recognition processing.

[0006] Based on the verification results and parsing time data, a customer service interaction method based on front-end agent combinations is determined. This method considers not only the number of front-end agent combinations with high consistency in verification results, but also the deviation in parsing time compared to all front-end agents. Therefore, when there are many front-end agent combinations with high consistency in verification results, the need to verify the semantic understanding results of the front-end agent combination using the semantic understanding results of all front-end agents is reduced. Thus, for some front-end agent combinations with significantly shorter parsing times and high consistency in verification results, when the parsing time of all front-end agents is too long, only the front-end agent combinations with high consistency in verification results are used for customer service interaction. This satisfies the user's satisfaction with the processing results and lays the foundation for further determining whether the front-end agent combination can match the user's needs.

[0007] Furthermore, the front-end intelligent agent is an intelligent agent used for identifying and processing user questions, and transmits the identification and processing results to the central intelligent agent for semantic understanding to obtain the understanding results.

[0008] Furthermore, the parsing time of the front-end intelligent agent in the process of identifying user questions is determined based on the time it takes for the front-end intelligent agent to obtain the identification and processing results during the process of identifying and processing user questions.

[0009] Furthermore, the parsing time deviation is determined based on the deviation in the time taken for different front-end intelligent agents to obtain the recognition and processing results.

[0010] Furthermore, it was determined that optimization processing of the agent composition was needed, specifically including: Based on the parsing time deviation, the difference in parsing time between the agent with the longest parsing time and the agent with the shortest parsing time is determined during the user question identification and processing process, and this difference is used as the parsing time difference. Based on the difference in parsing time during the user problem identification and processing process, identify and process the identification and processing process where the difference in parsing time is greater than a preset time threshold, and treat it as a delayed identification process. Based on the data from the delayed recognition process, it is determined whether optimization processing of the agent combination is required.

[0011] Furthermore, the method for determining the customer service interaction method based on the front-end intelligent agent combination is as follows: Based on the verification results, the interaction process that makes the understanding results of the front-end intelligent agents consistent is determined and used as the consistent interaction process. Based on the parsing time data of the front-end intelligent agent combination, the deviation between the parsing time of the front-end intelligent agent combination and the maximum parsing time of all front-end intelligent agents is determined in different interaction processes, and this deviation is used as the parsing time deviation. Based on the consistent interaction process data and the parsing time deviation, a customer service interaction method based on the front-end intelligent agent combination is determined.

[0012] In a second aspect, the present invention provides a computer system comprising: a memory and a processor connected in communication, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the aforementioned customer service interaction method based on multi-agent collaborative decision-making when running the computer program.

[0013] Other features and advantages will be set forth in the following description, and the objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0014] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0015] The above and other features and advantages of the present invention will become more apparent from a detailed description of exemplary embodiments thereof with reference to the accompanying drawings.

[0016] Figure 1 This is a flowchart of a customer service interaction method based on multi-agent collaborative decision-making; Figure 2 This is a flowchart for identifying agricultural products that do not exhibit matching bias. Figure 3This is a flowchart illustrating the method for determining the marketing association type of a marketing account for the agricultural product. Figure 4 This is a flowchart illustrating the method for determining the data processing approach used by marketing accounts to update user profiles. Detailed Implementation

[0017] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0018] Example 1 like Figure 1 As shown, this application provides a customer service interaction method based on multi-agent collaborative decision-making, specifically including: S1 uses the parsing data of the intelligent agent as a basis to determine the parsing time deviation of the front-end intelligent agent in the process of recognizing user questions. Based on the parsing time deviation, when it is determined that the intelligent agent combination needs to be optimized, the construction strategy of the front-end intelligent agent combination is determined by the ranking result of the parsing time of the front-end intelligent agent among all front-end intelligent agents and the parsing time deviation between it and other front-end intelligent agents. Furthermore, the front-end intelligent agent is an intelligent agent used for identifying and processing user questions, and transmits the identification and processing results to the central intelligent agent for semantic understanding to obtain the understanding results.

[0019] Furthermore, the parsing time of the front-end intelligent agent in the process of identifying user questions is determined based on the time it takes for the front-end intelligent agent to obtain the identification and processing results during the process of identifying and processing user questions.

[0020] Furthermore, the parsing time deviation is determined based on the deviation in the time taken for different front-end intelligent agents to obtain the recognition and processing results.

[0021] The core architecture concept is the Central LLM, which acts as a "commander," responsible for maintaining the dialogue context, understanding deep intents, perceiving user emotions, and clarifying ambiguous intents. It does not directly execute tasks but dynamically makes decisions and dispatches the most suitable "expert" agents based on cognitive results.

[0022] Functional Agents: A group of lightweight, single-function, highly cohesive specialized agents responsible for performing specific skills, interacting with business systems through APIs or RPAs, and feeding back the execution results to the central LLM for integration.

[0023] The technical chain of this solution can be divided into three major modules: cognitive and decision-making center, functional execution intelligent agent, and knowledge governance and learning loop.

[0024] Cognitive and Decision-Making Center Front-end intelligent agent: Before the central LLM processing, a front-end intelligent agent is introduced to perform preliminary analysis of user input. User intent recognition agent: This agent initially identifies the user's core intent. Specific examples could include: Multimodal Understanding Agent (Front-end Intelligent Agent): Processes multimodal information such as images and voice, such as identifying products or questions in images. Dialogue State Analysis Agent: Analyzes the current dialogue turn and historical interactions, maintaining the conversation context. Sentiment Analysis Agent: Analyzes user sentiment in real time, providing a basis for subsequent response strategies.

[0025] Multimodal understanding agent (front-end intelligent agent): processes multimodal information such as images and voice, such as identifying products or questions in images.

[0026] Dialogue State Analysis Agent (Front-end Intelligent Agent): Analyzes the current dialogue turn and historical interactions; Maintain the conversation context sentiment analysis agent (front-end intelligent agent): Analyze the user's emotional tendencies in real time to provide a basis for subsequent response strategies.

[0027] Central intelligent agent (commander): Makes core cognitive decisions based on the analysis results of the auxiliary intelligent agent and built-in memory (procedural memory, semantic memory, contextual / emotional memory).

[0028] Deep intent understanding and multi-turn interaction: Understanding complex intents and handling referential and topic-jumping issues. Proactive clarification and handling of ambiguous intents: Generating clarifying questions when intents are unclear.

[0029] Intelligent routing and scheduling: Dynamically select and combine one or more functional execution agents based on the understood intent.

[0030] Result Integration and Response Generation: Integrate the results of the executing agent into a natural language response to the user; Functional execution agents include CRM / RPA tool operation agents: interacting with systems such as CRM and ERP through RPA / API to perform back-end operations such as order modification, information query, and returns / exchanges.

[0031] Specifically, such as Figure 2 As shown, optimization processing for agent composition is required, specifically including: S11 Based on the parsing time deviation, determine the difference in parsing time between the agent with the longest parsing time and the agent with the shortest parsing time during the user problem identification and processing process, and use it as the parsing time difference. Specifically, the calculation of the parsing time difference is explained, along with a definition and example operation: Parsing time: refers to the time elapsed from when the user completes inputting the question until a front-end agent completes its specific analysis and outputs the result.

[0032] Operation (simulating the parsing time of each front-end agent in a single request): User intent recognition Agent: Processing the text "I want to return the goods...", taking 150ms.

[0033] Multimodal Understanding Agent: Processes images, performs object recognition and stain detection, taking 520ms.

[0034] Dialogue State Analysis Agent: Determines first round and urgent status, taking 100ms.

[0035] Sentiment Analysis Agent: Analyzed "anger" in 120ms.

[0036] Determine the longest and shortest parsing times: Longest: Multimodal understanding agent, 520ms.

[0037] Shortest: Dialogue State Analysis Agent, 100ms.

[0038] Calculate the difference in parsing time: longest - shortest = 520ms - 100ms = 420ms.

[0039] This is the first step in diagnosing performance bottlenecks. The difference (420ms) visually reflects the degree of speed imbalance among different "experts" when handling the same user problem. The larger the difference, the more pronounced the "weakest link" effect in the system's internal collaboration, with the overall response speed being dragged down by the slowest agent.

[0040] S12 uses the parsing time difference in the user problem identification and processing process to determine the identification and processing process where the parsing time difference is greater than the preset time threshold, and treats it as a delayed identification process. Delayed identification process: This refers to the entire process in a single user problem identification and processing where the calculated "parsing time difference" is greater than a preset threshold. Judgment: If the difference in this process is 420ms, and 420ms > T_diff (300ms), then this user problem identification and processing process is marked as a "delayed identification process".

[0041] This step is a qualitative assessment of the performance issue. The preset threshold (300ms) is an empirical "health line." Exceeding this line means that the internal latency of this processing has reached a level that requires attention and optimization. It provides a basic sample for subsequent aggregation analysis (to see if it is an occasional or normal occurrence).

[0042] S13 determines whether optimization processing of the agent combination is needed based on the delay recognition process data.

[0043] It is understandable that when the proportion of delayed recognition processes in all recognition processes is greater than the preset threshold for the proportion of delayed processes, the parsing time of different front-end agents will vary, which will lead to a longer parsing time for user problems. Therefore, by combining different front-end agents, it is determined whether the problem can be effectively solved when some front-end agents are used as input, thereby reducing the parsing time of user problems.

[0044] Specifically, the decision on whether intelligent agent combination optimization is needed, and the delayed recognition process data, refers to the set of all "delayed recognition processes" statistically recorded within a certain period of time (such as the past 1 hour).

[0045] Quantity percentage: Number of delayed identification processes / Total number of user problem identification and processing processes during this period.

[0046] Operations (simulating data from the past 30 days): Total number of processing steps: 100,000, of which 15,000 were marked as "delayed identification processes". Delayed process percentage = 0.15 (15%).

[0047] Judgment: 0.15 > P_delay (0.1). Since the proportion of the delay process (15%) exceeds the threshold (10%), it is determined that the agent combination optimization process needs to be performed.

[0048] This is a crucial step from case-by-case diagnosis to system optimization decision-making. Significance: If delays only occur occasionally (e.g., less than 10%), it may be due to accidental factors such as network jitter or excessively large images, and no architectural adjustments are necessary. Current situation: However, 15% of requests are experiencing significant delays, indicating that the "barrel effect" caused by the excessive time spent by the multimodal understanding agent in processing images has become a systemic performance bottleneck.

[0049] The optimization logic is: "By combining different front-end agents, we can determine whether the problem can be effectively solved with some of the front-end agents as input, thereby reducing the parsing time for user problems." This means that we need to explore whether, in certain scenarios, we can bypass the slowest agent or use the results of other agents to replace or assist the work of the slowest agent in order to accelerate the overall process.

[0050] Furthermore, the parsing time of the front-end intelligent agent is determined from the shortest to the longest parsing time of all front-end intelligent agents in the ranking results of all front-end intelligent agents, including the parsing time of the front-end intelligent agent in the recognition process.

[0051] Problem identified: In S13, it was determined that the proportion of "delayed recognition process" was too high (15%>10%), and agent combination optimization was required.

[0052] Continuing the analysis of four front-end agents in an e-commerce customer service system: User Intent Recognition (I, average 150ms), Multimodal Understanding (M, average 520ms), Dialogue State Analysis (D, average 100ms), and Sentiment Analysis (E, average 120ms). The goal has shifted from "alleviating the slowest bottleneck" to "leveraging the efficiency of the fastest agents to improve the overall balance and speed of the processing flow through combination." The aim is to allow other agents to receive assistance from "efficient agents" in decision-making through combination, thereby reducing reliance on slower agents or changing their working patterns.

[0053] Specifically, such as Figure 3 As shown, the method for determining the construction strategy of the front-end intelligent agent combination is as follows: S21 determines the highest ranking result of the front-end intelligent agent in the recognition process based on the parsing time of the front-end intelligent agent among all the front-end intelligent agents, and uses it as the matching recognition process. Definition: "Matching and Recognition Processing" here refers to the processes in a single user question recognition process where a specific front-end agent is the fastest (i.e., the speed champion) in parsing time. Example Operation: Analyze historical data and count the number of times each agent is the "fastest". Assume that in the past 1000 processes: the dialogue state analysis agent (D) is the fastest in 800 processes due to its simple logic; the sentiment analysis agent (E) is the fastest in 150 processes; the user intent recognition agent (I) is the fastest in 50 processes; and the multimodal understanding agent (M) has never been the fastest due to its complex processing. Therefore, for agent D, its "Matching and Recognition Processing" has occurred 800 times.

[0054] S22 determines the baseline agent among the front-end agents according to the matching and recognition process, and determines the parsing time deviation between the baseline agent and other front-end agents; The baseline agent is defined as the agent with the most "matching and recognition processing steps," meaning the agent that is most stable and frequently becomes the speed champion. Example operation: Based on the previous step, D (dialogue state analysis) becomes the baseline agent with an absolute advantage of 800 times. Next, the deviation of the baseline agent D from the parsing time of other agents in those processes where they are the slowest is calculated. This is not calculating the global average difference, but the performance difference in a specific scenario. For example, in the process where M is the slowest, D typically completes the analysis in only 100ms, while M takes 520ms, a deviation of 420ms. This deviation reveals that when the system is dragged down by slow agents, the baseline agent D can still maintain extremely high performance, highlighting its potential as a "cornerstone of efficiency."

[0055] S23 determines the construction strategy of the front-end agent combination based on the matching and recognition process data of the baseline agent and the parsing time deviation between it and other front-end agents.

[0056] It is understandable that, based on the matching and recognition processing data of the baseline agent and the parsing time deviation between it and other front-end agents, the construction strategy of the front-end agent combination is determined, specifically including: S231 Based on the proportion of the matching and identification process in all identification processes, determine the identification matching coefficient, and determine whether the identification matching coefficient of the benchmark agent is less than the preset identification matching coefficient threshold. If so, determine the construction strategy of the front-end agent combination as the first construction strategy; otherwise, proceed to the next step. Based on the data from D (the baseline agent), subsequent decision-making steps are executed to determine the combined construction strategy. Preset thresholds: matching coefficient threshold R_match = 0.6, matching coefficient preset value R_base = 0.1, agent number threshold N_agent = 2, agent number preset value N_candidate = 1, first duration threshold T1 = 50ms, second duration threshold T2 = 20ms (T1 > T2).

[0057] To determine the dominance of the baseline agent, we first calculate the recognition matching coefficient of D: 800 / 1000 = 0.8. We then check if this coefficient is less than the threshold of 0.6. Since 0.8 is not less than 0.6, the result is "no". This means that the baseline agent D has a very strong dominance and general advantage in speed. The problem is not the lack of a benchmark, but rather that the benchmark is too prominent. Therefore, we do not adopt the first construction strategy and proceed to the next step.

[0058] S232 takes the intelligent agents excluding the benchmark intelligent agent as other intelligent agents, and determines whether there are other intelligent agents whose identification matching coefficient is greater than the preset value of the matching coefficient. If not, the construction strategy of the front-end intelligent agent combination is determined to be the second construction strategy. If not, proceed to the next step. Detect the existence of significant secondary efficient agents, considering I, M, and E (excluding D) as "other agents". Determine if there are other agents with a matching coefficient greater than R_base (0.1). E's coefficient is 0.15 (>0.1), I's coefficient is 0.05 (<0.1), and M's coefficient is 0. Therefore, such an agent (E) exists. Proceed to the next step.

[0059] S233 Based on the identification matching coefficients of different other intelligent agents, determine the number of other intelligent agents whose identification matching coefficients are greater than the preset value of the matching coefficients, and determine whether the number of other intelligent agents whose identification matching coefficients are greater than the preset value of the matching coefficients is greater than the preset threshold of the number of intelligent agents. If yes, determine that the construction strategy of the front-end intelligent agent combination is the first construction strategy; otherwise, proceed to the next step. To determine the size of the efficient agent cluster, we identify that there is only one other agent, E, with a matching coefficient greater than 0.1, and its quantity is 1. We then determine whether it is greater than N_agent(2). Since 1 is not greater than 2, the result is "no". This indicates that although there is a secondary efficient agent (E), a large "efficient cluster" has not been formed. The system efficiency is highly dependent on a single star (D) and has a helper (E), but the efficiency of other members (I, M) varies greatly. The decision proceeds to the final step S234.

[0060] S234 takes the average parsing time of the front-end agent in different recognition processing processes as the average time of the agent, determines the average time difference of the front-end agent based on the difference between the average time of the front-end agent and the baseline agent, and determines the construction strategy of the front-end agent combination according to the average time difference of different front-end agents.

[0061] It is understandable that when the number of front-end agents with an average duration difference less than a preset duration threshold is greater than a preset number of agents, the construction strategy of the front-end agent combination is determined to be the first construction strategy; when the number of front-end agents with an average duration difference less than a preset duration threshold is not greater than a preset number of agents, the construction strategy of the front-end agent combination is determined to be the second construction strategy.

[0062] The final strategy is determined based on the average duration difference. First, the average duration difference between each agent and the baseline agent D is calculated: ID = 150-100=50ms, MD=520-100=420ms, ED=120-100=20ms. The system implicitly makes a judgment here: if there are many agents with similar speeds to the baseline agent (small difference) (forming a group), a more lenient and exploratory first construction strategy is adopted (using a large T1 to define the group); if the number is small, a more focused and strict second construction strategy is adopted (using a small T2 to define individuals). In the data, there are two agents (I, E) with a speed difference of less than 50ms from D, and only one agent (E) with a speed difference of less than 20ms. Considering N_candidate=1, if "number greater than 1" is used as the standard, then 2>1, triggering the first construction strategy.

[0063] It should be noted that the first construction strategy is to use front-end agents whose average duration difference is less than the first duration threshold as candidate agents, and to construct front-end agent combinations by freely combining the candidate agents based on the existence of the benchmark agent in different combinations of front-end agents.

[0064] It should be noted that the second construction strategy is to use front-end agents whose average duration difference is less than the second duration threshold as candidate agents, and to construct front-end agent combinations by freely combining the candidate agents based on the existence of the benchmark agent in different combinations of front-end agents.

[0065] The first construction strategy selects front-end agents whose average duration difference is less than a first duration threshold (T2 = 60ms) as candidate agents. Calculations show that only I and E meet the criteria. Therefore, the candidate agent pool is {I, E}. The construction strategy requires that, based on the existence of a baseline agent D in different combinations of front-end agents, the candidate agents are freely combined. Therefore, the combinations to be constructed and tested are [D, E], [D, I], and [D, I, E]. This unit may be used to quickly process certain types of user requests (such as emotionally agitated but with clear intentions), thereby attempting to bypass or simplify the dependence on slower agents I and M. The system will focus on verifying whether the three combinations [D, E], [D, I], and [D, I, E] can, in real-world interactions, make decisions by the central agent that are as correct as when using all four agents, using only information from these two fastest agents.

[0066] Specifically, the first duration threshold is greater than the second duration threshold.

[0067] S2 constructs a front-end intelligent agent combination based on the construction strategy. Based on the composition data of the front-end intelligent agents in the combination and the deviation of the composition data from other front-end intelligent agent combinations, the verification processing method of the front-end intelligent agent combination is determined. Based on the verification processing results and parsing time data, the customer service interaction method based on the front-end intelligent agent combination is determined.

[0068] This embodiment describes an e-commerce intelligent customer service system, which adopts an architecture of "central intelligent agent (commander) + functional front-end intelligent agents". The system continuously monitors the performance of four core front-end intelligent agents: User Intent Recognition Agent (I, average parsing time 150ms), Multimodal Understanding Agent (M, average 520ms), Dialogue State Analysis Agent (D, average 100ms), and Sentiment Analysis Agent (E, average 120ms). Initially, the system waits for all four agents (I, M, D, and E) to complete their analysis before summarizing the results and sending them to the central intelligent agent for decision-making. The overall response speed is severely constrained by the slowest agent, M. To address this, the system initiates an automated optimization process to build and validate an efficient combination of front-end intelligent agents, ultimately achieving dynamic and rapid customer service interaction.

[0069] Specifically, the method for determining the verification processing method of the front-end intelligent agent combination is as follows: S31 uses the composition data of the front-end intelligent agents in the front-end intelligent agent combination to determine the number of front-end intelligent agents in the combination. In the above steps, constituting data refers to the specific front-end intelligent agent members that make up a group of intelligent agents, and the number of front-end intelligent agents refers to the number of intelligent agents contained in the group.

[0070] This is the fundamental metric for validation and evaluation. The number of agents within a ensemble directly relates to its execution complexity, resource consumption, and potential uncertainties. A smaller number of agents in a ensemble is generally simpler and faster, but may provide incomplete information; a larger number of agents provides more comprehensive information, but may repeat the slow response of a "full ensemble." Determining the exact number is the primary basis for subsequent resource allocation (validation frequency) and risk grading (validation rigor).

[0071] The system reads the composition of three candidate combinations. Combination C1 [D, E] consists of two agents, D and E, with a total of 2 agents. Combination C2 [D, I] also consists of two agents, D and I, with a total of 2 agents. Combination C3 [D, I, E] consists of three agents, D, I, and E, with a total of 3 agents. These quantities will be used as key parameters input into the subsequent decision-making process.

[0072] S32 determines the ranking result of the average duration of all front-end agents in the front-end agent combination based on the deviation of the composition data between the front-end agent combination and other front-end agent combinations. Definitions: Bias in the composition of data: This refers to the differences in the composition of different agent combinations. However, the direct basis for ranking is not the differences in composition itself, but the theoretical performance indicators determined by the composition.

[0073] The average of the average time taken by all front-end agents: For a given agent combination, calculate the historical average parsing time of each agent, and then take the arithmetic mean of these averages. This value represents the average processing speed of the combination under ideal conditions.

[0074] Sorting results: Sort all candidate combinations in ascending order of the average of the above "average duration". The smaller the value, the higher the ranking, indicating a faster theoretical speed.

[0075] This involves theoretically predicting the performance of candidate combinations and prioritizing them. Before investing actual validation resources, an objective metric is needed to predict which combinations are more likely to deliver performance improvements. Calculations based on historical average durations provide a stable and comparable benchmark. By ranking the combinations, the system can identify the theoretically "optimal" combinations, thus prioritizing the validation of high-potential combinations when validation resources are limited.

[0076] Example operation: The system first calculates the "average of average duration" for each candidate combination.

[0077] For combination C1 [D, E]: D averages 100ms, E averages 120ms, and the average value is (100 + 120) / 2 = 110.0 ms.

[0078] For combination C2 [D, I]: D averages 100ms, I averages 150ms, and the average value is (100 + 150) / 2 = 125.0 ms.

[0079] For combination C3 [D, I, E]: D averages 100ms, I averages 150ms, E averages 120ms, and the average value is (100 + 150 + 120) / 3 ≈ 123.3ms.

[0080] Based on the calculation results, sorted from smallest to largest, the final ranking results are as follows: 1st place: combination C1 [D, E] (110.0 ms); 2nd place: combination C3 [D, I, E] (123.3 ms); 3rd place: combination C2 [D, I] (125.0 ms).

[0081] S33 determines the verification processing method for the front-end agent combination based on the number of front-end agents in the combination and the sorting result of the front-end agent combination.

[0082] It should be noted that the sorting result of the front-end agent combination is sorted from smallest to largest based on the average duration of all front-end agents in the combination.

[0083] Verification processing method: refers to the specific rules followed when testing and evaluating candidate agent combinations. It is mainly divided into two types: the first verification processing method (relaxed triggering) and the second verification processing method (strict triggering), as well as a special case: verification processing is performed in all interaction processes (full testing).

[0084] Verification requires additional computing resources and time, and it's impractical to perform undifferentiated full-scale comparative testing on all combinations and interactions, as that would be too costly. Therefore, it's essential to dynamically formulate differentiated verification strategies based on two key dimensions: the complexity (number) and performance potential (ranking) of the combinations. The core objective is to allocate the most verification resources to the combinations most likely to succeed and with manageable risks, thereby maximizing verification efficiency.

[0085] Example operation: The system independently runs the following decision sub-steps (S331 to S334) for each candidate combination (C1, C2, C3) to determine the respective verification processing method.

[0086] It is understood that, based on the number of front-end agents in the front-end agent combination and the ranking result of the front-end agent combination, the verification processing method for the front-end agent combination is determined, specifically including: S331 determines whether the number of the front-end intelligent agent combination is less than the preset threshold for the number of intelligent agent combinations. If yes, it determines that the verification processing method for the front-end intelligent agent combination is to perform verification processing in all interaction processing processes. If no, it proceeds to the next step. In the steps described above, the number of front-end agent combinations refers to the total number of all candidate combinations currently being validated, not the number of agents within a single combination. In this example, this total is 3.

[0087] Preset agent combination number threshold (C_total): an empirical value used to determine whether the verification task load is lightweight.

[0088] This is the first-level check of resource adequacy. If the total number of combinations to be verified is very small (e.g., less than 4), it means that the total verification overhead is very small and the system can fully bear it. In this case, the optimal strategy is to conduct full-volume, real-time A / B testing on all combinations, so as to accumulate verification data in all scenarios at the fastest speed, without complex sampling or triggering rules.

[0089] The system checks the total number of combinations to be verified: C1, C2, C3, a total of 3. It determines that 3<C_total (4), and for all three combinations (C1, C2, C3), the system directly determines their verification processing method as: "perform verification processing in all interaction processing procedures". This means that in subsequent real user interactions, whenever a request comes in, the system will run an experiment process (using the combination) and a control process (using the full set of agents [D, E, I, M]) in parallel for each of the three combinations.

[0090] S332 taking the ratio of the number of said front-end agent combinations to the number of said front-end agents as a combination ratio, determining whether said combination ratio is less than a preset ratio threshold, if yes, determining that the verification processing method of said front-end agent combinations is to perform verification processing in all interaction processing procedures, if no, proceeding to the next step; If the total number of combinations to be verified is 6 (greater than C_total=5), the combination ratio is: the number of candidate combinations to be verified / the total number of all possible front-end agent combinations. All possible combinations refer to all combinations formed by selecting any non-empty subset from all agents (D, E, I, M), and the number is C(4,1)+C(4,2)+C(4,3)+C(4,4)=15. Preset ratio threshold (Ratio_th): a threshold used to judge the stringency of screening.

[0091] This is the second-level check of resource adequacy and screening quality assessment. Even if the total number of candidate combinations is large, if they are carefully selected from all possible combinations (that is, the ratio is very small, such as 0.2), it indicates that the screening standard is very strict, and all remaining combinations are of extremely high potential. For these "elite" combinations, it is still worth investing resources to conduct comprehensive verification to ensure that the optimal solution is not missed.

[0092] Example operation (hypothetical): Assume that the combinations to be verified are still C1, C2, and C3, with a total of 3 combinations. The total number of all possible combinations is 15. Combination ratio = 3 / 15 = 0.2. It is determined that 0.2 < Ratio_th (0.25), which holds true. Therefore, the decision is still to perform full verification on these three combinations. If the ratio ≥ 0.25, the process proceeds to S333.

[0093] S333: Based on the sorting result of the agent combinations, determine whether the sorting result of the agent combination is before the preset sorting result. If yes, proceed to the next step; if not, determine that the verification processing method for the front-end agent combination is the second verification processing method; Before the preset sorting result: it usually means that the ranking number of the combination is less than or equal to the preset ranking threshold Rank_th. For example, if Rank_th=2, the combinations ranked 1st and 2nd meet the condition.

[0094] This is the preliminary screening of performance potential. When verification resources are limited, it is necessary to give priority to ensuring that combinations with the best theoretical performance are fully verified. Lower-ranked combinations (e.g., 3rd place and below) may have limited improvement effects, so the stricter and more resource-saving second verification processing method is adopted for them (verification samples are only recorded when their performance is extremely good), so as to avoid wasting too many resources on low-potential combinations.

[0095] Operation of the embodiment (hypothetical, taking combination C2 [D, I] as an example): The combination C2 is ranked 3rd. The preset ranking threshold Rank_th=2. It is determined that "the 3rd place is not before the top 2", therefore, for combination C2, the system will determine that its verification processing method is the second verification processing method. For C1 ranked 1st and C3 ranked 2nd, the process proceeds to the final step S334.

[0096] S334: Obtain the number of front-end agents of the agent combination, determine whether the number of front-end agents of the agent combination is less than the preset number threshold of front-end agent combinations. If yes, determine that the verification processing method for the front-end agent combination is the first verification processing method; if not, determine that the verification processing method for the front-end agent combination is the second verification processing method.

[0097] It can be understood that the first verification processing method is: when the average value of the resolution duration of the front-end agents in the interaction processing of the front-end agent combination is less than the average value of the resolution duration of all front-end agents, the verification processing of the front-end agent combination is performed.

[0098] Further, the second verification processing method is: when the parsing durations of all front-end agents in the front-end agent combination during the interaction processing are all less than the average value of the parsing durations of all front-end agents, the verification processing is performed on the front-end agent combination.

[0099] Number of front-end agents in the agent combination: refers to the number of agents included in a single combination (return to the concept of S31).

[0100] Preset front-end agent combination quantity threshold (Size_th): a threshold used to divide combination complexity.

[0101] This is the final decision for determining verification strictness based on combination complexity. The more agents there are in a combination, the more complex the collaborative relationship is, and the higher the probability of unexpected performance (such as a sudden slowdown or error of an agent). Therefore, for large complex combinations (quantity ≥ Size_th), the stricter second verification processing method must be adopted. This verification sample is only considered valid when all members of the combination perform excellently, so as to control evaluation noise. For small simple combinations (quantity < Size_th), the relatively lenient first verification processing method can be adopted, and verification can be triggered as long as the overall average performance is good, so that enough valid samples can be collected more quickly.

[0102] Example operation (hypothetical, taking combinations C1 and C3 as examples): for combination C1 [D, E], the number of agents is 2. After judging that 2 < Size_th (3), the first verification processing method is determined as the verification processing method for C1.

[0103] For combination C3 [D, I, E], the number of agents is 3. After judging that 3 is not less than Size_th (3), the second verification processing method is determined as the verification processing method for C3.

[0104] Execution definition of verification processing method First verification processing method: when the average value of the parsing durations of front-end agents of the front-end agent combination during a single interaction processing is less than the average value of the parsing durations of all front-end agents (Global_Avg = 222.5ms), the current verification processing for the combination is triggered.

[0105] Example: for combination C1 [D, E], in one interaction, D actually takes 105ms and E actually takes 115ms, with an average of 110ms. Since 110ms < 222.5ms, which meets the condition, the system will perform consistency comparison verification.

[0106] The second verification process is as follows: verification is triggered only when the parsing time of each front-end agent in a single interaction process is less than its own global historical average time.

[0107] Example: For the combination C3 [D, I, E], in a single interaction, the following conditions must be met simultaneously: D_actual < 100ms, I_actual < 150ms, E_actual < 120ms. Verification will only be triggered if all three conditions are met.

[0108] The core operations for verification processing: Regardless of the triggering method, the core verification operation is as follows: The agent recognition results of the experimental combination, along with the user question, are transmitted to the central agent to obtain understanding result A; simultaneously, the recognition results of all agents, along with the user question, are transmitted to another central agent instance to obtain the standard understanding result B. Finally, results A and B are compared to see if they are consistent in their core decision-making intent. By statistically analyzing the proportion of "consistency" across a large number of interactions, the reliability of the combination can be assessed.

[0109] It should be noted that the verification process for the aforementioned front-end intelligent agent combination specifically includes: The recognition and processing results of different front-end intelligent agents combined with user questions are transmitted to the central intelligent agent to obtain the understanding result. This is then compared with the understanding result obtained by transmitting the recognition and processing results of all front-end intelligent agents and user questions to the central intelligent agent to verify the combination of the front-end intelligent agents.

[0110] This embodiment builds upon the completion of the agent combination verification phase. It focuses on the optimal combination [D, E] (Dialogue State Analysis Agent D and Sentiment Analysis Agent E) that performed best in the verification. The system has accumulated a large amount of verification data about this combination by performing verification processing in all interaction processes. Now, based on this data, a rigorous decision-making process will be executed to determine whether and how to dynamically adopt this combination in real customer service interactions, thereby achieving a leap in response speed.

[0111] Specifically, such as Figure 4 As shown, the method for determining the customer service interaction method based on the combination of the front-end intelligent agents is as follows: The system first extracts two types of key data from the verification log: verification processing results (i.e., the decision consistency record between the combination [D, E] and the full combination [D, E, I, M]) and parsing time data (the actual time consumed by each agent in each interaction).

[0112] S41 Based on the verification processing result, determine the consistent interaction processing process of the understanding result of the front-end intelligent agent combination, and take it as the consistent interaction process. "Consistent interaction process" refers to an interaction process in which, in a certain interaction verification, the analysis results using only the combination of [D, E] and the analysis results using the full set of agents ultimately lead the central agent to make the same core decision (such as "urgent reassurance and initiation of rapid return").

[0113] This is the direct basis for assessing the reliability of portfolio decisions. Only those portfolios that have proven to maintain high consistency throughout history are eligible to be considered for use in real-world production environments.

[0114] Example Operation: The system analyzes all verification records within a past period (e.g., 24 hours). Assume there are a total of 1000 verification interactions, of which 850 yield consistent results. Therefore, the number of "consistent interaction processes" for combination [D, E] is recorded as 850.

[0115] S42 determines the deviation between the parsing time of the front-end agent combination and the maximum parsing time of all front-end agents in different interaction processes based on the parsing time data of the front-end agent combination, and uses it as the parsing time deviation. The "parsing time deviation" is used to quantify the speed gains from using a streamlined combination. Its calculation formula is: Deviation = Max(parsing time of all agents) - Max(parsing time of agents within the combination). Where Max(parsing time of all agents) represents the time required to wait for all agents (D, E, I, M) to complete (i.e., the time taken by the slowest agent); Max(parsing time of agents within the combination) represents the time required to wait for agents within that combination (such as D and E) to complete.

[0116] Consistency alone is not enough; it must be ensured that the combination delivers a significant performance improvement. The deviation directly reflects how much waiting time is saved. The larger the positive value, the more pronounced the speedup effect.

[0117] Taking a specific interaction as an example. In this interaction, the actual time taken by each agent is: D: 102ms, E: 118ms, I: 160ms, M: 480ms. Therefore, Max(full) = 480ms (M is the slowest), Max([D, E]) = 118ms (E is relatively slow). The parsing time deviation = 480ms - 118ms = 362ms. The system will calculate this value for each verification interaction.

[0118] S43 determines the customer service interaction method based on the consistent interaction process data and the parsing time deviation.

[0119] It should be noted that if, during the interaction process, the understanding result obtained by transmitting the recognition and processing results of different front-end intelligent agents combined with the user's question to the central intelligent agent is consistent with the understanding result obtained by transmitting the recognition and processing results of all front-end intelligent agents and the user's question to the central intelligent agent, then the interaction process is determined to be an interaction process with consistent understanding results.

[0120] Specifically, the parsing time deviation is determined by the difference between the maximum parsing time of all front-end agents and the parsing time of the front-end agent combination during the interaction process, wherein the parsing time of the front-end agent combination is the maximum parsing time of all front-end agents in the front-end agent combination.

[0121] Specifically, based on the consistent interaction process data and the parsing time deviation, a customer service interaction method based on the front-end intelligent agent combination is determined, including: S431 uses consistent interaction process data from different front-end agent combinations to determine the proportion of consistent interaction processes of the front-end agent combination in the number of interaction processes processed by the front-end agent combination, determines the identification matching factor of the front-end agent combination, and determines whether there is a front-end agent combination whose identification matching factor is greater than a preset matching factor threshold. If so, proceed to the next step; otherwise, continue to perform verification processing based on the verification processing method of the front-end agent combination. Under no circumstances should the front-end agent combination be used for customer service interaction processing. Assess the reliability threshold (identify the matching factor). The "identify the matching factor" is a quantitative indicator of portfolio reliability, calculated as: number of consistent interaction processes / total number of interaction processes involved in the validation of this portfolio. It represents the consistency rate between the portfolio's decisions and the overall decisions.

[0122] This is the first and most important safety valve before deploying the system. Only when a sufficiently high consistency rate is achieved can the system be proven reliable in the vast majority of cases and proceed to the next round of evaluation. Otherwise, further verification should continue, and it should never be deployed.

[0123] Example Operation: For combination [D, E], the consistent interaction process is 850 times, and the total verification interaction is 1000 times. Therefore, its identification matching factor = 850 / 1000 = 0.85 (or 85%). Assume that the system's preset matching factor threshold is 0.80. Since 0.85 > 0.80, the condition is met. Combination [D, E] passes the reliability test and is marked as a "usable agent combination," proceeding to the subsequent evaluation steps. If the identification matching factor of all combinations does not exceed 0.80, the system will decide to continue with full verification and will never use the front-end agent combination for customer service interaction processing in real customer service under any circumstances.

[0124] S432 identifies front-end agent combinations whose matching factor is greater than a preset matching factor threshold as available agent combinations. Only the verification process of available agent combinations needs to be performed to determine whether the number of available agent combinations is greater than a preset agent combination number threshold. If yes, proceed to the next step; otherwise, the verification process of available agent combinations is performed in different interaction processes. In any case, front-end agent combinations are not used for customer service interaction processing, thereby enabling a more comprehensive verification process of available agent combinations. This is a check on the generalizability of the optimization results. If there are many available combinations (e.g., more than a preset threshold), it means the system has found multiple efficient sub-processes, and can proceed to the formulation of more complex scheduling strategies. If there are few available combinations (e.g., only one), the system will adopt a more conservative strategy: continue to expand the validation sample to ensure its reliability.

[0125] Example Operation: Assume that currently only the combination [D, E] is marked as an "available agent combination," with a quantity of 1. The preset threshold for the number of agent combinations is set to N_available = 2. Since 1>2 is not true (i.e., 1 is not greater than 2), the system enters the "No" branch. According to the rules, the system decides at this point: to perform verification processing on available agent combinations in different interaction processes, and under no circumstances will the front-end agent combination be used for customer service interaction processing. This means that although [D, E] has been proven reliable, further verification processing is needed. The system chooses to continue verifying it in 100% of the interactions to observe its performance in longer periods and more diverse scenarios, and will not enable it until the number of verifications reaches a certain threshold before it can be put into use.

[0126] S433 determines the average deviation of the available agent combination based on the average parsing time deviation of the available agent combination in different interaction processes, and determines whether there is an available agent combination with an average deviation less than a preset time threshold. If yes, proceed to the next step; otherwise, verify the available agent combination in different interaction processes, and do not use the front-end agent combination for customer service interaction processing under any circumstances, so as to more comprehensively realize the verification processing of available agent combinations. Definition: "Mean Deviation" is the average of the "parse duration deviation" obtained across all verification interactions for the "available agent combination". It represents the average speed gain that can be achieved by using this combination.

[0127] Even if a combination is highly reliable, if the time savings are negligible (e.g., only 10ms faster on average), then its value in production is limited, and it increases system complexity. This step ensures that only combinations that deliver substantial performance improvements are further considered for dynamic scheduling.

[0128] The system calculates the parsing time deviation of the [D, E] combination across all 1000 verification interactions and takes its average. Assume the calculated average deviation is 350ms. A preset time threshold, such as T_significant = 100ms, is used. The system checks if 350ms < 100ms; if not (350ms is much greater than 100ms), the combination [D, E] is not only reliable but also provides a significant speed improvement, thus meeting the condition and proceeding to the next step. If the average deviation is less than 100ms, the system considers the speed improvement insignificant and chooses to continue verification, not enabling the combination.

[0129] S434 determines whether the average deviation of the available agent combination is less than a preset time threshold. If yes, if all the front-end agents of the available agent combination have been parsed, but some front-end agents have not been parsed, then the front-end agents of the available agent combination are directly used for customer service interaction processing. If no, if all the front-end agents of the available agent combination have been parsed, but some front-end agents have not been parsed, and if the parsing time is greater than the preset parsing time value, then the front-end agents of the available agent combination are directly used for customer service interaction processing. If the parsing time is not greater than the preset parsing time value, S435 will not use the combination of front-end intelligent agents for customer service interaction processing, thereby enabling a more comprehensive verification of available intelligent agent combinations.

[0130] This is the final decision-making step, which will formulate precise rules on when to enable the combination in real-time interaction based on a comparison of the "mean deviation" with another "preset duration threshold". The "preset duration threshold" here may be different from the threshold in S433, and it focuses more on defining a tolerable upper limit for response time.

[0131] This is key to implementing the "resilient degradation" or "bounded wait" strategy. The core idea is that in real-time processing, when a fast assembly is ready while a slow agent has not yet completed its work, the system should not wait indefinitely. It needs a set of rules to decide whether to immediately adopt the result of the fast assembly (sacrificing some possible completeness for speed) or to continue waiting for more complete information.

[0132] Example Operation: Assume the preset duration threshold here is T_wait = 200ms. Now, it is necessary to determine whether the average deviation of [D, E] (350ms) is not less than 200ms. Since the condition is "No", the rule of the "If No" branch in S434 is executed. This rule defines the core logic of dynamic scheduling: In real-time customer service interaction, if all members (D and E) of the available agent combination [D, E] have been parsed, but the other agents (I or M) in the full combination have not yet been parsed, the system checks the "parsing time" that has elapsed since the request began. If the parsing time is greater than the preset "parsing time preset value" (e.g., Wait_Limit=300ms), the system will not wait any longer and will directly use the analysis results of the ready [D, E] combination for customer service interaction processing.

[0133] Step S435 defines another case: If the parsing time is no greater than (i.e., less than or equal to) the waiting limit of 300ms, the system will choose not to use the [D,E] combination and will continue to wait for all agents to complete. Meanwhile, this interaction will serve as a new validation sample to further refine the evaluation of the [D,E] combination.

[0134] Example 2 In a second aspect, the present invention provides a computer system comprising: a memory and a processor connected in communication, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the aforementioned customer service interaction method based on multi-agent collaborative decision-making when running the computer program.

[0135] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0136] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0137] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A customer service interaction method based on multi-agent collaborative decision-making, characterized in that, Specifically, it includes: Based on the parsed data of the intelligent agent, the parsing time deviation of the front-end intelligent agent in the process of recognizing user questions is determined. When it is determined that the intelligent agent combination needs to be optimized based on the parsing time deviation, the construction strategy of the front-end intelligent agent combination is determined by the ranking result of the parsing time of the front-end intelligent agent among all front-end intelligent agents and the parsing time deviation between it and other front-end intelligent agents. Based on the construction strategy, the front-end intelligent agent combination is constructed. According to the composition data of the front-end intelligent agents in the combination and the deviation of the composition data of other front-end intelligent agent combinations, the verification processing method of the front-end intelligent agent combination is determined. According to the verification processing results and parsing time data, the customer service interaction method based on the front-end intelligent agent combination is determined. The method for determining the construction strategy of the front-end intelligent agent combination is as follows: Based on the parsing time of the front-end agent in the ranking of all front-end agents, the ranking result of the front-end agent in the recognition process is determined to be the highest, and it is used as the matching recognition process. Based on the matching and recognition process, a baseline agent is determined among the front-end agents, and the parsing time deviation between the baseline agent and other front-end agents is determined. Based on the matching and recognition process data of the baseline intelligent agent and the parsing time deviation between it and other front-end intelligent agents, the construction strategy of the front-end intelligent agent combination is determined. The system adopts a "central agent + front-end agent" architecture. For all user questions, the system must wait for the front-end agent to complete the analysis before summarizing the results and sending them to the central agent for decision-making. The aim is to build and verify an efficient combination of front-end agents and bypass some front-end agents. The construction strategy includes a first construction strategy and a second construction strategy. The first construction strategy is to use front-end agents whose average duration difference is less than a first duration threshold as candidate agents. Based on the fact that the benchmark agent exists in different combinations of front-end agents, the candidate agents are freely combined to construct a combination of front-end agents. The second construction strategy is to use front-end agents whose average duration difference is less than the second duration threshold as candidate agents, and to construct front-end agent combinations by freely combining the candidate agents based on the fact that the benchmark agent exists in different combinations of front-end agents. The method for determining the customer service interaction method based on the front-end intelligent agent combination is as follows: Based on the verification results, the interaction process that makes the understanding results of the front-end intelligent agents consistent is determined and used as the consistent interaction process. Based on the parsing time data of the front-end agent combination, the deviation between the parsing time of the front-end agent combination and the maximum parsing time of all front-end agents is determined in different interaction processes, and this deviation is used as the parsing time deviation. Based on the consistent interaction process data and the parsing time deviation, a customer service interaction method based on the front-end intelligent agent combination is determined. The front-end intelligent agent is used to identify and process user questions, and transmits the identification and processing results to the central intelligent agent for semantic understanding to obtain the understanding results.

2. The customer service interaction method based on multi-agent collaborative decision-making as described in claim 1, characterized in that, The parsing time of the front-end intelligent agent in the process of identifying user questions is determined based on the time it takes for the front-end intelligent agent to obtain the identification and processing results during the process of identifying and processing user questions.

3. The customer service interaction method based on multi-agent collaborative decision-making as described in claim 1, characterized in that, The deviation in parsing time is determined based on the deviation in the time taken for different front-end agents to obtain the recognition and processing results.

4. The customer service interaction method based on multi-agent collaborative decision-making as described in claim 1, characterized in that, The optimization process for agent composition needs to be determined, specifically including: Based on the parsing time deviation, the difference in parsing time between the agent with the longest parsing time and the agent with the shortest parsing time is determined during the user question identification and processing process, and this difference is used as the parsing time difference. Based on the difference in parsing time during the user problem identification and processing process, the identification and processing process where the difference in parsing time is greater than the preset time threshold is determined and treated as a delayed identification process. Based on the data from the delayed recognition process, it is determined whether optimization processing of the agent combination is required.

5. The customer service interaction method based on multi-agent collaborative decision-making as described in claim 4, characterized in that, When the proportion of delayed recognition processes in all recognition processes exceeds a preset threshold for the proportion of delayed processes, different front-end agents are combined to determine whether the problem can be effectively solved when some front-end agents are used as input.

6. The customer service interaction method based on multi-agent collaborative decision-making as described in claim 1, characterized in that, The parsing time of the front-end agent is determined by ranking all front-end agents, including the parsing time of the front-end agent during the recognition process, from shortest to longest.

7. The customer service interaction method based on multi-agent collaborative decision-making as described in claim 1, characterized in that, If, during the interaction process, the understanding result obtained by transmitting the recognition and processing results of different front-end intelligent agents combined with the user's question to the central intelligent agent is consistent with the understanding result obtained by transmitting the recognition and processing results of all front-end intelligent agents and the user's question to the central intelligent agent, then the interaction process is determined to be an interaction process with consistent understanding results.

8. A computer system, comprising: A memory and processor connected by communication, and a computer program stored in the memory and capable of running on the processor, characterized in that, when the processor runs the computer program, it executes a customer service interaction method based on multi-agent collaborative decision-making as described in any one of claims 1-7.

Images