Large model complaint data query strategy optimization processing method and system

Abstract

The invention discloses a big model complaint data query strategy optimization processing method and system, and relates to the technical field of data management, and the method comprises the steps: receiving a big model complaint data query request, generating an initial query intention feature vector, and obtaining a big model complaint data query strategy based on the interactive feedback of the initial query intention feature vector and a knowledge graph; generating a standardized query intention feature vector; matching the standardized query intention feature vector with a strategy template library to generate a preliminary weighted strategy recommendation abstract, and performing exploratory enhancement according to a confidence evaluation result of the preliminary weighted strategy recommendation abstract to generate a final weighted strategy recommendation abstract; and generating candidate meta query strategy descriptions based on the final weighted strategy recommendation abstract. According to the invention, weight adjustment and experience correction are carried out on the strategy template library according to the actual effect data of the case, an intelligent strategy optimization report is generated, and autonomous optimization and continuous evolution of the query strategy are realized.

Description

Technical Field

[0001] This invention relates to the field of data management technology, and in particular to a method and system for optimizing query strategies for large-scale complaint data. Background Technology

[0002] With the widespread application of large language models, the need for security supervision of the content they generate has led to massive and heterogeneous complaint data, scattered across multiple data sources such as logs and annotation platforms. Current technical solutions build retrieval systems based on rule engines and keyword indexes. Through predefined query templates and static rules, the query intent of analysts is compiled into specific instructions for the database, improving the level of intelligence. Machine learning classification models are introduced to pre-classify complaint content to assist in generating query conditions. This can effectively handle known and fixed-pattern complaint query tasks, and achieve preliminary structured retrieval of complaint data.

[0003] Existing technologies have limitations in addressing the dynamic evolution and exploratory needs of large-scale complaint queries. These limitations are mainly reflected in the fact that the generation and optimization of query strategies rely heavily on static prior knowledge, lacking the ability for autonomous evolution and closed-loop optimization. Existing systems are also insufficient in understanding complex, ambiguous, or novel query intents, typically only capable of simple template matching or label mapping. They cannot achieve accurate clarification and standardization of intents through deep interaction with domain knowledge graphs, making it difficult to automatically adapt to changes in data distribution and emerging complaint patterns. Furthermore, they cannot achieve dynamic balance among multiple objectives, resulting in unreliable query efficiency and accuracy when facing unknown or rapidly evolving complaint scenarios. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an optimized processing method for querying large-scale complaint data, which solves the problem that existing technologies, due to their reliance on static knowledge, cannot adaptively handle complex and dynamic complaint scenarios, resulting in rigid query strategies and unsustainable effects.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for optimizing the query strategy of large model complaint data, which includes receiving a query request for large model complaint data, generating an initial query intent feature vector, and generating a standardized query intent feature vector based on the interaction feedback between the initial query intent feature vector and the knowledge graph. The standardized query intent feature vector is matched with the strategy template library to generate a preliminary weighted strategy recommendation summary. Based on the confidence evaluation results of the preliminary weighted strategy recommendation summary, exploratory enhancement is performed to generate the final weighted strategy recommendation summary. Based on the final weighted strategy recommendation summary, candidate meta-query strategy descriptions are generated. Through static rule checking and dynamic sandbox simulation, a set of validated candidate strategies is output. Select meta-query strategy descriptions from the verified candidate strategy set, compile them into a distributed query task across data sources, and obtain the initial complaint dataset; Aggregate analysis and extract knowledge from the initial complaint dataset to generate a complaint analysis report, update the dynamic complaint knowledge graph, and form a strategy effect case study; Based on the actual effect data of strategy performance cases, the strategy template library is weighted and revised according to experience, and an intelligent strategy optimization report is generated.

[0007] As a preferred embodiment of the large-scale complaint data query strategy optimization processing method of the present invention, the following steps are included: receiving a large-scale complaint data query request and generating an initial query intent feature vector: The natural language processor receives the query request text of the large model complaint data, performs syntactic and semantic parsing on the large model complaint data query request text, and obtains the results of syntactic and semantic parsing. Based on the results of syntactic and semantic parsing, the core query entities, query targets, and constraints in the large model complaint data query requests are identified. Features are extracted from the core query entities, query targets, and constraints. The extracted features are then vectorized and encoded to generate an initial query intent feature vector.

[0008] As a preferred embodiment of the large-scale complaint data query strategy optimization processing method described in this invention, the following steps are included: Generating a standardized query intent feature vector based on the interaction feedback between the initial query intent feature vector and the knowledge graph: The intent clarification logic unit is used to interactively query the initial query intent feature vector with the dynamic complaint knowledge graph, and to obtain entities and potential ambiguity patterns associated with the initial query intent feature vector from the dynamic complaint knowledge graph. Based on entity and potential ambiguity patterns, the intent to clarify logic unit generates clarifying questions, presents the clarifying questions to the user, and obtains user feedback; By utilizing user feedback, the initial query intent feature vector is corrected and completed to obtain a standardized query intent feature vector.

[0009] As a preferred embodiment of the large-scale complaint data query strategy optimization processing method described in this invention, the method involves: matching standardized query intent feature vectors with a strategy template library to generate a preliminary weighted strategy recommendation summary, including the following steps: The similarity between the standardized query intent feature vector and the query intent feature vector of each historical success case stored in the strategy template library is calculated using a feature matcher. Sort all similarity results, select cases with similarity exceeding a preset threshold for historical success cases, and assign a matching weight to each selected historical success case. The meta-query strategy descriptions of all selected historical success cases are aggregated, and each meta-query strategy description is combined with its corresponding matching weight to form a preliminary weighted strategy recommendation summary.

[0010] As a preferred embodiment of the large-scale complaint data query strategy optimization processing method described in this invention, the method involves: conducting exploratory enhancement based on the confidence assessment results of the preliminary weighted strategy recommendation summary to generate the final weighted strategy recommendation summary, including the following steps: The confidence assessment unit is used to calculate the average time freshness and scene coverage dispersion of the historical successful cases on which the preliminary weighted strategy recommendation summary is based, and the confidence assessment results are generated. When the confidence assessment result indicates that the confidence level is lower than the preset confidence standard, the exploratory enhancement unit is triggered to query the dynamic complaint knowledge graph based on the standardized query intent feature vector to obtain the potential association structure in the graph; Complementary strategy recommendation hypotheses are synthesized using potential association structures, and exploratory enhancement units assign a preset exploration weight to the complementary strategy recommendation hypotheses; The supplementary policy recommendation hypothesis with exploration weights is combined with the preliminary weighted policy recommendation summary to generate the final weighted policy recommendation summary.

[0011] As a preferred embodiment of the large-scale complaint data query strategy optimization processing method described in this invention, the method includes the following steps: generating candidate meta-query strategy descriptions based on the final weighted strategy recommendation summary, and outputting a set of verified candidate strategies through static rule checking and dynamic sandbox simulation. The final weighted strategy recommendation summary is parsed to obtain the recommended data source combination and processing flow, and candidate meta-query strategy descriptions are generated. The static rule checker examines the logical conflicts and resource constraint violations in each candidate meta-query strategy description, marks the candidate meta-query strategy descriptions that pass the static rule check as candidate strategy descriptions that pass static verification, and generates static verification results. Based on a dynamic sandbox simulator, in a sandbox isolated from the production environment, recent data snapshots are used to tentatively execute candidate strategy descriptions that have passed static validation, evaluate the data access success rate and preliminary effect confidence of each candidate strategy description that has passed static validation, and generate dynamic validation results. Candidate strategy descriptions that pass static validation are marked as dynamically validated candidate strategy descriptions when the data access success rate and initial effect confidence level are simulated in a dynamic sandbox and reach the preset candidate criteria. By combining the states of candidate strategy descriptions that pass static validation and candidate strategy descriptions that pass dynamic validation, the meta-query strategy descriptions of all candidates that meet the conditions are output, and the set of validated candidate strategies is output.

[0012] As a preferred embodiment of the large-scale complaint data query strategy optimization processing method of the present invention, the following steps are included: selecting a meta-query strategy description from the verified candidate strategy set, compiling it into a distributed query task across data sources, and obtaining the initial complaint dataset: The strategy selector sorts the preliminary effect confidence based on the meta-query strategy description of each candidate in the verified candidate strategy set, and selects the meta-query strategy description of the candidate ranked first as the meta-query strategy description. The data source access instructions and logical operations in the meta-query strategy description are parsed, and the parsed data source access instructions and logical operations are converted into specific query statements for the dialogue log database, user feedback table and model inference tracing system. The query compilation engine organizes and coordinates specific query statements from the dialogue log database, user feedback table, and audit trail system into distributed query tasks across data sources. The initial complaint dataset is formed by query results obtained from the dialogue log database, user feedback forms, and model inference tracing system.

[0013] As a preferred embodiment of the large-scale complaint data query strategy optimization method described in this invention, the following steps are included: performing aggregation analysis and knowledge extraction on the initial complaint dataset to generate a complaint analysis report, and updating the dynamic complaint knowledge graph to form a strategy effectiveness case study: Clustering analysis algorithms are used to process the initial complaint dataset, identify complaint clusters, extract common patterns of complaint events from the initial complaint dataset, and use pattern recognition algorithms to scan for repetitive features in the initial complaint dataset. Based on complaint clusters, event sequence associations, and repetitive features, a complaint analysis report is generated, from which new complaint pattern entities, important relationships between entities, and statistical features are extracted. Add new complaint pattern entities, important relationships between entities, and statistical characteristics to the dynamic complaint knowledge graph; The standardized query intent feature vector, meta-query strategy description, and actual recall, precision, and time consumption data recorded in the complaint analysis report are packaged into a strategy effect case.

[0014] As a preferred embodiment of the large-scale complaint data query strategy optimization processing method of the present invention, the following steps are included: The strategy template library is weighted and empirically corrected based on the actual effect data of strategy effect cases, and an intelligent strategy optimization report is generated: The reinforcement learning processor reads the actual recall, precision, and latency data recorded in the policy performance cases, and locates historical successful case records in the policy template library that provide the basis for generating the preliminary weighted policy recommendation summary; The actual performance data is compared with the historical performance data stored in the historical success case records. When the actual performance data is better than the historical performance data, the reinforcement learning processor increases the weight and priority of the corresponding historical success case record. When the actual performance data is worse than the historical performance data, the reinforcement learning processor performs attribution analysis on the difference between the actual performance data and the historical performance data, and adds the attribution conclusion as a note to the corresponding historical success case record. Examine the generation process of strategy effectiveness cases, identify supplementary strategy recommendation hypotheses synthesized by exploratory enhancement units that have been validated in this query, and transform them into new historical success case templates; The report generator receives the final weighted strategy recommendation summary, static verification results, dynamic verification results, meta-query strategy description, complaint analysis report, strategy effect case, and weight adjustment and experience correction records, and compiles them to generate an intelligent strategy optimization report.

[0015] Secondly, the present invention provides a large model complaint data query strategy optimization processing system, including a standardization module, which receives a large model complaint data query request, generates an initial query intent feature vector, and generates a standardized query intent feature vector based on the interaction feedback between the initial query intent feature vector and the knowledge graph. The enhancement generation module matches the standardized query intent feature vector with the strategy template library to generate a preliminary weighted strategy recommendation summary. Based on the confidence evaluation results of the preliminary weighted strategy recommendation summary, exploratory enhancement is performed to generate the final weighted strategy recommendation summary. The candidate set generation module generates meta-query strategy descriptions for candidates based on the final weighted strategy recommendation summary. Through static rule checking and dynamic sandbox simulation, it outputs a set of verified candidate strategies. The execution module selects meta-query strategy descriptions from the verified candidate strategy set, compiles them into a distributed query task across data sources, and obtains the initial complaint dataset. The knowledge update module performs aggregation analysis and knowledge extraction on the initial complaint dataset to generate a complaint analysis report, and updates the dynamic complaint knowledge graph to form strategy effect cases; The optimization module adjusts the weights and makes empirical corrections to the strategy template library based on the actual effect data of strategy performance cases, and generates an intelligent strategy optimization report.

[0016] The beneficial effects of this invention are as follows: It receives query requests and generates an initial query intent feature vector, then combines this with interactive feedback from a dynamic complaint knowledge graph to generate a standardized query intent feature vector; it matches the standardized query intent feature vector with a strategy template library to generate a preliminary weighted strategy recommendation summary, and obtains a final recommendation summary after confidence evaluation and exploratory enhancement; it outputs a set of verified candidate strategies through a two-stage verification process involving static rule checking and dynamic sandbox simulation; it selects a strategy description and compiles it into a distributed query task across data sources to obtain an initial complaint dataset; it performs aggregation analysis and knowledge extraction on the dataset to generate a complaint analysis report and update the knowledge graph, forming strategy effectiveness cases; based on the actual effectiveness data of the cases, it adjusts the weights of the strategy template library and makes empirical corrections, generating an intelligent strategy optimization report, thus achieving autonomous optimization and continuous evolution of the query strategy. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart for optimizing the query strategy for large-scale complaint data.

[0019] Figure 2 A schematic diagram of a system for optimizing complaint data query strategies for large-scale models. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0023] Reference Figures 1-2 This is one embodiment of the present invention, which provides a method for optimizing the query strategy of large model complaint data, including the following steps: S1. Receive large model complaint data query requests and generate initial query intent feature vectors.

[0024] S1.1 Receive the large model complaint data query request text based on the natural language processor. The natural language processor performs syntactic and semantic parsing on the large model complaint data query request text to obtain the results of syntactic and semantic parsing.

[0025] Furthermore, after receiving the query request text for complaint data from the large model, the natural language processor first performs syntactic and semantic analysis on the text. Syntactic analysis identifies the syntactic structure of the text, delineating components such as subject, predicate, and object, as well as the boundaries of various phrases and clauses. For example, when processing a request to find all complaints related to generating false financial information last week, syntactic analysis clarifies that "find" is the core verb, "complaint" is the object, and "last week" and "related to generating false financial information" are attributive components modifying the object. Semantic analysis further understands the deeper meaning and relationships expressed by the syntactic components. By analyzing the semantics of words and context, it identifies "false financial information" as a composite conceptual entity with negative attributes, identifies "last week" as a clear time range constraint, and understands that the query goal expressed by all quantification words is to perform a full retrieval of records that meet the conditions rather than sampling.

[0026] Specifically, by transforming unstructured natural language input into structured, machine-interpretable linguistic representations and deconstructing sentence components and assigning them semantic labels, the core actions, objects, and limiting conditions in the query can be clearly extracted, avoiding misunderstandings caused by linguistic ambiguity or structural complexity in subsequent processing.

[0027] S1.2 Identify the core query entities, query targets, and constraints in the large model complaint data query request based on the results of syntax and semantic parsing.

[0028] Furthermore, the core query entity refers to the specific object or concept pointed to in the query request. For example, in the above parsing results, false financial information and complaints are identified as core query entities. They represent the content topic and record type involved in the complaint, respectively. The query target refers to the operational intent that the request wants to perform. From the search of all combinations, it can be identified that the query target is to perform a full search. The constraints are additional information that limits the scope or results of the query. For example, last week is identified as a time constraint. The generation involved indicates the semantic relationship constraint between the entity false financial information and the model behavior generation. The identification process is based on the structured representation of the semantic parsing output. Through predefined rules or pattern matching, different semantic roles are classified into the above three categories.

[0029] Specifically, linguistic representations are further abstracted into task-oriented logical elements, and complex natural language queries are mapped into a set of well-defined and interconnected metadata. This allows subsequent computer processing to operate directly on structured logical elements without needing to understand the natural language itself, thereby greatly improving the accuracy and operability of the processing and providing a direct basis for the accurate extraction of features.

[0030] S1.3 Extract features from the core query entity, query target and constraints, and vectorize the extracted features to generate the initial query intent feature vector.

[0031] Furthermore, for core query entities, their text identifiers, category classification in a predefined entity type graph, and semantic vector representations obtained through word embedding techniques are extracted. For query targets, their operation type codes are extracted, such as retrieval, statistics, and root cause analysis. For constraints, corresponding features are extracted according to their type, such as time constraints being converted into time interval codes, and relational constraints being converted into a combination of relational predicates and related entities. Heterogeneous features are integrated into a unified, fixed-dimensional numerical vector, i.e., the initial query intent feature vector, through a vectorization encoding process. Vectorization encoding can be performed by concatenation, weighted summation, or fusion through an encoding network.

[0032] Specifically, the structured logical elements are transformed into numerical feature representations that can be directly processed by machine learning models. Based on the refined logical structure obtained after deep analysis, targeted feature extraction and fusion are performed. This enables the generated initial query intent feature vector to more accurately and structurally carry the complete semantic intent of the query, stripping away redundant and noisy information in natural language. The feature construction method based on logical elements provides high-purity and high-information-density input for subsequent interaction and matching with the knowledge graph, which is a key foundation for achieving accurate understanding and standardization of query intent.

[0033] S2. Based on the interaction feedback between the initial query intent feature vector and the knowledge graph, a standardized query intent feature vector is generated.

[0034] S2.1. The intent clarification logic unit is used to perform interactive queries between the initial query intent feature vector and the dynamic complaint knowledge graph, and to obtain entities and potential ambiguity patterns associated with the initial query intent feature vector from the dynamic complaint knowledge graph.

[0035] Furthermore, the intent clarification logic unit uses the received initial query intent feature vector as input to interact with the dynamic complaint knowledge graph. It performs deep comparison and reasoning between the semantic information contained in the initial query intent feature vector and the entities, attributes, and relationship networks in the knowledge graph. The intent clarification logic unit maps the vector to the embedding space of the knowledge graph and searches for graph nodes with close semantic distance, representing related entities. For example, the initial query intent feature vector might imply an intent to commit financial fraud, while the graph returns specific complaint entities related to phishing emails, false investment advice, and identity theft. The intent clarification logic unit analyzes whether the initial query intent feature vector contains components pointing to multiple possible graph paths or with ambiguous concept definitions, thereby identifying potential ambiguity patterns. For example, unstable features output by the model might correspond to contradictory generated content, large fluctuations in response time, or significant performance differences in different contexts—all of which are identified as potential ambiguities requiring clarification.

[0036] Specifically, by leveraging the rich structured domain knowledge base of the dynamic complaint knowledge graph, the initial query intent is deeply expanded and ambiguity explored. By combining feature vectors with the semantic association and ambiguity identification of the graph, not only are clear related entities found, but more importantly, potential ambiguities in the intent are proactively discovered. Precise clarification provides clear goals and evidence, thereby transforming the initial user expression into a series of specific propositions or options that are closely related to domain knowledge and await confirmation, completing the transformation from vague semantics to clarifiable knowledge points.

[0037] S2.2 Based on entity and potential ambiguity patterns, the intent to clarify logic unit generates clarifying questions, presents the clarifying questions to the user, and obtains user feedback.

[0038] Furthermore, based on the associated entities and identified potential ambiguity patterns obtained from the dynamic complaint knowledge graph, the intent clarification logic unit constructs specific clarification questions. The construction of clarification questions is not random or templated, but rather specifically generated based on the strength of entity associations and the nature of ambiguity patterns. For strongly associated entities, clarification questions may aim to confirm whether the user's intent truly points to that specific entity or a class of entities. For identified potential ambiguity patterns, clarification questions aim to guide the user to eliminate ambiguity and make choices. For example, regarding the unstable multiple interpretation paths of the model output, the generated question may focus more on the self-contradictory logic of the generated content, or on the significant fluctuations in the quality of the answers across different questions, and this is presented to the user through the user interface.

[0039] Specifically, a closed loop of human-machine collaboration for precise intent is initiated, transforming the internal uncertainties inferred from the knowledge graph into externally guided questions for the user. Semantic ambiguities that the machine cannot decide autonomously are left to the user's ultimate intent holder for adjudication, respecting the subjectivity of the user's intent while greatly improving clarification efficiency. This is because the questions originate directly from the deep connections and ambiguity analysis of the knowledge graph, rather than general questions, ensuring that each interaction effectively promotes the clarification of intent. Open-ended, potentially lengthy natural language dialogues are optimized into targeted, limited-option rapid confirmation.

[0040] S2.3. Using user feedback, the initial query intent feature vector is corrected and completed to obtain a standardized query intent feature vector.

[0041] Furthermore, after obtaining user feedback on clarification questions, the intent clarification logic unit uses the feedback to modify and complete the initial query intent feature vector. Modification primarily targets potential ambiguity patterns, eliminating ambiguity in corresponding dimensions of the feature vector based on user selections or supplementary information. For example, if a user selects content that is logically contradictory, the feature weights pointing to logical consistency issues are strengthened, while features pointing to performance fluctuations are weakened or eliminated. Completion mainly targets related entities, integrating information about entities closely related to the knowledge graph, confirmed by the user or newly mentioned in the feedback, into the vector as structured features. For example, if a user confirms a false investment suggestion and adds a high-risk commitment, the high-risk commitment is integrated into the vector as a new feature dimension or the values ​​of related existing dimensions are enhanced. The integration process must maintain dimensional consistency and semantic space coherence in the feature vector. Finally, a standardized query intent feature vector is obtained.

[0042] Specifically, the standardization of query intent is achieved, producing a precise, unambiguous, and machine-understandable intent representation rich in verified domain knowledge. This realizes a closed-loop fusion between dynamic human-computer interaction and static feature vector representation. User feedback is not isolated information but is directly used to iteratively optimize the core feature vector. This ensures that the final standardized query intent feature vector not only contains the semantics of the initial parsing but also incorporates accurate information verified by the knowledge graph and confirmed by the user. As a result, it can more reliably drive the subsequent strategy matching and generation process, fundamentally reducing the risk of the entire query process failing due to intent comprehension bias.

[0043] S3. Match the standardized query intent feature vector with the strategy template library to generate a preliminary weighted strategy recommendation summary.

[0044] S3.1 Calculate the similarity between the standardized query intent feature vector and the query intent feature vector of each historical success case stored in the strategy template library using a feature matcher.

[0045] Furthermore, cosine similarity calculation does not directly compare the absolute length of vectors, but measures the alignment of two vectors in direction. This effectively eliminates the impact of differences in feature strength caused by varying levels of detail in query descriptions on the matching results, allowing the matching to focus more on semantic similarity at the intent level. The feature matcher treats the standardized query intent feature vector and the query intent feature vector of each historical successful case as points in a multi-dimensional space, and measures the cosine value of the angle between them through dot product operation. The similarity is at its maximum when the two vectors are completely aligned in direction, zero when the directions are orthogonal, and negative when the directions are opposite. Since feature values ​​are usually non-negative, the actual value range is between zero and one. Based on the measurement method of vector space direction consistency, it can accurately capture the semantic proximity of query intent in feature space. However, as long as their distribution directions in the feature space regarding risk type, involved fields, user impact, etc. are similar, a high similarity score can be obtained.

[0046] Specifically, the current standardized query intent feature vector is placed in a semantic space composed of historical experiences. The most relevant historical success experiences are found through quantified similarity values. By leveraging the characteristic of cosine similarity being sensitive to direction but insensitive to magnitude, a robust measure of the essential similarity of query intent is achieved, and preliminary semantic anchoring from the current query to the historical experience base is completed. The cosine similarity expression is: ; in, The current standardized query intent feature vector With the Query intent feature vectors from historical success cases The cosine similarity value between them For a standardized query intent feature vector, For the first Query intent feature vectors from historical success cases The dimension of the query intent feature vector. For standardized query intent feature vectors In the Eigenvalues ​​in dimension, For the first Query intent feature vectors from historical success cases In the Eigenvalues ​​in dimension, The dimension index of the feature vector. This serves as an index identifier for historical success stories.

[0047] S3.2 Sort all similarity results, select those with similarity exceeding the preset threshold of historical successful cases, and assign a matching weight to each selected historical successful case.

[0048] Furthermore, after obtaining all similarity results, the feature matcher sorts all calculated similarity values ​​in descending order to identify the set of historical successful cases that are closest to the current query intent. The sorting operation arranges all historical cases according to their semantic relevance to the current query, selecting all historical successful cases whose similarity exceeds a preset threshold. This threshold serves as a filter, ensuring that the selected cases have a sufficiently high semantic relevance to the current query and excluding noisy cases that, while having some overlapping features, deviate significantly from the overall intent.

[0049] Specifically, for each selected historical success case, the feature matcher assigns a matching weight based on its similarity value. Generally, the higher the similarity value, the greater the matching weight. This gives historical experiences more similar to the current query intent greater influence in subsequent aggregation processes. A highly relevant subset is selected from massive amounts of historical experience, and differentiated influence weights are assigned to them based on the strength of the relevance.

[0050] Specifically, the precision control of experience retrieval is achieved through sorting and threshold filtering, avoiding interference from irrelevant historical cases. The similarity of continuous values ​​is transformed into operable matching weights. The method of filtering and weighting based on quantified similarity makes the strategy recommendation process not a simple majority or random selection, but a refined experience sampling mechanism that is proportional to the closeness of the current query intent, effectively improving the quality and representativeness of the recommendation basis.

[0051] S3.3 Aggregate the meta-query strategy descriptions of all selected historical success cases, and combine each meta-query strategy description with its corresponding matching weight to form a preliminary weighted strategy recommendation summary.

[0052] Furthermore, the feature matcher aggregates the meta-query strategy descriptions of all selected historical success cases. This aggregation is not a simple listing or splicing; it combines the meta-query strategy description of each historical success case with its matching weight. Each historical success case's meta-query strategy description, whether it's the suggested data source access order, filter condition combination, or analysis algorithm call chain, is weighted according to its corresponding matching weight. Meta-query strategy descriptions of historical success cases with higher weights occupy a more dominant position in the aggregated summary, and their strategy elements are more likely to be retained or reproduced; those with lower weights have weaker influence. The weighted aggregation ultimately forms a preliminary weighted strategy recommendation summary. This summary is not a single strategy, but a comprehensive strategy blueprint with weighted indicators, integrating the essence of strategies from multiple highly relevant historical success cases.

[0053] Specifically, it integrates scattered, fragmented historical success experiences into a coherent set of strategy suggestions tailored to the current query intent, implying priority. This realizes the ensemble learning approach to strategy recommendation. Instead of recommending an optimal historical strategy in isolation, it weights and fuses descriptions of multiple related strategies to generate a new strategy summary that draws on the strengths of various approaches and balances multiple historical success paths. This avoids the adaptability problems that may arise from over-reliance on a single historical experience, enhances the robustness and generalization ability of the recommendation strategy, and ensures that the generated preliminary weighted strategy recommendation summary not only contains the strategy skeleton of what to do, but also implicitly includes weight information on the importance of different approaches.

[0054] S4. Based on the confidence assessment results of the preliminary weighted strategy recommendation summary, exploratory enhancements are performed to generate the final weighted strategy recommendation summary.

[0055] S4.1 Calculate the average time freshness and scene coverage dispersion of the historical successful cases on which the preliminary weighted strategy recommendation summary is based using the confidence assessment unit, and generate the confidence assessment results.

[0056] Furthermore, the confidence assessment unit generates confidence assessment results by calculating the average time freshness and scene coverage dispersion of historical successful cases on which the preliminary weighted strategy recommendation summary relies. The calculation of average time freshness first determines the individual time freshness of each historical successful case based on an exponential decay function, using the difference between the occurrence time of the historical case and the current time as a negative exponent to ensure that the longer the time interval, the more significant the decay of the individual freshness value. Then, the arithmetic mean of the individual time freshness of all relevant cases is taken to obtain the overall average time freshness. The calculation of scene coverage dispersion first extracts the scene feature vector of each relevant historical successful case. Typically, this is obtained by dimensionality reduction of the query intent feature vectors from historical successful cases or by selecting key dimensions. The Euclidean distance between all case pairs is calculated, and the average distance value is taken as the dispersion. The confidence assessment results comprehensively reflect the reliability of the historical experience relied upon in two dimensions: timeliness and scenario diversity. A lower average time freshness means that the relied-upon experience may be too outdated and cannot effectively reflect the latest state of the current data distribution or model behavior. On the other hand, too low or too high scenario coverage dispersion respectively suggests that the experience is too concentrated or too scattered. The former leads to insufficient generalization ability of the recommendation strategy, while the latter means a lack of focused experience for the current specific query scenario.

[0057] Specifically, it conducts quantitative quality diagnosis on the historical experience basis upon which the initial recommendations are based, abandoning single-dimensional assessment and constructing a composite assessment framework that integrates timeliness and diversity. This allows the confidence assessment results to more comprehensively reveal the degree of fit between historical experience and the current query context, thereby achieving an intelligent leap from simply relying on historical matching to conducting self-prudent assessment of matching quality.

[0058] The expression for average time freshness is: ; ; in, For average time freshness, The total number of historical success cases. for, For the first The time freshness of an individual historical success story The current time at the time of confidence assessment. For the first The timestamp of when a historical success case is successfully executed and recorded. The attenuation coefficient is... A traversal index for historical success cases.

[0059] The expression for scene coverage dispersion is: ; ; in, To cover the dispersion of scene coverage, For the first Feature vectors of scene names from historical success cases. For the first Feature vectors of scene profiles from historical success cases Feature vector index for creating scenario profiles of historical success cases. Let be the distance function. Feature vector index for the scene name of a historical success case.

[0060] S4.2 When the confidence assessment result indicates that the confidence level is lower than the preset confidence standard, the exploratory enhancement unit is triggered to query the dynamic complaint knowledge graph based on the standardized query intent feature vector and obtain the potential association structure in the graph.

[0061] Furthermore, when the confidence assessment results indicate that the overall reliability of the historical experience relied upon is lower than the preset confidence standard, the exploratory enhancement unit is triggered. This unit, based on the standardized query intent feature vector that triggered the current query, does not directly query strongly related entities that are already clearly present in the knowledge graph. Instead, it actively explores potential related structures in the dynamic complaint knowledge graph. Potential related structures may manifest as edges between entities in the graph that have not been fully verified by historical successful cases but have certain semantic or statistical associations. They may also manifest as clusters of entities that are adjacent to the current query intent feature vector in the graph embedding space but have not been directly used by historical queries. The current query intent involves the political bias of the generated content, while historical cases are mostly concentrated in culturally biased scenarios. The confidence assessment may issue a warning due to the low dispersion of scenario coverage. Specifically, the exploratory enhancement unit may discover potential correlation paths between politically biased entities and entities such as those exhibiting regional discrimination or historical narratives through graph queries. These paths are not explicitly invoked in the meta-query strategy descriptions of any historical success cases. The purpose of obtaining potential correlation structures is to break through the limitations of historical experience and discover new and potentially effective query strategy directions. The knowledge graph is viewed not only as a static experience base but also as a dynamic exploration space containing unexplored correlations. When historical experience is not sufficiently reliable, the unit actively mines adjacent possibilities from the graph to supplement strategy inspiration.

[0062] S4.3. Synthesize complementary strategy recommendation hypotheses using potential association structures, and assign a preset exploration weight to the complementary strategy recommendation hypotheses by the exploratory enhancement unit.

[0063] Furthermore, the exploratory enhancement unit utilizes the latent association structure obtained from the dynamic complaint knowledge graph to synthesize supplementary strategy recommendation hypotheses. The synthesis process is not random; rather, it constructs strategy fragments or complete strategy descriptions that differ from, or are entirely new, than the historically-based strategies in the initial weighted strategy recommendation summary, based on new query dimensions, new entity combinations, or new analytical paths implied by the latent association structure. For example, if the latent association structure suggests an underutilized correlation between user sentiment and model response length, a hypothesis might be synthesized suggesting the addition of an analysis of the user's question sentiment polarity to the query and cross-comparing it with the model-generated length.

[0064] Specifically, for each synthesized supplementary strategy recommendation hypothesis, the exploratory enhancement unit assigns it a preset exploration weight, which is usually lower than the strategy weight derived from high-confidence historical experience. This reflects a cautious attitude towards exploratory new strategies and gives them the opportunity to be considered in subsequent processes. It transforms the potential associations mined from the knowledge graph into specific and actionable strategy possibilities, establishing a generative pathway from knowledge associations to strategy hypotheses. This allows strategy recommendations to no longer be completely limited by past successful operations, but to reason and generate theoretically reasonable alternative strategies that have not yet been proven by history, based on the inherent logical structure of domain knowledge.

[0065] S4.4. Combine the supplementary strategy recommendation hypothesis with exploration weights with the preliminary weighted strategy recommendation summary to generate the final weighted strategy recommendation summary.

[0066] Furthermore, the exploratory enhancement unit merges supplementary strategy recommendation hypotheses with preset exploration weights with the preliminary weighted strategy recommendation summary. This merging is not a simple addition; rather, it integrates the supplementary strategy recommendation hypotheses as new, weighted components into the existing weighted strategy summary framework. Strategy elements derived from historical success stories in the preliminary weighted strategy recommendation summary retain their original weights, while the newly added supplementary strategy recommendation hypotheses participate with their respective exploration weights. The merged result is the final weighted strategy recommendation summary, which includes both historically validated strategy experience and potentially exploratory strategy directions derived from new domain knowledge. The difference in weights reflects different assessments of the credibility of strategies from different sources.

[0067] Specifically, it generates a more complete and balanced strategy recommendation blueprint that integrates historical experience with new exploration discoveries, realizing a robust experience-exploration hybrid strategy generation mechanism. It does not blindly explore when historical experience is reliable, nor is it helpless when historical experience is insufficient. Instead, it dynamically adjusts the composition of strategy recommendations based on the confidence assessment results: when the confidence of historical experience is high, the final summary is mainly based on historical experience; when the confidence of historical experience is low, it automatically introduces new hypotheses based on knowledge graph exploration as an important supplement.

[0068] S5. Based on the final weighted strategy recommendation summary, generate candidate meta-query strategy descriptions, and output a set of verified candidate strategies through static rule checks and dynamic sandbox simulation.

[0069] S5.1. Parse the final weighted strategy recommendation summary to obtain the recommended data source combination and processing flow, and generate candidate meta-query strategy descriptions.

[0070] Furthermore, the parsing process needs to identify different strategy elements in the weighted indicators within the summary. For example, the summary might suggest prioritizing access to user feedback logs, combining them with a knowledge graph for entity linking, and then invoking a sentiment analysis model. Based on the parsed data source access order, filtering conditions, aggregation operations, and analysis algorithm call chains, a structured description of candidate meta-query strategies is generated.

[0071] Specifically, by executing queries on dependencies and parameters, high-level, weighted strategy recommendations are transformed into actionable blueprints. This achieves the transformation from probabilistic, fusion-based strategy recommendations to deterministic, actionable task descriptions, paving the way for subsequent rigorous validation and execution, regardless of whether the source of the strategy recommendation summary is historical experience or exploratory hypotheses.

[0072] S5.2 The static rule checker examines the logical conflicts and resource constraint violations in each candidate meta-query strategy description, marks the candidate meta-query strategy descriptions that pass the static rule check as candidate strategy descriptions that pass static verification, and generates static verification results.

[0073] Furthermore, the static rule checker examines each generated candidate meta-query strategy description, primarily checking for logical conflicts and resource constraint violations. Logical conflict checks include identifying contradictory operational instructions within the strategy description, mutually exclusive filtering conditions applied to the same data field, or attempts to execute analytical operations dependent on that data without obtaining the necessary prior data.

[0074] Specifically, the resource constraint violation check assesses whether the estimated computational overhead, memory usage, and data access frequency of the strategy description exceed the predetermined resource quota or service rate limiting threshold. Candidate meta-query strategy descriptions that pass the static rule check are marked as candidate strategy descriptions that pass static verification, and static verification results are generated. The static verification results record the check conclusions, the violations found, and the correction suggestions.

[0075] S5.3. Based on a dynamic sandbox simulator, in a sandbox isolated from the production environment, recent data snapshots are used to tentatively execute candidate policy descriptions that have passed static validation, evaluate the data access success rate and preliminary effect confidence of each candidate policy description that has passed static validation, and generate dynamic validation results.

[0076] Furthermore, by monitoring the entire process of exploratory execution, the data access success rate and preliminary effect confidence of each candidate strategy description that has passed static verification are evaluated. The data access success rate reflects whether the data source interface to be accessed as planned in the strategy description is actually available, whether the permissions are sufficient, and whether the query syntax is correct. The preliminary effect confidence is based on the results returned by the simulation execution. Using predefined evaluation logic, the credibility of the core effect indicators such as recall and precision that the strategy may achieve in the real environment is estimated. The process and results of exploratory execution are recorded and dynamic verification results are generated. In a simulation environment that is infinitely close to reality but risk-free, the feasibility and potential effect of the strategy are empirically evaluated.

[0077] Specifically, a testbed for the strategy was built, which advances strategy verification from pure theoretical deduction and rule checking to an empirical simulation stage based on real data patterns. This allows us to discover problems that are correct in static logic but exposed due to the characteristics of the data itself, subtle differences in interfaces, or runtime states.

[0078] S5.4 Candidate strategy descriptions that pass static validation: When the data access success rate and the confidence level of the initial effect are simulated through dynamic sandbox and reach the preset candidate criteria, they are marked as candidate strategy descriptions that pass dynamic validation.

[0079] Furthermore, the preset candidate criteria define the minimum dynamic performance thresholds that a strategy must meet to enter the final candidate pool. For example, the data access success rate must be completely successful, and the confidence level of the initial effect must be higher than a basic threshold. Marking a strategy as having passed dynamic verification means that the strategy description is not only logically consistent and resource compliant, but also has been proven to have good executability and promising effect output in the simulation environment. The final quality certification is carried out on the strategies that have passed both dynamic and static verification, establishing a phased and progressive strategy admission mechanism.

[0080] Specifically, it does not require the strategy to achieve perfect results in the simulation, but rather sets a reasonable and achievable entry threshold to ensure that the strategies entering the final candidate set have basic usable qualities. At the same time, it allows for reasonable differences in the confidence level of the effect for subsequent selection, balancing the rigor of the validation with the diversity of the strategies. It avoids filtering out potentially innovative strategies due to excessively high validation standards, and also prevents low-quality strategies from being mixed into the candidate set due to excessively low standards. This makes the validated candidate strategy set a high-quality strategy resource pool that combines reliability and diversity.

[0081] S5.5. Combining the states of candidate strategy descriptions that have passed static verification and candidate strategy descriptions that have passed dynamic verification, output the meta-query strategy descriptions of all candidates that meet the conditions, and output the set of verified candidate strategies.

[0082] Furthermore, the candidate meta-query strategy descriptions that meet the criteria must at least satisfy the basic conditions for passing static verification. Descriptions that pass dynamic verification may have priority in subsequent selections due to their higher verification level. The verified candidate strategy set is a collection of multiple high-quality, rigorously tested strategy descriptions and their verification annotations. It summarizes and delivers the final results of all previous verification work, providing filtered and evaluated inputs for the next step of strategy selection and execution. It constructs a structured strategy product library, which not only contains the strategies themselves but also integrates rich metadata from static and dynamic verification results, such as logical check records, simulated execution performance, and effect predictions.

[0083] Specifically, the selection and execution process is no longer a black box strategy, but a more informed decision-making process based on detailed verification reports. The way the strategy is output along with its complete verification history greatly enhances the transparency, traceability, and sufficiency of the decision-making basis of the entire strategy optimization process, marking the formation of a rigorous and closed-loop quality control system for the strategy generation and verification process.

[0084] S6. Select a meta-query strategy description from the verified candidate strategy set, compile it into a distributed query task across data sources, and obtain the initial complaint dataset.

[0085] S6.1 Based on the strategy selector, the initial effect confidence is sorted according to the meta-query strategy description of each candidate in the verified candidate strategy set. The strategy selector selects the meta-query strategy description of the candidate ranked first as the meta-query strategy description.

[0086] Furthermore, the strategy selector extracts the preliminary effect confidence index from the dynamic validation results attached to the meta-query strategy description of each candidate in the validated candidate strategy set, and sorts all candidate meta-query strategy descriptions in descending order based on this index. The sorting operation places the candidate strategy with the highest preliminary effect confidence at the top of the sequence. The preliminary effect confidence, as a quantitative representation of the predicted strategy performance in the dynamic sandbox simulation, becomes the core basis for selecting the final execution strategy. The strategy selector then selects the meta-query strategy description of the first-ranked candidate as the meta-query strategy description to be executed.

[0087] Specifically, from multiple validated high-quality strategies, the one with the best current prediction effect is automatically selected as the execution plan. The strategy selection is based on empirical prediction data generated by dynamic simulation in the early stage, rather than random selection or selection based on simple rules. The optimization mechanism based on the ranking of simulation effects ensures that the strategy put into actual production is the one that has been proven in the simulation environment to be the most likely to achieve goals such as high recall and high precision. This makes the final decision-making stage of the entire strategy optimization process data-driven and effect-oriented, maximizing the value conversion efficiency from strategy generation to strategy execution.

[0088] S6.2 Parse the data source access instructions and logical operations in the meta-query strategy description, and convert the parsed data source access instructions and logical operations into specific query statements for the dialogue log database, user feedback table and model inference tracing system.

[0089] Furthermore, the selected meta-query strategy description is parsed to extract the defined data source access instructions and logical operation sequences. Data source access instructions specify the specific data storage entity to be queried, such as a specific table in a dialogue log database, a set of fields in a user feedback table, or a record stream in a model inference tracing system. Logical operations include filtering conditions, joins, aggregate functions, sorting rules, and call parameters for analysis algorithms. The parsing process needs to convert high-level, abstract operation instructions into specific query statements that can be understood by the execution engines of each target data source. For example, the instruction to search for conversations containing specific keywords in the dialogue log database is converted into the WHERE clause of the corresponding SQL query statement; the instruction to retrieve the model call chain within a certain time period for the model inference tracing system is converted into API call parameters or query language specific to that system.

[0090] Specifically, it translates platform-independent strategy descriptions into platform-specific execution code, acting as a bridge between abstract strategies and concrete data infrastructure. By parsing a unified strategy description language, it generates diverse query syntaxes required by heterogeneous data sources, thereby shielding the technical details of underlying data storage. This allows the same optimized strategy to seamlessly adapt to different databases, table structures, or service interfaces, greatly enhancing the universality and portability of the strategy execution process. It is the technical guarantee for the implementation of strategies across complex data environments.

[0091] S6.3 The query compilation engine organizes and coordinates specific query statements from the dialogue log database, user feedback table, and audit trail system into distributed query tasks across data sources.

[0092] Furthermore, the query compilation engine organizes and coordinates the individual query statements generated from the dialogue log database, user feedback table, and audit trail system to construct a unified distributed query task across data sources. The organization process requires identifying data dependencies and temporal relationships between different query statements. Based on these dependencies, the query compilation engine arranges the independent query statements into a directed acyclic graph-like task execution plan. Nodes in the graph represent specific query statements or data merging operations, and edges represent data flow dependencies. The coordination process allocates computing resources to this execution plan, determines the nodes for task distribution, and sets up fault handling and timeout control mechanisms. The resulting distributed query task across data sources is a complete work unit that can be scheduled and executed by the distributed computing framework, integrating discrete query operations across multiple data sources into an efficient, reliable, and ordered composite query process.

[0093] S6.4. The query results data obtained from the dialogue log database, user feedback table and model inference tracking system are used to form the initial complaint dataset.

[0094] Furthermore, orchestrated distributed query tasks across data sources are executed, concurrently or sequentially retrieving query results from three specified data sources: the conversation log database, the user feedback table, and the model inference tracing system. During task execution, each subquery is distributed to its corresponding data source for execution. The returned result dataset may include structured records, semi-structured log entries, or tracing reports in a specific format. Result data from different sources and with varying formats needs to be aggregated. The aggregation process typically includes data alignment, field mapping, format unification, and necessary data cleaning. For example, the conversation ID field from different sources is matched, timestamps are converted to a unified format, and obviously invalid or duplicate records are filtered out. The aggregated and initially integrated data set forms the initial complaint dataset.

[0095] Specifically, it achieves the physical acquisition and initial fusion of data, producing relatively well-organized raw data materials that can be used for subsequent in-depth analysis. It completes the final transformation from query plan to data entity. It is not just about simply executing queries and collecting results, but also about transforming scattered and heterogeneous raw data into an internally consistent and traceable analytical base through a structured aggregation process. The initial complaint dataset carries all information fragments extracted from multi-dimensional data sources that are directly related to the current complaint query intent, marking a key transition from strategy intelligence to data insight.

[0096] S7. Perform aggregation analysis and knowledge extraction on the initial complaint dataset to generate a complaint analysis report, and update the dynamic complaint knowledge graph to form a strategy effect case.

[0097] S7.1 Use clustering analysis algorithms to process the initial complaint dataset, identify complaint clusters, extract common patterns of complaint events from the initial complaint dataset, and use pattern recognition algorithms to scan for repetitive features in the initial complaint dataset.

[0098] Furthermore, clustering analysis algorithms are used to process the initial complaint dataset to discover naturally occurring groups or complaint clusters within the data. Based on similarity measures between data points, complaint events with similar characteristics are grouped into the same cluster. This ensures that events within the same cluster are highly homogeneous in terms of model behavior, user context, or question type, while different clusters show significant differences. For example, clustering might reveal that one cluster mainly contains complaints involving the generation of false financial information, while another cluster focuses on responses containing impolite language. Simultaneously or after clustering, common patterns of complaint events are extracted from the initial complaint dataset. This involves identifying stable feature combinations, frequently occurring term associations, or sets of co-occurring attributes across multiple complaint events.

[0099] Specifically, pattern recognition algorithms are used to scan for repetitive features in the initial complaint dataset. These features may include periodic complaint time distributions, strong correlations between specific user groups and specific complaint types, or certain input prompt templates consistently triggering similar problem model outputs. This process provides preliminary structured insights into the messy initial data, summarizing massive discrete complaint events into limited, meaningful groups, patterns, and features. By combining unsupervised machine learning methods with domain feature analysis, the inherent category structure and regularities are automatically extracted from the data itself, rather than relying on a predefined classification system.

[0100] S7.2 Based on complaint clusters, event sequence associations, and repetitive characteristics, generate a complaint analysis report and extract new complaint pattern entities, important relationships between entities, and statistical features from the complaint analysis report.

[0101] Furthermore, the complaint analysis report not only lists the findings but also provides explanations and integrations. For example, it points out that complaints about clusters of false financial information are often accompanied by the appearance of specific financial terms in user queries, and that such complaints peak during trading hours. From the structured complaint analysis report, new complaint pattern entities can be extracted, such as the generation of false financial information during trading hours; important relationships between entities, such as a strong correlation between specific financial terms and false information generated by the model; and statistical characteristics, such as the correlation coefficient between the daily average occurrence rate of such complaints and a specific version update of the model. The extraction process relies on the parsing of the report text or structured output to identify new concepts, new relationships, and new quantitative indicators represented therein.

[0102] Specifically, the discovery of data analysis is transformed into structured, computable knowledge elements. An automated pipeline from data insights to knowledge elements has been established. It is not only satisfied with generating an analysis summary for human readers, but also actively extracts knowledge triples and feature vectors that can be directly understood and processed by machines from the analysis conclusions. This makes the value produced by the analysis process no longer limited to a one-time report, but can be deposited as a reusable knowledge asset.

[0103] S7.3 Add new complaint pattern entities, important relationships between entities, and statistical features to the dynamic complaint knowledge graph.

[0104] Furthermore, adding a new complaint pattern entity means creating a concept node representing the new pattern in the knowledge graph, and potentially establishing category associations between it and existing entities such as financial complaints and generated content risks. Adding important relationships between entities involves creating or strengthening directed edges between related entity nodes, and may assign relationship weights or type labels. For example, creating an easily triggered relationship edge between a specific financial term node and a model-generated false information node. Statistical features can be attached as attributes to related entities or relationship edges. The addition process needs to follow the schema constraints of the knowledge graph and may trigger graph reasoning to derive new implicit relationships.

[0105] Specifically, this allows the dynamic complaint knowledge graph to continuously evolve, absorbing the latest analytical findings and feeding the results of data analysis back into the graph as the core knowledge carrier in real time. This enables online learning and incremental updates of the knowledge base, transforming it from a static background knowledge base into a living, evolving memory that grows and enriches with each query and analysis. Every successful complaint analysis expands and deepens the graph's understanding of the complaint domain, and the updated knowledge immediately serves subsequent stages such as clarifying query intent and enhancing strategy exploration, forming a virtuous cycle of analysis-driven knowledge growth and knowledge optimization for subsequent analyses. S7.4 Package the standardized query intent feature vector, meta-query strategy description, and actual recall, precision, and time consumption data recorded in the complaint analysis report into a strategy effect case.

[0106] Furthermore, the standardized query intent feature vector accurately represents the input intent of this query; the meta-query strategy describes the specific strategy scheme generated and executed in response to this intent; and the actual effect data in the complaint analysis report objectively records the execution performance of the strategy scheme in a real data environment. The packaging process organizes heterogeneous but logically related elements into a structured record, which may use a specific data format or storage entries to ensure that the reference relationships between them are complete and traceable. It fully records the key information of the entire query optimization process cycle, constructs a complete case for subsequent analysis, learning, and evaluation, and creates a closed-loop data unit covering the entire process of intent input, strategy execution, and effect output. The strategy effect case not only records what was done and what the result was, but also accurately anchors the result to the initial query requirement through the standardized query intent feature vector.

[0107] Specifically, this enables subsequent reinforcement learning processors to accurately evaluate the effectiveness of specific policies in dealing with specific types of intents, providing high-quality, context-rich training samples for weight adjustment and experience correction of the policy template library. The case-based experience encapsulation method transforms each query execution into a replayable, measurable, and learnable experiment, which is the core mechanism driving the entire policy optimization process to achieve experience accumulation and self-improvement.

[0108] S8. Based on the actual effect data of strategy effect cases, adjust the weights and make experience corrections to the strategy template library, and compile and generate an intelligent strategy optimization report.

[0109] S8.1 The reinforcement learning processor reads the actual recall, precision, and time data recorded in the policy effect cases, and locates historical successful case records in the policy template library that provide the basis for generating the preliminary weighted policy recommendation summary.

[0110] Furthermore, the reinforcement learning processor reads the actual recall, precision, and execution time data recorded in the policy effectiveness cases. These data constitute objective quantitative indicators for evaluating the effectiveness of the current query strategy. The reinforcement learning processor then locates historical successful case records in the policy template library that provided the basis for generating the preliminary weighted strategy recommendation summary used in this query. The location process is accomplished by comparing the standardized query intent feature vector used in this query with the historical query intent feature vectors stored in the historical successful case records, and by tracing back the similarity matching and weight allocation records performed in the stage of generating the preliminary weighted strategy recommendation summary. This allows for the precise identification of historical experience sources that directly influenced the ultimately adopted strategy.

[0111] Specifically, establishing a direct causal link between the current strategy execution results and historical experience contributions enables targeted feedback of strategy effects to historical experience sources. This ensures that subsequent weight adjustments are not blind or global, but rather precisely strengthen or correct specific historical experience items that truly participated in the decision-making process and influenced the final outcome. This transforms the actual effect of each strategy execution into targeted learning signals for specific items in the strategy template library.

[0112] S8.2 Compare the actual performance data with the historical performance data stored in the historical success case records. When the actual performance data is better than the historical performance data, the reinforcement learning processor increases the weight and priority of the corresponding historical success case record.

[0113] Furthermore, when the actual performance data comprehensively surpasses or, without significantly increasing time consumption, the key metrics outperform historical performance data, the reinforcement learning processor determines that the current strategy execution validates the high effectiveness of relevant historical experience in the current context, thereby increasing the weight and priority of the corresponding historical success case record. Increased weight makes this historical case more influential in future similar query matching, and increased priority may place it higher in the ranking or display of the strategy recommendation summary.

[0114] Specifically, positive reinforcement of proven effective historical experiences mimics the use-and-disuse principle in biological learning. By rewarding historical strategy patterns that succeed again in real-world scenarios, the attention and trust of the strategy template library are focused on those experiences that have been repeatedly proven effective. This allows the historical experience library to automatically and continuously converge to a higher quality and more reliable set of strategy patterns, thereby continuously improving the baseline quality of future strategy recommendations. This is the key positive cycle for achieving self-purification and performance improvement of the strategy library.

[0115] S8.3 When the actual performance data is worse than the historical performance data, the reinforcement learning processor performs attribution analysis on the difference between the actual performance data and the historical performance data, and adds the attribution conclusion as a note to the corresponding historical success case record.

[0116] Furthermore, when the actual performance data is worse than the historical performance data stored in the historical success case records, the reinforcement learning processor initiates an attribution analysis process. Attribution analysis does not simply determine that historical experience has failed, but rather delves into the reasons for the discrepancy between the actual and historical performance data. The analysis may examine changes in the query context, shifts in data distribution, interference from external environmental factors, or overfitting or measurement bias in the performance data of the historical case records themselves. Attribution conclusions, such as an abnormally high response latency in data source A during the current query, leading to an increase in the total strategy execution time, and not a logical flaw in the strategy itself, are added as structured notes to the corresponding historical success case records.

[0117] Specifically, instead of simply punishing strategies that fail to meet expectations, diagnostic learning is employed. A failed or diminished-performance query is transformed into valuable diagnostic information. By adding attribution notes, historical experience is no longer just a single performance score, but a rich body of knowledge including effective conditions, environmental dependencies, and potential limitations. This allows future strategy matching to not only refer to historical performance data but also consider the risks or constraints indicated in the attribution notes, resulting in more prudent strategy recommendations that better fit the current environment. S8.4 Examine the generation process of strategy effectiveness cases, identify the supplementary strategy recommendation hypotheses synthesized by the exploratory enhancement unit and validated in this query, and transform them into new historical success case templates based on the valid supplementary strategy recommendation hypotheses.

[0118] Furthermore, for these identified effective supplementary strategy recommendation hypotheses, the reinforcement learning processor initiates a template transformation process. This process extracts the core strategy logic, associated query intent features, and performance data from the current execution, and constructs a new, complete historical success case template according to the format requirements of the strategy template library. This new template records the successful strategy experience inspired by potential associations in the knowledge graph and validated in practice. This solidifies the successful results of exploratory innovation into the experience base.

[0119] Specifically, it opens up a pathway from exploratory conjecture to solidified experience. It ensures that the system can not only utilize historical experience, but also promptly absorb new and effective strategy patterns discovered through its own exploration into new historical experience, thereby continuously expanding the coverage and innovation capabilities of the strategy template library. This makes the strategy template library a living organism that can grow and enrich itself, rather than just a static archive of past records.

[0120] S8.5. Based on the report generator, receive the final weighted strategy recommendation summary, static verification results, dynamic verification results, meta-query strategy description, complaint analysis report, strategy effect case, and weight adjustment and experience correction records, and compile them to generate an intelligent strategy optimization report.

[0121] Furthermore, the report generator receives multiple key intermediate and final outputs from the entire query optimization process, including the initial and final weighted strategy recommendation summary, static verification results reflecting strategy logic and resource compliance, dynamic verification results reflecting strategy simulation execution performance, a description of the final selected and executed meta-query strategy, a complaint analysis report generated after in-depth analysis of complaint data, strategy effect cases encapsulating full-link information, and weight adjustment and experience correction records generated by the reinforcement learning processor. The report generator sorts, integrates, and formats heterogeneous but logically closely related information to generate a structured intelligent strategy optimization report.

[0122] Specifically, the entire black-box operation of the intelligent processing is made transparent and documented, producing a comprehensive record that can be audited, analyzed, and continuously improved. This constructs an interpretable archive covering the entire strategy lifecycle. The intelligent strategy optimization report is not merely a summary of results, but a detailed explanation of why the result was obtained and what was learned from the query. It makes each query process a complete experiment that can be reviewed, audited, and disseminated, thus enhancing user trust in the strategy optimization process.

[0123] This embodiment also provides a large model complaint data query strategy optimization processing system, including: a standardization module, which receives large model complaint data query requests, generates an initial query intent feature vector, and generates a standardized query intent feature vector based on the interaction feedback between the initial query intent feature vector and the knowledge graph; The enhancement generation module matches the standardized query intent feature vector with the strategy template library to generate a preliminary weighted strategy recommendation summary. Based on the confidence evaluation results of the preliminary weighted strategy recommendation summary, exploratory enhancement is performed to generate the final weighted strategy recommendation summary. The candidate set generation module generates meta-query strategy descriptions for candidates based on the final weighted strategy recommendation summary. Through static rule checking and dynamic sandbox simulation, it outputs a set of verified candidate strategies. The execution module selects meta-query strategy descriptions from the verified candidate strategy set, compiles them into a distributed query task across data sources, and obtains the initial complaint dataset. The knowledge update module performs aggregation analysis and knowledge extraction on the initial complaint dataset to generate a complaint analysis report, and updates the dynamic complaint knowledge graph to form strategy effect cases; The optimization module adjusts the weights and makes empirical corrections to the strategy template library based on the actual effect data of strategy performance cases, and generates an intelligent strategy optimization report.

[0124] This embodiment also provides a computer device applicable to the large-model complaint data query strategy optimization processing method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the large-model complaint data query strategy optimization processing method proposed in the above embodiment.

[0125] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0126] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the optimization processing method for large-scale complaint data query strategy proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0127] In summary, this invention receives query requests and generates an initial query intent feature vector, then combines this with interactive feedback from a dynamic complaint knowledge graph to generate a standardized query intent feature vector. The standardized query intent feature vector is matched with a strategy template library to generate a preliminary weighted strategy recommendation summary, which is then subjected to confidence evaluation and exploratory enhancement to obtain a final recommendation summary. A two-stage verification process—static rule checking and dynamic sandbox simulation—outputs a set of verified candidate strategies. A strategy description is selected and compiled into a distributed query task across data sources to obtain an initial complaint dataset. The dataset is then aggregated and analyzed for knowledge extraction to generate a complaint analysis report and update the knowledge graph, forming strategy effectiveness cases. Based on the actual effectiveness data of the cases, the strategy template library is weighted and empirically corrected to generate an intelligent strategy optimization report, thus achieving autonomous optimization and continuous evolution of the query strategy.

[0128] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for optimizing the query strategy of large-scale complaint data, characterized by: This includes receiving query requests for large model complaint data, generating an initial query intent feature vector, and generating a standardized query intent feature vector based on the interaction feedback between the initial query intent feature vector and the knowledge graph. The standardized query intent feature vector is matched with the strategy template library to generate a preliminary weighted strategy recommendation summary. Based on the confidence evaluation results of the preliminary weighted strategy recommendation summary, exploratory enhancement is performed to generate the final weighted strategy recommendation summary. Based on the final weighted strategy recommendation summary, candidate meta-query strategy descriptions are generated. Through static rule checking and dynamic sandbox simulation, a set of validated candidate strategies is output. Select meta-query strategy descriptions from the verified candidate strategy set, compile them into a distributed query task across data sources, and obtain the initial complaint dataset; Aggregate analysis and extract knowledge from the initial complaint dataset to generate a complaint analysis report, update the dynamic complaint knowledge graph, and form a strategy effect case study; Based on the actual effect data of strategy performance cases, the strategy template library is weighted and revised according to experience, and an intelligent strategy optimization report is generated.

2. The method for optimizing the query strategy of large-scale complaint data as described in claim 1, characterized in that: Receive large-scale model complaint data query requests and generate initial query intent feature vectors, including the following steps: The natural language processor receives the query request text of the large model complaint data, performs syntactic and semantic parsing on the large model complaint data query request text, and obtains the results of syntactic and semantic parsing. Based on the results of syntactic and semantic parsing, the core query entities, query targets, and constraints in the large model complaint data query requests are identified. Features are extracted from the core query entities, query targets, and constraints. The extracted features are then vectorized and encoded to generate an initial query intent feature vector.

3. The method for optimizing the query strategy of large-scale complaint data as described in claim 2, characterized in that: Based on the interaction feedback between the initial query intent feature vector and the knowledge graph, a standardized query intent feature vector is generated, including the following steps: The intent clarification logic unit is used to interactively query the initial query intent feature vector with the dynamic complaint knowledge graph, and to obtain entities and potential ambiguity patterns associated with the initial query intent feature vector from the dynamic complaint knowledge graph. Based on entity and potential ambiguity patterns, the intent to clarify logic unit generates clarifying questions, presents the clarifying questions to the user, and obtains user feedback; By utilizing user feedback, the initial query intent feature vector is corrected and completed to obtain a standardized query intent feature vector.

4. The method for optimizing the query strategy of large-scale complaint data as described in claim 3, characterized in that: The standardized query intent feature vector is matched with a policy template library to generate a preliminary weighted policy recommendation summary, including the following steps: The similarity between the standardized query intent feature vector and the query intent feature vector of each historical success case stored in the strategy template library is calculated using a feature matcher. Sort all similarity results, select cases with similarity exceeding a preset threshold for historical success cases, and assign a matching weight to each selected historical success case. The meta-query strategy descriptions of all selected historical success cases are aggregated, and each meta-query strategy description is combined with its corresponding matching weight to form a preliminary weighted strategy recommendation summary.

5. The method for optimizing the query strategy of large-scale complaint data as described in claim 4, characterized in that: Based on the confidence assessment results of the preliminary weighted strategy recommendation summary, exploratory enhancement is performed to generate the final weighted strategy recommendation summary, including the following steps: The confidence assessment unit is used to calculate the average time freshness and scene coverage dispersion of the historical successful cases on which the preliminary weighted strategy recommendation summary is based, and the confidence assessment results are generated. When the confidence assessment result indicates that the confidence level is lower than the preset confidence standard, the exploratory enhancement unit is triggered to query the dynamic complaint knowledge graph based on the standardized query intent feature vector to obtain the potential association structure in the graph; Complementary strategy recommendation hypotheses are synthesized using potential association structures, and exploratory enhancement units assign a preset exploration weight to the complementary strategy recommendation hypotheses; The supplementary policy recommendation hypothesis with exploration weights is combined with the preliminary weighted policy recommendation summary to generate the final weighted policy recommendation summary.

6. The method for optimizing the query strategy of large-scale complaint data as described in claim 5, characterized in that: Based on the final weighted strategy recommendation summary, candidate meta-query strategy descriptions are generated. Through static rule checking and dynamic sandbox simulation, a set of validated candidate strategies is output, including the following steps: The final weighted strategy recommendation summary is parsed to obtain the recommended data source combination and processing flow, and candidate meta-query strategy descriptions are generated. The static rule checker examines the logical conflicts and resource constraint violations in each candidate meta-query strategy description, marks the candidate meta-query strategy descriptions that pass the static rule check as candidate strategy descriptions that pass static verification, and generates static verification results. Based on a dynamic sandbox simulator, in a sandbox isolated from the production environment, recent data snapshots are used to tentatively execute candidate strategy descriptions that have passed static validation, evaluate the data access success rate and preliminary effect confidence of each candidate strategy description that has passed static validation, and generate dynamic validation results. Candidate strategy descriptions that pass static validation are marked as dynamically validated candidate strategy descriptions when the data access success rate and initial effect confidence level are simulated in a dynamic sandbox and reach the preset candidate criteria. By combining the states of candidate strategy descriptions that pass static validation and candidate strategy descriptions that pass dynamic validation, the meta-query strategy descriptions of all candidates that meet the conditions are output, and the set of validated candidate strategies is output.

7. The method for optimizing the query strategy of large-scale complaint data as described in claim 6, characterized in that: The process involves selecting a meta-query strategy description from the validated candidate strategy set, compiling it into a distributed query task across data sources, and obtaining the initial complaint dataset, including the following steps: The strategy selector sorts the preliminary effect confidence based on the meta-query strategy description of each candidate in the verified candidate strategy set, and selects the meta-query strategy description of the candidate ranked first as the meta-query strategy description. The data source access instructions and logical operations in the meta-query strategy description are parsed, and the parsed data source access instructions and logical operations are converted into specific query statements for the dialogue log database, user feedback table and model inference tracing system. The query compilation engine organizes and coordinates specific query statements from the dialogue log database, user feedback table, and audit trail system into distributed query tasks across data sources. The initial complaint dataset is formed by query results obtained from the dialogue log database, user feedback forms, and model inference tracing system.

8. The method for optimizing the query strategy of large-scale complaint data as described in claim 7, characterized in that: The initial complaint dataset is aggregated and analyzed, and knowledge is extracted to generate a complaint analysis report. The dynamic complaint knowledge graph is then updated to create a strategy effectiveness case study. This process includes the following steps: Clustering analysis algorithms are used to process the initial complaint dataset, identify complaint clusters, extract common patterns of complaint events from the initial complaint dataset, and use pattern recognition algorithms to scan for repetitive features in the initial complaint dataset. Based on complaint clusters, event sequence associations, and repetitive features, a complaint analysis report is generated, from which new complaint pattern entities, important relationships between entities, and statistical features are extracted. Add new complaint pattern entities, important relationships between entities, and statistical characteristics to the dynamic complaint knowledge graph; The standardized query intent feature vector, meta-query strategy description, and actual recall, precision, and time consumption data recorded in the complaint analysis report are packaged into a strategy effectiveness case.

9. The method for optimizing the query strategy of large-scale complaint data as described in claim 8, characterized in that: Based on the actual performance data of strategy case studies, the strategy template library is weighted and revised according to experience, and an intelligent strategy optimization report is generated, including the following steps: The reinforcement learning processor reads the actual recall, precision, and latency data recorded in the policy performance cases, and locates historical successful case records in the policy template library that provide the basis for generating the preliminary weighted policy recommendation summary; The actual performance data is compared with the historical performance data stored in the historical success case records. When the actual performance data is better than the historical performance data, the reinforcement learning processor increases the weight and priority of the corresponding historical success case record. When the actual performance data is worse than the historical performance data, the reinforcement learning processor performs attribution analysis on the difference between the actual performance data and the historical performance data, and adds the attribution conclusion as a note to the corresponding historical success case record. Examine the generation process of strategy effectiveness cases, identify supplementary strategy recommendation hypotheses synthesized by exploratory enhancement units that have been validated in this query, and transform them into new historical success case templates; The report generator receives the final weighted strategy recommendation summary, static verification results, dynamic verification results, meta-query strategy description, complaint analysis report, strategy effect case, and weight adjustment and experience correction records, and then compiles them to generate an intelligent strategy optimization report.

10. A large-scale complaint data query strategy optimization and processing system, based on the large-scale complaint data query strategy optimization and processing method according to any one of claims 1 to 9, characterized in that: include, The standardization module receives query requests for complaint data from large models, generates an initial query intent feature vector, and generates a standardized query intent feature vector based on the interaction feedback between the initial query intent feature vector and the knowledge graph. The enhancement generation module matches the standardized query intent feature vector with the strategy template library to generate a preliminary weighted strategy recommendation summary. Based on the confidence evaluation results of the preliminary weighted strategy recommendation summary, exploratory enhancement is performed to generate the final weighted strategy recommendation summary. The candidate set generation module generates meta-query strategy descriptions for candidates based on the final weighted strategy recommendation summary. Through static rule checking and dynamic sandbox simulation, it outputs a set of verified candidate strategies. The execution module selects meta-query strategy descriptions from the verified candidate strategy set, compiles them into a distributed query task across data sources, and obtains the initial complaint dataset. The knowledge update module performs aggregation analysis and knowledge extraction on the initial complaint dataset to generate a complaint analysis report, and updates the dynamic complaint knowledge graph to form strategy effect cases; The optimization module adjusts the weights and makes empirical corrections to the strategy template library based on the actual effect data of strategy performance cases, and generates an intelligent strategy optimization report.

Images