Intelligent interaction management and decision-making system and method based on large model

Abstract

The invention provides an intelligent interaction management and decision-making system and method based on a large model, and belongs to the technical field of large models, and the method specifically comprises the steps: carrying out the determination of an interruption decision-making strategy of a user through the update data of an interruption interaction problem, and in combination with the recognition deviation conditions of large models of different problems and the matching data of a recognition method, obtaining a decision-making strategy of the user; based on the abortion decision-making strategy, abortion management processing of interaction problems of different users is carried out, according to interaction processing data of different users in the abortion interaction problems, and in combination with abortion interaction data of different users, an identification processing target in the abortion interaction problems is determined, and on the basis of the identification processing target, the interaction problems of different users are determined. And in combination with the similarity of unidentified users of different identification processing targets, the interaction management strategy of the large model in the identification processing targets is determined, so that the reliability of interaction processing of the large model is improved.

Description

Technical Field

[0001] This invention belongs to the field of large model technology, and in particular relates to an intelligent interactive management and decision-making system and method based on large models. Background Technology

[0002] Traditional Knowledge Management Systems (KMS) can only manage structured documents and cannot understand unstructured content specific to the construction industry (such as CAD drawings, construction logs, change orders, etc.). General chatbots lack expertise in the construction field and have low accuracy in their responses. Existing Management Information Systems (MIS) have rigid functionalities and cannot adapt to complex changes in business scenarios.

[0003] To address the aforementioned technical problems, the large-scale question-answering model in invention patent application CN202510068783.5, "An Intelligent Question-Answering Method and System in the Construction Field," is obtained through fine-tuning of sample question information and a knowledge base containing professional knowledge in the construction field. The knowledge base assists the large-scale question-answering model in generating responses based on the sample question information, thereby achieving highly reliable responses in intelligent question-answering scenarios within the construction field. However, the above technical solution has the following technical problems: When using large models for interactive processing, especially during periods of port shortage, directly terminating the interaction for questions with a high probability of answer recognition will lead to a large overall data processing burden. Therefore, identifying and processing interaction termination issues and determining targeted termination decision strategies for users in these situations has become an urgent technical problem to be solved.

[0004] Therefore, there is an urgent need for an intelligent interactive management and decision-making system and method based on large models. Summary of the Invention

[0005] To achieve the objectives of this invention, the following technical solution is adopted: Specifically, this application provides an intelligent interactive management and decision-making method based on a large model, which includes: S1 uses the interaction data of the large model as a basis to determine the shortage period of the interaction port of the large model. Based on the distribution data of the shortage period, it determines the identification method of the large model's interaction termination problem. Based on the identification method, it performs the update processing of the interaction termination problem. Based on the updated data of the interaction termination problem, and combined with the identification deviation of the large model for different problems and the matching data of the identification method, it determines the user's termination decision strategy. S2 manages the termination of different user interaction issues based on the termination decision strategy. Based on the interaction processing data of different users in the termination interaction issues, and combined with the termination interaction data of different users, the identification and processing target in the termination interaction issues is determined. S3 determines the interaction management strategy of the large model in the identification and processing target based on the identification and processing target and in combination with the similarity of unidentified users of different identification and processing targets.

[0006] The beneficial effects of this invention are as follows: The higher the system's workload (manifested as more frequent and severe service shortages), the lower its fault tolerance and the higher its stability priority. A more conservative interaction strategy is needed, employing a lower "identification bias probability" threshold to determine whether to abort interactions for various user-submitted questions, thereby quickly reducing the pressure on the system from low-determinism requests. Conversely, when the system is relatively idle, it can tolerate more risk, using a higher threshold to allow the model to attempt to answer more questions with lower certainty, thus improving service coverage and exploratory learning. This method assesses system workload by quantitatively analyzing recent shortage period data and maps this workload to dynamically changing decision thresholds, thereby achieving an optimal balance between system load and service quality while ensuring the stability of core services.

[0007] By analyzing the total number of interrupted interaction issues that occur within a complete system cycle, the temporal distribution characteristics of various issues, and their proximity to the system's security threshold (risk structure), the breadth and granularity of the current circuit breaker mechanism can be assessed. When the number of interrupted interaction issues is excessive, their distribution is dense, or a large number of issues are about to become interrupted interaction issues, it indicates that the circuit breaker mechanism may be too broad and stringent. A more lenient exemption strategy should be adopted, that is, after a certain amount of interruption processing, data processing can continue for some interrupted interaction issues. This achieves macro-level optimization and adjustment of the system's circuit breaker strategy from the perspective of issue type, balancing system protection and user service breadth, and avoiding the occurrence of technical problems that significantly impact user experience due to frequent interruption processing for a single user over a long period.

[0008] Furthermore, the interaction data of the large model includes the number of idle interaction ports of the large model in different time periods.

[0009] Furthermore, the shortage period of the interaction port in the large model is the period during which there are no idle interaction ports.

[0010] Furthermore, the issue of terminating the interaction refers to the issue of directly terminating the interaction.

[0011] Furthermore, the method for determining the identification method of the large model's interaction termination problem is as follows: Based on the distribution data of the aforementioned shortage periods, the duration of different shortage periods is determined; Based on the duration, the shortage period with a duration greater than a preset duration threshold is determined and taken as the period of shortage impact; Based on the shortage impact period and shortage period data within the most recent preset time period, a method for identifying the interruption of interaction problems in the large model is determined.

[0012] Furthermore, the method for determining the interaction management strategy of the large model in the identification and processing target is as follows: Based on the identification and processing target, determine the proportion of the identification and processing target in the interaction termination problem, and use the proportion of the identification and processing target in the interaction termination problem as the identification ratio; Users who do not interact with the large model in the identification and processing targets are identified as non-interactive users. Based on the overlap data between the non-interactive users and other identification and processing targets, the associated identification targets of different non-interactive users are determined. Based on the identification ratio, the unidentified users of the identification processing target, and the associated identification targets of different non-interactive users, the interaction management strategy of the large model in the identification processing target is determined.

[0013] Secondly, this application provides an intelligent interactive management and decision-making system based on a large model, employing the aforementioned intelligent interactive management and decision-making method based on a large model, specifically including: Stop control module, identification and processing module, interactive association module; The termination control module is responsible for determining the user's termination decision strategy. The identification processing module is responsible for determining the identification processing target in the termination interaction problem; The interaction association module is responsible for determining the interaction management strategy of the large model in the recognition and processing target.

[0014] 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.

[0015] 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

[0016] 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.

[0017] Figure 1 This is a flowchart of an intelligent interactive management and decision-making method based on a large model; Figure 2 This is a flowchart illustrating the method for determining the identification method of the termination interaction problem in large models; Figure 3 This is a flowchart illustrating the method for determining the user's termination decision strategy; Figure 4 This is a framework diagram of an intelligent interactive management and decision-making system based on a large model. Detailed Implementation

[0018] 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.

[0019] Example 1 like Figure 1 As shown, this application provides an intelligent interactive management and decision-making method based on a large model, specifically including: S1 uses the interaction data of the large model as a basis to determine the shortage period of the interaction port of the large model. Based on the distribution data of the shortage period, it determines the identification method of the large model's interaction termination problem. Based on the identification method, it performs the update processing of the interaction termination problem. Based on the updated data of the interaction termination problem, and combined with the identification deviation of the large model for different problems and the matching data of the identification method, it determines the user's termination decision strategy. S2 manages the termination of different user interaction issues based on the termination decision strategy. Based on the interaction processing data of different users in the termination interaction issues, and combined with the termination interaction data of different users, the identification and processing target in the termination interaction issues is determined. S3 determines the interaction management strategy of the large model in the identification and processing target based on the identification and processing target and in combination with the similarity of unidentified users of different identification and processing targets.

[0020] Furthermore, the interaction data of the large model includes the number of idle interaction ports of the large model in different time periods.

[0021] Furthermore, the shortage period of the interaction port in the large model is the period during which there are no idle interaction ports.

[0022] Furthermore, the issue of terminating the interaction refers to the issue of directly terminating the interaction.

[0023] Specifically, such as Figure 2 As shown, the method for determining the identification method of the large model's interaction termination problem is as follows: The core objective of this embodiment is to dynamically adjust the criteria for "interaction termination issues" based on the real-time busy level of the large model service system, thereby achieving intelligent adjustment of system load and resource optimization. The core logic is: the higher the system's busy level (manifested as more frequent and severe service shortage events), the worse its fault tolerance and the higher the priority of stability. Therefore, a more conservative interaction strategy is needed, i.e., using a lower "identification bias probability" threshold to determine whether interaction should be terminated for various user-submitted questions, thus quickly reducing the pressure on the system caused by low-determinism requests. Conversely, when the system is relatively idle, it can tolerate more risk, using a higher threshold to allow the model to attempt to answer more questions with low certainty, thereby improving service coverage and exploratory learning. This method assesses system busyness by quantitatively analyzing recent shortage period data and maps this busyness to dynamically changing judgment thresholds, thereby achieving an optimal balance between system load and service quality while ensuring the stability of core services.

[0024] S11 uses the distribution data of the shortage periods to determine the duration of different shortage periods; "Shortage periods" refer to the time periods during which the large-scale model service is unable to process user interaction requests normally or requires significant delays in processing due to resource overload (such as request queue congestion or the absence of available interaction ports). The system accurately records the start and end timestamps of each shortage event, thus obtaining a series of specific time period data. "Duration of different shortage periods," that is, the duration of each shortage event, is the basic indicator for measuring the severity of a single overload.

[0025] This step involves refined data collection and quantification of system load anomalies. Its significance lies in transforming vague operational perceptions such as "system busy" or "service unavailable" into structured time-series data that can be analyzed by the program. Simply knowing that a shortage has occurred is insufficient; the duration of the shortage directly reflects the depth of overload and the extent of disruption to user interaction. Brief periods of resource contention and prolonged service outages fundamentally impact subsequent system recovery strategies and circuit breaker strength decisions. This is the cornerstone of achieving scientific and precise load control.

[0026] S12 determines the shortage period with a duration greater than a preset duration threshold based on the duration, and uses it as the shortage impact period; "Shortage Impact Period" is a key subset of "Shortage Period," specifically referring to shortage events that last longer than a "preset duration threshold." This threshold is used to distinguish between minor, transient service fluctuations (such as second-level response delays or jitter) and severe service outages that may substantially impair user experience and indicate that the system may be approaching its stability limits.

[0027] This step aims to filter out significant events from all load fluctuations that characterize deep system stress or potential failures. Its significance lies in differentiating the severity of failure impacts. Brief delays may be tolerable, but prolonged unresponsiveness typically leads to user task failures or session timeouts. By focusing on these "shortage impact periods," the system can more accurately assess the severity of the current load condition, providing crucial information for deciding the appropriate circuit breaker intervention (mild current limiting or aggressive circuit breaking), avoiding overreacting to minor fluctuations and preventing under-response to serious problems.

[0028] S13 determines the identification method for the interruption of interaction problem of the large model based on the shortage impact period and the shortage period data in the most recent preset time period.

[0029] The "method for identifying aborted interaction problems" specifically refers to a decision logic that dynamically evolves based on the system's real-time workload. Its core output is a dynamically changing "identification bias probability threshold." This threshold determines whether the current user's question should be categorized as an "interaction abort problem." The "identification bias probability" is another key input to this method, defined as the proportion of times a user reported "unresolved" or "the answer has a problem" in historical interaction records for a specific type of problem, relative to the total number of interactions for that type of problem. It quantifies the unreliability of the model's answer in that domain. An "interaction abort problem" means that once the system identifies it as such, a pre-defined interruption process will be triggered, outputting a standardized response such as "no learning at the moment," and terminating subsequent model calls and interactions.

[0030] This step is the core of the intelligence in this embodiment, creatively coupling system-level health status monitoring with problem-level quality risk assessment in a dynamic manner. Its significance lies in constructing an adaptive system with negative feedback adjustment capabilities. Instead of using static, uniform rules, it allows the safety boundary (judgment threshold) to elastically change with system load: the busier and more vulnerable the system, the tighter the safety boundary (threshold decreases), prioritizing overall system stability and the experience of remaining high-value requests at the cost of sacrificing some service scope (circuiting out more problems); the less busy and more robust the system, the wider the safety boundary (threshold increases), allowing the model to explore capability boundaries and serve more diverse needs within acceptable risk ranges. This achieves a refined and dynamic trade-off between the conflicting goals of service quality and system stability under resource-constrained conditions.

[0031] It should be noted that if the number of shortage-affected periods in the most recent preset time period is greater than the preset threshold for the number of affected periods, then the problem will be treated as an interaction termination problem as long as the identification deviation probability of the problem is greater than the preset probability threshold.

[0032] In this embodiment, the "recent preset time period" is extended to a rolling monthly window (e.g., the past 30 days). The "preset threshold for the number of affected periods" is set for this monthly window and is used to determine whether the system is in a "monthly high-frequency severe shortage phase". When the total number of "shortage-affected periods" in the past 30 days exceeds this threshold, the system is determined to be in a high-pressure period.

[0033] This sub-step is used to identify systemic and persistent periods of high load. Its significance lies in confirming that the system is in a prolonged high-pressure period, requiring a stable and stringent full-cycle circuit breaker strategy. Using a lower "preset probability threshold" (e.g., 5%) means remaining vigilant throughout the month for any issues with even a slight risk of deviation, continuously and stably controlling system load and ensuring the stability of basic services throughout the entire critical business cycle.

[0034] Additionally, it should be noted that if the number of shortage-affected periods in the most recent preset time period is not greater than the preset threshold for the number of affected periods, then in step S131, the interval between adjacent shortage-affected periods is obtained, and it is determined whether there are shortage-affected periods with an interval length less than the preset interval length threshold. If so, then as long as the identification deviation probability of the problem is greater than the preset probability threshold, the problem is treated as a problem to terminate interaction. Otherwise, step S132 is entered. Based on monthly analysis, "interval duration" refers to the time difference between adjacent "shortage impact periods" within a month. The "preset interval duration threshold" reflects the density of the distribution between shortage impact periods (e.g., 10 minutes). When multiple severe shortage clusters with very short intervals are found within a month, it is determined that a "dense stress cluster" exists.

[0035] This sub-step aims to capture concentrated high-pressure shocks during specific short periods within the monthly cycle. Its significance lies in the fact that even if the overall monthly load does not exceed the limit, if there are localized and dense pressure clusters, decisive measures need to be taken, using a lower "preset probability threshold" to ensure the reliability of interactive processing.

[0036] S132 determines whether there is a shortage period in different dates. If yes, proceed to step S133. If no, the problem is regarded as an interaction termination problem as long as the identification deviation probability of the problem is greater than the second preset probability threshold. "Whether shortage periods exist on different dates" from a monthly perspective refers to analyzing whether shortage periods exist every day. After identifying such patterns, the level of busy interaction processing is determined. The method for identifying interaction problems is adaptively adjusted, and the specific "second preset probability threshold" is a relatively high circuit breaker standard preset for such regular busy patterns.

[0037] This sub-step enables intelligent adaptation to predictable periodic busy periods. Its significance lies in optimizing user experience and resource utilization. If there are not shortage periods every day, the system's idle state is sufficient; therefore, a larger second preset probability threshold is used to identify and handle interaction termination issues.

[0038] It should be noted that the preset probability threshold is less than the second preset probability threshold.

[0039] S133 determines the interaction idle coefficient of the large model based on the number of shortage-affected periods and the number of shortage periods in the most recent preset time period, and determines the identification method of the interruption interaction problem of the large model based on the interaction idle coefficient.

[0040] The "Interactive Idle Coefficient" is calculated based on monthly window data. The formula must reflect the overall load level within a month, for example: 1 - (Total duration of shortage periods in the month / Total service duration in the month) * (Number of shortage-affected periods in the month / Total number of shortage periods). The closer this coefficient is to 1, the lighter the monthly load. The "Preset Idle Coefficient Threshold" is used to distinguish between "Monthly Normal Load" and "Monthly Idle Load" states.

[0041] This sub-step is the system's monthly baseline adjuster. When the specific patterns mentioned above are not met, the system determines the overall resource sufficiency based on the monthly comprehensive load coefficient. Its significance lies in providing a default, smooth global strategy. If the entire month is relatively busy (low coefficient), a stricter threshold is used by default; if the entire month is relatively idle (high coefficient), a more lenient threshold is used by default. This ensures that the system strategy matches the overall monthly resource situation, achieving long-term resource and service balance.

[0042] It is understandable that the method for identifying the termination of interaction problems in the large model based on the aforementioned interaction idle coefficient specifically includes: Determine whether the interaction idle coefficient is less than a preset idle coefficient threshold. If so, the problem is designated as an interaction termination problem as long as the identification deviation probability of the problem is greater than a preset probability threshold. If not, the problem is designated as an interaction termination problem as long as the identification deviation probability of the problem is greater than a second preset probability threshold.

[0043] Taking an enterprise-level large-scale model as an example, we demonstrate its decision-making process based on monthly data analysis. The preset parameters are adjusted to a monthly perspective: preset duration threshold = 1 hour, monthly preset impact period number threshold = 20 times, preset interval duration threshold = 10 minutes, preset probability threshold = 5%, second preset probability threshold = 15%, monthly preset idle coefficient threshold = 0.85.

[0044] Scenario 1: Identify monthly high-pressure cycles and activate strict circuit breaker: The system analyzed data from the past 30 days and recorded 25 instances of "shortage impact periods" exceeding one hour last month, surpassing the monthly threshold of 20. The system determined that this month as a whole is a "high-pressure period."

[0045] Time: On any future date, if a user asks a question of the type "deep log analysis," the historical identification bias probability is 12%. Decision: A strict "preset probability threshold" (5%) is adopted globally. 12% > 5%.

[0046] Execution: The system replied: "We are currently unable to process this type of in-depth analysis task. We recommend using a dedicated log analysis platform or contacting the architecture team." In the future, we will continue to filter requests with high computational complexity and medium to high uncertainty to ensure the stability of high-frequency core functions such as code completion and API queries.

[0047] Scenario 2: Coping with the pressure of a dense release schedule within a month: During the second week of last month, there were three consecutive days with shortage impact periods less than 10 minutes apart from the previous short shortage impact period, forming a "dense stress cluster". Decision: In the future, a strict "preset probability threshold" (5%) will be adopted.

[0048] When the data processing pressure of a large model is concentrated in a particularly concentrated period, a relatively strict preset probability threshold is still used to identify the problem of interrupted interaction, thereby reducing the system's operating load.

[0049] Scenario 3: Identifying whether there are shortage periods every day: Data from last month showed that shortages did not occur every day, resulting in relatively idle system operation. Therefore, a higher second preset probability threshold (e.g., 15%) was used to identify and handle interaction interruption issues.

[0050] Scenario 4: Routine adjustments based on overall monthly load (corresponding to S134) Analysis of last month's data: The monthly interactive idle coefficient was calculated to be 0.88 (>threshold 0.85), indicating that the overall load was relatively light last month.

[0051] Time: In future dates, due to the high idle rate of the previous month, a more lenient "second preset probability threshold" (15%) will be used by default. Execution: The large model is invoked, code is generated, and analysis is performed. In months with ample resources, the system more actively serves various issues, improving user satisfaction and model training opportunities. If the idle rate is low (e.g., 0.7), a strict threshold of 5% will be used by default, and the above issues will be suspended.

[0052] For questions with the keyword "support structure, safety factor calculation": there were a total of 800 interactions, 80 of which were reported as "problem exists", and the identification deviation probability was 10%. When the question "Calculate the safety factor Fs of the support structure using the Swedish circular arc method" was encountered, the second preset probability threshold was used. Therefore, questions with the keyword "support structure, safety factor calculation" did not fall under the category of questions that caused the interaction to stop.

[0053] In this application, the identification and processing of the interaction termination problem is performed according to the operating load, thereby avoiding the technical problem of excessive waste of computing resources in problems with a high probability of identification error, which leads to poor reliability of interaction processing. This realizes the dynamic adjustment and optimization of the identification method for the interaction termination problem.

[0054] Specifically, such as Figure 3 As shown, the method for determining the user's termination decision strategy is as follows: The core objective of this embodiment is to dynamically adjust the unified termination exemption strategy applicable to all users based on the distribution characteristics of the types of termination interaction problems that occur in the large-scale model service system. Its core logic is: during the circuit breaker process, the system encounters various types of interaction problems, and the degree of impact of termination on user experience and business continuity varies depending on the type of problem. By analyzing the total number of termination interaction problems that occur in a complete cycle, the temporal distribution characteristics of various problems, and their proximity to the system's security threshold (risk structure), the breadth and granularity of the current circuit breaker mechanism are evaluated. When the number of termination interaction problems is excessive, densely distributed, or there are a large number of problems that are about to become termination interaction problems, it indicates that the circuit breaker mechanism may be too broad and strict, and a more lenient exemption strategy should be adopted. When termination interaction problems are concentrated, sparsely distributed, and mainly high-risk types far from the threshold, it indicates that the circuit breaker mechanism is accurate and effective, and a strict strategy should be maintained. This achieves macro-level optimization and adjustment of the system's circuit breaker strategy from the perspective of problem type, balancing system protection and the breadth of user service.

[0055] S21 uses the updated data of the interrupted interaction issues to determine the number of interrupted interaction issues and the number of updates on different dates; "Update data for interrupted interaction issues" refers to the complete log set generated by processes S11-S13, recording all users identified as having "interaction interrupted issues." This step pays particular attention to the "issue type" dimension. "Number of interrupted interaction issues" specifically refers to the total number of different interrupted interaction issue types that occurred within the set statistical period; that is, the deduplicated type count, not the number of individual issue instances. "Number of updates on different dates" refers to the distribution of these different interrupted interaction issues based on the date they first appeared or appeared on that day, reflecting the temporal diffusion characteristics of type diversity.

[0056] This step aims to assess the breadth and potential impact of the circuit breaker mechanism from a macro perspective of "problem type diversity." Its significance lies in the fact that the total number of interrupted interactions directly reflects how many different problem scenarios the system's circuit breaker mechanism addresses. A larger number of types means a wider impact from the circuit breaker mechanism, potentially leading to greater restrictions on diverse user needs. Analyzing the distribution of types across different dates reveals whether this diversity is a slow accumulation or a concentrated eruption, helping to determine whether the circuit breaker rules are overly sensitive and constantly blocking new types, or whether changes in system load patterns have led to the emergence of new problem types. This is the foundation for understanding the overall impact of the circuit breaker mechanism from the perspective of problem type, reflecting the breadth of the strategy more effectively than simply counting the number of instances.

[0057] S22 determines the difference between the probability threshold corresponding to the identification method and the identification deviation probability of the problem based on the identification deviation of different problems and large models and the matching data of the identification method, and uses the difference as the deviation probability difference. The "probability threshold corresponding to the identification method" is the dynamic probability threshold used in step S13 when the system determines that a certain type of problem should be terminated. The "probability of problem identification deviation" refers to the inherent historical average deviation rate of each specific problem type (e.g., "calculating the safety factor Fs"). The "deviation probability difference" is calculated for each terminated interaction problem, subtracting the probability threshold that triggers its termination from the identification deviation probability of that type of problem. This difference measures the "strictness" or "risk buffer" of the problem being circuit-broken; a smaller difference indicates that the problem is closer to the system's safety boundary and thus has a higher risk of developing into a terminated interaction problem later.

[0058] This step aims to deeply analyze the inherent risk characteristics of the various types of problems that were suspended and the rationality of the suspension decision. Its significance lies in identifying which types of problems have a lower risk of developing into suspended interaction problems later (larger difference), and which have a higher risk of developing into suspended interaction problems later (smaller difference). If a large number of suspended problem types have very small deviation probability differences (e.g., <5%), it indicates that the risk of developing into suspended interaction problems later is higher, thus requiring a greater degree of risk control.

[0059] S23 determines the user's termination decision strategy based on the number of questions requiring interaction termination, the number of updates on different dates, and the difference in the deviation probability of different questions.

[0060] It is understood that the above steps include the following: S231 Obtain the number of the interrupted interaction issues, and determine whether the number of the interrupted interaction issues is greater than a preset threshold for the number of interaction issues. If yes, determine that the interruption decision strategy for all users is the first interruption strategy. If no, proceed to step S232. The "Preset Threshold for the Number of Interactive Question Types" is a pre-set critical value used to determine whether the total number of interactive questions that need to be stopped has reached a "high breadth" level that requires special attention.

[0061] This is the most direct trigger point for intervention. If the number of interrupted interactions within a cycle is very large (exceeding this threshold), it indicates that the circuit breaker mechanism has a very wide impact, which may severely limit the types of services available to users and pose a high risk of impaired user experience. In this case, no more complex analysis is needed; the most lenient strategy (first-line interruption strategy) should be adopted directly to quickly alleviate this widespread limitation and avoid user churn or a significant drop in satisfaction due to delayed strategy adjustments.

[0062] The preset threshold is 15 types. The total number of identified interaction termination issues is 12, which does not exceed the threshold (12 < 15). Therefore, the first strategy is not triggered directly, and the process proceeds to the next step S232.

[0063] S232 determines whether the number of interrupted interaction issues is within a preset number range. If yes, proceed to step S233. If no, determine that the interruption decision strategy for all users is the third interruption strategy. The "preset quantity range" is a range value used to classify different levels of the total number of interactive issues that have been terminated. If the number is within this range, it indicates that the number of types is at a "medium breadth" level and requires further analysis; if it is below this range, it indicates that the number of types is small and the impact is limited; if it is above this range, it has already been handled in S231.

[0064] This step provides a refined entry point for handling "moderately broad" scenarios. It avoids simply categorizing all situations that do not reach high thresholds into the same group. When the number of termination types is at a moderate level, the impact can be quite complex, requiring a comprehensive assessment based on time distribution and risk structure to make more precise strategy choices.

[0065] The preset quantity range is [8, 15]. The total number of termination types, 12, falls within this range, therefore proceeding to the next step S233 for in-depth analysis.

[0066] S233. Based on the number of updates to the interrupted interaction problem on different dates, determine whether the proportion of the number of dates with interrupted interaction problem updates is greater than a preset threshold for the proportion of update dates. If yes, determine that the user's interruption decision strategy is the first interruption strategy. If no, proceed to step S234. Determining the time distribution density of the interaction termination problem: "Average number of updates on different dates" refers to the average number of aborted interactive issues added within the statistical period. "Preset update date number percentage threshold" is a critical value used to determine whether aborted interactive issues are "updated frequently".

[0067] This step assesses the frequency and rate of updates to aborted interaction issues. A high update frequency means that new aborted interaction issues are generated almost daily. A high update frequency of aborted interaction issues leads to a greater impact of these issues on the user's interaction process, resulting in poor experience continuity and strong frustration. In this case, even if the total number of types does not reach the maximum threshold, the experience damage caused by the dense distribution can still be significant, warranting intervention with a lenient strategy.

[0068] Calculations show that there were 10 days in September with update issues that resulted in interrupted interaction, representing 10 / 30 = 0.33 of the total number of such days. The preset threshold for the percentage of update days is 0.3 types / day. Since 0.33 > 0.3, the condition is met, and therefore the first interruption strategy is adopted for all users. This means that although the total number of types (12) did not exceed the highest threshold, their dense occurrence triggered a lenient strategy.

[0069] S234 uses the deviation probability threshold to identify problems where the deviation probability threshold is less than the preset deviation probability value as the risk problems of termination identification. It determines whether the proportion of the number of interactions of the risk problems of termination of interaction on different dates is greater than the preset proportion threshold. If so, the user's termination decision strategy is determined to be the first termination strategy. If not, proceed to step S235. S235 determines the total percentage by the percentage of interactions with the risk of termination on different dates and the sum of the percentages of interactions with the risk of termination on different dates, and determines the user's termination decision strategy based on the total percentage.

[0070] Sub-steps S234 & S235: Comprehensive judgment based on the proportion of risk types: When the update frequency of interaction termination issues is low, the system enters a deeper analysis. "Abortion Identification Risk Issues" refer to marginal interaction termination issues where the deviation probability difference is less than the "deviation probability preset value (e.g., 0.05)". "Interaction Count Percentage" refers to the proportion of this type of marginal termination to the total number of user interactions (or total terminations). The "Preset Percentage Threshold" and "Percentage Preset Value" are critical values ​​used to assess the depth of impact of marginal risks. "Total Percentage" is a comprehensive indicator combining the percentage of marginal types and the overall termination percentage.

[0071] These sub-steps aim to more precisely quantify the actual depth of the impact of "edge-level circuit breakers" on user experience. Even if the update frequency of interrupted interaction issues is not high, there are still a large number of issues at the edge of the circuit breaker, so the probability of new interrupted interaction issues arising later is higher, thus the risk of impacting users is higher, and the experience remains poor. By checking whether the proportion of edge-level interruptions remains high (S234), or calculating the comprehensive impact index (S235), it is possible to more accurately determine whether a lenient strategy is still needed to compensate for this part of the user experience loss.

[0072] It is understandable that the system determines whether the total percentage is greater than a preset percentage value. If so, the user's termination decision strategy is determined to be the first termination strategy; otherwise, the user's termination decision strategy is determined to be the second termination strategy.

[0073] It should be noted that the first termination strategy is that if the proportion of interactions caused by the termination problem in the user's most recent preset number of interactions (e.g., 10 times) is greater than a preset proportion threshold (e.g., 0.3), then when the user recognizes the termination problem, the interaction processing will no longer be terminated, and the interaction processing will continue to be carried out using the large model.

[0074] It should be noted that the second termination strategy is that if the proportion of interactions caused by the termination problem in the user's recent preset number of interactions is greater than the second preset proportion threshold (0.2), then when the user recognizes the termination problem, the interaction processing will no longer be terminated, and the interaction processing will continue to be carried out using the large model.

[0075] It should be noted that the third termination strategy is as follows: if the proportion of interactions caused by the termination problem in the user's recent preset number of interactions is greater than the third preset proportion threshold (0.1), then when the user recognizes the termination problem, the interaction processing will no longer be terminated, and the interaction processing will continue to be carried out using the large model.

[0076] It is understood that the preset ratio threshold is greater than the second preset ratio threshold, and the second preset ratio threshold is greater than the third preset ratio threshold.

[0077] Specifically, the method for determining the identification and processing target in the aforementioned interaction termination problem is as follows: The core objective of this embodiment is to identify and filter out interaction problem types that might be "misjudged" by the system as requiring termination, but which the large model is actually capable of reliably resolving. These are prioritized for verification and optimization. The core logic is that not all types judged by the system as "interaction termination problems" are unsolvable by the large model. By analyzing two key data points: the degree of user impact (the frequency with which users are affected by a certain type of termination problem) and the problem-solving potential (the actual performance of the large model in resolving this type of problem in a few exempted or specially handled cases), the system can identify problem types that, although terminated, may have high resolution potential and a wide impact on users. These types are listed as "identification and processing targets" for specific verification. If the verification passes (confirming that the large model can reliably resolve them), they are removed from the "interaction termination problem" list, thereby dynamically optimizing the circuit breaker rules. While ensuring system stability, this gradually expands the effective service scope of the system and improves the overall user experience.

[0078] S31 uses the user's interaction processing data in the interaction termination problem to determine the number of interactions with abnormal feedback results for different users in the interaction termination problem, and takes them as the number of abnormal interactions; "User interaction processing data in cases of aborted interaction" refers to the user interaction history logs recorded by the system, specifically the event data that was previously identified as an "interaction aborted issue," particularly user feedback data that was not aborted (e.g., due to the execution of an exemption policy) or was processed through other channels. "Abnormal feedback result" specifically refers to users who, after processing the issue, explicitly stated that "there is a problem," "unresolved," or gave a negative evaluation. "Number of abnormal interactions" refers to the total number of times all users gave negative feedback after processing a specific type of aborted interaction issue.

[0079] This step aims to gather limited but valuable evidence of the "handling effect" on the "interaction termination problem." Its significance lies in the fact that while a certain type of problem is typically terminated, there are always a few exceptions (such as when a user triggers an exemption) where the large model attempts to handle it. User feedback in these cases provides direct evidence of the large model's potential to solve the problem. If a certain type of termination problem receives very few instances of "abnormal" user feedback after handling it, it indicates that the model performs reasonably well in handling that type of problem, and the current termination decision may be overly conservative, warranting further verification.

[0080] S32 uses the interaction termination data of different users to determine the number of interaction processing times of the user with interaction termination problems, and uses the proportion of the number of interaction processing times of the user with interaction termination problems to determine the proportion of the user's termination impact. "Number of interactions where a user's interaction was interrupted" refers to the total number of times a user's submitted requests were determined by the system to fall under any type of "interaction interruption issue" within the statistical period. "Percentage of impact from interruption" is the proportion of this number to the user's total number of interactions, calculated as: Percentage of impact from interruption = Number of interrupted interactions / Total number of interactions. This metric quantifies the intensity of the circuit breaker mechanism's interference with a single user's daily user experience.

[0081] This step aims to quantify the impact of the circuit breaker mechanism from an individual user perspective. Its significance lies in identifying which users are "heavily affected" by the circuit breaker mechanism. If a user's majority of interactions are blocked due to various interruption issues, their workflow and experience will be severely impacted. Simultaneously, combining user-level impact data with problem type-level resolution potential data (S31) helps prioritize identifying problem types that affect a large number of users and are potentially prone to misjudgment, ensuring that optimized resources are invested in directions that best improve overall satisfaction.

[0082] S33 determines whether the interruption of interaction is a target for identification and processing based on the proportion of interruption impact among different users and the number of abnormal interactions among different users.

[0083] "Identified Processing Target" refers to the "Interaction Aborted Issue" type that has been screened by the system and requires a dedicated verification process. The purpose of verification is to confirm whether the large model can reliably solve this type of problem. If the verification passes, the issue type will be removed from the "Interaction Aborted Issue" list and will no longer be aborted by default in the future; if the verification fails, the aborted status will remain.

[0084] This step is the decision-making hub, responsible for translating impact analysis into concrete optimization actions. Its significance lies in establishing a data-driven, prudent problem-type "unblocking" mechanism. The system will not blindly broaden all termination types, but rather, through a multi-layered filtering system (breadth of user impact, depth of user impact, and indications of problem-solving potential), accurately identify the problem types most likely to be successfully "unblocked" and that will bring the greatest user experience benefits after unblocking. This allows for the effective expansion of the system's service capabilities with minimal risk and cost.

[0085] Specifically, determining whether the interrupted interaction issue is a target for identification and processing includes: S331 determines whether the percentage of the user's interruption impact is less than the preset impact percentage threshold. If so, it is determined that all interruption interaction issues are not the target of identification and processing. If not, proceed to the next step. The "Preset Impact Percentage Threshold" is a percentage value used to determine whether the impact of the suspension on a single user is "minor".

[0086] This step serves as the first filter, aiming to quickly eliminate situations that have minimal impact on any user. If the probability of all users being affected by the interruption is extremely low (the impact percentage is less than the threshold), then the user experience does not need optimization. In this case, investing resources for verification is not cost-effective, so it can be directly determined that all interruption interaction issues are not priority optimization targets, and resources should be reserved for other more important types.

[0087] Specific example: The impact of interrupted interaction on all users is 0.1% (preset threshold: 5%). This means that interrupted interaction has almost no impact on any user's daily use. The system determines that all interrupted interaction issues are not within the target of identification and processing, and will not initiate verification at this time.

[0088] S332 identifies users whose interruption impact ratio is not less than a preset impact ratio threshold as affected users, and determines whether the proportion of affected users among users interacting with the agent is less than a preset impact user ratio threshold. If so, it is determined that all interruption interaction issues do not belong to the identification and processing target; otherwise, it proceeds to the next step. "Affected users" refers to users whose "percentage of impact aborted" is not lower than a preset threshold. "Percentage of affected users" refers to the proportion of such users to the total number of users interacting with the agent. The "preset threshold for percentage of affected users" is used to determine whether the breadth of affected users is sufficiently large.

[0089] This step assesses the "scope of impact" of the problem. Even if some users are severely affected (not filtered out by S331), if the number of such users is very small (the proportion is below the threshold), it indicates that the problem may only affect a small number of users with special needs or specific roles. Although the impact on the individual experience of these users is significant, from the perspective of the overall system benefits, the impact on the overall users is relatively small. Therefore, to avoid delays in interaction processing, all interrupted interaction issues are determined not to be included in the identification and processing targets.

[0090] In one possible implementation, if the percentage of users affected by the interruption is less than 1% (with a preset threshold of 2%), the impact is considered significant. The system determines that the scope of affected users is too narrow and will not initially target all interrupted interaction issues for identification and processing.

[0091] S333 Based on the number of abnormal interactions in different users regarding the interrupted interaction problem, determine the total number of abnormal interactions for the interrupted interaction problem, and determine whether the proportion of the total number of abnormal interactions in the interrupted interaction problem is less than a preset proportion threshold. If so, determine that the interrupted interaction problem belongs to the identification and processing target, that is, determine whether it can be reliably resolved, and thus remove it from the interrupted interaction problem. If not, proceed to the next step. "Total number of abnormal interactions" refers to the total number of negative feedback messages from all users regarding the issue of aborted interaction, as counted in S31. "Interaction frequency ratio" refers to the proportion of this total number of abnormal interactions to all interactions of this issue type that were actually processed (not aborted). "Preset ratio threshold" is used to determine whether negative feedback is too frequent.

[0092] This step is crucial for verifying the "solution potential." If, in cases where a certain type of aborted problem is handled in small numbers, the proportion of negative user feedback is very high (exceeding a threshold), for example, exceeding 30%, it indicates that the current reliability of the large model in handling this type of problem is indeed very low, and the previous abortion decision was likely correct. In this case, instead of "unblocking," the abortion strategy should be maintained or even strengthened. Only those problem types with a low proportion of negative feedback (below the threshold) indicate the possibility of "misjudgment" and are worth identifying and further validating as targets for processing.

[0093] Specific example (continuous): The interaction suspension issue "B" was suspended. In the 20 cases of historical exemption processing, the number of times users reported "there is a problem" was only 2 (abnormal interaction ratio = 10%), which is lower than the preset threshold (e.g., 25%). This indicates that the model performs reasonably well in handling this type of problem and has the initial potential to "unblock" it. Therefore, the interaction suspension issue "B" was determined to be an identification and processing target.

[0094] S334 determines the identification requirement coefficient of the interrupted interaction problem by taking the proportion of the total number of abnormal interactions in the interrupted interaction problem, and combining it with the proportion of users who affect the interaction between the user and the agent. Based on the identification requirement coefficient, it determines whether the interrupted interaction problem belongs to the identification and processing target.

[0095] It should be noted that the identification demand coefficient is related to the proportion of the total number of abnormal interactions in the interaction termination problem, and in combination with the proportion of users who affect the user's interaction with the agent. The lower the proportion of the total number of abnormal interactions in the interaction termination problem, and the higher the proportion of users who affect the user's interaction with the agent, the larger the identification demand coefficient will be.

[0096] It is understood that when the identification demand coefficient of the interaction termination problem is greater than the preset demand coefficient threshold, the interaction termination problem is determined to be an identification and processing target.

[0097] The "Identification Demand Coefficient" is a comprehensive quantitative indicator used for final consideration when S333 is not passed directly. This coefficient is positively correlated with two factors: 1) The lower the "total number of abnormal interactions as a percentage of the number of interactions in the terminated interaction problem" (i.e., the greater the potential for resolution), the higher the demand coefficient; 2) The higher the "proportion of users affected" (i.e., the wider the scope of impact), the higher the demand coefficient.

[0098] This step deals with "gray area" cases. Some issue types have a negative feedback rate slightly higher than the S333 threshold, but affect a very wide range of users; or their feedback rate is low, but the impact on users is not particularly wide. A single threshold may not be sufficient for optimal decision-making. The "Identify Demand Coefficient" provides a more comprehensive evaluation perspective by weighting and combining the two dimensions of "Solution Potential" and "Breadth of Impact." A higher demand coefficient means that "unblocking" it may bring greater overall user experience benefits (because more people are affected) and the risks are relatively controllable (because the solution potential is still acceptable), thus making it more worthy of being listed as an identification and processing target for final verification.

[0099] Specific example (continuous): The negative feedback rate for the issue type "T - Automatic Generation of Project Progress Reports" is 20% (slightly higher than the S333 threshold of 18%), but it affects as many as 40% of users (i.e., 40% of project managers frequently encounter this issue and are subsequently suspended). Its identification demand coefficient is calculated as (40% + (1 - 20%)) / 2 = 0.6, which is relatively high and exceeds the preset demand coefficient threshold (e.g., 0.45). The system ultimately determines that, given its extremely wide impact and the only slightly high negative feedback rate, it is still listed as an identification and processing target, and a special verification process is initiated.

[0100] In this application, dynamic optimization and refinement of the circuit breaker rules are achieved: This method makes the "interaction suspension issue" list no longer a static blacklist, but a "list to be observed" that can be dynamically adjusted based on the actual processing effect and user impact data. During the iterative update process, the large model can actively identify and remove "false judgment" items, making the circuit breaker mechanism more and more accurate.

[0101] Intelligent expansion of service capability boundaries driven by data: By systematically verifying and recovering problem types that have the potential to be reliably solved, the system gradually and safely expands its effective service scope while ensuring stability, thereby enhancing the overall capabilities and competitiveness of the product.

[0102] Maximizing the return on investment for improving user experience: Through multi-level screening (depth of impact, breadth of impact, and potential for resolution), we ensure that limited verification and optimization resources are concentrated on the types of issues that, once "unblocked," will bring significant experience improvements to the widest range of users, thus achieving optimal resource allocation.

[0103] Constructing a closed-loop feedback for model capability assessment: The verification process of identifying the processing target itself is a concentrated assessment of the large model's capabilities in a specific domain. Regardless of whether the verification result is pass (unblocking) or fail (maintaining and stopping), it provides clear direction and high-quality feedback data for iterative optimization of the model (such as targeted supplementation of training data).

[0104] Enhance the system's self-learning and adaptive capabilities: The entire process forms a complete closed loop of "execute circuit breaker -> collect feedback -> analyze impact -> verify potential -> optimize rules", enabling the system to continuously improve itself and adapt to business changes, evolving from a "static rule enforcer" to a "dynamic rule optimizer".

[0105] Specifically, the method for determining the interaction management strategy of the large model in the identification and processing target is as follows: This embodiment develops differentiated interaction management strategies for "identification and processing targets" selected from "interaction termination issues" to securely and efficiently verify the reliable problem-solving capabilities of a large model on these target issues. Its core logic is that different "identification and processing targets" vary in proportion to the overall termination issues, the range of users who have not attempted them, and the degree of overlap between these unattended issues. This reflects the complexity, priority, and potential risks of the verification work. This method intelligently allocates different verification strategies by quantitatively analyzing these characteristics: for issues with a small proportion or high user overlap, a fast and broad verification strategy (first preset duration) is adopted to quickly collect data; for issues with a large proportion and many verified users (few unattended users), a robust and in-depth verification strategy (second preset duration) is adopted to ensure the reliability of the conclusions. This achieves scientific scheduling of verification resources, maximizing the security of the verification process and the credibility of the conclusions while controlling verification costs.

[0106] S41 Based on the identification and processing target, determine the proportion of the identification and processing target in the interaction termination problem, and use the proportion of the identification and processing target in the interaction termination problem as the identification proportion; "Identification Ratio" refers to the proportion of the number of issue types identified as "Identification and Processing Targets" out of the total number of "Interaction Termination Issues". This metric reflects the relative size of the target scope that needs to be verified and optimized relative to the entire circuit breaker boundary.

[0107] This step aims to assess the scale and urgency of the verification work from a macro perspective. Its significance lies in the fact that a high identification rate indicates the system believes many currently suspended problem types may be "false positives," resulting in a massive verification workload and a wide range of potential changes to the existing circuit breaker mechanism, requiring a more cautious, phased approach. Conversely, a low identification rate indicates a concentrated set of verification targets, allowing for a more proactive and rapid strategy to quickly complete the optimization loop and promptly narrow the scope of circuit breakers.

[0108] S42 identifies users who have not interacted with the large model in the identification processing targets as non-interactive users, and determines the associated identification targets for different non-interactive users based on the overlap data between the non-interactive users and other identification processing targets. "Non-interactive users" refer to users who have never actually interacted with the large model to process a specific "identification processing target" problem in the historical records. "Associated identification target" means that for a given non-interactive user, they simultaneously belong to multiple sets of non-interactive users for different "identification processing targets," and these targets are associated with each other because of that user.

[0109] This step aims to build a two-dimensional analytical view of "issue-user," with a particular focus on "uncovered" user areas. Its significance lies in identifying which users represent "blank spots" in validation data and the degree of overlap in user coverage across different validation objectives. If a large number of users haven't tried multiple objectives, this user group represents a valuable source of validation data, but simultaneously pushing multiple new unblocking issues to them could be risky. Analyzing this overlap helps design more reasonable user sampling and issue push strategies, avoiding subjecting individual users to excessive uncertainty.

[0110] S43 determines the interaction management strategy of the large model in the identification processing target based on the identification ratio, the unidentified users of the identification processing target, and the associated identification targets of different uninterrupted users.

[0111] The "Interaction Management Strategy of Large Model in Identifying and Processing Targets" defines how the system should arrange the subsequent actual interaction verification process for a certain type or group of identification and processing targets. The core variables are the verification window period (first or second preset duration) and the triggering conditions (whether limited to non-interactive users). The purpose of the strategy is to allow the large model to actually process the target problem under specified conditions and collect the processing results and feedback to evaluate its reliable resolution.

[0112] This step serves as the decision engine for validating the plan. Its significance lies in transforming the aforementioned quantitative characteristics into an executable validation plan. By setting different durations (e.g., a "first preset duration" of 1 month, a "second preset duration" of 2 months) and user scopes, the system can flexibly control the intensity, speed, and risk of validation. For high-risk, high-complexity validation objectives, a long-term, limited-user strategy is adopted, prioritizing stability; for low-risk, well-defined objectives, a short-term, broad-based strategy is employed to achieve rapid results. This ensures that the entire validation optimization process is controllable, gradual, and risk-aware.

[0113] It should be noted that the associated identification target is the identification processing target that is simultaneously a non-interactive user among the non-interactive users.

[0114] It is understandable that if the recognition ratio is less than the preset recognition ratio threshold, then because there are a large number of recognition and processing targets, for all recognition and processing targets, that is, within the first preset time period in the future, as long as the large model is recognized, the large model will be used for interactive processing to determine the reliability of the recognition and processing of the large model.

[0115] Additionally, the following content should also be noted: S431 If the recognition ratio is not less than the preset recognition ratio threshold, then determine the unrecognized user data of the recognition processing target, and determine whether the proportion of the unrecognized users of the recognition processing target among all users is less than the preset user proportion threshold. If so, then for the recognition processing target, in its historical recognition processing results, the number of users who performed interactive processing was relatively large, so the number of unrecognized users was relatively small, resulting in a low proportion. However, at the same time, its reliability is relatively high. That is, the interaction management strategy of the large model in the recognition processing target is a reliable interaction processing strategy. That is, within the second preset time period in the future, as long as the large model is recognized, the large model will be used for interactive processing, thereby determining the reliability of the recognition processing of the large model. If not, proceed to step S432. It should be noted that the second preset duration is longer than the first preset duration.

[0116] Preliminary decision-making based on recognition ratio and judgment of non-interactive user ratio: The "preset recognition ratio threshold" is the critical value for judging whether the number of targets to be recognized and processed is "too many" or "too few". The "preset user ratio threshold" is the critical value for judging whether the range of non-interactive users is "wide" or "narrow". The "reliable interaction processing strategy" is a more lenient and proactive verification strategy, which usually triggers verification for a wider range of users within a longer verification window period (the second preset duration).

[0117] This step is the first branch of the strategy decision tree. First, if there are few identified targets (the proportion is less than the threshold), it means that even if all identified targets are optimized and no longer treated as interaction termination issues, the impact on users will be minimal. The fastest strategy (short cycle, extensive validation) can be adopted to complete the process quickly. Second, if there are many identified targets, it is necessary to further examine the "popularity" of a specific target: if a target has already been tried by many users with good results (indicated by a very low proportion of non-interactive users), then although it is a newly added target, its historical implicit reliability data is already sufficient. A proactive and slightly longer-term "reliable interaction processing strategy" can be adopted for final confirmation. This reflects respect for and utilization of historical data.

[0118] Specific example: The system has 100 types of abort issues, of which 5 types are identified as targets for processing, with an identification rate of 5%. Since 5% is less than a preset threshold (e.g., 10%), the system decides: for all 5 types of targets, within the next month (the first preset duration), as long as the large model identifies the relevant request, it will be processed directly to quickly collect verification data.

[0119] S432 determines whether the proportion of the identification processing target of the reliable interaction processing strategy in all identification processing targets is greater than the preset identification target proportion threshold. If so, for the identification processing target, the interaction management strategy of the large model in the identification processing target is other interaction processing strategies. That is, within the first preset time period in the future, as long as the large model identifies the identification processing target among the unidentified users of the identification processing target, the large model will be used for interaction processing to determine the reliability of the identification processing of the large model. If not, proceed to step S433. Decision-making based on the proportion of reliable targets: "Reliable interaction processing strategy identification targets" refers to those targets that are determined in S431 to be eligible for this strategy. "Preset identification target proportion threshold" is used to determine whether such "reliable" targets dominate among all identified targets.

[0120] This step deals with "mixed cases." When some targets are considered "reliable" and can be handled with long-term strategies, but the proportion of these targets is not high (not exceeding the threshold), it indicates that the reliability of the overall target group's interaction processing reliability is poor under the current interaction management strategy. In this case, it is necessary to proceed to the next step.

[0121] Assuming the identification rate is 15% (above the threshold), proceed to step S431. Analyze target TA; its non-interactive user ratio is only 8% (below the user ratio threshold), so a "reliable interaction processing strategy" (long cycle) is initially adopted for TA. However, further statistics show that "reliable" targets like TA only account for 30% of all identified targets, which does not exceed the preset ratio threshold (e.g., 50%). Therefore, for the identified targets that have been excluded from the reliable interaction processing strategy, proceed to the next step for further analysis.

[0122] S433 determines the average number of target overlaps of the identification processing target based on the average number of associated identification targets of the identification processing target in different non-interactive users, and determines the interaction management strategy of the large model in the identification processing target based on the average number of target overlaps.

[0123] It should be noted that if the average number of overlapping targets is greater than a preset overlap threshold, then for the identification and processing target, the interaction management strategy of the large model in the identification and processing target is another interaction processing strategy. That is, within a first preset time period in the future, as long as the large model identifies the identification and processing target among the unidentified users of the identification and processing target, it will use the large model for interaction processing to determine the reliability of the large model's identification and processing. Otherwise, for the identification and processing target, the interaction management strategy of the large model in the identification and processing target is a second interaction processing strategy. That is, within a second preset time period in the future, as long as the large model identifies the identification and processing target among the unidentified users of the identification and processing target, it will use the large model for interaction processing to determine the reliability of the large model's identification and processing.

[0124] "Average Target Overlap Count" calculates, on average, how many other target non-interactive users also represent a specific target. For example, if target TB has 100 non-interactive users, and these users are also on average in the non-interactive user lists of two other targets, then its average target overlap count is 2. The "Preset Overlap Count Threshold" is used to determine whether this overlap level is "high."

[0125] This step is the final and most refined decision-making level, focusing on verifying "user-side risks." If a target's uninteracted users also happen to be uninteracted users of many other targets, it means that when verification requests are pushed to these users, they may simultaneously face multiple "new unblocking" issues, increasing the uncertainty and potential risks of their experience. For such highly overlapping targets, the system should adopt a more conservative strategy (short cycle) to quickly pass the verification period and reduce the long-term impact on this user group. For targets with low overlap, whose uninteracted users are relatively "specific" and the risks are more dispersed, a longer-term strategy can be adopted for more robust observation.

[0126] Continuing with the previous example, we analyze the target TC. We calculate that, on average, non-interactive users are associated with 1.8 other identified targets. The preset overlap threshold is 2.0. Since 1.8 < 2.0, the system determines that the target TC has a low degree of association overlap. Therefore, a second interaction processing strategy is adopted for TCs: within the next two months (the second preset duration), the large model will only be invoked when these non-interactive users raise TC-type questions, ensuring the breadth of verification processing for the large model's recognition reliability across different user groups.

[0127] Example 2 Secondly, such as Figure 4As shown, this application provides an intelligent interactive management and decision-making system based on a large model, employing the aforementioned intelligent interactive management and decision-making method based on a large model, specifically including: Stop control module, identification and processing module, interactive association module; The termination control module is responsible for determining the user's termination decision strategy. The identification processing module is responsible for determining the identification processing target in the termination interaction problem; The interaction association module is responsible for determining the interaction management strategy of the large model in the recognition and processing target.

[0128] 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.

[0129] 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.

[0130] 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 method for intelligent interactive management and decision-making based on a large model, characterized in that, Specifically, it includes: Based on the interaction data of the large model, the shortage period of the interaction port of the large model is determined. Based on the distribution data of the shortage period, the identification method of the large model's interaction termination problem is determined. Based on the identification method, the interaction termination problem is updated. Based on the updated data of the interaction termination problem, and combined with the identification deviation of the large model for different problems and the matching data of the identification method, the user's termination decision strategy is determined. Based on the aforementioned termination decision strategy, termination management is performed for different user interaction issues. Based on the interaction processing data of different users in the termination interaction issues, and combined with the termination interaction data of different users, the identification and processing targets in the termination interaction issues are determined. Based on the identification and processing target, and combined with the similarity of unidentified users to different identification and processing targets, the interaction management strategy of the large model in the identification and processing target is determined.

2. The intelligent interactive management and decision-making method based on a large model as described in claim 1, characterized in that, The interaction data of the large model includes the number of idle interaction ports of the large model in different time periods.

3. The intelligent interactive management and decision-making method based on a large model as described in claim 1, characterized in that, The shortage period of the interaction port in the large model is the period when there is no idle interaction port.

4. The intelligent interactive management and decision-making method based on a large model as described in claim 1, characterized in that, The issue of terminating interaction refers to the issue of directly terminating interaction.

5. The intelligent interactive management and decision-making method based on a large model as described in claim 1, characterized in that, The method for determining the identification method of the large model's interaction termination problem is as follows: Based on the distribution data of the aforementioned shortage periods, the duration of different shortage periods is determined; Based on the duration, the shortage period with a duration greater than a preset duration threshold is determined and taken as the period of shortage impact; Based on the shortage impact period and shortage period data within the most recent preset time period, a method for identifying the interruption of interaction problems in the large model is determined.

6. The intelligent interactive management and decision-making method based on a large model as described in claim 5, characterized in that, If the number of shortage-affected periods within the most recent preset time period is greater than the preset threshold for the number of affected periods, then the problem will be treated as an interaction termination problem as long as the identification deviation probability of the problem is greater than the preset probability threshold.

7. The intelligent interactive management and decision-making method based on a large model as described in claim 1, characterized in that, The method for determining the user's termination decision strategy is as follows: Based on the updated data of the interrupted interaction issues, determine the number of interrupted interaction issues and the number of updates on different dates; Based on the identification deviation of different problems and large models and the matching data of the identification method, the difference between the probability threshold corresponding to the identification method and the identification deviation probability of the problem is determined, and the difference is used as the deviation probability difference. Based on the number of questions that lead to the termination of interaction, the number of updates on different dates, and the difference in the probability of deviation for different questions, the user's termination decision strategy is determined.

8. The intelligent interactive management and decision-making method based on a large model as described in claim 7, characterized in that, The number of questions that would cause the user to stop interacting is obtained. When the number of questions that would cause the user to stop interacting is greater than a preset threshold for the number of questions that would cause the user to stop interacting, the user's decision-making strategy is determined to be the first stop strategy.

9. The intelligent interactive management and decision-making method based on a large model as described in claim 1, characterized in that, The method for determining the interaction management strategy of the large model in the identification and processing target is as follows: Based on the identification and processing target, determine the proportion of the identification and processing target in the interaction termination problem, and use the proportion of the identification and processing target in the interaction termination problem as the identification ratio; Users who do not interact with the large model in the identification and processing targets are identified as non-interactive users. Based on the overlap data between the non-interactive users and other identification and processing targets, the associated identification targets of different non-interactive users are determined. Based on the identification ratio, the unidentified users of the identification processing target, and the associated identification targets of different non-interactive users, the interaction management strategy of the large model in the identification processing target is determined.

10. A large-scale model-based intelligent interactive management and decision-making system, employing the large-scale model-based intelligent interactive management and decision-making method described in any one of claims 1-9, characterized in that, Specifically, it includes: Stop control module, identification and processing module, interactive association module; The termination control module is responsible for determining the user's termination decision strategy. The identification processing module is responsible for determining the identification processing target in the termination interaction problem; The interaction association module is responsible for determining the interaction management strategy of the large model in the recognition and processing target.

Images