An agent explainable search path generation system and verification method

Abstract

The application relates to the technical field of an explainable retrieval path generation system for an intelligent agent, and in particular to an explainable retrieval path generation system for an intelligent agent and a verification method. The method comprises the following steps: obtaining intelligent agent association knowledge retrieval data through association knowledge retrieval, and constructing an initial node multi-path structure based on the data; determining a path node causal reasoning chain and node semantic explanation information, and generating structured explainable path data; detecting explainable path deviation by using the structured path data, implementing path structure simulation correction based on the deviation identification result, obtaining path structure simulation correction data, and combining the correction data; optimizing an explainable retrieval path strategy, generating optimized path retrieval strategy data, performing intelligent agent explainable retrieval path update processing on the structured path data, and obtaining an updated intelligent agent explainable retrieval path; and the intelligent agent explainable retrieval path generation is used to realize more accurate path retrieval.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of explainable retrieval path generation for agents, and in particular to an explainable retrieval path generation system for agents and a verification method. BACKGROUND

[0002] Agents are increasingly widely used in complex decision-making scenarios such as question-answering systems, recommendation systems, and dialogue systems. However, when traditional agents perform path decision tasks, they often lack the explainability of path decision-making, i.e., they cannot clearly present why the agent chooses a certain path in the retrieval process, the behavior motivation of each node in the path, and the semantic association logic. This black-box path decision mechanism seriously restricts the application expansion of agents in security-sensitive or high-trust-demand scenarios, and there is a clear need for traceability and verification of the logic chain behind the path behavior of agents. In existing technologies, some solutions attempt to introduce graph structures, attention mechanisms, or semantic labels to label paths, but there are three common shortcomings: first, the ability to structure the causal reasoning relationship between path nodes is lacking, and the triggering logic and behavior dependency between nodes cannot be accurately described; second, semantic explanation is limited to static annotation at the node level, and does not form a continuous semantic expression of the overall path, resulting in fragmented and scattered explainability; third, there is a lack of dynamic correction mechanism and feedback learning closed loop for path behavior results once they deviate. However, traditional explainable retrieval path generation for agents has the problem of inaccurate data detection for explainable path deviation identification, and the problem of inaccurate data determination for path node causal reasoning chain. SUMMARY

[0003] Therefore, it is necessary to provide an explainable retrieval path generation system for agents and a verification method to solve at least one of the above technical problems.

[0004] To achieve the above-mentioned purpose, an explainable retrieval path verification method for agents includes the following steps:

[0005] Step S1: Obtain agent multi-modal input data; perform associated knowledge retrieval processing based on the agent multi-modal input data to obtain agent associated knowledge retrieval data;

[0006] Step S2: Construct node multi-path initial structure data based on the agent semantic associated knowledge retrieval data; determine path node causal reasoning chain data based on the node multi-path initial structure data; extract path node semantic explanation information based on the multi-path node initial path structure data; determine structured explainable path data based on the path node causal reasoning chain data and the path node semantic explanation information;

[0007] Step S3: detecting explainable path deviation identification data based on the structured explainable path data; performing path structure simulation correction processing on the explainable path deviation identification data to obtain path structure simulation correction data;

[0008] Step S4: performing explainable search path strategy optimization processing according to the path structure simulation correction data to obtain explainable path search strategy optimization data; performing agent explainable search path update processing on the structured explainable path data according to the explainable path search strategy optimization data to obtain agent explainable search path update data.

[0009] The present application introduces a multi-modal data processing mechanism, fuses heterogeneous input information such as text, image, voice, etc., improves the understanding ability of the agent to the complex input context, and ensures that the associated knowledge retrieval process has higher semantic matching accuracy and context awareness ability. In the path structure construction stage, through semantic pointing feature extraction and path node topology modeling, the integrity and expansibility of the path structure are strengthened, and multi-dimensional coverage and high-quality connection relationship construction of the candidate path in the semantic space are realized. The construction of the causal reasoning chain ensures the systematic identification of the trigger mechanism and behavior dependence logic between the path nodes, so that the path has clear execution basis and traceability. After generating structured expression combined with path node semantic explanation information, the path semantics have the ability of coherence, readability and structured output, which significantly improves the explainability of the retrieval path to the end user. In the deviation detection and simulation correction stage, with the help of path semantic jump analysis and logical consistency deviation calculation, the semantic interruption position can be accurately located and the path semantic strength attenuation condition can be evaluated, and further dynamic correction is performed through structure simulation, which ensures the continuity and robustness of the path in complex tasks. Finally, through the extraction of correction feedback and the verification of path strategy rationality, the adaptive adjustment and closed-loop evolution ability of the path optimization mechanism are realized, so that the agent path update process has the characteristics of feedback driving, strategy stability and dynamic enhancement, which improves the decision accuracy and controllability of the system in a high-complexity semantic environment. Therefore, the present application is an optimization processing of the traditional agent explainable search path generation, solves the problems of inaccurate detection of explainable path deviation identification data and inaccurate determination of path node causal reasoning chain data in the traditional agent explainable search path generation, improves the accuracy of detection of explainable path deviation identification data and the accuracy of determination of path node causal reasoning chain data.

[0010] The present application also provides an agent explainable search path generation system for executing the agent explainable search path verification method as described above, which comprises:

[0011] The associated knowledge retrieval processing module is configured to acquire multi-modal input data of the agent, and perform associated knowledge retrieval processing according to the multi-modal input data of the agent to obtain associated knowledge retrieval data of the agent.

[0012] The interpretable path determination module is configured to construct node multi-path initial structure data according to the semantic associated knowledge retrieval data of the agent, determine path node causal reasoning chain data according to the node multi-path initial structure data, extract path node semantic explanation information according to the multi-path node initial path structure data, and determine structured interpretable path data based on the path node causal reasoning chain data and the path node semantic explanation information.

[0013] The path structure simulation correction module is configured to detect interpretable path deviation identification data based on the structured interpretable path data, and perform path structure simulation correction processing on the interpretable path deviation identification data to obtain path structure simulation correction data.

[0014] The retrieval path update module is configured to perform interpretable retrieval path strategy optimization processing according to the path structure simulation correction data to obtain interpretable path retrieval strategy optimization data, and perform agent interpretable retrieval path update processing on the structured interpretable path data according to the interpretable path retrieval strategy optimization data to obtain agent interpretable retrieval path update data.

[0015] The agent interpretable retrieval path generation system of the present application can realize the agent interpretable retrieval path verification method of the present application, and is used as a medium for joint operation and signal transmission between modules to complete the agent interpretable retrieval path verification method. The internal modules of the system cooperate with each other, and through the fusion of multi-modal understanding, causal chain reasoning, semantic structure annotation and deviation correction optimization, the high interpretability, strong logic and self-adaptive updating capability of the agent retrieval path are realized, thereby significantly improving the path generation precision and stability of the agent in a complex semantic environment. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 It is a step flowchart of an agent interpretable retrieval path verification method.

[0017] Figure 2 It is a step flowchart of an agent interpretable retrieval path verification method. Figure 1 It is a detailed implementation step flowchart of step S3.

[0018] Figure 3 It is an associated knowledge retrieval processing diagram in the present application.

[0019] The implementation of the present application, functional characteristics and advantages will be further described with reference to the accompanying drawings. DETAILED DESCRIPTION

[0020] The technical method of the present application will be described clearly and completely in combination with the drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0021] In addition, the drawings are only schematic illustrations of the present application, and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and thus repeated description thereof will be omitted. Some block diagrams shown in the drawings are functional entities, which do not necessarily have to correspond to physically or logically independent entities. The functional entities can be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0022] It should be understood that although the terms "first", "second" and the like can be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, without departing from the scope of the exemplary embodiments, a first element can be called a second element, and similarly a second element can be called a first element. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0023] To achieve the above-mentioned purposes, please refer to Figures 1 to 3 An agent interpretable retrieval path verification method, comprising the following steps:

[0024] Step S1: Obtain the multi-modal input data of the agent; and perform associated knowledge retrieval processing according to the multi-modal input data of the agent to obtain associated knowledge retrieval data of the agent;

[0025] In the first stage of the path verification method, the input information base of the agent is obtained by constructing a multi-modal input data acquisition and processing flow. This step calls independent acquisition modules for sensory perception and text processing to obtain input information from three modal channels of image, speech, and text. Image input is collected by a high-definition industrial camera with a resolution of 1024x1024, speech input is completed by a far-field microphone array with a sampling rate of 16kHz, and text input is received through a natural language recording interface. All modal data is encoded in Tensor format. To eliminate redundant interference between modalities, ResNet-50 network is used to extract visual semantic feature vectors for images, STFT (Short-Time Fourier Transform) is used to convert speech input into a frequency spectrum graph, and then input it into the VGGish model to obtain audio semantic vectors, and BERT-base model is used to obtain semantic embedding for text. Subsequently, cross-modal attention weighting mechanism (Cross-Modal Attention Aggregator) is used to fuse three types of modal vectors, and the fused feature representation is sent to the embedding space alignment module to perform linear mapping and norm constraint, so that it is normalized to a semantic vector set with a dimension of 768. Based on the fused semantic representation data, the constructed semantic vector index engine (based on FAISS library) is called to perform Top-k semantic retrieval operation, and the retrieval range is limited in the entity and relationship index associated with the structured knowledge graph to output the agent associated knowledge retrieval data. This data is stored in JSON structured format, and the fields include: semantic query vector ID, matching entity ID, matching confidence, and matching path keyword, etc., which provides semantic support for subsequent path structure construction.

[0026] Step S2: constructing node multi-path initial structure data according to the agent semantic association knowledge retrieval data; determining path node causal reasoning chain data according to the node multi-path initial structure data; extracting path node semantic explanation information according to the multi-path node initial path structure data; determining structured interpretable path data based on the path node causal reasoning chain data and the path node semantic explanation information;

[0027] In the embodiment of the application, after completing semantic retrieval data extraction, enter the path structure construction phase. This phase takes the semantic association knowledge retrieval data of the agent as input, extracts all semantic matching entities and their corresponding context relationship information therefrom, and represents the semantic guiding relationship existing between the entities in the retrieval results by constructing a semantic pointing feature matrix. A rule-driven semantic abstraction algorithm is used to convert these semantic entities into path candidate nodes, wherein each node contains a node ID, a semantic topic label, a context fragment summary, and a confidence score. Subsequently, based on the edge relationship set of the knowledge graph, the semantic connection relationship between the candidate nodes is identified, and a BFS (Breadth-First Search) traversal is used to limit the connection path to a maximum depth of 4, thereby obtaining node semantic connection relationship data. Filtering constraint processing based on semantic correlation is performed on all connection paths, and only connection structures with a semantic consistency score greater than 0.8 are retained to form path effective connection structure data. On this basis, the NetworkX library is called to establish a multi-path topological structure network model, output a topological connection graph between nodes, and initialize the semantic importance weight of each edge to calculate initial weight data. The initial weight data and the connection structure are merged to construct node multi-path initial structure data. Subsequently, in the node structure graph, the control output field (such as execution behavior, reasoning target) and the response input field (such as pre-state, trigger condition) of each node are parsed, and the node control output field data and the response input field data are extracted. Based on the mapping relationship between the input and output fields, the control dependency chain existing in the path is identified. Time series analysis and logical rule induction are performed on these control dependency chains to form the node path behavior evolution law. Through the combination of dependency chain analysis and behavior law mapping, a causal trigger relationship identification operation is performed to construct path node causal basic units and combine to obtain path node causal reasoning chain data. Finally, the path node semantic explanation information is constructed in combination with the semantic label, context summary, and associated entity description information of the node. Based on the path node causal reasoning chain data and the path node semantic explanation information, the node order optimization and semantic field alignment processing are performed in accordance with the chain causal order reconstruction mechanism, the BILOU sequence labeling model is used to generate the path node semantic labeling structure, and the semantic continuity analysis is completed according to the syntactic continuity and semantic transition logic to identify the trusted semantic paragraph interval to generate structured path expression template data, and unified encoding output structured interpretable path data, which is packaged in a hierarchical path structure JSON file format, and the fields include node ID sequence, causal label, semantic role, time sequence position, and trusted level.

[0028] Step S3: detecting interpretable path deviation identification data based on the structured interpretable path data; performing path structure simulation correction processing on the interpretable path deviation identification data to obtain path structure simulation correction data;

[0029] In the embodiment of the application, when the structured interpretable path data is generated, the path deviation detection and correction stage is immediately entered. This step performs node semantic jump analysis for each structured path, traverses the path node sequence, extracts node semantic category and context semantic similarity score, and if the continuous node semantic similarity is lower than 0.4, it is determined as a semantic jump section. A path structure logical consistency deviation matrix is constructed, the semantic jump information and semantic field continuity data are compared, the deviation paragraph interval is extracted, and semantic interruption position clustering (using the DBSCAN clustering algorithm) is performed, and the path semantic interruption clustering data is output. On this basis, the semantic interpretation strength between path nodes (determined by the complexity of the interpretation template structure and the semantic consistency score weighted) is linearly regressed to obtain an interpretation strength decay sequence. By identifying the paragraph position with a significant drop in the sequence, the path semantic structure offset point is located. The path semantic structure offset data is used to perform deviation weight normalization processing, and the semantic deviation score is standardized to the range of [0, 1]. The path semantic deviation comprehensive data is synthesized, and the interpretable path deviation identification data is generated according to the comprehensive data, which includes jump point position, deviation intensity score, and node ID information. Then, the rule library and the benchmark sample path database are called to perform deviation cause analysis processing, and the deviation type (such as semantic mutation, causal rupture, etc.) and its cause characteristics are identified. According to the cause type, the pre-defined deviation correction strategy mapping table is called to output the corresponding correction strategy scheme, including node interpolation, path reconstruction, or edge weight adjustment. The path structure simulation correction module is executed, and the deviation paragraph in the path structure is simulated and modified at the structure level according to the correction strategy, to ensure that the output path structure simulation correction data is consistent with the structured interpretable path data, and only the deviation area is updated.

[0030] Step S4: According to the path structure simulation correction data, the interpretable retrieval path strategy optimization processing is performed to obtain interpretable path retrieval strategy optimization data; according to the interpretable path retrieval strategy optimization data, the agent interpretable retrieval path update processing is performed on the structured interpretable path data to obtain agent interpretable retrieval path update data.

[0031] In this embodiment of the invention, after the path structure simulation correction data is generated, the system enters the strategy optimization and update stage for interpretable retrieval paths. Correction node positions, path span variation ranges, and node semantic weight changes are extracted from the simulated correction path structure to form path simulation correction feedback data. This data is input into the correction strategy rationality verification module, which performs a multi-indicator comprehensive scoring calculation based on the path optimization objective function (such as minimum jump rate and maximum semantic consistency), outputting path correction strategy rationality data. For strategy schemes with scores higher than 0.9, strategy optimization data integration is performed, using this data to update the path matching rules and weight configurations in the original retrieval strategy library, generating interpretable path retrieval strategy optimization data. Finally, the path update processing module is called to update and reconstruct relevant fields (such as node order, path structure, and causal relationship mapping) in the structured interpretable path data according to the optimized retrieval strategy, outputting agent-interpretable retrieval path update data. The updated path data is repackaged into a standard path expression template for the agent's next retrieval decision, thus completing the closed-loop iterative operation of the path verification method.

[0032] Please see, Figure 3 This is a schematic diagram of the related knowledge retrieval process in this invention;

[0033] Preferably, step S1 includes the following steps:

[0034] Step S11: Acquire multimodal input data of the agent;

[0035] In the embodiment of the application, a multi-modal data acquisition system of the agent is constructed, which includes multiple heterogeneous sensing channels to ensure complete input information from three modalities of vision, hearing and text. The visual modality data is acquired by an industrial-grade high-definition camera with a resolution of 1920x1080 pixels, which is fixed at a preset position of the agent to observe the environment, and continuously acquires image sequences at a frame rate of 30 fps. The audio modality data is acquired by a 16-channel array microphone, with a sampling rate of 44.1 kHz, to acquire environmental sound and voice information, and the data format is a PCM-encoded raw audio stream. The text modality data is accessed through an interface to access text input sources, including structured database query text and unstructured text log files, and the text encoding adopts the UTF-8 standard. After the acquisition of each modality data, a timestamp synchronization mechanism is used to uniformly label each data to ensure the consistency of the corresponding time sequence of each modality data in the subsequent fusion. The original image data is converted into an RGB three-channel pixel matrix, the audio signal is pre-filtered and framed to generate a mel-frequency cepstral coefficient (MFCC) feature vector, and the text data is processed by word segmentation and word embedding to form a sequence feature vector. All the pre-processed multi-modal feature data is packaged into a standardized tensor format as the input basis for cross-modal fusion. After this step, the generated multi-modal input data set includes image feature tensors, audio feature tensors and text feature sequences, which are used in the next step.

[0036] Step S12: performing cross-modal attention adjustment processing according to the multi-modal input data of the agent, to obtain multi-modal weighted fusion data of the agent;

[0037] In the embodiment of the present application, based on the multi-modal feature tensor generated in step S11, a cross-modal attention mechanism is designed to perform weighted fusion processing. In specific implementation, a cross-modal attention module is constructed, and the module structure includes a multi-head attention mechanism (Multi-Head Attention) for performing parallel self-attention calculation on the feature vectors of the visual, audio and text modalities respectively. Taking the visual modality feature as a query (Query), and the audio and text modality features as keys (Key) and values (Value), a weighted attention score matrix is calculated to capture the correlation weight between modalities. After normalization by a Softmax function, the weight distribution of the interaction between each modality is generated. The weight distribution is multiplied by the corresponding modality feature to complete the weighted summation fusion, and a cross-modal fusion feature representation is output. In order to enhance the stability of the feature expression, residual connection (Residual Connection) and layer normalization (Layer Normalization) are applied to standardize the fusion result. Subsequently, a feedforward neural network (Feedforward Neural Network) is used to perform nonlinear transformation on the fusion feature to enhance the abstract level of the output. The agent multi-modal weighted fusion data is a multi-modal representation tensor with unified dimensions and high fusion degree, and the format is a unified three-dimensional tensor (batch_size x feature dimension x time step), which is used as the input of subsequent semantic fusion.

[0038] Step S13: performing semantic representation fusion processing according to the agent multi-modal weighted fusion data, so as to obtain input fusion semantic representation data;

[0039] In the embodiment of the present application, based on the multi-modal weighted fusion data output in step S12, a multi-layer stacked bidirectional long short-term memory network (BiLSTM) is used for time sequence semantic fusion processing. The network is composed of three layers of BiLSTM units, and the number of hidden units in each layer is set to 512. A Dropout layer is used between layers to prevent overfitting at a rate of 0.3. The multi-modal fusion tensor is input into the first layer of BiLSTM units to capture the context dependency within the time step. The feature vectors propagated in both directions are merged, and the hidden state sequence is output. It is sequentially transmitted to the subsequent BiLSTM layer for feature abstraction and semantic enhancement. Then, the attention pooling mechanism is used to weight the sum of the output sequence of BiLSTM, focusing on extracting the time step information that contributes most to the overall semantic representation, generating a fixed-length fusion semantic representation vector with a fixed dimension of 1024. The fusion semantic representation vector contains cross-modal temporal and spatial semantic information, and the structure is a one-dimensional vector, which meets the input specification of the subsequent semantic retrieval engine. The fusion process is implemented through the TensorFlow framework, and all weights are initialized as normal distribution and adjusted using the Adam optimization algorithm. At the end of this step, the input fusion semantic representation data is obtained as the basis for semantic query of associated knowledge retrieval.

[0040] Step S14: According to the input fusion semantic representation data, the associated knowledge retrieval process is performed, and the agent associated knowledge retrieval data is obtained.

[0041] In the embodiment of the present application, the input fusion semantic representation vector obtained in step S13 is input into the semantic associated knowledge retrieval module. The module is built-in with a high-dimensional vector index database based on the vector retrieval engine Faiss, and the database contains the semantic embedding vectors of entities and relationships in the pre-constructed structured knowledge graph. The retrieval process calculates the cosine similarity between the input fusion semantic vector and all vectors in the knowledge base, and performs Top-k (k=10) retrieval through a fast neighbor search algorithm (such as IVF+PQ index), and returns the most relevant entity and relationship candidate set. Then, based on the preliminary retrieval results, a secondary screening based on semantic consistency is performed, and the matching degree of the entity context description and the query semantics is combined to eliminate low confidence matching items, and the agent associated knowledge retrieval data is constructed, and the data structure includes: candidate entity ID, entity name, matching score, corresponding relationship path, context abstract text, etc. The data is packaged in JSON format to meet the input requirements of subsequent node multi-path structure construction, and ensure the accuracy of semantic information and causal reasoning chain construction. After completing this step, the agent associated knowledge retrieval data is used as the key input of the path verification method, and is transmitted to the path structure construction module.

[0042] Preferably, step S14 includes the following steps:

[0043] Step S141: semantic query representation extraction is performed according to the input fusion semantic representation data, and input semantic query representation data is obtained;

[0044] In the embodiment of the application, based on the input fusion semantic representation vector obtained in step S13, a semantic query representation extraction process is performed, and a fixed-structure multi-layer perception (MLP) network is used to map the input fusion semantic vector. The MLP network includes three fully connected layers, and the number of nodes in each layer is 1024, 512 and 256 respectively. The activation function uses ReLU (Rectified Linear Unit). After the input fusion semantic vector is mapped by the first layer, it is activated by ReLU nonlinearity, and then it is continuously mapped by the second layer and the third layer, and finally a 256-dimensional semantic query representation vector is output. The purpose of the mapping process is to convert the fusion semantic data into a query-specific low-dimensional representation, highlighting the semantic core features, facilitating subsequent similarity calculation and keyword retrieval. To avoid overfitting, a batch normalization (Batch Normalization) layer is added to the network, and L2 regularization is used to control the weight amplitude. After this step, the generated input semantic query representation data is a 256-dimensional fixed-length vector, which is used in the next step.

[0045] Step S142: semantic similarity calculation is performed according to the input semantic query representation data, and input semantic similarity data is obtained;

[0046] In the embodiment of the application, based on the input semantic query representation data obtained in step S141, semantic similarity calculation is performed for each knowledge vector in the knowledge base. The knowledge vectors in the knowledge base are 256-dimensional vectors with the same dimension as the input semantic query representation. The calculation process uses the cosine similarity algorithm, which is specifically: the input semantic query representation vector and each vector in the knowledge base are multiplied, and then divided by the product of the module lengths of the two vectors to calculate the cosine similarity value, which is in the range of [-1, 1]. Utilizing the high-performance matrix operation library BLAS (Basic Linear Algebra Subprograms) parallelization calculation acceleration, the similarity of millions of knowledge vectors is processed. After the calculation is completed, all knowledge vectors are sorted according to the similarity value, and the top N (such as N=50) high similarity data is filtered out to form a list of input semantic similarity data. The list includes the knowledge vector ID and its corresponding similarity score, which serves as the basis for subsequent keyword retrieval and candidate selection. After this step, the input semantic similarity data is output to support the precise matching of subsequent semantic retrieval.

[0047] Step S143: keyword retrieval processing is performed according to the input semantic similarity data and the input semantic query representation data, and input semantic keyword retrieval data is obtained;

[0048] In the embodiment of the present application, the high similarity knowledge vector obtained in step S142 and the semantic query representation of step S141 are used to perform keyword retrieval to extract keywords from the original text or vector corresponding to the input semantic query representation. The keyword extraction adopts the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm to calculate the word frequency weight, combines the TextRank algorithm to construct the keyword priority ranking, and selects the top 20 keywords as the retrieval keyword set. Then, for the high similarity knowledge base entries screened out in step S142, keyword matching is performed on each text content. The keyword matching adopts the Boolean retrieval logic, that is, the entries containing multiple keywords in the retrieval text have higher priority. The inverted index structure is combined to accelerate the keyword matching process, and the retrieval results are ranked according to the comprehensive score of the number of matched keywords and the keyword weight. The input semantic keyword retrieval data obtained by screening includes the matched knowledge entry ID, the corresponding keyword matching score, the keyword position and context information, and the data format is a structured table. The data is used as the input of the candidate knowledge fragment screening to support further semantic consistency evaluation.

[0049] Step S144: determining input semantic candidate knowledge fragment data based on input semantic keyword retrieval data and input semantic similarity data;

[0050] In the embodiment of the present application, the keyword matching result obtained in step S143 and the semantic similarity data of step S142 are combined to calculate the comprehensive score of the candidate knowledge fragment. The comprehensive score is calculated by weighted average, and the weight coefficients are respectively given to the keyword matching score and the semantic similarity score. The specific weight proportion is 0.6 and 0.4. For each knowledge entry, after calculating the comprehensive score, the knowledge fragments are ranked from high to low according to the score, and the top 30% of the knowledge fragments in the comprehensive score are selected as the input semantic candidate knowledge fragment data. Each candidate knowledge fragment data includes the knowledge entry ID, the text content, the comprehensive score, the keyword matching details and the corresponding semantic similarity. The screening ensures the balance of the knowledge fragment in the semantic relevance and the keyword coverage, and meets the subsequent context consistency evaluation requirement. The data storage adopts the relational database structure, which is convenient for subsequent indexing and query operation. After completing the step, the input semantic candidate knowledge fragment data is output, which is used for the next step of semantic context consistency evaluation.

[0051] Step S145: performing context sentence consistency evaluation on the input semantic candidate knowledge fragment data to obtain semantic context sentence consistency data;

[0052] In the embodiment of the application, the candidate knowledge fragments screened in step S144 are subjected to context sentence consistency evaluation. A natural language processing technology based on syntax dependency analysis is adopted, and a dependency syntax analyzer (such as the SpaCy toolkit) is used to perform sentence segmentation and syntax structure analysis on the text of each candidate knowledge fragment, to identify core components such as subject-predicate-object and inter-sentence relationships. Then, an inter-sentence similarity calculation algorithm is used to calculate the semantic similarity between adjacent sentences based on a sentence vector model (such as GloVe word embedding), to determine the semantic continuity between sentences. For multi-sentence text, a semantic adjacency matrix is constructed, and the semantic consistency interval in the text is divided according to a similarity threshold. A graph clustering algorithm (such as spectral clustering) is used to cluster the semantic adjacency matrix, to determine the sentence cluster with the highest consistency in the text, to quantify the sentence consistency score of each knowledge fragment, and to generate semantic context sentence consistency data, including sentence cluster distribution, continuity score and consistency label. The data is stored in a structured JSON format, facilitating subsequent screening. After this step, semantic context sentence consistency data is output, providing a basis for knowledge fragment screening.

[0053] Step S146: knowledge fragment screening processing based on semantic context sentence consistency data, to obtain input semantic knowledge fragment data;

[0054] In the embodiment of the application, the semantic context sentence consistency data obtained in step S145 is used to screen the candidate knowledge fragments. The screening rule is based on a consistency score threshold setting, and the threshold is set to 0.75. Knowledge fragments with a sentence consistency score higher than the threshold are selected as input semantic knowledge fragments. During screening, the comprehensive score (step S144) and the consistency score are combined to perform weighted composite score calculation, with the weighting ratio being 0.7 for the comprehensive score and 0.3 for the consistency score. The composite scores are sorted, and the top 50 knowledge fragments are selected to ensure the depth and coherence of the content semantics. Each knowledge fragment in the screening result contains text content, score details, sentence cluster information and context label. The screening operation adopts a batch processing mode, and SQL filtering and sorting are performed in a relational database to ensure high efficiency and stability. After screening is completed, input semantic knowledge fragment data is generated, serving as the input basis for the next knowledge provenance tracking.

[0055] Step S147: knowledge provenance tracking processing according to the input semantic knowledge fragment data, to obtain knowledge provenance tracking data;

[0056] In the embodiment of the present application, for the screened input semantic knowledge fragments, knowledge provenance tracking is performed. The provenance tracking searches the source identifier of the knowledge fragments, relies on the pre-constructed knowledge graph node database, wherein each knowledge node is associated with its creator, timestamp, version number and reference relationship metadata. Using the knowledge graph traversal algorithm, starting from each knowledge fragment node, recursively tracking the upstream associated nodes, the depth is limited to 3 layers. Record all knowledge nodes and their metadata on each tracking path to form a provenance chain. Through the node attribute and time sequence sorting of the provenance chain, the provenance path structure of the knowledge fragment is established. The provenance chain data is stored in a graph database (such as Neo4j) to realize efficient path query and relationship visualization. During the provenance tracking process, the node data is verified for version consistency, and the invalid or conflicting nodes are removed to output the knowledge provenance tracking data, which includes the provenance path node list, node attribute and version information. The data provides trust and context support for the next associated knowledge retrieval.

[0057] Step S148: According to the knowledge provenance tracking data and the input semantic knowledge fragment data, associated knowledge retrieval processing is performed to obtain agent semantic associated knowledge retrieval data.

[0058] In the embodiment of the present application, based on the knowledge provenance tracking data of step S147 and the input semantic knowledge fragment data of step S146, the associated knowledge retrieval is performed. This process combines the semantic knowledge fragments with their provenance chains to construct a complete knowledge context association network. Using knowledge graph relationship reasoning technology, path search algorithms (such as shortest path algorithm and graph convolution network) are applied to analyze the cause and effect and semantic relationship between knowledge nodes, and determine the association strength between knowledge. Then, all associated paths are integrated, and the weighted comprehensive score is calculated according to the provenance trust degree, semantic relevance and path coherence of the nodes. According to the comprehensive score sorting, the agent semantic associated knowledge retrieval data is output, and the data format includes knowledge fragment ID, text content, provenance chain information, associated path strength and context summary. The data is stored in JSON format, supporting subsequent node multi-path structure construction and causal reasoning. After this step, the agent semantic associated knowledge retrieval data forms a complete knowledge association graph foundation, ensuring the accuracy and explainability of the retrieval path verification.

[0059] Preferably, the step S2 of constructing the node multi-path initial structure data according to the agent semantic associated knowledge retrieval data comprises:

[0060] According to the agent semantic associated knowledge retrieval data, semantic pointing feature data is extracted;

[0061] In this embodiment of the invention, the semantic association knowledge retrieval data of the intelligent agent output in step S148 is used as the basis. This data includes information such as knowledge fragment text, tracing chain information, and association path strength. For this data, natural language processing (NLP) technology is used to extract semantic features from the knowledge fragment text. Specifically, word embedding technology (such as Word2Vec or GloVe) is used to convert the text into a vector representation. Combined with named entity recognition (NER) and keyword extraction algorithms, semantic pointing features pointing to specific entities, concepts, and actions in the text are extracted. This processing step includes syntactic analysis to identify the dependency relationships between verbs, nouns, and modifiers, forming directed semantic edges. By constructing a semantic pointing graph, representing each entity and the pointing relationships between them, semantic pointing feature data is output. This data structure is stored in the form of a graph, where nodes represent entities or concepts, edges represent semantic pointing relationships, and edge weights represent semantic association strength.

[0062] Semantic node abstraction is performed based on semantic pointing feature data to obtain path candidate node data;

[0063] In this embodiment of the invention, the aforementioned semantic pointing feature data is used to perform node abstraction processing to simplify and aggregate semantic information. This step employs a clustering algorithm (such as hierarchical clustering or density clustering DBSCAN) to aggregate the node feature vectors in the semantic pointing graph, merging semantically similar or functionally similar nodes into abstract nodes. During the abstraction process, clustering is determined based on the similarity of the semantic embedding vectors of the nodes and the connectivity between nodes. Each clustering result represents a path candidate node, reflecting the semantic induction of multiple semantic entities. The generated path candidate node data includes the node ID, the abstracted semantic description, and the original node set. This data facilitates the reduction of redundancy and improvement of computational efficiency during subsequent path construction. The node abstraction operation utilizes graph theory toolkits (such as NetworkX) to achieve the merging of graph nodes and attribute updates.

[0064] Semantic connection relationship recognition is performed based on the candidate node data of the path to obtain node semantic connection relationship data;

[0065] In the embodiment of the present application, based on the path candidate node data, the potential semantic connection relationship between nodes is identified, the semantic similarity matrix between abstract nodes is calculated, and the cosine similarity or Euclidean distance is used to measure the similarity of node semantic feature vectors. Then, the connection intention between nodes is analyzed by using a rule engine combined with pre-defined semantic connection rules (such as causal relationship, time sequence, functional dependence), and the effective connection edge is confirmed. In order to ensure the effectiveness of the connection relationship, the semantic path mining algorithm is used to extract the multi-hop associated path between nodes, and the node semantic connection relationship data is generated by verifying the context consistency on the path to describe the connection direction and connection strength weight between nodes in the form of directed edge. The data structure is stored in an adjacency list, which includes the starting node ID, the terminal node ID, the connection weight and the connection type label.

[0066] The node semantic connection relationship data is subjected to path constraint screening processing, so as to obtain path effective connection structure data;

[0067] In the embodiment of the present application, the node semantic connection relationship data contains a large number of connection candidates, in order to ensure the rationality and effectiveness of the path, the path constraint screening processing is performed. This step is based on the set constraint conditions, such as the upper limit of the path length (such as not more than 6 hops), the semantic connection strength threshold (eliminate the edges with a weight less than 0.3), the node type compatibility rule (ensure that the connection node type conforms to the business semantic logic) and the like. The screening method uses graph filtering technology, iterates the connection relationship, eliminates the connection edges that do not meet the constraints, and automatically maintains the connectivity of the graph. After the screening is completed, the path effective connection structure data is obtained, which contains the node connection subset and the corresponding structure topology relationship that meet the constraint conditions. The data is saved in the form of the adjacency matrix of the graph, which is convenient for subsequent topology network construction and weight initialization.

[0068] Based on the path effective connection structure data, the multi-path topology structure network is constructed, so as to obtain the multi-path node topology structure data;

[0069] In the embodiment of the present application, the path effective connection structure data is used to construct the multi-path topology structure network. The network shows the node and connection relationship in the form of a graph structure, supports the coexistence of multiple paths, reflects the semantic diversity and causal diversity. The construction process includes initializing the graph database node, labeling the abstract node attribute, inserting the effective connection edge in turn, and maintaining the weight and direction of the edge. In order to support efficient query and subsequent operation, the topology structure network adopts double storage of adjacency list and edge weight matrix. The connectivity and integrity of the topology structure are verified by the graph traversal algorithm (such as depth-first search and breadth-first search), so as to ensure that there is no isolated node and broken path, and the multi-path node topology structure data is output, which contains the node set, the edge set and the corresponding weight, and forms the complete semantic topology network foundation.

[0070] The multi-path node topology structure data is subjected to weight initialization processing to obtain initial weight data of the node topology structure;

[0071] In the embodiment of the application, for a multi-path node topology structure network, the weight parameters of nodes and edges are initialized. The node weight initialization is based on the centrality index of the node in the semantic direction feature, such as degree centrality, betweenness centrality, etc., and is calculated by a graph analysis algorithm. The edge weight initialization uses the previously screened semantic connection strength data, which is subjected to standardization processing and is uniformly mapped to the 0~1 interval. In order to enhance the weight expression capability, a weighted normalization strategy is introduced to ensure that the sum of the node weights is 1, and the edge weight reflects the connection strength and directionality. After the calculation is completed, the node and edge weights are stored in the weight matrix and the node attribute table to form the initial weight data of the node topology structure. This data structure has the basis to support subsequent path reasoning and weight updating.

[0072] The initial weight data of the node topology structure and the path effective connection structure data are used to construct initial node multi-path structure data.

[0073] In the embodiment of the application, the initial weight data of the node topology structure and the path effective connection structure data are combined to construct complete initial node multi-path structure data. This structure data is based on a graph data model, integrates node attributes (such as semantic abstract information, weight values) and edge attributes (such as connection types, connection weights), and forms a multi-dimensional semantic path structure. Graph database tools (such as Neo4j) are used for data import and management to support multi-path parallel storage and efficient access. The structure data includes node sets, edge sets, weight matrices, and path index information, which lays a foundation for subsequent path causal reasoning and semantic interpretation. After the construction is completed, the initial node multi-path structure data is output as the result of step S2 for use in step S3.

[0074] Preferably, the step S2 of determining the path node causal reasoning chain data according to the initial node multi-path structure data comprises:

[0075] The node control output field data is obtained by performing node control output field analysis according to the initial node multi-path structure data.

[0076] In the embodiment of the application, the node multi-path initial structure data is taken as input, which contains a node set, node attributes and connection relationships between nodes. For each node, a field parsing operation is performed, and a control output field of the node is identified. The control output field is defined as a key information field of the node affecting subsequent nodes or elements in a path, including but not limited to operation instructions, trigger event identifiers, action states, signal transmission and the like semantic tags. In the parsing process, a structured data parsing technology is used to traverse the node attribute table and extract fields related to the "control output". By using text analysis combined with rule matching, the field values are semantically classified, and the field types and scope of action are labeled. The parsing result forms node control output field data, which includes node ID, control output field name, field content and corresponding semantic tags, for subsequent causal reasoning.

[0077] According to the node multi-path initial structure data, node response input field parsing is performed, so as to obtain node response input field data;

[0078] In the embodiment of the application, response input field parsing is performed for each node based on the same node multi-path initial structure data. The response input field indicates control information or trigger signals received by the node from the outside or the predecessor node, including state feedback, signal acceptance identifier, condition satisfaction flag and the like. The same structured field parsing technology is used to scan the node attributes and the connection edge attributes of adjacent nodes, and all fields indicating the "response input" function are extracted. The semantic rule library is applied to classify and label the field content, so as to accurately identify the input signal types and trigger conditions. The parsing result is organized into node response input field data, which includes node ID, response input field name, field content and semantic tags, to form basic data that can be used for control dependency relationship construction.

[0079] Based on the node response input field data and the node control output field data, path node control dependency data is determined;

[0080] In the embodiment of the application, the control dependency relationship between nodes is analyzed by combining the node control output field data and the node response input field data. The control dependency relationship is defined as a direct or indirect influence of the control output field of node A on the response input field of node B. The matching analysis method is used to compare the node control output field with the node response input field one by one, and the dependency association is determined by the similarity of the field name, semantic tag and field content. The Boolean logic matching rule is used to determine the causal relationship between the fields, for example, the control dependency is determined when the output field "start signal" corresponds to the response input field "start trigger". The result forms path node control dependency data, which includes control node ID, controlled node ID, corresponding control output field and response input field pair and dependency relationship strength indicator. The data provides a basis for subsequent path behavior rule analysis.

[0081] Determine node path behavior evolution law according to path node control dependency data;

[0082] In the embodiment of the application, based on path node control dependency data, node path behavior evolution law analysis is carried out. The evolution law refers to the dynamic change characteristics of node control behavior and its dependency chain triggered by time or event. By using time series analysis method, combined with the event trigger time sequence of control dependency, a node behavior state transition model is constructed. Specifically, state machine or directed graph dynamic analysis technology is adopted to map the output state change of the control node and its influence on the input of the response node in time sequence. Through behavior trajectory tracking, the evolution path and evolution frequency of the node behavior are determined. The node path behavior evolution law data is output, which specifically describes the state transition sequence of each node in the path, the trigger timestamp and its influence mode on the node behavior, providing time sequence basis for inter-node causal reasoning.

[0083] According to the node path behavior evolution law, analyze the trigger dependency relationship between nodes, and obtain node interdependence trigger data;

[0084] In the embodiment of the application, the trigger dependency relationship between nodes is analyzed by using the node path behavior evolution law. The trigger dependency relationship is defined as the state change of one node triggering the response action of another node. By using trigger event capture technology, combined with node behavior state transition data, the trigger condition and response mechanism between nodes are identified. Through event sequence association algorithm, the sequence and trigger logic of node behavior are analyzed, the interference of incidental events is excluded, and the stable dependency trigger mode is extracted. The node interdependence trigger data is constructed, including trigger node ID, response node ID, trigger condition description, trigger time interval and trigger intensity measurement, supporting accurate construction of causal chain.

[0085] Identify path node causal basic unit data based on node interdependence trigger data;

[0086] In the embodiment of the application, based on the node interdependence trigger data, the basic unit constituting the path causal reasoning is identified. The causal basic unit is defined as a pair of node trigger pairs with clear causal relationship, including cause node, effect node and trigger mechanism. By combining trigger data, applying graph mining algorithm, all direct causal node pairs and multi-level indirect causal chains are identified. The causal basic unit meeting the theoretical time sequence and semantic logic is screened out, and invalid or abnormal trigger pairs are excluded. The path node causal basic unit data is generated, including causal unit ID, cause node information, effect node information, causal relationship description and trigger parameters, forming the smallest element for constructing complete causal chain.

[0087] According to the path node causal basic unit data, determine the path node causal reasoning chain data.

[0088] In the embodiment of the application, the path node causal reasoning chain is constructed according to the causal relationship order by using the path node causal basic unit data. The chain construction adopts a directed acyclic graph structure, and the units are sequentially connected to form a complete chain according to the causal connection relationship of the basic units. The node order is determined by a topological sorting algorithm to ensure the logical continuity and time sequence consistency of the reasoning chain. The process integrates the trigger timestamp and the dependence strength between nodes, and assigns weights and explanations to the chain nodes. The path node causal reasoning chain data is output, including chain ID, node sequence, causal relationship description, weight distribution and time sequence annotation, which provides direct data support for the generation of subsequent structured interpretable path data.

[0089] Preferably, the step S2 of determining the structured interpretable path data based on the path node causal reasoning chain data and the path node semantic explanation information comprises:

[0090] The chain order reconstruction processing is performed according to the path node causal reasoning chain data to obtain path node chain order reconstruction data.

[0091] In the embodiment of the application, the input data is path node causal reasoning chain data, including a node ID set, causal relationships between nodes, and timestamp or event sequence information triggered by the nodes. The chain order reconstruction adopts a directed acyclic graph (DAG) topological sorting method, and performs topological structure analysis on the nodes and their causal relationships in the input data. The specific operation process is as follows: a node adjacency table structure is constructed to store the out-edge and in-edge information of the nodes; then all nodes are traversed, and the nodes without predecessors are sequentially added to the sequence as starting points; then the out-edges of the nodes already added to the sequence are recursively deleted, the adjacency table state is updated, and nodes without predecessors are continuously extracted. The result of the topological sorting is the linear order of the nodes, reflecting the execution or trigger order of the nodes in the causal chain. The path node chain order reconstruction data is output, including a node sequence list and the order constraint relationship between nodes, providing a basis for subsequent time sequence positioning.

[0092] The path node time sequence positioning data is determined based on the path node chain order reconstruction data.

[0093] In the embodiment of the present application, the path node chain order reconstruction data provides the node execution order timing positioning data generation step. By introducing the timestamp inference and event synchronization technology, combined with the node event occurrence time or trigger interval data, the positioning of each node on the time axis is accurately calculated. The specific operation includes: for each node in the chain order, according to the known node trigger time data or estimated time interval, the start time and end time of the node are accumulated to form a time period label; for the missing timestamp node, the adjacent node time is used for interpolation processing to ensure the time continuity. The path node timing positioning data output includes node ID and corresponding time interval, which clearly shows the timing position of the node in the causal chain, and supports the subsequent semantic field alignment processing.

[0094] The path node timing positioning data and the path node semantic interpretation information data are field-aligned to obtain semantic field alignment data.

[0095] In the embodiment of the present application, the path node timing positioning data and the path node semantic interpretation information data are combined to complete the alignment matching at the field level. The input includes the timing information of the node and the semantic interpretation text or structured semantic label of each node. According to the node ID correspondence, the timing positioning data and the semantic interpretation information are mapped at the node level. Then, for the field structure inside the node, the field name and the semantic label are matched, the timing field and the semantic field are one-to-one corresponding according to the logical association principle, and the alignment table at the field level is formed. The alignment process adopts strict field name matching and semantic similarity calculation technology to eliminate irrelevant or repeated fields, and ensures the accurate correspondence of the data. The output semantic field alignment data structure includes node ID, time field, semantic field and their corresponding relationship, which is used as the basic data for subsequent semantic structure labeling processing.

[0096] The semantic field alignment data is processed by path node semantic structure labeling to obtain path node semantic labeling structure data.

[0097] In the embodiment of the present application, based on the semantic field alignment data, the semantic structure labeling is performed to clearly show the semantic role and hierarchical relationship of each field in the path node. The labeling process includes defining labeling tag system such as entity tag, action tag and condition tag, and performing semantic classification and hierarchical assignment on each field. The rule engine combined with the semantic analysis method is used to deeply analyze the aligned field content, identify the logical dependence and combination mode between fields, and form a hierarchical semantic structure. Through labeling, not only the semantic meaning of the field is clear, but also the context association between fields is revealed, ensuring the integrity and consistency of the semantic structure. The output path node semantic labeling structure data includes node ID, field name, labeling tag and semantic hierarchical structure description, which supports subsequent continuity analysis.

[0098] According to the semantic annotation structure data of the path node, semantic continuity analysis is performed to obtain path semantic continuity data;

[0099] In the embodiment of the application, the semantic continuity analysis is performed on the semantic hierarchy and field sequence in the annotated structure to detect the continuity and logical continuity in the semantic chain. The analysis process uses sequence analysis technology and a semantic network diagram to check whether the logical relationship between the semantic annotations of adjacent nodes is continuous, and to identify abnormalities such as semantic jumps, breaks or semantic ambiguities. The specific operation includes calculating the semantic similarity of the annotated fields of adjacent nodes, evaluating the cause-and-effect and time sequence connection between the fields, and marking the semantic continuous region and the breakpoint. The analysis outputs path semantic continuity data, and the data structure includes node pairs, continuity scores, abnormality identification and cause description, which assists subsequent trusted section identification.

[0100] Based on the path semantic continuity evaluation data, semantic trusted section identification is performed to obtain path semantic trusted section structure data;

[0101] In the embodiment of the application, the semantic trusted section identification is based on the semantic continuity score to divide the section in the path where the semantic expression is stable and consistent. The identification uses clustering analysis and threshold determination method to classify the node sequence with higher continuity score into the same trusted section, and the segmentation point is used as the boundary with lower trustworthiness. The process analyzes the semantic relationship strength between the nodes, combines the time continuity and logical consistency, and generates multiple semantic trusted sections, and the semantic expression in each section is complete and has no obvious break. The output path semantic trusted section structure data includes the start and end nodes of the trusted section, the section length, the trustworthiness index and the internal structure description, which provides semantic guarantee for path expression template generation.

[0102] The path semantic trusted section structure data and the path node semantic annotation structure data are merged to generate structured path expression template data;

[0103] In the embodiment of the application, the semantic trusted section structure data and the path node semantic annotation structure data are fused to form a comprehensive structured path expression template. The fusion process nests and combines the node annotation structure in each semantic trusted section according to the node ID and the semantic annotation hierarchy, and maintains the semantic hierarchy and time sequence. The data structure mapping method is used to ensure that the structured path expression template formed by seamlessly connecting the section boundary and the node semantic annotation contains multi-level semantic structure, time sequence information and section division label, and can completely express the semantic logic and time sequence relationship in the path. The data lays a data foundation for subsequent path structure unified output.

[0104] Based on the structured path expression template data, path structure unified output processing is performed to generate structured interpretable path data.

[0105] In the embodiments of the present application, according to the structured path expression template data, the unified formatting and output of the path structure are performed. The unified processing includes data format conversion, semantic label standardization and path structure normalization. The unified data model is used to define the path nodes, edges, semantic attributes and time information, so as to ensure that the output structure conforms to the predetermined specification interface format (such as JSON, XML, etc.). At the same time, data integrity check is performed to ensure the consistency of the path node causal relationship, semantic annotation and time sequence. The generated structured interpretable path data includes node sequence, causal chain, semantic annotation and credible section information, which are used for subsequent steps of agent path update and deviation identification.

[0106] Preferably, step S3 comprises the following steps:

[0107] Step S31: detecting interpretable path deviation identification data based on the structured interpretable path data;

[0108] In the embodiments of the present application, the structured interpretable path data is taken as input, and the path deviation detection is carried out for the multi-dimensional information such as the causal order, semantic annotation and time positioning of the path. Based on the path node chain order and time positioning data, the sequence comparison algorithm is used to compare the actual path sequence with the expected path sequence, and to identify the deviation such as node order abnormality, missing or inserted node. Secondly, the semantic annotation structure is used for node semantic consistency check, and the semantic label of the path node is compared with the reference standard to detect the abnormality such as semantic label mismatch and semantic logic break. Thirdly, the semantic continuity and credible section data of the path are combined to identify the poor semantic continuity and logic breakpoint position in the path. The above detection results are combined to form the path deviation identification report, which is specifically composed of fields such as deviation node set, deviation type, deviation position and deviation severity, to generate the interpretable path deviation identification data for subsequent deviation cause analysis.

[0109] Step S32: performing deviation cause analysis processing according to the interpretable path deviation identification data to obtain interpretable path deviation cause feature data;

[0110] In the embodiment of the application, according to the obtained explainable path deviation identification data, cause analysis is carried out on the deviation nodes and abnormal types. The analysis process includes: for node sequence abnormality, a dependency relationship backtracking method is used to search for upstream causal nodes and trigger conditions of the deviation nodes, and the specific reasons for the causal chain breakage or error triggering are located; for semantic mismatch, a semantic hierarchy is used to compare the context semantic environment of the deviation nodes, and the specific fields of the label error or semantic explanation deviation are identified; for semantic continuity breakage, combined with time positioning data, technical factors such as data acquisition delay, timestamp error or event omission are analyzed. The cause analysis includes a deviation cause feature vector, including a causal chain abnormality flag, a semantic consistency index and a time synchronization abnormality parameter, to form structured explainable path deviation cause feature data, which provides a basis for the correction strategy.

[0111] Step S33: According to the explainable path deviation cause feature data, an explainable path deviation correction strategy is generated, and the explainable path deviation correction strategy is obtained.

[0112] In the embodiment of the application, the deviation cause feature data output in step S32 is used to develop a targeted correction strategy. The correction strategy generation is based on the deviation type and the cause feature, adopts a rule-driven method combined with structured data analysis, and forms a rule set including sequence adjustment rules, semantic label correction rules and time synchronization correction rules. The sequence adjustment rules are defined for the causal chain breakage, and the operation instructions of node insertion, node sequence rearrangement and node deletion are defined; the semantic label correction rules are based on the semantic hierarchy and the context consistency, and the label replacement, field supplement and semantic re-labeling scheme are defined; the time synchronization correction rules use the timestamp correction, interpolation completion and time window adjustment strategy to solve the time asynchronous problem. The explainable path deviation correction strategy contains a plurality of rule instruction sets, clearly defines the correction object, the correction method and the execution order, and outputs the structured deviation correction strategy data for path structure simulation correction.

[0113] Step S34: The explainable path deviation correction strategy is used to perform path structure simulation correction processing on the explainable path deviation identification data, and path structure simulation correction data is obtained.

[0114] In the embodiment of the present application, the deviation correction strategy generated in step S33 is used to correct the deviation identification data of step S31. The correction process executes the rule instructions in the correction strategy in sequence, adjusts the path structure of the deviation nodes, including node order rearrangement, missing node supplement, abnormal node elimination and semantic label correction. The specific operation first reads the correction strategy instructions, modifies the path node chain data and semantic annotation structure in sequence, and ensures the consistency of the cause-effect order and semantic logic. The time stamp synchronization realizes the continuity of the node time sequence through interpolation and time window adjustment. During the correction process, the semantic continuity index of the path and the credible section division are updated in real time to ensure that the corrected path structure meets the logical requirements. After the correction is completed, the path structure simulation correction data containing the node adjustment record, the semantic label change log and the time synchronization correction information are output as the input of the subsequent path strategy optimization and update.

[0115] Preferably, step S31 comprises the following steps:

[0116] Step S311: based on the structured interpretable path data, analyze the semantic jump situation of the nodes to obtain path node semantic jump data;

[0117] In the embodiment of the present application, based on the structured interpretable path data, the semantic jump situation analysis is implemented for the semantic sequence of the path nodes. The specific operation includes: extracting the semantic annotation structure and time sequence positioning data of the path nodes from the structured path expression template data to construct the semantic sequence of the path nodes. The semantic similarity measurement method is used to continuously test the semantic labels of adjacent nodes, and the vector distance or correlation coefficient of the semantic labels of each pair of adjacent nodes is calculated to identify the semantic jump points exceeding the predetermined threshold. The semantic jump point is defined as a pair of nodes with too large semantic difference, resulting in the break of semantic continuity. In the analysis process, the dynamic window sliding strategy is applied to gradually scan the entire node sequence, and the jump frequency and jump intensity are counted to output the path node semantic jump data, including the jump node pair index, the jump intensity value and the jump position time sequence information, which provides the basis data for the subsequent path logic consistency comparison.

[0118] Step S312: according to the path node semantic interpretation information and the path node semantic jump data, perform comparison processing to obtain path structure logic consistency deviation data;

[0119] In the embodiment of the application, the logical consistency of the path structure is evaluated by comparing the path node semantic interpretation information with the path node semantic jump data obtained in step S311. Specifically, the semantic labels and their context semantic relationships in the path node semantic interpretation information are extracted, and a semantic correlation matrix is constructed to express the semantic dependence and context connection strength between nodes. In combination with the jump node pairs identified in the semantic jump data, for each jump point, the weak connection degree index of the jump point in the semantic correlation matrix is calculated to determine whether the jump causes a logical break in the path semantics. For the logical break nodes, they are marked as structural logical consistency deviation points, and the deviation types (such as semantic incoherence, context absence, semantic contradiction, etc.) are counted to form path structure logical consistency deviation data, and the data structure includes fields such as deviation node position, deviation category, and context semantic break strength, which are used as input for path semantic interruption clustering and subsequent analysis.

[0120] Step S313: Perform path semantic interruption position clustering processing based on the path structure logical consistency deviation data to obtain path semantic interruption clustering data;

[0121] In the embodiment of the application, for multiple deviation points in the path structure logical consistency deviation data, clustering processing is performed to identify semantic interruption clusters. The clustering method uses a density clustering algorithm based on time sequence distance and semantic deviation strength, and the specific operation process includes: calculating the time interval distance matrix between deviation nodes according to the time sequence positions of the deviation nodes; secondly, a weighted distance matrix is constructed in combination with the deviation strength index to enhance the sensitivity of clustering to serious deviation nodes. Then, using a density clustering algorithm such as DBSCAN or OPTICS, a node cluster with similar time sequence and deviation strength is identified as a semantic interruption section. The clustering result is output in the form of semantic interruption clustering data, including the position of each clustering center, the set of deviation nodes included in the clustering, and the length of the clustering section, which provides a basis for path explanation strength attenuation and structure deviation analysis.

[0122] Step S314: Perform path explanation strength attenuation evaluation based on the path structure logical consistency deviation data to obtain path explanation strength attenuation data;

[0123] In the embodiment of the present application, the degree of attenuation of the path explanation strength is evaluated for the semantic interruption and structural logical deviation in the path. In specific implementation, the path explanation strength index is defined, and the semantic consistency, context coherence and node importance weight of the path node are comprehensively considered. The local attenuation coefficient of the overall explanation strength of the path is calculated by using the deviation node and its deviation strength in the path structural logical consistency deviation data. The specific method includes: in the path node sequence, for each deviation point, the explanation strength value of the position and its adjacent nodes is proportionally reduced according to the deviation strength thereof. The overall path explanation strength is obtained by weighted accumulation of the node explanation strength, and the attenuation data is expressed as the dynamic change curve of the explanation strength of each node in the path. The evaluation result is saved in the form of path explanation strength attenuation data, which contains the node number, explanation strength value and strength change trend, and is used for subsequent semantic structure deviation determination.

[0124] Step S315: path semantic structure deviation determination based on path explanation strength attenuation data, to obtain path semantic structure deviation data;

[0125] In the embodiment of the present application, the deviation degree of the path semantic structure is determined based on the path explanation strength attenuation data obtained in step S314. The specific operation includes: trend analysis is performed on the explanation strength change trend of the path node, and the sliding average filtering and inflection point detection techniques are used to identify the section with obvious strength attenuation. Within the identified section, the deviation position and deviation amplitude of the path semantic structure compared with the standard path are determined by combining the path node time sequence positioning and semantic labeling information. The deviation amplitude is calculated to generate path semantic structure deviation data by measuring the change degree of the node semantic label and the order variation degree of the node in the section, and the data contains the deviation start and end nodes, the deviation amplitude value, the influence node list and the deviation type, which are used for subsequent deviation weight normalization processing.

[0126] Step S316: deviation weight normalization processing according to the path semantic structure deviation data, to obtain path semantic deviation comprehensive data;

[0127] In the embodiment of the present application, the normalization processing of the deviation weight is implemented for the deviation amplitude and influence node recorded in the path semantic structure deviation data. The specific method is as follows: first, the deviation weight is proportionally mapped to the [0, 1] interval according to the deviation amplitude, to ensure the uniformity of the weight scale; second, the normalized weight is weighted averaged by combining the number of nodes in the deviation section and the node importance index (such as the criticality of the node in the path, the connectivity, etc.), to form a comprehensive deviation weight value. The normalization process is realized by using numerical calculation tools, to ensure the uniformity and comparability of the deviation weight. The path semantic deviation comprehensive data is output, which contains the normalized weight of each deviation section, the detailed list of influence nodes and weight distribution information, to provide a quantitative basis for deviation identification.

[0128] Step S317: detecting explainable path deviation identification data according to path semantic deviation comprehensive data and path semantic structure deviation data.

[0129] In the embodiment of the application, the path semantic deviation comprehensive data and the path semantic structure deviation data are integrated to detect the comprehensive deviation of the path. The specific operation includes: combining the normalized deviation weight and the structure deviation amplitude, setting a plurality of deviation threshold values, and screening out path sections with significant deviation; secondly, for the screened deviation sections, the deviation weight, the deviation node information and the structure deviation characteristics are integrated to construct a structured deviation identification record, and the record fields include the deviation section number, the deviation severity level, the influence node sequence and the deviation type description. The deviation identification data is saved through a database or a data structure, which is used as basic data for subsequent deviation cause analysis and path correction, and realizes comprehensive quantitative identification of path deviation.

[0130] Especially important is that step S33 includes the following steps:

[0131] Step S331: performing path deviation type classification processing according to explainable path deviation cause feature data to obtain path deviation type classification data;

[0132] In the embodiment of the application, the field structure in the explainable path deviation cause feature data is standardized, and all path deviation data are composed into a standardized feature vector matrix according to the contents of "deviation node number", "semantic jump position", "structure fracture position", "explanation redundancy field" and the like. Then, a fixed rule pattern matching method is used for classification judgment, which includes the following rules: if the semantic correlation degree before and after the deviation node is less than 0.2, and the number of jump nodes in the path is greater than 2, it is classified as "semantic jump type deviation"; if there is no continuous dependence path between the output node and the target node in the cause-effect reasoning chain, it is classified as "cause-effect chain type deviation"; if there is a repeated explanation section or the explanation field information amount entropy value is too low in the path, it is determined as "semantic redundancy type deviation". Each type of deviation is marked with an independent label, and finally the path deviation type classification data is obtained.

[0133] Step S332: performing deviation trigger scene rule extraction processing according to the path deviation type classification data to obtain deviation trigger scene rule data;

[0134] In the embodiment of the application, after obtaining the path deviation type classification data, the deviation trigger scene rule extraction processing is performed. This step takes each deviation record in the classification data as input, and sequentially performs context rule screening processing on its path structure context information, node timestamp distribution, node source context and target context matching degree and other parameters. Rule extraction calls a rule set based on trigger type coding, for example, when the deviation type is "causal chain type deviation", the rule template group ID-C04 is called to extract the matching failure reason of the path interruption position and its front and rear node response / control fields. For "semantic jump type deviation", the rule template ID-J02 is called to extract the difference degree of entity distribution before and after the semantic disconnection jump node and the coverage rate of upper and lower attribute. The context state fields involved in each rule are retrieved through the records in the path structure simulation correction data and the path node semantic annotation structure data. The deviation trigger scene rule data is finally output.

[0135] Step S333: path reconstruction mode extraction processing based on the deviation trigger scene rule data is performed to obtain path reconstruction mode configuration data;

[0136] In the embodiment of the application, the path reconstruction mode extraction processing is carried out for the deviation trigger scene rule data generated in step S332. In this processing flow, first, a mapping table of deviation type and executable reconstruction mode is constructed. The mapping table is preconfigured with multiple path reconstruction means, including "path breakpoint logic bridging", "node semantic reconfiguration insertion", "interpretation window sliding expansion", "redundant node elimination and merging" and other structure adjustment means. Then, for each trigger rule data, according to the combination result of its scene classification label and path deviation type classification data, the executable reconstruction mode is screened in the mapping table. The screening criteria include structure connectivity maintenance rules, causal relationship preservation criteria and interpretation information integrity constraint logic. The matched path reconstruction mode is output through a standard configuration template to generate path reconstruction mode configuration data.

[0137] Step S334: strategy adaptability range evaluation processing based on the path reconstruction mode configuration data is performed to obtain path deviation correction strategy evaluation data;

[0138] In the embodiment of the present application, after obtaining the path reconstruction mode configuration data, a strategy adaptability range evaluation process is performed. This process takes each path reconstruction mode defined in the configuration data as input and evaluates its adaptability under different structure deviation conditions in turn. The evaluation indicators include: structure coherence score, semantic coverage ratio, explanation strength recovery rate, and path causal consistency index. The calculation of each indicator depends on the path node semantic annotation structure data and the original pre-and post-deviation node causal chain information, and a fixed weight weighting method is used for comprehensive evaluation. In the evaluation process, if a reconstruction mode is below the preset threshold (such as semantic coverage ratio less than 0.6) in more than three indicators, it is marked as "not adaptive", and the rest enter the strategy packaging process. The final path deviation correction strategy evaluation data is obtained.

[0139] Step S335: According to the strategy boundary packaging process, an interpretable path deviation correction strategy is generated.

[0140] In the embodiment of the present application, after completing the strategy adaptability range evaluation, the path reconstruction mode that meets the adaptation condition is executed. The strategy boundary packaging process. This processing step first reads the entries marked as "adaptive" in the strategy evaluation data, extracts the evaluation indicator range, logical constraint condition, action node range, and execution order parameters, and constructs a complete strategy template. Then, for each strategy, configure the control parameter structure, including path modification step, explanation window start and end points, causal verification threshold, and maximum number of node replacement control items. All parameters are packaged in the form of a data table. After completing the parameter filling, the strategy is packaged into a JSON format data structure, and the fields include: strategy ID, action path ID, strategy control parameter, adaptability label, binding deviation type, and trigger rule number. Finally, the output is an interpretable path deviation correction strategy.

[0141] Especially important is that step S34 includes the following steps:

[0142] Step S341: According to the interpretable path deviation identification data, a path deviation node clustering process is performed to obtain path deviation node clustering data;

[0143] In the embodiment of the application, after completing path deviation identification and obtaining interpretable path deviation identification data, path deviation node clustering processing is performed. First, each deviation node information recorded in the path deviation identification data is structured and organized, including node number, path position index, semantic category label, deviation type code, explanation intensity attenuation value, node semantic weight, context causal interruption level, and the like. Based on the above feature data, a deviation node feature vector set is constructed, and a K-Medoids clustering algorithm with parallel constraints of Euclidean distance and semantic classification label is used for clustering analysis processing. In the clustering process, random initialization is not introduced, but the position point with the maximum explanation intensity attenuation value is used as the initial center to improve the semantic saliency of the clustering center. After the clustering algorithm is run, the center node index of each cluster, the list of member node numbers in the cluster, the cluster label, and the semantic deviation mean index of the cluster are output, forming path deviation node clustering data as the index basis for local structure replacement.

[0144] Step S342: Perform path local structure replacement processing based on the path deviation node clustering data to obtain path structure local correction data;

[0145] In the embodiment of the application, path local structure replacement processing is performed based on path deviation node clustering data. First, according to the center node number of each cluster, the node and its front and rear 2-hop nodes are extracted from the structured interpretable path data to form a local path segment, forming a local path window. Then, according to the path deviation type classification data and the "recommended replacement mode" provided in the correction strategy, structural fragments with a similarity greater than 0.85 in semantic label and behavior mode corresponding to the path segment are retrieved from the semantic structure library to construct a candidate structure replacement fragment set. Subsequently, structure replacement matching degree evaluation is performed, and the highest scoring fragment is selected to replace the original node structure in the local path window according to the continuity of the causal reasoning chain, the semantic compatibility of the upstream and downstream nodes, and the original path length retention rate. After the structure replacement is completed, the node list before and after the modification, the structure change range, and the replacement strategy number are recorded for each path, and path structure local correction data is output.

[0146] Step S343: Perform structure splicing verification processing according to the path structure local correction data to obtain path structure splicing consistency verification data;

[0147] In the embodiment of the application, the structure splicing consistency verification processing is performed according to the local modified data of the path structure. First, the first and last connection node numbers of the replaced paragraph in each path are read, and the output field of the connection front node and the input field of the connection rear node are extracted based on the causal reasoning chain data to establish the field matching mapping relationship respectively. Then, the field semantic alignment degree evaluation operation is performed, which uses the Jaccard semantic similarity algorithm to calculate the ratio of the intersection and union of the semantic bag of words of the front and rear node fields, and if the similarity is lower than 0.5, it is marked as splicing anomaly and the original modified segment identifier is fed back. At the same time, whether the path structure forms a new closed loop or jump is verified, and whether an isolated subgraph or a non-directional closed path appears in the path is detected by using the structure graph traversal algorithm, and if the above situations appear, it is also marked as splicing conflict. Finally, the path structure splicing consistency verification data is formed, including the field corresponding relationship before and after splicing of each path, the semantic alignment score, the structure connectivity state identifier and the consistency evaluation result.

[0148] Step S344: performing semantic relationship reconstruction processing according to the path structure splicing consistency verification data to obtain path semantic relationship reconstruction data;

[0149] In the embodiment of the application, after completing the structure splicing consistency verification, the semantic relationship reconstruction processing is performed based on the splicing consistency verification data. The processing flow first locates the context semantic category and semantic function type field of the two nodes before and after the splicing segment of each path, analyzes whether the semantic span change or explanation direction mutation is introduced due to the splicing operation. For the connection point judged as “semantic transition interruption”, a transition explanation node is inserted, which is called from the structured semantic template library, and is required to have a semantic overlap degree greater than 0.7 with the front and rear nodes. Then the semantic dependency relationship between the splicing segment and the adjacent nodes is re-established, and the “explanation pre-mark bit” and “causal leader number” between the path nodes are updated, and the semantic relationship chain is reconstructed. The path node sequence after reconstruction is reorganized as a semantic structure graph to form path semantic relationship reconstruction data, and the data format includes: the reconstructed node directed edge table, the semantic label update log, the explanation function conduction path and the splicing operation record list.

[0150] Step S345: performing path structure simulation modification processing based on the path semantic relationship reconstruction data to obtain path structure simulation modification data.

[0151] In the embodiment of the present application, after completing the semantic relationship reconstruction, the path structure simulation correction processing is performed based on the path semantic relationship reconstruction data. First, the node graph structure in the semantic relationship reconstruction data is converted into a topological structure table, and is compared with the original structured interpretable path data to identify the structure evolution paragraphs and the unchanged paragraphs. Then, taking the evolution paragraph as the center, the path simulation engine module is called to simulate the response of the modified path structure under the condition of typical input semantic query. The simulation process adopts a three-step mode of semantic query, explanation expansion, and path backtracking, and records the explanation path length, node activation number, explanation redundancy rate, and end point arrival rate in each simulation round. When the simulation result meets the conditions of semantic coherence and structure convergence, the path is marked as "correction successful", otherwise the correction strategy rollback is continued. Finally, the path structure simulation correction data is generated, including path number, final correction structure graph, semantic explanation path node sequence, simulation verification index table, and structure replacement success flag, which is used for subsequent feedback extraction and strategy optimization stage calling.

[0152] Preferably, step S4 comprises the following steps:

[0153] Step S41: performing path simulation correction feedback extraction according to the path structure simulation correction data, so as to obtain path simulation correction feedback data;

[0154] In the embodiment of the present application, based on the path structure simulation correction data, the path simulation correction feedback extraction operation is implemented. The specific operation process comprises: obtaining the path structure simulation correction data, which contains the corrected path node sequence, the connection relationship adjustment between nodes, and the corresponding semantic label change record. By comparing the original structured interpretable path data and the simulation corrected data, the path structure change indicators such as the number of added and deleted nodes, the connection edge change frequency, and the semantic label adjustment amplitude are calculated. Further, through the node topology analysis tool, the key change points in the corrected path are identified, including the newly added semantic jump, the connection missing, and the sequence adjustment. Based on the above analysis results, the path simulation correction feedback data is extracted, and the data format contains the correction point coordinates, the change type, the change amplitude, and the corresponding timestamp, which is used as the input basis for the subsequent correction strategy rationality verification, so as to realize the quantitative feedback of the correction effect.

[0155] Step S42: performing correction strategy rationality verification according to the path simulation correction feedback data, so as to obtain path correction strategy rationality data;

[0156] In the embodiment of the application, the modification strategy rationality verification is implemented for path simulation modification feedback data. The specific operation process is: key modification indicators such as modification point distribution density, modification amplitude and modification frequency are extracted from the path simulation modification feedback data. The modified path structure is subjected to connectivity detection and cyclic dependency identification by using a structural integrity detection algorithm, to determine whether the modification leads to path structure fracture or abnormal loop. Combined with semantic coherence detection, the semantic consistency of the modified node is evaluated to ensure the smoothness of the node semantics after modification. The multi-index scoring mechanism is used to calculate the modification strategy rationality score according to the distribution rationality of the modification point, the path structure integrity and the semantic coherence. The scoring results are output in the form of path modification strategy rationality data, and the data content includes rationality score value, abnormal modification area identifier and related modification strategy parameters, supporting subsequent strategy optimization processing.

[0157] Step S43: perform explainable search path strategy optimization processing based on the path modification strategy rationality data to obtain explainable path search strategy optimization data.

[0158] In the embodiment of the application, the optimization processing of the explainable search path strategy is performed based on the path modification strategy rationality data. The specific operation process includes: receiving the rationality score and abnormal identifier in the path modification strategy rationality data, and adjusting the path modification strategy parameters such as node priority weight, path connection constraint and semantic continuity threshold according to the score result. The optimization algorithm (such as heuristic search or iterative optimization algorithm) is used to adjust the path search space and node connection rules, to strengthen the semantic coherence and structural integrity in the path. By constructing a multi-scheme comparison framework, the influence of different strategy parameter combinations on the path structure and semantic coherence is evaluated, and the optimal strategy parameter combination is selected to generate explainable path search strategy optimization data, including optimized strategy parameter settings, adjustment range and strategy effect evaluation index, for subsequent path update.

[0159] Step S44: perform agent explainable search path update processing on the structured explainable path data according to the explainable path search strategy optimization data to obtain agent explainable search path update data.

[0160] In the embodiment of the present application, based on the explainable path retrieval strategy, the data is optimized, and the explainable retrieval path update processing of the agent is implemented. The specific operation is: using the optimized strategy parameters, the node weight, connection relationship and semantic label in the structured explainable path data are re-adjusted. Through the path update module, the node priority is adjusted according to the optimization strategy, the path node is increased or decreased, the connection between nodes is re-built, and the balance of the path in the semantic coherence and structural integrity is ensured. The path update process relies on the path topology construction tool, and the causal logic chain between nodes is maintained in real time, and the semantic structure of the path node is re-labeled. After the update is completed, the explainable retrieval path update data of the agent is output, the data structure includes the updated path node list, connection topology relationship, semantic labeling result and update strategy parameter, which ensures the traceability and explainability of the path update, and supports the subsequent retrieval and verification operation of the agent.

[0161] The present application also provides an explainable retrieval path generation system of an agent for executing the explainable retrieval path verification method of an agent as described above, which comprises:

[0162] The associated knowledge retrieval processing module is configured to acquire the multi-modal input data of the agent, and perform associated knowledge retrieval processing according to the multi-modal input data of the agent to obtain the associated knowledge retrieval data of the agent.

[0163] The explainable path determination module is configured to construct node multi-path initial structure data according to the semantic associated knowledge retrieval data of the agent, determine path node causal reasoning chain data according to the node multi-path initial structure data, extract path node semantic explanation information according to the multi-path node initial path structure data, and determine structured explainable path data based on the path node causal reasoning chain data and the path node semantic explanation information.

[0164] The path structure simulation correction module is configured to detect explainable path deviation identification data based on the structured explainable path data, and perform path structure simulation correction processing on the explainable path deviation identification data to obtain path structure simulation correction data.

[0165] The retrieval path update module is configured to perform explainable retrieval path strategy optimization processing according to the path structure simulation correction data to obtain explainable path retrieval strategy optimization data, and perform explainable retrieval path update processing of the agent on the structured explainable path data according to the explainable path retrieval strategy optimization data to obtain explainable retrieval path update data of the agent.

[0166] The foregoing is considered as illustrative only of the principles of the application. Numerous modifications and changes will readily occur to those skilled in the art, and it is intended to embrace all such modifications and changes that fall within the scope of the application. Accordingly, the application is not to be restricted in scope to the specific embodiments disclosed herein but is to be accorded the full scope that the principles and novel features request appropriately granted.

Claims

1. An agent explainable retrieval path verification method, characterized in that, The method comprises the following steps: Step S1: obtaining multi-modal input data of an agent, including image, voice and text modalities; calling a constructed semantic vector index engine to perform associated knowledge retrieval processing in entity and relationship indexes associated with a structured knowledge graph according to the multi-modal input data of the agent, and obtaining agent associated knowledge retrieval data; Step S2: constructing node multi-path initial structure data according to the agent semantic associated knowledge retrieval data; determining path node causal reasoning chain data according to the node multi-path initial structure data, wherein the path node causal reasoning chain data comprises: performing node control output field analysis according to the node multi-path initial structure data, thereby obtaining node control output field data; performing node response input field analysis according to the node multi-path initial structure data, thereby obtaining node response input field data; determining path node control dependency data based on the node response input field data and the node control output field data; determining node path behavior evolution law according to the path node control dependency data; performing inter-node trigger dependency relationship analysis according to the node path behavior evolution law, thereby obtaining inter-node dependency relationship trigger data; identifying path node causal basic unit data based on the inter-node dependency relationship trigger data, wherein the causal basic unit is defined as a pair of node trigger pairs with clear causal relationship, including a cause node, an effect node and a trigger mechanism thereof; determining path node causal reasoning chain data according to the path node causal basic unit data; extracting path node semantic interpretation information according to the multi-path node initial path structure data; determining structured interpretable path data based on the path node causal reasoning chain data and the path node semantic interpretation information, wherein the structured interpretable path data comprises: performing chain order reconstruction processing according to the path node causal reasoning chain data, thereby obtaining path node chain order reconstruction data; determining path node time sequence positioning data based on the path node chain order reconstruction data; performing field alignment processing on the path node time sequence positioning data and the path node semantic interpretation information data, thereby obtaining semantic field alignment data; performing path node semantic structure annotation processing on the semantic field alignment data, thereby obtaining path node semantic annotation structure data; performing semantic continuity analysis according to the path node semantic annotation structure data, thereby obtaining path semantic continuity data; performing semantic trust segment identification based on the path semantic continuity evaluation data, thereby obtaining path semantic trust segment structure data; merging the path semantic trust segment structure data and the path node semantic annotation structure data to generate structured path expression template data; performing path structure unified output processing based on the structured path expression template data, thereby generating structured interpretable path data; Step S3: detecting interpretable path deviation identification data based on the structured interpretable path data, wherein the interpretable path deviation identification data comprises: performing node semantic jump situation analysis based on the structured interpretable path data, thereby obtaining path node semantic jump data; performing comparison processing according to the path node semantic interpretation information and the path node semantic jump data, thereby obtaining path structure logic consistency deviation data; The path semantic break position clustering processing is performed based on the path structure logical consistency deviation data, and path semantic break clustering data is obtained. The path explanation strength attenuation evaluation is performed based on the path structure logical consistency deviation data, and path explanation strength attenuation data is obtained. The path semantic structure offset determination is performed based on the path explanation strength attenuation data, and path semantic structure offset data is obtained. The deviation weight normalization processing is performed according to the path semantic structure offset data, and path semantic deviation comprehensive data is obtained. The intelligible path deviation identification data is detected according to the path semantic deviation comprehensive data and the path semantic structure offset data; the path structure simulation correction processing is performed on the intelligible path deviation identification data, and path structure simulation correction data is obtained. Step S4: The intelligible retrieval path strategy optimization processing is performed according to the path structure simulation correction data, and intelligible path retrieval strategy optimization data is obtained; the agent intelligible retrieval path update processing is performed on the structured intelligible path data according to the intelligible path retrieval strategy optimization data, and agent intelligible retrieval path update data is obtained. 2.The agent explainable retrieval path verification method of claim 1, wherein, Step S1 includes the following steps: Step S11: Obtain the agent multi-modal input data, including image, voice and text modalities; Step S12: The cross-modal attention adjustment processing is performed according to the agent multi-modal input data, so as to obtain agent multi-modal weighted fusion data; Step S13: The semantic representation fusion processing is performed according to the agent multi-modal weighted fusion data, so as to obtain input fusion semantic representation data; Step S14: The associated knowledge retrieval processing is performed according to the input fusion semantic representation data, so as to obtain agent associated knowledge retrieval data. 3.The agent explainable retrieval path verification method of claim 2, wherein, Step S14 includes the following steps: Step S141: The semantic query representation extraction is performed according to the input fusion semantic representation data, and input semantic query representation data is obtained; Step S142: The semantic similarity calculation is performed according to the input semantic query representation data, and input semantic similarity data is obtained; Step S143: The keyword retrieval processing is performed according to the input semantic similarity data and the input semantic query representation data, and input semantic keyword retrieval data is obtained; Step S144: The input semantic candidate knowledge segment data is determined based on the input semantic keyword retrieval data and the input semantic similarity data; Step S145: The context sentence consistency evaluation is performed on the input semantic candidate knowledge segment data, so as to obtain semantic context sentence consistency data; Step S146: The knowledge segment screening processing is performed based on the semantic context sentence consistency data, and input semantic knowledge segment data is obtained; Step S147: The knowledge provenance tracking processing is performed according to the input semantic knowledge segment data, and knowledge provenance tracking data is obtained; Step S148: The associated knowledge retrieval processing is performed according to the knowledge provenance tracking data and the input semantic knowledge segment data, so as to obtain agent semantic associated knowledge retrieval data. 4.The agent explainable retrieval path verification method of claim 1, wherein, In step S2, the node multi-path initial structure data is constructed according to the agent semantic associated knowledge retrieval data, including: The semantic pointing feature data is extracted according to the agent semantic associated knowledge retrieval data; The semantic node abstraction processing is performed based on the semantic direction feature data, so as to obtain path candidate node data; The semantic connection relationship identification processing is performed according to the path candidate node data, so as to obtain node semantic connection relationship data; The path constraint screening processing is performed on the node semantic connection relationship data, so as to obtain path effective connection structure data; The multi-path topological structure network is constructed based on the path effective connection structure data, so as to obtain multi-path node topological structure data; The weight initialization processing is performed on the multi-path node topological structure data, so as to obtain node topological structure initial weight data; The node multi-path initial structure data is constructed according to the node topological structure initial weight data and the path effective connection structure data. 5.The agent explainable retrieval path verification method of claim 1, wherein, Step S3 includes the following steps: Step S31: detecting interpretable path deviation identification data based on structured interpretable path data; Step S32: performing deviation cause analysis processing according to the interpretable path deviation identification data, to obtain interpretable path deviation cause feature data; Step S33: generating an interpretable path deviation correction strategy according to the interpretable path deviation cause feature data, to obtain the interpretable path deviation correction strategy; Step S34: performing path structure simulation correction processing on the interpretable path deviation identification data using the interpretable path deviation correction strategy, to obtain path structure simulation correction data. 6.The agent explainable retrieval path verification method of claim 1, wherein, Step S4 includes the following steps: Step S41: performing path simulation correction feedback extraction according to the path structure simulation correction data, so as to obtain path simulation correction feedback data; Step S42: performing correction strategy rationality verification according to the path simulation correction feedback data, to obtain path correction strategy rationality data; Step S43: performing interpretable retrieval path strategy optimization processing based on the path correction strategy rationality data, to obtain interpretable path retrieval strategy optimization data; Step S44: performing agent interpretable retrieval path update processing on the structured interpretable path data according to the interpretable path retrieval strategy optimization data, to obtain agent interpretable retrieval path update data.

7. An agent explainable search path generation system, characterized by, An agent interpretable retrieval path generation system for performing the agent interpretable retrieval path verification method as claimed in claim 1, the agent interpretable retrieval path generation system comprising: An associated knowledge retrieval processing module configured to acquire agent multi-modal input data; perform associated knowledge retrieval processing according to the agent multi-modal input data, to obtain agent associated knowledge retrieval data; An interpretable path determination module configured to construct node multi-path initial structure data according to the agent semantic associated knowledge retrieval data; determine path node causal reasoning chain data according to the node multi-path initial structure data; extract path node semantic explanation information according to the multi-path node initial path structure data; determine structured interpretable path data based on the path node causal reasoning chain data and the path node semantic explanation information; A path structure simulation correction module configured to detect interpretable path deviation identification data based on the structured interpretable path data; perform path structure simulation correction processing on the interpretable path deviation identification data, to obtain path structure simulation correction data; The retrieval path updating module is configured to perform an interpretable retrieval path strategy optimization process according to the path structure simulation correction data, to obtain interpretable path retrieval strategy optimization data; and perform an agent interpretable retrieval path updating process on the structured interpretable path data according to the interpretable path retrieval strategy optimization data, to obtain agent interpretable retrieval path updating data.

Images