An agent consistency analysis method, device and equipment

Abstract

The present application relates to the technical field of agent analysis, and particularly relates to an agent consistency analysis method, device and equipment; a target resource is accessed from a preset target resource object according to a compliance label set and an authorized use tool type; an available channel is determined according to the target resource, a preset resource graph topology, a preset Bayesian analysis method, a preset Thompson sampling method and a preset three-target utility function; in an available channel execution state; original unstructured documents and semi-structured documents are analyzed based on a pre-trained graph neural network model and a preset consistency test condition to obtain a consistency analysis result; with the help of the pre-trained graph neural network combined with the consistency test condition, multiple types of documents are deeply analyzed, the consistency and uncertainty of the quantifiable result are quantified, low-confidence content is identified, the reliability of document verification is improved, traceable data support is provided for knowledge transformation and semantic reasoning, and the stability and compliance of a business process are ensured.

Description

Technical Field

[0001] This invention relates to the field of agent analysis technology, and in particular to an agent consistency analysis method, apparatus and equipment. Background Technology

[0002] In recent years, intelligent agent technology driven by large language models has developed rapidly, becoming a support for business automation with its powerful reasoning and tool invocation capabilities. However, its enterprise-level implementation still faces prominent problems such as multimodal fragmentation, rigid process orchestration, and fragmented development systems. Existing intelligent agents are mostly text-processing-centric, lacking native semantic parsing and topological association mining capabilities for various unstructured and semi-structured multimodal documents, easily leading to knowledge fragmentation and causing content analysis bias and insufficient credibility. Traditional multi-agent processes rely heavily on static rule orchestration, failing to combine resource topology and performance costs for dynamic routing and adaptive scheduling, resulting in poor channel adaptation and task execution stability. Furthermore, conventional document verification methods struggle to quantify the consistency and uncertainty of analysis results, cannot effectively identify low-credibility data, and lack end-to-end compliance verification and traceability capabilities. This makes it difficult to support the integrated application requirements of intelligent agent process orchestration, resource routing, document retrieval, and compliance auditing, hindering the large-scale application of intelligent agents in complex enterprise business scenarios. Summary of the Invention

[0003] In order to overcome the shortcomings of the prior art, the purpose of this invention is to provide an agent consistency analysis method, apparatus and device.

[0004] A method for intelligent agent consistency analysis includes: acquiring user interaction messages and analyzing them to obtain a set of compliance tags and authorized tool types; accessing target resources from preset target resource objects based on the set of compliance tags and authorized tool types; determining available channels based on the target resources, a preset resource graph topology, a preset Bayesian analysis method, a preset Thompson sampling method, and a preset three-objective utility function; executing available channels, and collecting original unstructured documents and semi-structured documents while the available channels are in execution state; and analyzing the original unstructured documents and semi-structured documents based on a pre-trained graph neural network model and preset consistency check conditions to obtain consistency analysis results.

[0005] Further, the step of accessing the target resource from the preset target resource object based on the compliance tag set and the authorized tool type includes: analyzing the compliance tag set and the authorized tool type according to preset resource access rules to obtain rule analysis results; if the rule analysis results satisfy the resource access rules, then collecting the currently invoked tool type; performing compliance analysis on the user interaction message and the compliance tag set according to the preset first sufficient and necessary condition and the currently invoked tool type to obtain a first compliance analysis result; if the first compliance analysis result satisfies the first sufficient and necessary condition, then accessing the target resource from the target resource object according to the authorized tool type and the preset resource access conditions.

[0006] Furthermore, the step of determining available channels based on target resources, a preset resource graph topology, a preset Bayesian analysis method, a preset Thompson sampling method, and a preset three-objective utility function includes: analyzing the resource graph topology based on the target resources and a preset second necessary and sufficient condition to obtain a second compliance analysis result; if the second compliance analysis result indicates that the target resource satisfies the second necessary and sufficient condition in the resource graph topology, then collecting all passable information channels to obtain a channel pool; and filtering the channel pool based on the Bayesian analysis method, the Thompson sampling method, and the three-objective utility function to obtain available channels.

[0007] Furthermore, the process of filtering the channel pool based on Bayesian analysis, Thompson sampling, and a three-objective utility function to obtain usable channels includes: estimating and analyzing the channel pool using Bayesian analysis to obtain channel quality score features, latency score features, and cost score features; sampling the channel quality score features and latency score features based on Thompson sampling and cost score features to obtain sampling features; calculating the sampling features according to the three-objective utility function to obtain utility score features; and filtering the channel pool based on the utility score features to obtain usable channels.

[0008] Furthermore, the analysis of the original unstructured and semi-structured documents based on the pre-trained graph neural network model and preset consistency check conditions to obtain consistency analysis results includes: reasoning on the original unstructured and semi-structured documents based on the graph neural network model to obtain a center embedding vector, a positive embedding vector, and a negative embedding vector; generating a dense embedding vector based on the center embedding vector, the positive embedding vector, and the negative embedding vector; and performing consistency analysis on the dense embedding vector based on the consistency check conditions, a preset consistency threshold, and a preset uncertainty threshold to obtain consistency analysis results.

[0009] Furthermore, the step of reasoning on the original unstructured and semi-structured documents based on the graph neural network model to obtain the center embedding vector, positive embedding vector, and negative embedding vector includes: constructing a document knowledge base based on the original unstructured and semi-structured documents; transforming the document knowledge base based on a pre-trained large language model to obtain a structured knowledge representation; extracting entities from the structured knowledge representation based on a preset named entity recognition model to obtain category entity features; extracting relations from the category entity features based on a preset joint extraction method to obtain entity semantic relation features; and reasoning on the entity semantic relation features based on the graph neural network model to obtain the center embedding vector, positive embedding vector, and negative embedding vector.

[0010] Furthermore, the step of generating dense embedding vectors based on the center embedding vector, positive embedding vector, and negative embedding vector includes: calculating the center embedding vector, positive embedding vector, and negative embedding vector based on a preset loss function to obtain a loss value; optimizing the graph neural network model based on a preset gradient descent optimization method and the loss value to obtain an optimized graph neural network model; and inferring entity semantic relationship features based on the optimized graph neural network model to obtain dense embedding vectors.

[0011] Furthermore, the step of performing consistency analysis on the dense embedding vector based on the consistency check conditions, the preset consistency threshold, and the preset uncertainty threshold to obtain the consistency analysis result includes: verifying the dense embedding vector according to the preset consistency checker to obtain a consistency score. The dense embedding vector is analyzed according to the pre-set regression model to obtain the uncertainty index; the consistency score and uncertainty index are analyzed according to the consistency test conditions, consistency threshold and uncertainty threshold to obtain the consistency analysis results.

[0012] Furthermore, an intelligent agent consistency analysis device includes: a first analysis module for acquiring user interaction messages and analyzing them to obtain a set of compliance tags and authorized tool types; an access module for accessing target resources from a preset target resource object based on the set of compliance tags and authorized tool types; a channel determination module for determining available channels based on the target resource, a preset resource graph topology, a preset Bayesian analysis method, a preset Thompson sampling method, and a preset three-objective utility function; an execution module for executing available channels and, while the available channels are in execution state, collecting original unstructured documents and semi-structured documents; and a second analysis module for analyzing the original unstructured documents and semi-structured documents based on a pre-trained graph neural network model and preset consistency check conditions to obtain consistency analysis results.

[0013] Furthermore, an agent consistency analysis device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the agent consistency analysis device to perform the various steps of an agent consistency analysis method as described above.

[0014] In this invention, user interaction messages are collected and parsed to extract a set of compliance tags and authorized tool types, thus completing the pre-defined permission division and laying the foundation for full-process compliance control. Then, by matching compliance tags with authorized tool types, compliant access to target resources is achieved, permission boundaries are defined, and the risk of unauthorized access is prevented. Combining the resource graph topology, Bayesian analysis, Thompson sampling, and a three-objective utility function are integrated to filter available channels, optimize channel resource allocation, and ensure stable and compliant transmission links. Under normal operation of available channels, various types of unstructured and semi-structured documents are collected in a standardized manner to ensure the integrity and authenticity of the data source. Finally, relying on a pre-trained graph neural network model and consistency verification conditions, the semantic and topological features of documents are deeply mined to avoid fragmented analysis bias, quantify content credibility, and identify low-credibility data. This entire process forms a system of permission control, resource access, channel optimization, document collection, and intelligent verification, effectively improving business processing compliance, channel scheduling efficiency, and document analysis accuracy, providing stable and reliable support for downstream knowledge application and semantic reasoning. Attached Figure Description

[0015] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a first flowchart of an agent consistency analysis method provided in an embodiment of the present invention; Figure 2 This is a second flowchart of an agent consistency analysis method provided in an embodiment of the present invention; Figure 3 This is a third flowchart of an agent consistency analysis method provided in an embodiment of the present invention; Figure 4 This is a fourth flowchart of an agent consistency analysis method provided in an embodiment of the present invention; Figure 5 The fifth flowchart of an agent consistency analysis method provided in an embodiment of the present invention; Figure 6 The sixth flowchart of an agent consistency analysis method provided in an embodiment of the present invention; Figure 7 The seventh flowchart of an agent consistency analysis method provided in an embodiment of the present invention; Figure 8 The eighth flowchart of an agent consistency analysis method provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of an agent consistency analysis device provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of the structure of an intelligent agent consistency analysis device provided in an embodiment of the present invention. Detailed Implementation

[0016] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0017] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the agent consistency analysis method in this invention includes: 101. Obtain user interaction messages and analyze them to obtain a set of compliance tags and authorized tool types; In this embodiment, by collecting and parsing user interaction messages, a set of compliance tags and authorized tool types are extracted to establish a basis for permission division. This provides a preliminary judgment basis for subsequent resource access verification, channel matching and compliance control, standardizes the scope of access subject permissions, and ensures that subsequent business processes are carried out in a compliant and orderly manner. 102. Obtain the target resource from the preset target resource object based on the compliance tag set and authorized tool type; In this embodiment, based on the set of compliance tags and the constraints of authorized tool types, targeted compliant access to target resource objects is carried out, and a standardized resource access control mechanism is established. By matching and verifying compliance tags and tool authorizations, the boundaries of resource access permissions are defined, effectively preventing security risks such as unauthorized access and unauthorized resource retrieval. At the same time, the resource retrieval process is standardized to ensure the legality and orderliness of resource access, providing a compliant and reliable source of resource data for subsequent channel screening, document collection and intelligent analysis, and laying a solid foundation for the security and compliance of the entire business process. 103. Determine available channels based on target resources, preset resource map topology, preset Bayesian analysis method, preset Thompson sampling method, and preset three-objective utility function; In this embodiment, available channels are selected based on resource graph topology constraints combined with Bayesian analysis, Thompson sampling, and ternary utility functions. Channel adaptation standards are standardized at the resource graph topology level, balancing channel compliance and path rationality. A multi-dimensional evaluation mechanism matches business access requirements, effectively avoiding invalid, low-quality, and high-cost channels, and optimizing channel resource allocation logic. Simultaneously, it provides stable and reliable channel support for subsequent document collection and semantic reasoning, ensuring data transmission link security and compliance, and improving resource access adaptation efficiency and overall scheduling stability. 104. Execute available channels. While the available channels are in execution state, collect raw unstructured documents and semi-structured documents. In this embodiment, the original unstructured and semi-structured documents include common file types such as PDF (including scanned copies and text files), Microsoft Word / Excel, PPT, TXT, HTML, and PNG / JPG (including OCR support). Under the premise of stable operation of available channels, the original unstructured and semi-structured documents are collected in a standardized manner, providing a real and complete data source for subsequent model inference and consistency verification. Collection through reliable channels avoids problems such as missing or distorted data, ensuring the complete preservation of original document features, adapting to the needs of graph neural network semantic analysis, laying the foundation for consistency judgment, feature extraction, and compliance verification, and improving the authenticity of subsequent data analysis and the overall process coherence. 105. Based on a pre-trained graph neural network model and preset consistency check conditions, analyze the original unstructured documents and semi-structured documents to obtain consistency analysis results; In this embodiment, relying on the deep semantic perception and reasoning capabilities of a pre-trained graph neural network model, and combined with preset consistency verification conditions, a comprehensive consistency analysis is conducted on the original unstructured and semi-structured documents to generate reliable consistency analysis results. This effectively integrates the semantic information and topological association features of the documents, avoids the analysis bias caused by knowledge fragmentation, and realizes the verification and credibility determination of document information. At the same time, it quantifies the consistency and uncertainty of the document analysis results, identifies low-credibility content, provides a clear basis for process optimization, improves the reliability of document analysis, and provides high-quality and traceable support for subsequent downstream tasks such as knowledge transformation and semantic matching, ensuring the stability and compliance of the overall process. In this embodiment, user interaction messages are collected and parsed to extract a set of compliance tags and authorized tool types, completing the pre-delineation of permissions and laying the foundation for full-process compliance control. Then, by matching compliance tags with authorized tool types, compliant access to target resources is achieved, permission boundaries are defined, and the risk of unauthorized access is prevented. Combining the resource graph topology, Bayesian analysis, Thompson sampling, and a three-objective utility function are integrated to screen available channels, optimize channel resource allocation, and ensure stable and compliant transmission links. Under normal operation of available channels, various types of unstructured and semi-structured documents are collected in a standardized manner to ensure the integrity and authenticity of the data source. Finally, relying on a pre-trained graph neural network model and combining consistency verification conditions, the semantic and topological features of documents are deeply mined to avoid fragmented analysis bias, quantify content credibility, and identify low-credibility data. The entire process forms a system of permission control, resource access, channel optimization, document collection, and intelligent verification, effectively improving business processing compliance, channel scheduling efficiency, and document analysis accuracy, providing stable and reliable support for downstream knowledge application and semantic reasoning.

[0018] Please see Figure 2 A second embodiment of the agent consistency analysis method in this invention includes: 201. Analyze the set of compliance tags and the types of authorized tools according to the preset resource access rules to obtain the rule analysis results; In this embodiment, the preset resource access rules set multiple access constraints to match and analyze the compliance tag set with the authorized tool types. The rule determination logic determines whether to allow tool t to access target resource r, which requires all conditions to be met simultaneously: the compliance tag set must contain at least one valid compliance tag c, and the compliance tag c must have an authorization compatibility relationship with tool t, meaning that the tool t to be called belongs to the authorized tool category corresponding to compliance tag c. By comparing and verifying the coverage of compliance tags and the tool authorization compatibility relationship one by one, the pre-access qualification screening is completed, and the rule analysis results are output, providing a pre-access basis for subsequent tool call collection and compliance determination. 202. If the rule analysis result shows that the resource access rule is satisfied, then collect the type of the currently invoked tool; In this embodiment, when the rule analysis results determine that the current set of compliance tags and the type of tool used fully meet the preset resource access rule access requirements, it indicates that the user has the basic qualifications for tool invocation and resource access. The collection process is then triggered to capture and record the type of tool invoked in real time corresponding to the user's current request, and to lock the parameters of the tool object to be verified. This provides tool type input parameter support for subsequent compliance analysis of user interaction messages and compliance tags based on the first sufficient and necessary conditions, ensuring the pertinence and accuracy of subsequent compliance verification. 203. Perform compliance analysis on user interaction messages and compliance tag sets based on the preset first sufficient and necessary conditions and the current tool type to obtain the first compliance analysis result; In this embodiment, a preset first necessary and sufficient condition serves as the judgment criterion for compliance verification of user interaction messages, performing a full-domain compliance analysis on user interaction messages and their associated compliance tags. The first necessary and sufficient condition for passing the compliance check of user interaction messages is limited to: traversing each compliance tag carried by the user interaction message, all of which must satisfy the corresponding authorization constraints, i.e., each compliance tag c can authorize and adapt to the currently invoked tool type, and the currently invoked tool belongs to the scope of approved tools preset by the corresponding compliance tag; through tag-by-tag and tool-by-tool matching and verification, compliance screening at the message dimension and tag dimension is completed, generating the first compliance analysis result, which provides a judgment basis for the final granting of resource access permissions. 204. If the first compliance analysis result satisfies the first sufficient and necessary condition, then the target resource is obtained from the target resource object according to the authorized tool type and the preset resource access conditions; In this embodiment, the first compliance analysis result satisfies the first necessary and sufficient condition, indicating that the user interaction message and the set of compliance tags it carries have passed the compliance verification at the tool authorization adaptation level and are qualified to enter the resource access authorization verification stage. The resource access condition adds a privacy level matching constraint to the resource access. This constraint is a necessary access requirement for resource access. The specific judgment standard is: the privacy level of the target resource itself must not be higher than the privacy access level explicitly required by the currently calling tool. Based on this, the system securely retrieves and accesses the target resource from the target resource object according to the authorized use of the tool. Through privacy level matching control, the compliance and security of resource access are achieved, and the risk of privacy leakage is prevented. In this embodiment, a preliminary matching analysis is first performed on the set of compliant tags and the types of authorized tools using preset resource access rules to complete the initial screening of resource access qualifications. After meeting the requirements of the resource access rules, the type of currently invoked tool is collected in real time to provide parameter support for subsequent compliance verification. Based on the set first sufficient and necessary condition, authorization adaptation verification is performed on user interaction messages and the set of compliant tags, strictly controlling the compliance of messages and tags. After the verification is passed, matching constraints of resource privacy level and tool access level are superimposed. Only after multi-layer linkage review can the target resource be legally accessed. This solution constructs a layered access control system of tag authorization, rule screening, message compliance, and privacy level control. By setting restrictions at each layer and verifying at each level, it effectively regulates tool invocation and resource access behavior, avoids unauthorized operations and illegal access risks, strengthens data privacy protection capabilities, and improves the standardization, rigor, and security of resource permission control.

[0019] Please see Figure 3A third embodiment of an agent consistency analysis method according to the present invention includes: 301. Analyze the resource graph topology based on the target resources and the preset second necessary and sufficient conditions to obtain the second compliance analysis results; In this embodiment, the preset second necessary and sufficient condition serves as the criterion for determining the reachability of the target resource. Based on the connectivity characteristics of the resource graph topology, a full topology traversal and logical analysis are performed on the target resource to be accessed. Specifically, the second necessary and sufficient condition for resource reachability determination is: in the resource graph topology labeled g, there must exist a complete connected path from the starting access node to the target resource r for the target resource to be considered to have the basic conditions for topology reachability. By traversing and verifying the node associations and link connectivity of the resource graph one by one, the compliance determination of the target resource topology reachability is completed, and the second compliance analysis result is output, providing a preliminary topology basis for subsequent collection of passable channels and construction of the channel pool. 302. If the second compliance analysis result is that the target resource satisfies the second necessary and sufficient condition in the resource graph topology, then collect all passable information channels to obtain the channel pool. In this embodiment, when the second compliance analysis result determines that the target resource satisfies the second necessary and sufficient condition in the resource graph topology, i.e., there exists an effective connected path from the starting point to the target resource, it indicates that the target resource has a topological reachability. Then, a full channel collection process is triggered, traversing all passable information channels on the connected path in the resource graph topology, uniformly collecting and summarizing them to form a standardized channel pool. By filtering effective passable channels within the topological path and eliminating broken or failed link channels, a complete and effective set of candidate channels is provided for subsequent channel estimation analysis, sampling optimization, and optimal selection. 303. The channel pool is filtered based on Bayesian analysis, Thompson sampling method and three-objective utility function to obtain usable channels; In this embodiment, the constructed channel pool is used as the processing object. A multi-layer channel selection mechanism is formed by integrating Bayesian analysis, Thompson sampling, and a three-objective utility function. First, Bayesian analysis is used to perform probability estimation modeling on each channel in the channel pool, and the three types of score features—channel quality score, latency score, and cost score—are obtained by quantitative solution to achieve channel performance uncertainty assessment. Then, Thompson sampling is used to perform adaptive sampling based on the probability distribution of each channel's score, balancing channel exploration and utilization characteristics, and avoiding high-cost ineffective sampling by combining cost feature constraints, generating multi-dimensional sampling features. Subsequently, the three-objective utility function is introduced to perform multi-dimensional weighted fusion calculation on the sampling features to calculate the comprehensive utility score of each channel. The channel pool is sorted and optimized according to the utility score, and the available channels with the best comprehensive performance are selected to achieve global balance optimization of channel quality, transmission latency, and usage cost, providing reliable channel support for secure and efficient resource access. In this embodiment, the reachability of target resources is screened through resource graph topology compliance verification and node connectivity determination. A pre-access determination is performed based on the second necessary and sufficient condition, effectively eliminating unreachable resource links and simplifying subsequent processing. A standardized candidate channel pool is constructed by traversing the resource graph's connected paths to collect passable channels, laying the foundation for channel selection. The channel pool is then hierarchically screened using Bayesian analysis, Thompson sampling, and multi-objective utility functions. Combined with resource graph node connectivity and topology characteristics, quantitative evaluation of channel performance and selection of optimal paths are achieved. The overall resource graph topology verification and channel selection mechanism effectively avoids invalid channels and broken links, ensuring the compliance and stability of resource access channels. It also improves channel selection efficiency and resource matching, providing reliable support for subsequent channel performance evaluation, path optimization, and resource scheduling, efficiently completing reasonable channel selection and optimal configuration.

[0020] Please see Figure 4 A fourth embodiment of an agent consistency analysis method according to the present invention includes: 401. Based on Bayesian analysis, estimate and analyze the channel pool to obtain channel quality score characteristics, latency score characteristics, and cost score characteristics; In this embodiment, the operational status of each communication channel in the channel pool is probabilistically estimated and quantitatively analyzed based on the Bayesian posterior probabilistic inference mechanism. An initial prior probability distribution is set for the transmission quality, transmission delay, and resource cost of each communication channel in the channel pool. Then, real-time operational sample observation data of the channels is continuously introduced, and the confidence interval of the parameters is iteratively corrected based on Bayesian inference rules to complete the uncertainty estimation of various performance indicators of the channels, obtaining the posterior probability distribution. On this basis, the expectation solution and quantification normalization processing of the posterior probability distribution are performed, and the channel quality score feature, delay score feature, and cost score feature corresponding to each channel are calculated and output respectively. Through Bayesian probabilistic modeling, the uncertainty of channel performance is quantitatively evaluated, characterizing the three dimensions of channel transmission quality, transmission delay, and resource usage cost. This provides standardized and quantifiable basic feature support for subsequent channel sampling, utility calculation, and optimal selection. 402. Based on the Thompson sampling method and cost score features, the channel quality score features and latency score features are sampled to obtain the sampled features; In this embodiment, the Thompson sampling method is used as the sampling strategy, and the acquired cost score features are used as constraints to perform adaptive probability sampling on the channel quality score features and latency score features. Taking advantage of the exploration and utilization balance characteristics of Thompson sampling, random sampling is performed according to the probability distribution of each channel score, taking into account both the reuse of high-quality channels and the exploration of unknown potential channels. At the same time, cost score features are introduced for sampling constraints to avoid blind sampling of high-cost channels. Finally, sampling features containing information on quality, latency, and cost are generated, providing standardized sampling input data for subsequent multi-objective utility calculations. 403. Calculate the sampling features based on the three-objective utility function to obtain the utility score features; In this embodiment, a three-objective utility function is constructed that integrates quality, latency, and cost dimensions. The sampling features obtained by Thompson sampling are substituted into the three-objective utility function as a whole, and a weighted fusion operation is performed from the three dimensions of channel transmission quality, real-time transmission latency, and resource usage cost. The comprehensive evaluation score of each candidate channel under the current operating state is quantitatively calculated, and the utility score corresponding to each channel is generated to form a utility score feature, providing a unified quantitative evaluation standard for channel selection. 404. Filter the channel pool based on utility score characteristics to obtain available channels; In this embodiment, based on the utility score characteristics calculated for each channel, all candidate channels in the entire channel pool are sorted and selected for optimization. The channel with the highest utility score is selected for actual invocation according to the sorting rule from highest to lowest utility score. The system monitors the overall business operation budget consumption and the SLA indicator status of each channel in real time. The SLA indicator status specifically covers response time, task execution pass rate, and service availability. Once a business operation budget overrun is detected, or any SLA indicator exceeds the preset compliance threshold or an anomaly occurs, the alternative channel linkage switching mechanism is immediately triggered. Through real-time status monitoring and anomaly triggering logic, the system avoids business operation risks caused by single-channel performance failures and budget overruns, ensuring that business transmission is always in a compliant and stable operating state, and providing anomaly fallback guarantee for dynamic optimization scheduling of channels. In this embodiment, Bayesian analysis is used to probabilistically estimate and model the channel pool. Based on prior distribution and real-time observation data, the posterior probability is iteratively updated. After expectation calculation and normalization quantization, three types of score features are output: channel quality score, latency score, and cost score, enabling the assessment of channel performance uncertainty. Combined with the Thompson sampling method, the quality and latency features are adaptively sampled using the cost score feature as a constraint, balancing channel exploration and utilization while avoiding high-cost ineffective sampling. A three-objective utility function is constructed to perform multi-dimensional weighted fusion calculation of the sampled features, forming a unified quantitative utility score feature. This solves the problem of the one-sidedness of single-indicator evaluation. Based on the utility score feature, the channel pool is optimized, prioritizing the channel with the best overall performance. Budget consumption and SLA operation indicators are monitored in real time, and channel switching is automatically triggered when budget limits are exceeded or SLA violations occur. The overall solution achieves quantitative channel evaluation, intelligent sampling, multi-objective optimization, and dynamic fault-tolerant scheduling, effectively balancing transmission quality, latency, and cost requirements, and improving channel resource allocation efficiency, service operation stability, and service compliance assurance capabilities.

[0021] Please see Figure 5 The fifth embodiment of an agent consistency analysis method in this invention includes: 501. Based on a graph neural network model, reason about the original unstructured and semi-structured documents to obtain the center embedding vector, positive embedding vector, and negative embedding vector; In this embodiment, a graph neural network model is used to perform deep reasoning on the original unstructured and semi-structured documents, generating center embedding vectors, positive embedding vectors, and negative embedding vectors. This effectively integrates the semantic information and topological association features of the documents, avoiding the problem of knowledge fragmentation. It not only realizes the structural transformation of unstructured and semi-structured documents, but also clearly distinguishes between related and unrelated entities, strengthens the semantic representation capability of vectors, provides high-quality vector support for subsequent semantic matching and relationship mining, reduces the risk of misjudgment, improves the efficiency and accuracy of knowledge utilization, and provides a reliable data foundation for downstream tasks, ensuring the stability of the overall process. 502. Generate a dense embedding vector based on the center embedding vector, positive embedding vector, and negative embedding vector; In this embodiment, dense embedding vectors are generated based on the central embedding vector, positive embedding vector, and negative embedding vector. This can integrate the entity's own features and topological relationships, enhance the vector's semantic representation capability, effectively optimize the vector space distribution, shorten the distance between related entities and distance irrelevant entities, improve the degree of subsequent semantic matching, similarity calculation and model training, ensure high-quality output of embedding vectors, and stably support various downstream tasks of knowledge retrieval and intelligent reasoning. 503. Perform consistency analysis on dense embedding vectors based on consistency test conditions, preset consistency thresholds, and preset uncertainty thresholds to obtain consistency analysis results; In this embodiment, joint consistency analysis is performed on densely embedded vectors through consistency test conditions, consistency thresholds, and uncertainty thresholds to identify low-confidence and high-uncertainty generated results. By quantifying consistency scores and uncertainty indicators, the credibility of the results is standardized, providing a clear basis for triggering the reflection process, ensuring that the quality of generated results is controllable and traceable, improving the reliability and stability of the agent's reasoning results, and providing reliable support for subsequent task execution. In this embodiment, a graph neural network model is used to perform deep reasoning on unstructured and semi-structured documents, generating center embedding vectors, positive embedding vectors, and negative embedding vectors. This effectively integrates the semantic and topological features of the documents, achieving a structured transformation of document information, avoiding the problem of knowledge fragmentation, and distinguishing between related and unrelated entities. Then, based on the fusion of the three types of vectors, dense embedding vectors are generated, optimizing the vector space distribution, enhancing semantic representation capabilities, and providing high-quality vector support for semantic matching and model training. Finally, by combining consistency check conditions, consistency thresholds, and uncertainty thresholds, joint consistency analysis is performed on the dense embedding vectors to quantify the credibility of the results, identify low-credibility and high-uncertainty content, and provide a basis for judgment for process iteration and optimization. The overall solution is progressive, effectively reducing the risk of reasoning misjudgment, improving knowledge utilization efficiency and result credibility, and providing stable and reliable technical support for downstream businesses such as knowledge retrieval and intelligent reasoning.

[0022] Please see Figure 6 The sixth embodiment of an agent consistency analysis method in this invention includes: 601. A document knowledge base is constructed based on the original unstructured and semi-structured documents; In this embodiment, the document knowledge base is used to uniformly store and manage the original unstructured or semi-structured documents uploaded by users, providing basic data support for subsequent knowledge transformation, entity extraction, and relational reasoning. It has multi-format compatibility, automatic parsing, and vectorized indexing capabilities. The specific implementation process is as follows: Multi-format document parsing: The system has a built-in format adapter that can automatically recognize and parse common file types such as PDF (including scanned documents and text files), Microsoft Word / Excel, PPT, TXT, HTML, and PNG / JPG (with OCR support); for image files, it uses an optical character recognition (OCR) engine to extract text content; for table files, it fully preserves row and column structure information and converts it into structured text descriptions, ensuring that standardized text extraction can be completed for all types of documents; Slicing Strategy: To balance retrieval granularity and contextual integrity, the system provides three optional or combined text slicing modes: Slicing by paragraph, dividing based on natural paragraph boundaries such as line breaks and heading levels, suitable for documents with clear chapters; Slicing by fixed length, sliding the slicing in units of a preset number of characters such as 512 tokens, supporting overlapping windows to avoid semantic truncation; Slicing by semantics, calculating semantic similarity between sentences based on pre-trained language models such as BERT and Sentence-BERT, dynamically identifying semantic boundaries through clustering or segmentation point detection algorithms to ensure semantic coherence within slices, suitable for high-density semantic texts such as technical documents and contracts; The parsed and sliced ​​text content and original document metadata are input into a multilingual embedding model (such as text-embedding-3-large) to generate dense vectors, which are then uniformly stored in the database, completing the structured construction of the document knowledge base and forming a searchable, manageable, and reprocessable basic resource library; 602. Based on a pre-trained large language model, the document knowledge base is transformed to obtain a structured knowledge representation; In this embodiment, relying on the deep semantic perception capability of the pre-trained Large Language Model (LLM), semantic parsing and knowledge extraction are performed on unstructured text slices in the document knowledge base; through prompt word constraints and semantic pattern matching, the document content in natural language form is mapped to a high-dimensional feature space, completing the transformation of unstructured information into structured knowledge representation, sorting out the logical connections of knowledge, providing a standardized and normalized knowledge carrier for subsequent entity extraction and relationship recognition, and supporting complex knowledge relationship reasoning and question-answering interaction; 603. Based on a pre-defined named entity recognition model, entity extraction is performed on the structured knowledge representation to obtain category entity features; In this embodiment, a named entity recognition model (such as LLaMA-finetunedNER) is used to extract entities from structured knowledge representations. Entities of predefined categories such as people, organizations, products, technical terms, time, and location are identified and extracted from document slices. The category and recognition confidence of each entity are labeled simultaneously to achieve classification and feature extraction of different types of entities, forming standardized category entity features, which lays the foundation for subsequent mining of semantic relationships between entities. 604. Based on a preset joint extraction method, perform relation extraction on the categorical entity features to obtain entity semantic relation features; In this embodiment, a joint extraction or pipeline extraction method is used to perform relation extraction (RE). Based on the extracted category entity features, semantic associations between different entities are identified. The relationship type between entities is determined by a relation classification model, such as "[Product A]-[Applicable to]-[Scenario B]" or "[Company X]-[Established in]-

[2020] ". It is compatible with both open domain relation recognition and predefined relation template matching modes, flexibly adapting to the relation mining needs of general scenarios and specific domains. Finally, the semantic relation features of entities are output, clearly representing the logical associations between entities. 605. Based on the graph neural network model, reason about the semantic relationship features of entities to obtain the center embedding vector, positive embedding vector and negative embedding vector; In this embodiment, a graph neural network model is used to perform deep reasoning on the semantic relationship features of entities, generating center embedding vectors, positive embedding vectors, and negative embedding vectors. This provides a high-quality vector foundation for subsequent loss calculation, model optimization, and semantic matching. The specific implementation process is as follows: Triple generation and normalization: Combine the (head entity, relation, tail entity) obtained through entity extraction and relation identification to generate standard RDF triples; simultaneously perform entity normalization processing, including alias merging, unit standardization, time format unification, etc., and assign a unique entity identifier (EntityID) to each entity to effectively avoid the problem of knowledge fragmentation and ensure the integrity, consistency and standardization of the knowledge graph; Graph Storage and Reasoning: Normalized RDF triples are stored in a graph database to construct a knowledge graph in the form of a directed attribute graph. The system supports two types of core reasoning capabilities: first, multi-hop queries based on graph traversal, which can realize cross-level and relational queries such as "the location of the headquarters of a supplier of a certain product"; second, logical reasoning based on rules or graph neural networks (GNN), such as transitive reasoning: "A is a subsidiary of B, B is a subsidiary of C, and A is a grandchild of C", which fully explores the deep relationships hidden between entities. Neighbor Sampling and Aggregation: For a target node u in the knowledge graph, randomly sample K-order neighbor nodes, and fuse the feature information of the neighbor nodes to the target node u through the GraphSAGE aggregation function to complete the integration of local topological information, so that the features of the target node simultaneously include its own attributes and neighbor association information; Embedding mapping: The aggregated node features are nonlinearly transformed through a fully connected layer and the ReLU activation function to generate a fixed-dimensional center embedding vector. This completes the mapping of entity features to a high-dimensional semantic space, representing the entity's own attributes and topological relationship features; among them, the center embedding vector It is generated by aggregating the target node u and its neighbor information using GraphSAGE; positive embedding vector The negative embedding vector is taken from the neighboring node v in the knowledge graph that has a real connection relationship with the target node u. Nodes vn are generated by randomly sampling from the knowledge graph and having no real connection with the target node u; In this embodiment, a document knowledge base is constructed through multi-format document parsing, intelligent slicing, and vectorized indexing to achieve standardized storage and efficient retrieval of unstructured and semi-structured documents. A large language model is used to transform unstructured text into structured knowledge representation. Named entity recognition and joint relation extraction are combined to extract entity and semantic relationship features. Furthermore, a graph neural network is used for deep reasoning of entity relationships, generating center embedding vectors, positive embedding vectors, and negative embedding vectors to complete the construction of a knowledge graph and high-quality vector representation. This effectively solves problems such as knowledge fragmentation, low utilization of unstructured information, and insufficient entity association mining, ensuring the integrity, consistency, and standardization of knowledge. It provides reliable support for downstream tasks such as semantic retrieval, knowledge reasoning, and result verification, improving the agent's knowledge understanding, semantic matching, and complex reasoning capabilities, and enhancing the efficiency and stability of knowledge application.

[0023] Please see Figure 7 The seventh embodiment of an agent consistency analysis method in this invention includes: 701. Calculate the center embedding vector, positive embedding vector, and negative embedding vector based on the preset loss function to obtain the loss value; In this embodiment, the loss function is used to quantify the representation error of the graph neural network on entity relations in the knowledge graph. The loss value is calculated by substituting the center embedding vector, positive embedding vector, and negative embedding vector pair into the loss function. ( )= , In the formula, ( ) represents the loss function value corresponding to the center node u; σ is the Sigmoid activation function, which maps the vector dot product result to the (0,1) interval to represent the association probability; To obtain the expected value based on the negative sample distribution, The negative sample weighting coefficient controls the balance of loss between positive and negative samples. The vector inner product represents the similarity between the center node u and the positive sample neighbor node v; For negative sample distribution functions, random sampling based on the noise distribution of global node degrees in the knowledge graph is typically used. 702. Optimize the graph neural network model according to the preset gradient descent optimization method and loss value to obtain an optimized graph neural network model; In this embodiment, the loss value is used as the optimization objective. The stochastic gradient descent (SGD) optimizer is used to iteratively update the weights of the graph neural network model, and the loss function value is minimized through the backpropagation mechanism. During the optimization process, positive and negative sample pairs are continuously substituted to calculate the loss value, and the model parameters are continuously corrected. This makes the distance between adjacent entities with actual relationships in the knowledge graph closer in the high-dimensional vector space, and the distance between unrelated entities farther away, thereby strengthening the model's ability to represent the semantic relationships of entities. Finally, a converged optimized graph neural network model is obtained, which ensures the semantic accuracy and relevance of entity embedding vectors. 703. Based on the optimized graph neural network model, reason about the semantic relationship features of entities to obtain dense embedding vectors; In this embodiment, the optimized graph neural network model has the ability to capture the semantic relationships of entities in the knowledge graph, and can perform global feature reasoning and association representation of each entity in the graph; during the model reasoning stage, a vector solidification operation is performed, traversing all entities in the knowledge graph and outputting the dense embedding vector corresponding to each entity; this dense embedding vector integrates the entity's own attributes and global topological associations, and after being stored in the vector database, it can directly support the high-dimensional semantic matching task of the retrieval subsystem, providing a high-quality, structured entity vector foundation for subsequent consistency verification, evidence retrieval and other links; In this embodiment, the model's ability to learn topological relationships and semantic features is enhanced by fusing the center embedding vector, positive embedding vector, and negative embedding vector through a loss function. This is achieved through a two-way constraint that brings related entities closer together and distances them from irrelevant entities. A negative sample sampling strategy based on global node degree is adopted to ensure the effectiveness and representativeness of training samples and avoid representational bias. After iterative optimization using stochastic gradient descent, the graph neural network model can output a dense embedding vector that integrates entity attributes and global association information after convergence. This provides a high-quality semantic index for the vector database, supports the retrieval subsystem in achieving high-dimensional semantic matching, lays a solid foundation for subsequent consistency verification, evidence retrieval, and other processes, improves the accuracy and stability of agent knowledge reasoning and result verification, and ensures the reliability and efficiency of task execution in complex scenarios.

[0024] Please see Figure 8 The eighth embodiment of an agent consistency analysis method in this invention includes: 801. The dense embedding vector is validated according to the preset consistency checker to obtain a consistency score; In this embodiment, a consistency checker is used to perform consistency checks on the dense embedding vector and output an overall consistency score, which is used to quantify the degree of consistency among multiple candidate results; the dense embedding vector includes multiple candidate results. Consistency vote score : For each candidate result: , Consistency score: In the formula, m is the number of candidate results. The similarity score between the i-th candidate result and the j-th candidate result is calculated. By cross-comparing pairwise similarity scores, the consistency of individual candidates and the overall consistency are calculated, thereby achieving a quantitative evaluation of the consistency among multiple generated results. For consistency score; 802. Analyze the dense embedding vectors according to the preset regression model to obtain the uncertainty index; In this embodiment, the uncertainty index is used to characterize the confidence and unreliability of the current generated result. It is obtained by comprehensively analyzing the features of dense embedding vectors through a regression model. The regression model can be linear regression, neural network, or XGBoost. The model input includes consistency statistics, embedding distribution features, model confidence, and retrieval matching score. Through feature fusion and nonlinear fitting, a normalized uncertainty index u is output, providing a quantitative basis for the uncertainty dimension for subsequent consistency verification. 803. Based on the consistency test conditions, consistency threshold, and uncertainty threshold, perform consistency analysis on the consistency score and uncertainty index to obtain the consistency analysis results; In this embodiment, consistency scores and uncertainty indicators are jointly determined by a consistency threshold and an uncertainty threshold to identify low-confidence and high-uncertainty generated results, providing a clear basis for triggering the reflection process; the thresholds are set as follows: Consistency threshold: , Uncertainty threshold: , In the formula, The average historical consistency score. The standard deviation of the historical consistency score; The average of historical uncertainty. κ represents the historical uncertainty standard deviation; κ is the adjustment coefficient (usually 1–2); the above statistics are derived from the system's historical runtime dataset or validation set and can be dynamically updated according to business scenarios to ensure that the thresholds adapt to different tasks and data distributions; consistency check judgment rules: regeneration and verification: when v < or u≥ When the time comes, the reflection process is triggered to regenerate the candidate set and calculate the new consistency score and uncertainty index, and iteratively optimize the result quality; failure handling: if the verification conditions are still not met after triggering the reflection process multiple times, the current task is judged to be high-risk, and the system is automatically switched to manual intervention mode to avoid erroneous results and ensure system reliability. In this embodiment, the consistency among multiple candidate results is quantified by a consistency voting score formula. A regression model is used to integrate consistency statistical features, embedding distribution features, model confidence, and retrieval matching scores to characterize result uncertainty and avoid bias from a single dimension. A dynamic threshold setting mechanism is used to adaptively update consistency and uncertainty thresholds based on historical data, adapting to different business scenarios and data distributions and enhancing the universality of verification. When the consistency analysis results are insufficient or the uncertainty is too high, a reflection process is automatically triggered for regeneration and optimization. If continuous verification fails, manual intervention is switched, effectively reducing model illusions and erroneous outputs, ensuring controllable and traceable quality of generated results. This provides reliable quality assurance for tasks such as intelligent agent reasoning and knowledge question answering, improving the overall system operating efficiency and service credibility.

[0025] The above describes an agent consistency analysis method according to an embodiment of the present invention. The following describes an agent consistency analysis device according to an embodiment of the present invention. Please refer to [link / reference]. Figure 9 One embodiment of the intelligent agent consistency analysis device of the present invention includes: The first analysis module 1 is used to acquire user interaction messages and analyze them to obtain a set of compliance tags and authorized tool types. Access module 2 is used to access the target resource from a preset target resource object based on the set of compliance tags and the type of authorized tool. Channel determination module 3 is used to determine available channels based on target resources, preset resource graph topology, preset Bayesian analysis method, preset Thompson sampling method, and preset three-objective utility function; Execution module 4 is used to execute available channels. When the available channels are in execution state, it collects raw unstructured documents and semi-structured documents. The second analysis module 5 is used to analyze the original unstructured documents and semi-structured documents based on the pre-trained graph neural network model and preset consistency check conditions to obtain consistency analysis results. In this embodiment, user interaction messages are collected and parsed to extract a set of compliance tags and authorized tool types, completing the pre-delineation of permissions and laying the foundation for full-process compliance control. Then, by matching compliance tags with authorized tool types, compliant access to target resources is achieved, permission boundaries are defined, and the risk of unauthorized access is prevented. Combining the resource graph topology, Bayesian analysis, Thompson sampling, and a three-objective utility function are integrated to screen available channels, optimize channel resource allocation, and ensure stable and compliant transmission links. Under normal operation of available channels, various types of unstructured and semi-structured documents are collected in a standardized manner to ensure the integrity and authenticity of the data source. Finally, relying on a pre-trained graph neural network model and combining consistency verification conditions, the semantic and topological features of documents are deeply mined to avoid fragmented analysis bias, quantify content credibility, and identify low-credibility data. The entire process forms a system of permission control, resource access, channel optimization, document collection, and intelligent verification, effectively improving business processing compliance, channel scheduling efficiency, and document analysis accuracy, providing stable and reliable support for downstream knowledge application and semantic reasoning.

[0026] Figure 10 This is a schematic diagram of the structure of an agent consistency analysis device 900 provided in an embodiment of the present invention. The agent consistency analysis device 900 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 910 (e.g., one or more processors) and a memory 920, and one or more storage media 930 (e.g., one or more mass storage devices) for storing application programs 933 or data 932. The memory 920 and storage media 930 can be temporary or persistent storage. The program stored in the storage media 930 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the agent consistency analysis device 900. Furthermore, the processor 910 may be configured to communicate with the storage media 930 and execute a series of instruction operations in the storage media 930 on the agent consistency analysis device 900 to implement the steps of the agent consistency analysis method provided in the above-described method embodiments.

[0027] An agent consistency analysis device 900 may further include one or more power supplies 940, one or more wired or wireless network interfaces 950, one or more input / output interfaces 960, and / or one or more operating systems 931, such as Windows Server, MacOSX, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 10 The illustrated structure of an agent consistency analysis device does not constitute a limitation on an agent consistency analysis device. It may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0028] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0029] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for agent consistency analysis, characterized in that, include: Acquire user interaction messages and analyze them to obtain a set of compliance tags and authorized tool types; The target resource is obtained by accessing the preset target resource object based on the set of compliance tags and the type of authorized tool. Available channels are determined based on target resources, a preset resource map topology, a preset Bayesian analysis method, a preset Thompson sampling method, and a preset three-objective utility function. Execute the available channel; while the available channel is in execution state, collect raw unstructured documents and semi-structured documents. The original unstructured documents and semi-structured documents are analyzed based on a pre-trained graph neural network model and preset consistency check conditions to obtain consistency analysis results.

2. The agent consistency analysis method as described in claim 1, characterized in that, The step of accessing the target resource from the preset target resource object based on the compliance tag set and the authorized tool type includes: The set of compliance tags and the types of authorized tools are analyzed according to the preset resource access rules to obtain the rule analysis results; If the rule analysis result indicates that the resource access rule is met, then the type of the currently invoked tool is collected; Based on the preset first sufficient and necessary conditions and the current tool type, a compliance analysis is performed on the user interaction messages and compliance tag set to obtain the first compliance analysis result; If the first compliance analysis result satisfies the first necessary and sufficient condition, then the target resource is obtained from the target resource object according to the authorized tool type and the preset resource access conditions.

3. The agent consistency analysis method as described in claim 1, characterized in that, The process of determining available channels based on target resources, a preset resource graph topology, a preset Bayesian analysis method, a preset Thompson sampling method, and a preset three-objective utility function includes: The resource graph topology is analyzed based on the target resources and the pre-set second necessary and sufficient conditions to obtain the second compliance analysis results; If the second compliance analysis result is that the target resource satisfies the second necessary and sufficient condition in the resource graph topology, then all passable information channels are collected to obtain the channel pool; The channel pool is filtered using Bayesian analysis, Thompson sampling, and a three-objective utility function to obtain usable channels.

4. The agent consistency analysis method as described in claim 3, characterized in that, The method of filtering the channel pool based on Bayesian analysis, Thompson sampling, and a three-objective utility function to obtain usable channels includes: The channel pool is estimated and analyzed based on Bayesian analysis to obtain channel quality score characteristics, latency score characteristics, and cost score characteristics. The channel quality score and latency score are sampled based on the Thompson sampling method and cost score features to obtain the sampled features; The sampling features are calculated based on the three-objective utility function to obtain the utility score features; The channel pool is filtered based on utility score characteristics to obtain available channels.

5. The agent consistency analysis method as described in claim 1, characterized in that, The pre-trained graph neural network model and preset consistency check conditions are used to analyze the original unstructured and semi-structured documents to obtain consistency analysis results, including: The graph neural network model is used to reason about the original unstructured and semi-structured documents to obtain the center embedding vector, positive embedding vector and negative embedding vector. Generate a dense embedding vector based on the center embedding vector, positive embedding vector, and negative embedding vector; Consistency analysis is performed on dense embedding vectors based on consistency test conditions, preset consistency thresholds, and preset uncertainty thresholds to obtain consistency analysis results.

6. The agent consistency analysis method as described in claim 5, characterized in that, The graph neural network model is used to infer the original unstructured and semi-structured documents to obtain the center embedding vector, positive embedding vector, and negative embedding vector, including: A document knowledge base is constructed based on the original unstructured and semi-structured documents; The document knowledge base is transformed based on a pre-trained large language model to obtain a structured knowledge representation; Based on a pre-defined named entity recognition model, entity extraction is performed on structured knowledge representation to obtain category entity features; Based on a pre-defined joint extraction method, relation extraction is performed on the categorical entity features to obtain entity semantic relation features; The semantic relationship features of entities are inferred based on the graph neural network model to obtain the center embedding vector, positive embedding vector and negative embedding vector.

7. The agent consistency analysis method as described in claim 6, characterized in that, The process of generating dense embedding vectors based on the center embedding vector, positive embedding vector, and negative embedding vector includes: The center embedding vector, positive embedding vector, and negative embedding vector are calculated based on a preset loss function to obtain the loss value; The graph neural network model is optimized based on the preset gradient descent optimization method and loss value to obtain an optimized graph neural network model; The entity semantic relationship features are inferred based on the optimized graph neural network model to obtain dense embedding vectors.

8. The agent consistency analysis method as described in claim 5, characterized in that, The process of performing consistency analysis on dense embedding vectors based on consistency test conditions, a preset consistency threshold, and a preset uncertainty threshold to obtain consistency analysis results includes: The dense embedding vector is validated according to the preset consistency checker to obtain a consistency score; The dense embedding vectors are analyzed based on a pre-defined regression model to obtain an uncertainty index. Consistency analysis is performed on consistency scores and uncertainty indices based on consistency test conditions, consistency thresholds, and uncertainty thresholds to obtain consistency analysis results.

9. An agent consistency analysis device, characterized in that, include: The first analysis module is used to acquire user interaction messages and analyze them to obtain a set of compliance tags and authorized tool types. The access module is used to access and obtain target resources from a preset target resource object based on the set of compliance tags and the type of authorized tools. The channel determination module is used to determine available channels based on the target resource, the preset resource graph topology, the preset Bayesian analysis method, the preset Thompson sampling method, and the preset three-objective utility function. The execution module is used to execute available channels, and in the available channel execution state, it collects raw unstructured documents and semi-structured documents; The second analysis module is used to analyze the original unstructured and semi-structured documents based on a pre-trained graph neural network model and preset consistency check conditions to obtain consistency analysis results.

10. An agent consistency analysis device, characterized in that, The intelligent agent consistency analysis device includes: a memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the agent consistency analysis device to perform the steps of the agent consistency analysis method as described in any one of claims 1-8.

Images