Intelligent agent interpretable retrieval path generation system and verification method

Abstract

The invention relates to the technical field of intelligent agent interpretable retrieval path generation, in particular to an intelligent agent interpretable retrieval path generation system and a verification method. The method comprises the following steps: acquiring agent associated knowledge retrieval data through associated knowledge retrieval, and constructing a node multi-path initial structure based on the data; determining a path node causal reasoning chain and node semantic interpretation information, and generating structured interpretable path data; detecting explainable path deviation by using the structured path data, implementing path structure simulation correction based on a deviation identification result to obtain path structure simulation correction data, and combining the correction data; carrying out interpretable retrieval path strategy optimization to generate optimized path retrieval strategy data, and carrying out intelligent agent interpretable retrieval path updating processing on the structured path data to obtain an updated intelligent agent interpretable retrieval path; retrieval path generation can be explained through the intelligent agent, so that path retrieval is more accurate.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent agent interpretable retrieval path generation, and in particular to an intelligent agent interpretable retrieval path generation system and a verification method. Background Art

[0002] Intelligent agents are increasingly being used in complex decision-making scenarios, such as question-answering systems, recommendation systems, and dialogue systems. However, traditional intelligent agents often lack interpretability when performing path decision-making tasks. This lack of clarity reveals why the agent selected a particular path during the search process, the behavioral motivations for each node in the path, and the semantic association logic. This black-box path decision-making mechanism severely restricts the application and expansion of intelligent agents in security-sensitive or high-trust scenarios, requiring clear traceability and verifiability of the logical chain behind the agent's path behavior. In the existing technology, some solutions attempt to introduce graph structures, attention mechanisms or semantic labels to annotate paths, but there are generally three shortcomings: first, there is a lack of the ability to structuredly model the causal reasoning relationship between path nodes, and it is impossible to accurately characterize the triggering logic and behavioral dependencies between nodes; second, the semantic interpretation only stays at the static annotation level of the node, failing to form a continuous semantic expression of the overall path, resulting in fragmented and scattered interpretability; third, once the path behavior results deviate, there is a lack of dynamic correction mechanism and feedback learning loop. However, traditional intelligent agent interpretable retrieval path generation has the problem of inaccurate detection of interpretable path deviation identification data, and inaccurate determination of path node causal reasoning chain data. Summary of the Invention

[0003] Based on this, it is necessary to provide an agent-interpretable retrieval path generation system and verification method to solve at least one of the above technical problems.

[0004] To achieve the above object, an agent-interpretable retrieval path verification method is provided, comprising the following steps: Step S1: Acquire multimodal input data of the intelligent agent; perform associated knowledge retrieval processing based on the multimodal input data of the intelligent agent to obtain intelligent agent associated knowledge retrieval data; Step S2: constructing node multi-path initial structure data based on the agent semantic association knowledge retrieval data; determining path node causal reasoning chain data based on the node multi-path initial structure data; extracting path node semantic interpretation information based on 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; 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; Step S4: Perform interpretable retrieval path strategy optimization processing on the path structure simulation correction data to obtain interpretable path retrieval strategy optimization data; perform intelligent agent interpretable retrieval path update processing on the structured interpretable path data according to the interpretable path retrieval strategy optimization data to obtain intelligent agent interpretable retrieval path update data.

[0005] The present invention introduces a multimodal data processing mechanism to integrate heterogeneous input information such as text, images, and voice, thereby improving the intelligent agent's ability to understand complex input contexts and ensuring that the associated knowledge retrieval process has higher semantic matching accuracy and context perception capabilities. In the path structure construction stage, the integrity and scalability of the path structure are enhanced through semantic pointing feature extraction and path node topology modeling, and multi-dimensional coverage and high-quality connection relationship construction of candidate paths in the semantic space are achieved. The construction of the causal reasoning chain ensures that the trigger mechanism and behavior dependency logic between path nodes are systematically identified, so that the path has a clear execution basis and traceability. After generating a structured expression based on the semantic interpretation information of the path nodes, the path semantics have coherence, readability, and structured output capabilities, which significantly improves the interpretability 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 status can be evaluated. Further dynamic correction is performed through structural simulation to ensure the continuity and robustness of the path in complex tasks. Finally, by correcting feedback extraction and verifying the rationality of path strategies, the adaptive adjustment and closed-loop evolution capabilities of the path optimization mechanism are achieved, thereby making the agent path update process feedback-driven, strategy-robust, and dynamically enhanced, improving the system's decision-making accuracy and controllability in highly complex semantic environments. Therefore, the present invention optimizes the traditional agent-interpretable retrieval path generation, solving the problems of inaccurate detection of interpretable path deviation identification data and inaccurate determination of path node causal reasoning chain data in traditional agent-interpretable retrieval path generation, and improving the accuracy of detecting interpretable path deviation identification data and determining the accuracy of path node causal reasoning chain data.

[0006] The present invention also provides an agent-interpretable retrieval path generation system for executing the agent-interpretable retrieval path verification method described above. The agent-interpretable retrieval path generation system includes: The associated knowledge retrieval processing module is used to obtain multimodal input data of the intelligent agent; perform associated knowledge retrieval processing based on the multimodal input data of the intelligent agent to obtain intelligent agent associated knowledge retrieval data; The interpretable path determination module is used to construct node multi-path initial structure data based on the agent semantic association knowledge retrieval data; determine path node causal reasoning chain data based on the node multi-path initial structure data; extract path node semantic interpretation information based on 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 interpretation information; A path structure simulation and correction module is used to detect interpretable path deviation identification data based on structured interpretable path data; perform path structure simulation and correction processing on the interpretable path deviation identification data to obtain path structure simulation and correction data; The retrieval path update module is used to perform interpretable retrieval path strategy optimization processing on the path structure simulation correction data to obtain interpretable path retrieval strategy optimization data; and perform intelligent agent interpretable retrieval path update processing on the structured interpretable path data according to the interpretable path retrieval strategy optimization data to obtain intelligent agent interpretable retrieval path update data.

[0007] The intelligent agent interpretable retrieval path generation system of the present invention can implement the interpretable retrieval path verification method of any intelligent agent of the present invention, and is used to combine the operations and signal transmission media between various modules to complete the intelligent agent interpretable retrieval path verification method. The internal modules of the system cooperate with each other and achieve high interpretability, strong logic and adaptive update capabilities of the intelligent agent retrieval path by integrating multimodal understanding, causal chain reasoning, semantic structure annotation and deviation correction optimization, thereby significantly improving the path generation accuracy and stability of the intelligent agent in complex semantic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 A flowchart of a method for verifying an agent-interpretable retrieval path is provided; Figure 2 for Figure 1 Detailed implementation steps of step S3 in FIG. Figure 3 Schematic diagram of the associated knowledge retrieval process in the present invention; The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0009] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. It is obvious that the embodiments described are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts are within the scope of protection of the present invention.

[0010] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0011] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0012] To achieve this, please refer to Figures 1 to 3 , an agent-interpretable retrieval path verification method, comprising the following steps: Step S1: Acquire multimodal input data of the intelligent agent; perform associated knowledge retrieval processing based on the multimodal input data of the intelligent agent to obtain intelligent agent associated knowledge retrieval data; In this embodiment of the present invention, in the first phase of the execution path verification method, a multimodal input data acquisition and processing process is constructed to obtain the agent's input information foundation. This step invokes independent acquisition modules for sensor perception and text processing to acquire input information from three modal channels: image, voice, and text. Image input is acquired by a high-definition industrial camera with a 1024×1024 resolution, voice input is acquired via a far-field microphone array with a 16kHz sampling rate, and text input is received via a natural language recording interface. All modal data is uniformly encoded in Tensor format. To eliminate redundant interference between modalities, a ResNet-50 network is used to extract visual semantic feature vectors for images. Voice input is converted to a spectrogram using an STFT (Short-Time Fourier Transform) and then fed into a VGGish model to obtain audio semantic vectors. A BERT-based model is used to obtain semantic embeddings for text. Subsequently, the three modal vectors are fused through a cross-modal attention weighting mechanism (Cross-Modal Attention Aggregator), and the fused feature representation is fed into an embedding space alignment module, where linear mapping and norm constraints are applied to normalize them into a set of semantic vectors with a dimension of 768. Based on this fused semantic representation data, the constructed semantic vector indexing engine (built based on the FAISS library) is called to perform a Top-k semantic search operation. The search scope is limited to the entity and relationship indexes associated with the structured knowledge graph, and the output of the agent-related knowledge retrieval data is stored in a JSON structured format. The fields include: semantic query vector ID, matching entity ID, matching confidence, matching path keywords, etc., providing a semantic support foundation for the subsequent path structure construction.

[0013] Step S2: constructing node multi-path initial structure data based on the agent semantic association knowledge retrieval data; determining path node causal reasoning chain data based on the node multi-path initial structure data; extracting path node semantic interpretation information based on 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; In this embodiment of the present invention, after completing semantic retrieval data extraction, the path structure construction phase begins. This phase uses the agent's semantically associated knowledge retrieval data as input, extracting all semantically matching entities and their corresponding contextual relationship information. A semantically directed feature matrix is ​​constructed to represent the semantically guided relationships between entities in the retrieval results. A rule-driven semantic abstraction algorithm is used to convert these semantic entities into candidate path nodes, each containing a node ID, a semantic topic label, a contextual snippet summary, and a confidence score. Semantic connections between candidate nodes are then identified based on the knowledge graph's edge relationship set. Breadth-first search (BFS) is used to traverse connection paths with a maximum depth of four, thereby obtaining node semantic connection relationship data. Semantic relevance-based filtering constraints are applied to all connection paths, retaining only those connection structures with a semantic consistency score greater than 0.8 to form the valid path connection structure data. Based on this, the NetworkX library is used to build a multi-path topology network model, output a node topology connection graph, initialize the semantic importance weights for each edge, and calculate initial weight data. This initial weight data is then merged with the connection structure to construct the node multi-path initial structure data. Subsequently, within the node structure diagram, the control output fields (e.g., execution behavior, inference target) and response input fields (e.g., pre-condition, trigger condition) of each node are parsed, extracting the node control output field data and response input field data. Based on the mapping relationship between the input and output fields, the control dependency chains within the path are identified. Time series analysis and logical rule induction are performed on these control dependency chains to develop the behavioral evolution patterns of the node paths. By combining dependency chain analysis with behavioral pattern mapping, causal trigger relationship identification is performed, and causal basic units of path nodes are constructed. These units are then combined to obtain causal reasoning chain data for the path nodes. Finally, semantic interpretation information for the path nodes is constructed by combining the node's semantic label, context summary, and associated entity description information. Based on the path node causal reasoning chain data and path node semantic interpretation information, node order optimization and semantic field alignment processing are performed according to the chain causal order reconstruction mechanism. The BILOU sequence annotation model is used to generate the path node semantic annotation structure. According to the syntactic continuity and semantic transition logic, semantic continuity analysis is completed, and credible semantic paragraph intervals are identified to generate structured path expression template data. The structured interpretable path data is uniformly encoded and output. The data is encapsulated in the hierarchical path structure JSON file format. The fields include: node ID sequence, causal label, semantic role, temporal position and credibility level.

[0014] 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; In this embodiment of the present invention, once structured interpretable path data is generated, the path deviation detection and correction phase immediately begins. This step performs node semantic jump analysis on each structured path. The path node sequence is traversed, and node semantic categories and context semantic similarity scores are extracted. If the semantic similarity of consecutive nodes is less than 0.4, the segment is identified as a semantic jump. A path structure logical consistency deviation matrix is ​​constructed. Semantic jump information and semantic field continuity data are compared to extract deviating segment intervals. Semantic break location clustering (using the DBSCAN clustering algorithm) is performed, and path semantic break clustering data is output. Based on this, a linear regression is performed on the semantic interpretation strength between path nodes (determined by the weighted interpretation template structural complexity and semantic consistency score) to generate an interpretation strength decay sequence. By identifying the segments with significant decreases in this sequence, path semantic structure deviation points are located. Deviation weight normalization is performed using the path semantic structure deviation data, normalizing the semantic deviation score to the range [0, 1]. This is then used to synthesize path semantic deviation comprehensive data. Based on this comprehensive data, interpretable path deviation identification data is generated, which contains information such as the jump point location, deviation strength score, and the node ID to which it belongs. Next, the rule library and the benchmark sample path database are used to perform deviation cause analysis, identifying the deviation type (such as semantic mutation, causal break, etc.) and its causal characteristics. Based on the cause type, a predefined deviation correction strategy mapping table is called to output the corresponding correction strategy, which can be node interpolation, path reconstruction, or edge weight adjustment. The path structure simulation correction module is executed to perform structural simulation modifications to the deviation sections in the path structure based on the correction strategy. This ensures that the output path structure simulation correction data is consistent with the structured interpretable path data, with only the deviation areas being updated.

[0015] Step S4: Perform interpretable retrieval path strategy optimization processing on the path structure simulation correction data to obtain interpretable path retrieval strategy optimization data; perform intelligent agent interpretable retrieval path update processing on the structured interpretable path data according to the interpretable path retrieval strategy optimization data to obtain intelligent agent interpretable retrieval path update data.

[0016] In this embodiment of the present invention, after the path structure simulation and correction data is generated, the interpretable retrieval path strategy optimization and update phase begins. The simulated and corrected path structure contains the corrected node positions, path span variation range, and node semantic weight changes, generating the path simulation correction feedback data. This data is input into the correction strategy rationality verification module, which calculates a multi-metric comprehensive score based on the path optimization objective function (e.g., minimum jump rate, maximum semantic consistency), and outputs path correction strategy rationality data. Strategy solutions with scores above 0.9 are then subjected to strategy optimization data integration, which is used 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 invoked to update and reconstruct relevant fields in the structured interpretable path data (e.g., node order, path structure, and causal relationship mapping) based on the optimized retrieval strategy, outputting updated interpretable retrieval path data for the agent. The updated path data is repackaged into a standard path expression template for use by the agent in its next retrieval decision, thus completing the closed-loop iterative operation of the path verification method.

[0017] Please see, Figure 3 Schematic diagram of the associated knowledge retrieval process in the present invention; Preferably, step S1 includes the following steps: Step S11: Obtaining agent multimodal input data; In an embodiment of the present invention, a multimodal data acquisition system for an intelligent agent is constructed. The system includes multiple heterogeneous sensing channels to ensure that complete input information is obtained from the three modalities of vision, hearing, and text. Visual modality data is acquired by configuring an industrial-grade high-definition camera with a resolution of 1920×1080 pixels. The camera is fixed at a preset position in the intelligent agent's observation environment and continuously captures image sequences at a frame rate of 30fps. Audio modality data is collected by a 16-channel array microphone with a sampling rate set to 44.1kHz. It collects environmental sounds and voice information, and the data format is a PCM-encoded raw audio stream. Text modality data is connected to a text input source through an interface, including structured database query text and unstructured text log files. The text encoding adopts the UTF-8 standard. After the acquisition of each modality data is completed, a timestamp synchronization mechanism is used to uniformly annotate each data to ensure the consistency of the corresponding time sequence of each modality data during subsequent fusion. The raw image data is converted into a three-channel RGB pixel matrix. The audio signal is processed through pre-emphasis filtering and frame segmentation to generate Mel-Frequency Cepstral Coefficient (MFCC) feature vectors. The text data is processed through word segmentation and word embedding to form sequence feature vectors. All preprocessed multimodal feature data is encapsulated into a standardized tensor format as the input for cross-modal fusion. After completing this step, the generated multimodal input dataset contains image feature tensors, audio feature tensors, and text feature sequences for use in the next step.

[0018] Step S12: performing cross-modal attention adjustment processing based on the agent multimodal input data, thereby obtaining agent multimodal weighted fusion data; In an embodiment of the present invention, based on the multimodal feature tensor generated in step S11, a weighted fusion process is performed by designing a cross-modal attention mechanism. In specific implementation, a cross-modal attention module is constructed, and the module structure includes a multi-head attention mechanism (Multi-Head Attention), which performs parallel self-attention calculations on the feature vectors of the visual, audio, and text modalities respectively. Using the visual modality feature as the query (Query), and the audio and text modality features as the key (Key) and value (Value), the weighted attention score matrix is ​​calculated to capture the correlation weights between the modalities. After normalization by the Softmax function, a weight distribution of the interaction between each modality is generated. The weight distribution is multiplied by the corresponding modal feature to complete the weighted sum fusion and output the cross-modal fusion feature representation. In order to enhance the stability of the feature expression, the residual connection (Residual Connection) and layer normalization (Layer Normalization) are applied to standardize the fusion results. Subsequently, a feedforward neural network is used to perform nonlinear transformation on the fused features, and the multimodal weighted fusion data of the agent output at the enhanced feature abstraction level is converted into a multimodal representation tensor with unified dimension and high fusion degree. The format is a unified three-dimensional tensor (batch_size × feature dimension × time step). This data serves as the input for subsequent semantic fusion.

[0019] Step S13: performing semantic representation fusion processing on the agent multimodal weighted fusion data, thereby obtaining input fusion semantic representation data; In this embodiment of the present invention, based on the multimodal weighted fusion data output from step S12, a multi-layer stacked bidirectional long short-term memory (BiLSTM) network is used to perform temporal semantic fusion. This network consists of three layers of BiLSTM units, with 512 hidden units in each layer. Dropout layers with a ratio of 0.3 are used between layers to prevent overfitting. The multimodal fusion tensor is input to the first layer of BiLSTM units, capturing contextual dependencies within the time step. After bidirectional propagation, the feature vectors are merged to output a hidden state sequence. This sequence is then passed to subsequent BiLSTM layers for feature abstraction and semantic enhancement. Subsequently, an attention pooling mechanism is used to perform a weighted summation of the BiLSTM output sequence, focusing on extracting the time step information that contributes most to the overall semantic representation. This generates a fixed-length fused semantic representation vector with a fixed dimension of 1024. This fused semantic representation vector incorporates cross-modal temporal and spatial semantic information and is structured as a one-dimensional vector, meeting the input specifications of subsequent semantic search engines. The fusion process is implemented using the TensorFlow framework. All weights are initialized to a normal distribution and adjusted using the Adam optimization algorithm. At the end of this step, the input fusion semantic representation data is obtained, which serves as the semantic query basis for associated knowledge retrieval.

[0020] Step S14: performing associated knowledge retrieval processing based on the input fusion semantic representation data, thereby obtaining agent associated knowledge retrieval data.

[0021] In this embodiment of the present invention, the input fused semantic representation vector obtained in step S13 is input into the semantic association knowledge retrieval module. This module incorporates a high-dimensional vector index database built based on the vector search engine Faiss. The database contains semantic embedding vectors for entities and relationships in a pre-constructed structured knowledge graph. The retrieval process calculates the cosine similarity between the input fused semantic vector and all vectors in the knowledge base. A top-k (k=10) search is performed using a fast nearest neighbor search algorithm (such as IVF+PQ indexing) to return the most relevant entity and relationship candidate sets. Subsequently, a secondary screening based on semantic consistency is performed on the initial search results. Low-confidence matches are eliminated based on the match between the entity context description and the query semantics. The agent-associated knowledge retrieval data is constructed. The data structure includes fields such as candidate entity ID, entity name, match score, corresponding relationship path, and context summary text. This data is encapsulated in JSON format to meet the input requirements for the subsequent node multi-path structure construction, ensuring the accuracy of the semantic information and causal reasoning chain construction. After completing this step, the agent-associated knowledge retrieval data serves as the key input for the path verification method and is passed to the path structure construction module.

[0022] Preferably, step S14 includes the following steps: Step S141: extracting semantic query representation based on the input fusion semantic representation data to obtain input semantic query representation data; In this embodiment of the present invention, semantic query representation extraction is performed based on the input fused semantic representation vector obtained in step S13. A fixed-structure multi-layer perceptron (MLP) network is used to map the input fused semantic vector. This MLP network consists of three fully connected layers, with 1024, 512, and 256 nodes in each layer, respectively, and uses the Rectified Linear Unit (ReLU) activation function. After the first layer of mapping, the input fused semantic vector undergoes a ReLU nonlinear activation, followed by continuous mapping in the second and third layers, ultimately outputting a 256-dimensional semantic query representation vector. This mapping process aims to convert the fused semantic data into a query-specific low-dimensional representation that highlights the core semantic features and facilitates subsequent similarity calculation and keyword retrieval. To prevent overfitting, a batch normalization layer is added to the network, and L2 regularization is used to control the weight magnitude. After this step is completed, the generated input semantic query representation data is a 256-dimensional fixed-length vector for use in the next step.

[0023] Step S142: performing semantic similarity calculation based on the input semantic query representation data to obtain input semantic similarity data; In this embodiment of the present invention, semantic similarity is calculated for each knowledge vector in the knowledge base based on the input semantic query representation data obtained in step S141. The knowledge vectors in the knowledge base are 256-dimensional vectors of the same dimensionality as the input semantic query representation. The calculation process utilizes the cosine similarity algorithm. Specifically, the inner product operation is performed on the input semantic query representation vector and each vector in the knowledge base, and then divided by the product of the modulo lengths of the two vectors to calculate the cosine similarity value, which is in the range [-1, 1]. Parallel computing is accelerated using the high-performance matrix operation library BLAS (Basic Linear Algebra Subprograms), allowing the similarity of millions of knowledge vectors to be processed. After the calculation is complete, all knowledge vectors are sorted by similarity value, and the top N (e.g., N = 50) highly similar data are selected to form a list of input semantic similarity data. This list includes the knowledge vector IDs and their corresponding similarity scores, which serve as the basis for subsequent keyword searches and candidate screening. After completion of this step, the input semantic similarity data is output to support precise matching in subsequent semantic searches.

[0024] Step S143: performing keyword search processing based on the input semantic similarity data and the input semantic query representation data to obtain input semantic keyword search data; In this embodiment of the present invention, keyword search is performed using the high-similarity knowledge vectors obtained in step S142 and the semantic query representation in step S141. Keywords are extracted from the original text or vector corresponding to the input semantic query representation. Keyword extraction uses the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm to calculate word frequency weights, combined with the TextRank algorithm to establish a keyword priority ranking, and the top 20 keywords are selected as the search keyword set. Subsequently, keyword matching is performed on each text content of the high-similarity knowledge base entries filtered out in step S142. Keyword matching uses Boolean search logic, meaning that entries containing multiple keywords in the search text are given higher priority. An inverted index structure is used to accelerate the keyword matching process. Search results are ranked by a comprehensive score based on the number of matching keywords and keyword weights. The input semantic keyword search data obtained by filtering includes the matching knowledge item ID, the corresponding keyword match score, the keyword occurrence location, and contextual information. The data is formatted as a structured table. This data serves as input for candidate knowledge fragment screening and supports further semantic consistency assessment.

[0025] Step S144: determining input semantic candidate knowledge segment data based on the input semantic keyword retrieval data and the input semantic similarity data; In an embodiment of the present invention, based on the keyword matching results obtained in step S143 and the semantic similarity data in step S142, the comprehensive score of the candidate knowledge fragment is comprehensively calculated. The comprehensive score is calculated by weighted average, and the weight coefficients are assigned to the keyword matching score and the semantic similarity score respectively, with the specific weight ratios being 0.6 and 0.4. For each knowledge item, after calculating the comprehensive score, the knowledge fragments with the top 30% comprehensive scores are sorted from high to low according to the scores, and the knowledge fragments ranked in the top 30% are screened out as input semantic candidate knowledge fragment data. Each candidate knowledge fragment data contains the knowledge item ID, text content, comprehensive score, keyword matching details and corresponding semantic similarity. This screening ensures the balance of semantic relevance and keyword coverage of the knowledge fragment, and meets the requirements of subsequent context consistency evaluation. The data storage adopts a relational database structure to facilitate subsequent indexing and query operations. After completing this step, the input semantic candidate knowledge fragment data is output for the next step of semantic context consistency evaluation.

[0026] Step S145: performing context sentence consistency evaluation on the input semantic candidate knowledge fragment data, thereby obtaining semantic context sentence consistency data; In this embodiment of the present invention, contextual sentence consistency assessment is performed on the candidate knowledge segments selected in step S144. Using natural language processing technology based on syntactic dependency analysis, a dependency syntax analyzer (such as the SpaCy toolkit) is used to segment and parse the syntactic structure of each candidate knowledge segment, identifying core components such as subject, predicate, and object, as well as inter-sentence relationships. Next, 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 inter-sentence semantic continuity. For multi-segment text, a semantic adjacency matrix is ​​constructed, and semantically consistent intervals within the text are divided based on a similarity threshold. A graph clustering algorithm (such as spectral clustering) is used to cluster the semantic adjacency matrix, identifying the most consistent sentence clusters within the text. The sentence consistency score within each knowledge segment is quantified, generating semantic context sentence consistency data, including sentence cluster distribution, continuity score, and consistency label. The data is stored in a structured JSON format to facilitate subsequent screening. After completion of this step, the semantic context sentence consistency data is output, providing a basis for knowledge segment screening.

[0027] Step S146: performing knowledge fragment screening based on the semantic context sentence consistency data to obtain input semantic knowledge fragment data; In an embodiment of the present invention, the semantic context sentence consistency data obtained in step S145 is used to screen candidate knowledge fragments. The screening rule is based on the consistency score threshold setting, and the threshold is set to 0.75. Knowledge fragments with sentence consistency scores higher than the threshold are screened as input semantic knowledge fragments. During screening, the comprehensive score (step S144) and the consistency score are combined to perform a weighted composite score calculation, with a weighted ratio of 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 deep coherence and relevance of the content semantics. In the screening results, each knowledge fragment contains text content, score details, sentence cluster information and context labels. The screening operation adopts a batch processing method to execute SQL filtering and sorting in the relational database to ensure high efficiency and stability. After the screening is completed, the input semantic knowledge fragment data is generated as the input basis for the next step of knowledge traceability.

[0028] Step S147: performing knowledge traceability processing based on the input semantic knowledge fragment data to obtain knowledge traceability data; In an embodiment of the present invention, knowledge traceability is performed on the input semantic knowledge fragments after screening. The traceability tracing retrieves the source identifier of the knowledge fragment and relies on a pre-built knowledge graph node database, in which each knowledge node is associated with metadata such as its creator, timestamp, version number and reference relationship. Using the knowledge graph traversal algorithm, starting from the node of each knowledge fragment, recursively trace the upstream associated nodes, with a depth limited to 3 layers. All knowledge nodes and their metadata on each tracing path are recorded to form a traceability chain. By sorting the node attributes and time series of the traceability chain, the traceability path structure of the knowledge fragment is established. A graph database (such as Neo4j) is used to store the traceability chain data to achieve efficient path query and relationship visualization. During the traceability tracing process, the node data is checked for version consistency, and invalid or conflicting nodes are eliminated to output knowledge traceability data, the structure of which includes a list of traceability path nodes, node attributes and version information. This data provides trust and context support for the next step of associated knowledge retrieval.

[0029] Step S148: Perform associated knowledge retrieval processing based on the knowledge traceability data and the input semantic knowledge fragment data, thereby obtaining intelligent agent semantic associated knowledge retrieval data.

[0030] In an embodiment of the present invention, associated knowledge retrieval is performed based on the knowledge traceability tracking data of step S147 and the input semantic knowledge fragment data of step S146. This process combines the semantic knowledge fragment with its traceability chain to construct a complete knowledge context association network. Utilizing knowledge graph relational reasoning technology, path search algorithms (such as the shortest path algorithm and graph convolutional network) are applied to analyze the causal and semantic connections between knowledge nodes and determine the strength of the association between knowledge. Subsequently, all associated paths are integrated and a weighted comprehensive score is calculated based on the node's traceability trust, semantic relevance, and path coherence. Sorted according to the comprehensive score, the agent semantic associated knowledge retrieval data is output. The data format includes the knowledge fragment ID, text content, traceability chain information, associated path strength, and context summary. The data is stored in JSON format to support subsequent node multi-path structure construction and causal reasoning. After this step is completed, the agent semantic associated knowledge retrieval data forms the basis of a complete knowledge association graph, ensuring the accuracy and interpretability of the retrieval path verification.

[0031] Preferably, in step S2, constructing node multi-path initial structure data based on agent semantic association knowledge retrieval data includes: Extract semantically pointing feature data based on the agent's semantically related knowledge retrieval data; In an embodiment of the present invention, the intelligent agent semantic association knowledge retrieval data output in step S148 is used as the basis, and the data includes information such as knowledge fragment text, traceability chain information, and association path strength. For this data, natural language processing (NLP) technology is used to extract semantic features of the knowledge fragment text. Specifically, word embedding technology (such as Word2Vec or GloVe) is used to convert the text into a vector representation, and 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 to form directional semantic edges. By constructing a semantic pointing graph, each entity and the pointing relationship between them is represented, and semantic pointing feature data is output. The data structure is stored in the form of a graph, with nodes representing entities or concepts, edges representing semantic pointing relationships, and edges containing weight values ​​representing the strength of semantic associations.

[0032] Perform semantic node abstraction processing based on semantic pointing feature data to obtain path candidate node data; In an embodiment of the present invention, the above-mentioned semantic pointing feature data is used to perform node abstraction processing to simplify and aggregate semantic information. This step uses a clustering algorithm (such as hierarchical clustering or density clustering DBSCAN) to aggregate the node feature vectors in the semantic pointing graph and merge nodes with similar semantics or functions into abstract nodes. During the abstraction process, clustering judgment is made based on the similarity of the semantic embedding vectors of the nodes and the connectivity between the 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 is convenient for reducing redundancy and improving computational efficiency during subsequent path construction. The node abstraction operation uses a graph theory toolkit (such as NetworkX) to implement the merging and attribute updating of graph nodes.

[0033] Perform semantic connection relationship recognition processing based on the path candidate node data to obtain node semantic connection relationship data; In an embodiment of the present invention, based on the path candidate node data, the potential semantic connection relationship between nodes is identified, and by calculating the semantic similarity matrix between abstract nodes, the cosine similarity or Euclidean distance is used to measure the similarity of the node semantic feature vectors. Then, the rule engine is used in combination with predefined semantic connection rules (such as causality, temporal sequence, functional dependency) to analyze the connection intention between nodes and confirm the valid connection edges. To ensure the validity of the connection relationship, the semantic path mining algorithm is used to extract the multi-hop association path between nodes, and the node semantic connection relationship data is generated by verification combined with the path context consistency, which describes the connection direction and connection strength weight between nodes in the form of directed edges. The data structure is stored in an adjacency table, which contains the starting node ID, the ending node ID, the connection weight and the connection type label.

[0034] Perform path constraint screening on node semantic connection relationship data to obtain path effective connection structure data; In an embodiment of the present invention, the node semantic connection relationship data contains a large number of connection candidates. In order to ensure the rationality and effectiveness of the path, path constraint screening is performed. This step is based on set constraints, such as the upper limit of the path length (such as no more than 6 hops), the semantic connection strength threshold (eliminating edges with a weight lower than 0.3), the node type compatibility rule (ensuring that the connection node type conforms to the business semantic logic), etc. The screening method uses graph filtering technology to traverse the connection relationship one by one, eliminate the connection edges that do not meet the constraints, and automatically maintain the connectivity of the graph. After the screening is completed, the effective connection structure data of the path is obtained, which includes the node connection subset that meets the constraints and the corresponding structural topological relationship. The data is saved in the form of an adjacency matrix of the graph, which is convenient for subsequent topological network construction and weight initialization.

[0035] Constructing a multi-path topology network based on the effective path connection structure data, thereby obtaining multi-path node topology structure data; In an embodiment of the present invention, a multi-path topology network is constructed using path effective connection structure data. The network uses a graph structure to represent the relationship between nodes and connections, supports the coexistence of multiple paths, and reflects semantic diversity and causal multiplicity. The construction process includes initializing the graph database nodes, annotating the nodes with abstract node attributes, inserting valid connection edges in sequence, and maintaining the weights and directions of the edges. To support efficient queries and subsequent operations, the topology network uses dual storage of adjacency lists and edge weight matrices. The connectivity and integrity of the topology structure are verified through graph traversal algorithms (such as depth-first search and breadth-first search) to ensure that there are no isolated nodes and broken paths. The multi-path node topology structure data is output, including node sets, edge sets and corresponding weights, forming the basis of a complete semantic topology network.

[0036] Performing weight initialization processing on multi-path node topology structure data to obtain initial weight data of node topology structure; In an embodiment of the present invention, for a multi-path node topology network, the weight parameters of the nodes and edges are initialized. The node weight initialization is calculated by a graph analysis algorithm based on the centrality indicators of the nodes in the semantic directional features, such as degree centrality, betweenness centrality, etc. The edge weight initialization uses the previously screened semantic connection strength data, performs normalization processing, and uniformly maps it to the range of 0~1. In order to enhance the weight expression ability, a weighted normalization strategy is introduced to ensure that the sum of the node weights is 1, and the edge weights reflect 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 for supporting subsequent path reasoning and weight updates.

[0037] Node multi-path initial structure data is constructed based on node topology structure initial weight data and path effective connection structure data.

[0038] In this embodiment of the present invention, the initial weight data of the node topology structure and the effective path connection structure data are combined to construct complete node multi-path initial structure data. This structure data is based on a graph data model and integrates node attributes (such as semantic abstraction information and weight values) with edge attributes (such as connection type and connection weight) to form a multi-dimensional semantic path structure. Graph database tools (such as Neo4j) are used for data import and management, supporting multi-path parallel storage and efficient access. The structure data includes node sets, edge sets, weight matrices, and path index information, laying the foundation for subsequent path causal reasoning and semantic interpretation. After construction is complete, the initial node multi-path structure data is output as the result of step S2 for use in step S3.

[0039] Preferably, determining the path node causal reasoning chain data according to the node multi-path initial structure data in step S2 includes: Parsing the node control output field according to the node multi-path initial structure data to obtain the node control output field data; In an embodiment of the present invention, the node multi-path initial structure data is used as input, and the data includes a node set, node attributes, and the connection relationship between nodes. For each node, a field parsing operation is performed, focusing on identifying the control output field of the node. The control output field is defined as a key information field that affects the node on subsequent nodes or elements in the path, including but not limited to semantic labels such as operation instructions, trigger event identifiers, action status, and signal transmission. During parsing, structured data parsing technology is used to traverse the node attribute table and extract fields related to "control output". Text analysis is combined with rule matching to semantically classify field values ​​and mark field types and scopes. The parsing result forms node control output field data, and the data structure includes node ID, control output field name, field content, and corresponding semantic labels for subsequent causal reasoning.

[0040] Parsing the node response input field according to the node multi-path initial structure data to obtain the node response input field data; In an embodiment of the present invention, response input field parsing is performed for each node based on the multi-path initial structure data of the same node. The response input field indicates the control information or trigger signal received by the node from the outside or the predecessor node, including status feedback, signal acceptance identification, condition satisfaction flag, etc. The same structured field parsing technology is used to scan the node attributes and the edge attributes connecting it with adjacent nodes, and extract all fields representing the "response input" function. The semantic rule library is applied to classify and label the field content to ensure accurate identification of the input signal type and its triggering conditions. The parsing results are organized into node response input field data, including node ID, response input field name, field content and semantic label, forming basic data that can be used to control the construction of dependency relationships.

[0041] determining path node control dependency data based on node response input field data and node control output field data; In an embodiment of the present invention, the control dependency between nodes is analyzed in combination with the node control output field data and the node response input field data. The control dependency is defined as the direct or indirect impact of the control output field of node A on the response input field of node B. A 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 label and field content. The Boolean logic matching rule is used to determine the causal relationship between the fields. For example, the output field "start signal" corresponds to the response input field "start trigger", which determines the control dependency. The result forms the path node control dependency data, and the data structure includes the control node ID, the controlled node ID, the corresponding control output field and the response input field pair, and the dependency strength index. This data provides a basis for the subsequent path behavior law analysis.

[0042] Determine the node path behavior evolution law based on the path node control dependency data; In an embodiment of the present invention, an analysis of the evolution law of node path behavior is carried out based on the control dependency data of the path nodes. The evolution law refers to the dynamic change characteristics of the node control behavior over time or event triggering and its dependency chain. A node behavior state transition model is constructed by using a time series analysis method in combination with the event triggering timing of the control dependency. Specifically, a state machine or directed graph dynamic analysis technology is used to perform a time-series mapping of the output state changes of the control node and its impact on the input of the response node. By tracking the behavior trajectory, the evolution path and evolution frequency of the node behavior are determined. This step outputs the node path behavior evolution law data, which specifically describes the state transition sequence, trigger timestamp and impact pattern of each node in the path, and provides a time-series basis for causal reasoning between nodes.

[0043] Analyze the inter-node trigger dependency relationship based on the node path behavior evolution law to obtain the inter-node dependency trigger data; In an embodiment of the present invention, the evolution law of node path behavior is used to analyze the trigger dependency relationship between nodes. The trigger dependency relationship is defined as the state change of one node triggering the response action of another node. The trigger event capture technology is used, combined with the node behavior state transfer data, to identify the trigger conditions and response mechanisms between nodes. Through the event sequence association algorithm, the sequence of node behavior and the trigger logic are analyzed to eliminate the interference of accidental events and extract stable dependency trigger patterns. The node dependency trigger data is constructed, including the trigger node ID, the response node ID, the trigger condition description, the trigger time interval and the trigger intensity measurement, to support the precise construction of the causal chain.

[0044] Identify the causal basic unit data of path nodes based on the trigger data of the dependency relationship between nodes; In an embodiment of the present invention, the basic units constituting path causal reasoning are identified based on the trigger data of the dependency relationship between nodes. A causal basic unit is defined as a pair of node triggers with a clear causal relationship, including a cause node, an effect node, and their triggering mechanism. By combining the trigger data and applying a graph mining algorithm, all direct causal node pairs and multi-level indirect causal chains are identified. The causal basic units that conform to the theoretical timing and semantic logic are screened out, and invalid or abnormal trigger pairs are eliminated. The path node causal basic unit data is generated, and the structure includes the causal unit ID, cause node information, effect node information, causal relationship description, and trigger parameters, forming the minimum element for constructing a complete causal chain.

[0045] The path node causal reasoning chain data is determined based on the path node causal basic unit data.

[0046] In an embodiment of the present invention, the path node causal basic unit data is used to construct a path node causal reasoning chain in the order of causal relationships. The chain construction adopts a directed acyclic graph structure, and the units are connected in sequence to form a complete chain based on the causal connection relationship of the basic units. The node order is determined by a topological sorting algorithm to ensure the logical continuity and temporal consistency of the reasoning chain. This process integrates the trigger timestamps and dependency strengths between nodes, and assigns weights and explanations to the chain nodes. The path node causal reasoning chain data is output, including the chain ID, node sequence, causal relationship description, weight distribution and temporal annotation, providing direct data support for the subsequent generation of structured interpretable path data.

[0047] Preferably, determining the structured interpretable path data based on the path node causal reasoning chain data and the path node semantic interpretation information in step S2 includes: Perform chain sequence reconstruction processing based on the path node causal reasoning chain data to obtain path node chain sequence reconstruction data; In this embodiment of the present invention, the input data is path node causal reasoning chain data, including a set of node IDs, causal relationships between nodes, and timestamps or event sequence information of node triggering. Chain sequence reconstruction uses a directed acyclic graph (DAG) topological sorting method to perform 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 node outgoing and incoming edge information; then, by traversing all nodes, nodes without predecessor nodes are added to the sequence as starting points; then, outgoing edges of nodes already added to the sequence are recursively deleted, the adjacency table status is updated, and nodes without predecessor nodes are extracted again. The result of the topological sorting is the linear order of the nodes, reflecting the execution or triggering order of the nodes in the causal chain. This step outputs the path node chain sequence reconstruction data, which includes a node sequence list and the order constraints between nodes, providing a basis for subsequent time series positioning.

[0048] Determine the path node time sequence positioning data based on the path node chain sequence reconstruction data; In the embodiment of the present invention, the path node chain sequence reconstruction data provides the execution order of the node. The timing positioning data generation step accurately calculates the position of each node on the time axis by introducing timestamp inference and event synchronization technology, combined with the event occurrence time or trigger interval data of the node. The specific operations include: for each node in the chain sequence, based on the known node trigger time data or the estimated time interval, the start time and end time of the node are accumulated to form a time period label; for nodes with missing timestamps, the time of adjacent nodes is used for interpolation processing to ensure time continuity. This step outputs the path node timing positioning data, including the node ID and the corresponding time interval, to clarify the temporal position of the node in the causal chain and support subsequent semantic field alignment processing.

[0049] Perform field alignment processing on the path node temporal positioning data and the path node semantic interpretation information data to obtain semantic field alignment data; In an embodiment of the present invention, the path node timing positioning data and the path node semantic interpretation information data are combined to complete the field-level alignment and matching. The input includes the timing information of the node and the semantic interpretation text or structured semantic label of each node. According to the correspondence between the node ID, the timing positioning data and the semantic interpretation information are mapped at the node level. Subsequently, for the field structure inside the node, the field name and the semantic label are used to match, and the timing field and the semantic field are matched one-to-one according to the principle of logical association to form a field-level alignment table. The alignment process uses strict field name matching and semantic similarity calculation technology to eliminate irrelevant or repeated fields to ensure accurate correspondence of the data. The output semantic field alignment data structure contains the node ID, time field, semantic field and their correspondence, which serves as the basic data for subsequent semantic structure annotation processing.

[0050] Perform path node semantic structure annotation processing on the semantic field alignment data to obtain path node semantic annotation structure data; In an embodiment of the present invention, semantic structure annotation is performed based on semantic field alignment data to clarify the semantic role and hierarchical relationship of each field in the path node. The annotation process includes defining an annotation label system, such as entity labels, action labels, conditional labels, etc., and performing semantic classification and hierarchical assignment on each field. A rule engine is used in combination with a semantic analysis method to deeply analyze the content of the aligned fields, identify the logical dependencies and combination patterns between fields, and form a hierarchical semantic structure. Through annotation, not only the semantic meaning of the field is clarified, but also the contextual association between fields is revealed to ensure the integrity and consistency of the semantic structure. The output path node semantic annotation structure data includes node ID, field name, annotation label and semantic hierarchical structure description to support subsequent continuity analysis.

[0051] Perform semantic continuity analysis based on the semantic annotation structure data of the path nodes to obtain path semantic continuity data; In an embodiment of the present invention, semantic continuity analysis targets the semantic hierarchy and field sequence in the annotation structure to detect the coherence and logical continuity in the semantic chain. The analysis process uses sequence analysis technology and semantic network diagrams to check whether the logical relationship between the semantic annotations of adjacent nodes is coherent, and to identify anomalies such as semantic jumps, breaks, or semantic ambiguity. Specific operations include calculating the semantic similarity of adjacent node annotation fields, evaluating the causal and temporal connections between fields, and marking semantic continuity areas and breakpoints. The analysis outputs path semantic continuity data, and the data structure includes node pairs, continuity scores, anomaly identifiers, and cause descriptions to assist in subsequent trusted segment identification.

[0052] Based on the path semantic continuity evaluation data, semantically credible segments are identified to obtain path semantically credible segment structure data; In an embodiment of the present invention, semantically credible segment identification is based on semantic continuity scores, dividing the path into segments with stable and consistent semantic expressions. Cluster analysis and threshold judgment methods are used for identification, and node sequences with higher continuity scores are classified as the same credible segment, with the segmentation points serving as boundaries with lower credibility. This process analyzes the strength of the semantic relationship between nodes, combines temporal continuity and logical consistency, and generates multiple semantically credible segments, each with complete semantic expression and no obvious breaks. The path semantically credible segment structure data is output, including the start and end nodes of the credible segment, segment length, credibility index, and internal structure description, providing semantic support for the generation of path expression templates.

[0053] Merge the path semantic trusted segment structure data with the path node semantic annotation structure data to generate structured path expression template data; In an embodiment of the present invention, semantically trusted segment structure data is fused with path node semantic annotation structure data to form a comprehensive structured path expression template. This fusion process nests and combines the node annotation structures within each semantically trusted segment based on the node ID and semantic annotation level, maintaining the semantic hierarchy and temporal order. A data structure mapping method is employed to ensure seamless integration of segment boundaries and node semantic annotations. The resulting structured path expression template contains a multi-level semantic structure, temporal information, and segment division labels, fully expressing the semantic logic and temporal relationships within the path. This data provides the foundation for the subsequent unified output of the path structure.

[0054] Based on the structured path expression template data, the path structure is uniformly output and processed to generate structured interpretable path data.

[0055] In this embodiment of the present invention, unified formatting and output of the path structure is performed based on structured path expression template data. This unified processing includes data format conversion, semantic label standardization, and path structure normalization. A unified data model is used to define path nodes, edges, semantic attributes, and time information, ensuring that the output structure conforms to a predefined standard interface format (such as JSON, XML, etc.). Data integrity verification is also performed to ensure consistency in path node causal relationships, semantic annotations, and chronological order. The resulting structured, interpretable path data contains node sequences, causal chains, semantic annotations, and trusted segment information, which can be used by the agent in subsequent retrieval steps such as path updates and deviation identification.

[0056] Preferably, step S3 includes the following steps: Step S31: detecting interpretable path deviation identification data based on the structured interpretable path data; In an embodiment of the present invention, structured interpretable path data is used as input, and path deviation detection is carried out for multi-dimensional information such as the causal order, semantic annotation, and time sequence positioning of the path. Based on the chain order of the path nodes and the time positioning data, a sequence comparison algorithm is used to compare the actual path execution sequence with the expected path sequence to identify deviations such as abnormal node order, missing or inserted nodes. Secondly, the semantic annotation structure is used to perform node semantic consistency verification. By comparing the semantic labels of the path nodes with the reference standard, anomalies such as semantic label mismatch and semantic logic breakage are detected. Thirdly, the semantic continuity of the path and the trusted segment data are combined to identify the poor semantic continuity and logical breakpoint locations in the path. Based on the above detection results, a path deviation identification report is formed, which is specifically composed of fields such as deviation node set, deviation type, deviation location and deviation severity, and generates interpretable path deviation identification data for subsequent deviation cause analysis.

[0057] Step S32: performing deviation cause analysis based on the interpretable path deviation identification data to obtain interpretable path deviation cause feature data; In an embodiment of the present invention, a cause analysis is carried out on the deviation nodes and anomaly types based on the obtained interpretable path deviation identification data. The analysis process includes: for node sequence anomalies, a dependency backtracking method is used to retrieve the upstream causal nodes of the deviation node and their triggering conditions, and locate the specific reasons for the causal chain break or error triggering; for semantic mismatches, the semantic hierarchy structure is used to compare the contextual semantic environment of the deviation node to identify the specific fields with labeling errors or semantic interpretation deviations; for semantic continuity breaks, the time positioning data is combined to analyze the existing technical factors such as data collection delays, timestamp errors or event omissions. The cause analysis constructs a deviation cause feature vector, which includes causal chain anomaly flags, semantic consistency indicators, time synchronization anomaly parameters, etc., to form structured interpretable path deviation cause feature data, providing a basis for the formulation of correction strategies.

[0058] Step S33: generating an explainable path deviation correction strategy based on the explainable path deviation cause feature data to obtain an explainable path deviation correction strategy; In an embodiment of the present invention, the deviation cause characteristic data outputted in step S32 is used to formulate a targeted correction strategy. The correction strategy is generated based on the deviation type and cause characteristics, and a rule-driven method is combined with structured data analysis to form a sequence adjustment rule, a semantic label correction rule, and a time synchronization correction rule. The sequence adjustment rule defines the operation instructions of node insertion, node reordering, and node deletion for the break of the causal chain; the semantic label correction rule specifies label replacement, field supplementation, and semantic re-labeling schemes based on semantic hierarchy and context consistency; the time synchronization correction rule uses timestamp correction, interpolation completion, and time window adjustment strategies to solve the time asynchrony problem. The generated interpretable path deviation correction strategy contains multiple rule instruction sets, clearly defines the correction object, correction method, and execution order, and outputs structured deviation correction strategy data for path structure simulation correction.

[0059] 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.

[0060] In an embodiment of the present invention, the deviation correction strategy generated in step S33 is used to perform structural simulation correction on the deviation identification data of step S31. The correction process executes the rule instructions in the correction strategy in sequence, and adjusts the path structure for the deviation nodes, including node order re-arrangement, missing node supplementation, abnormal node removal and semantic label correction. The specific operation first reads the correction strategy instructions, and modifies the path node chain data and semantic annotation structure in sequence to ensure the restoration of the consistency of causal order and semantic logic. Timestamp synchronization achieves the continuity of node timing through interpolation and time window adjustment. During the correction process, the semantic continuity index and trusted segment division of the path 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 node adjustment records, semantic label change logs and time synchronization correction information is output as input for subsequent path strategy optimization and update.

[0061] Preferably, step S31 includes the following steps: Step S311: Analyze node semantic jumps based on the structured interpretable path data to obtain path node semantic jump data; In an embodiment of the present invention, a semantic jump analysis is performed on the semantic sequence of path nodes based on structured interpretable path data. The specific operations include: extracting the semantic annotation structure and temporal positioning data of the path nodes from the structured path expression template data, and constructing a semantic sequence of path nodes. The semantic similarity measurement method is used to perform continuity testing on the semantic labels of adjacent nodes, and the semantic jump points exceeding a predetermined threshold are identified by calculating the vector distance or correlation coefficient of each pair of semantic labels of adjacent nodes. The semantic jump point is defined as a node pair with excessive semantic differences between adjacent nodes, resulting in a break in semantic continuity. During the analysis process, a dynamic window sliding strategy is applied to gradually scan the entire node sequence, and the jump frequency and jump strength are counted to output the semantic jump data of the path nodes, which includes the jump node pair index, jump strength value and jump position timing information, providing basic data for subsequent path logic consistency comparison.

[0062] Step S312: performing a comparison process based on the path node semantic interpretation information and the path node semantic jump data to obtain path structure logic consistency deviation data; In an embodiment of the present invention, the logical consistency of the path structure is evaluated by comparing the semantic interpretation information of the path nodes with the semantic jump data of the path nodes obtained in step S311. The specific operation is: extracting the semantic labels and their contextual semantic relationships in the semantic interpretation information of the path nodes, and expressing the semantic dependency and contextual connection strength between nodes by constructing a semantic association matrix. In combination with the jump node pairs identified in the semantic jump data, for each jump point, the weak connectivity index in the semantic association matrix is ​​calculated to determine whether the jump causes a logical break in the semantics of the path. For logical break nodes, they are marked as structural logical consistency deviation points, and the deviation types (such as semantic incoherence, contextual missing, semantic contradiction, etc.) are statistically analyzed to form path structure logical consistency deviation data. The data structure includes fields such as deviation node position, deviation category, contextual semantic break strength, etc., which serve as input for path semantic interruption clustering and subsequent analysis.

[0063] Step S313: performing path semantic interruption location clustering processing based on the path structure logic consistency deviation data to obtain path semantic interruption clustering data; In an embodiment of the present invention, clustering processing is implemented for multiple deviation points in the path structure logical consistency deviation data to identify the concentrated segments of semantic interruption. The clustering method adopts a density clustering algorithm based on temporal distance and semantic deviation strength. The specific operation process includes: according to the temporal position of the deviation node, the time interval distance matrix between the deviation nodes is calculated; secondly, combined with the deviation strength index, a weighted distance matrix is ​​constructed to enhance the clustering sensitivity to serious deviation nodes. Subsequently, a density clustering algorithm such as DBSCAN or OPTICS is used to identify node clusters with similar temporal sequences and similar deviation strengths, and determine them as semantic interruption segments. The clustering results are output in the form of semantic interruption clustering data, including the position of each cluster center, the set of deviation nodes contained in the cluster, and the length of the cluster segment, providing a basis for path interpretation strength attenuation and structural offset analysis.

[0064] Step S314: performing a path interpretation strength attenuation evaluation based on the path structure logic consistency deviation data to obtain path interpretation strength attenuation data; In an embodiment of the present invention, the attenuation degree of the path interpretation strength is evaluated for semantic interruptions and structural logic deviations in the path. During specific implementation, a path interpretation strength index is defined, and the semantic consistency, contextual coherence and node importance weights of the path nodes are comprehensively considered. The deviation nodes and their deviation strengths in the path structural logic consistency deviation data are used to calculate the local attenuation coefficient of the overall path interpretation strength. The specific method includes: in the path node sequence, for each deviation point, the interpretation strength value of the position and its adjacent nodes is proportionally reduced according to its deviation strength. The overall path interpretation strength is obtained by weighted accumulation of the node interpretation strength, and the attenuation data is expressed as a dynamic change curve of the interpretation strength of each node in the path. The evaluation result is saved in the form of path interpretation strength attenuation data, which includes the node number, interpretation strength value and strength change trend, and is used for subsequent semantic structure offset determination.

[0065] Step S315: determining the path semantic structure offset based on the path interpretation strength attenuation data to obtain path semantic structure offset data; In an embodiment of the present invention, the degree of offset of the path semantic structure is determined based on the path interpretation strength attenuation data obtained in step S314. The specific operations include: performing trend analysis on the interpretation strength change trend of the path nodes, and using sliding average filtering and inflection point detection technology to identify sections with obvious strength attenuation. Within the identified section, the offset position and offset amplitude of the path semantic structure compared to the standard path are determined in combination with the path node temporal positioning and semantic annotation information. The offset amplitude is calculated by measuring the degree of change of the node semantic labels and the degree of node sequence variation in the section to generate the path semantic structure offset data. The data includes the offset start and end nodes, the offset amplitude value, the affected node list and the offset type, which are used for subsequent deviation weight normalization processing.

[0066] Step S316: performing deviation weight normalization processing based on the path semantic structure offset data to obtain path semantic deviation comprehensive data; In an embodiment of the present invention, a normalization process of the deviation weight is implemented for the deviation amplitude and the affected nodes recorded in the path semantic structure deviation data. The specific method is as follows: first, the deviation weight is mapped to the interval [0,1] in a linear proportion according to the deviation amplitude to ensure the uniformity of the weight scale; secondly, the normalized weight is weighted averaged based on the number of nodes in the offset segment and the node importance index (such as the criticality and connectivity of the node in the path) to form a comprehensive deviation weight value. This normalization process is implemented using numerical calculation tools to ensure the uniformity and comparability of the deviation weight. The path semantic deviation comprehensive data is output, which includes the normalized weight of each deviation segment, a detailed list of affected nodes, and weight distribution information, providing a quantitative basis for deviation identification.

[0067] Step S317: Detect interpretable path deviation identification data based on the path semantic deviation comprehensive data and the path semantic structure offset data.

[0068] In an embodiment of the present invention, the path semantic deviation comprehensive data and the path semantic structural offset data are integrated to perform comprehensive deviation detection on the path. The specific operations include: combining the normalized deviation weight with the structural offset amplitude, setting multiple deviation thresholds, and screening out the path segments with significant deviations step by step; secondly, for the screened deviation segments, integrating the deviation weight, deviation node information and structural offset characteristics, and constructing a structured deviation identification record. The record fields include the deviation segment number, deviation severity level, affected node sequence and deviation type description. The deviation identification data is saved through a database or data structure as the basic data for subsequent deviation cause analysis and path correction, thereby realizing comprehensive quantitative identification of path deviations.

[0069] It is particularly important that step S33 includes the following steps: Step S331: performing path deviation type classification processing based on the explainable path deviation cause feature data to obtain path deviation type classification data; In an embodiment of the present invention, a normalization operation is performed on the field structure in the characteristic data of the cause of explainable path deviation, and all path deviation data are composed of a standardized feature vector matrix according to the contents such as "deviation node number", "semantic jump position", "structural break position", and "explanatory redundant field". Then, a fixed rule-based pattern matching method is used for classification and judgment, which includes the following rules: if the semantic correlation between the front and back of the deviation node is less than 0.2, and the number of jump nodes in the middle of the path is greater than 2, it is classified as a "semantic jump type deviation"; if there is no continuous dependency path between the output node and the target node in the causal reasoning chain, it is classified as a "causal chain break type deviation"; if there are repeated explanation segments in the path or the entropy value of the explanation field information is too low, it is determined to be a "semantic redundancy type deviation". Each of the above deviation types is marked with an independent label, and finally the path deviation type classification data is obtained.

[0070] Step S332: performing deviation triggering scenario rule extraction processing based on the path deviation type classification data to obtain deviation triggering scenario rule data; In an embodiment of the present invention, after obtaining the path deviation type classification data, the deviation trigger scenario rule extraction process is performed. This step takes each deviation record in the classification data as input, and sequentially performs context rule screening on its path structure context information, node timestamp distribution, node source context and target context matching parameters. The rule extraction is called based on the rule set encoded by the trigger type. For example, when the deviation type is a "causal chain break type deviation", the rule template group ID-C04 is called to extract the cause of the failure to match the path interruption position with the response / control fields of its previous and subsequent nodes; for a "semantic jump type deviation", the rule template ID-J02 is called to extract the degree of difference in entity distribution before and after the jump of the semantically disjointed jump node and the coverage of upper and lower attributes. The context status 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 final output is the deviation trigger scenario rule data.

[0071] Step S333: performing path reconstruction mode extraction processing based on the deviation triggering scenario rule data to obtain path reconstruction mode configuration data; In an embodiment of the present invention, path reconstruction method extraction processing is carried out for the deviation trigger scenario rule data generated in step S332. In this processing flow, a mapping table of deviation types and executable reconstruction methods is first constructed. The mapping table presets a variety of path reconstruction means, including "path breakpoint logic bridging", "node semantic reconfiguration insertion", "interpretation window sliding expansion", "redundant node elimination and merging" and other structural adjustment means. Then, for each trigger rule data, according to the combination of its scenario classification label and path deviation type classification data, the executable reconstruction method is screened in the mapping table. The screening criteria include structural connectivity maintenance rules, causal relationship retention criteria, and interpretation information integrity constraint logic. The matched path reconstruction method is output through a standard configuration template to generate path reconstruction method configuration data.

[0072] Step S334: performing strategy adaptability range evaluation processing based on the path reconstruction mode configuration data to obtain path deviation correction strategy evaluation data; In an embodiment of the present invention, after obtaining the path reconstruction method configuration data, a strategy adaptability range evaluation process is performed. This process takes each path reconstruction method defined in the configuration data as input, and evaluates its adaptability under different structural deviation conditions in turn. The evaluation indicators include: Structure-Coherence Score, Semantic Coverage Ratio, Explanation Strength Recovery Rate and Causal Consistency Index. The calculation of each indicator depends on the semantic annotation structure data of the path node and the causal chain information before and after the original deviation node, and a fixed weight weighting method is used for comprehensive evaluation. In the evaluation process, if a reconstruction method is lower than the preset threshold in more than three indicators (such as semantic coverage is less than 0.6), it is marked as "unsuitable", and the rest enter the strategy encapsulation process. The final path deviation correction strategy evaluation data is obtained.

[0073] Step S335: Generate an explainable path deviation correction strategy based on the policy boundary encapsulation processing.

[0074] In an embodiment of the present invention, after completing the strategy adaptability range evaluation, the strategy boundary encapsulation processing is performed on the path reconstruction method that meets the adaptation conditions. This processing step first reads the entry marked as "adaptive" in the strategy evaluation data, extracts its evaluation index range, logical constraints, action node range and execution order parameters, and constructs a complete strategy template. Then, for each strategy, a control parameter structure is configured, including control items such as the path modification step, the start and end points of the interpretation window, the causal verification threshold value and the maximum number of node replacements. All parameters are encapsulated in the form of a data table. After completing the parameter filling, the strategy is encapsulated into a data structure in JSON format. The fields include: strategy ID, action path ID, strategy control parameters, adaptability label, binding deviation type and trigger rule number, and the final output is an interpretable path deviation correction strategy.

[0075] It is particularly important that step S34 includes the following steps: Step S341: performing path deviation node clustering processing based on the interpretable path deviation identification data to obtain path deviation node clustering data; In an embodiment of the present invention, after completing the path deviation identification and obtaining the interpretable path deviation identification data, the path deviation node clustering process is performed. First, the information of each deviation node recorded in the path deviation identification data is structured and organized, including the node number, path position index, semantic category label, deviation type code, interpretation strength attenuation value, node semantic weight, context causal interruption level and other fields of each deviation node. Based on the above feature data, a set of deviation node feature vectors is constructed, and a K-Medoids clustering algorithm with parallel constraints of Euclidean distance and semantic classification labels is used for clustering analysis. Random initialization is not introduced in the clustering process, but the position point with the largest interpretation strength attenuation value is used as the initial center to improve the semantic significance of the cluster center. After the clustering algorithm is completed, the center node index of each cluster cluster, the list of member node numbers in the cluster, the cluster label and the semantic offset mean index of the cluster are output to form path deviation node clustering data as the index basis for local structure replacement.

[0076] Step S342: performing path local structure replacement processing based on the path deviation node clustering data to obtain path structure local correction data; In an embodiment of the present invention, a path local structure replacement process is performed based on path deviation node clustering data. First, according to the central node number of each cluster, the node and its two preceding and following nodes are extracted from the structured interpretable path data to form a local path segment, forming a local path window. Then, based on the path deviation type classification data and the "recommended replacement mode" provided in the correction strategy, structural fragments with semantic labels corresponding to the path segment and behavioral pattern similarity greater than 0.85 are retrieved from the semantic structure library to construct a set of candidate structural replacement fragments. Subsequently, a structural replacement matching evaluation is performed, and scores are given according to three indicators: causal reasoning chain continuity, semantic compatibility of upstream and downstream nodes, and original path length retention rate. The fragment with the highest score is selected to replace the original node structure in the local path window. After the structural replacement is completed, the node list before and after the modification, the structural change range, and the replacement strategy number are recorded for each path, and the local correction data of the path structure is output.

[0077] Step S343: performing structure splicing verification processing based on the local correction data of the path structure to obtain path structure splicing consistency verification data; In an embodiment of the present invention, a structural splicing consistency check is performed based on the local correction data of the path structure. First, the head and tail connection node numbers of the replaced paragraphs in each path are read, and the output field of the node before the connection and the input field of the node after the connection are extracted based on the causal reasoning chain data, and field matching mapping relationships are established respectively. Then, a field semantic alignment evaluation operation is performed. This operation uses the Jaccard semantic similarity algorithm to calculate the ratio of the intersection and union of the semantic word bags of the front and back node fields. If the similarity is lower than 0.5, it is marked as a splicing anomaly, and the original correction segment identifier is fed back. At the same time, it is checked whether the path structure forms a new closed loop or jump, and the structural graph traversal algorithm is used to detect whether an isolated subgraph or an undirected closed path appears in the path. If the above situation occurs, it is also marked as a splicing conflict. Finally, the path structure splicing consistency verification data is formed, including the correspondence between the fields before and after the splicing of each path, the semantic alignment score, the structural connectivity state identifier and the consistency evaluation result.

[0078] Step S344: performing semantic relationship reconstruction processing based on the path structure splicing consistency verification data to obtain path semantic relationship reconstruction data; In an embodiment of the present invention, after completing the structural splicing consistency check, a semantic relationship reconstruction process is performed based on the splicing consistency check data. The processing flow first locates the contextual semantic category and semantic function type field of the two nodes before and after each path splicing segment, and analyzes whether the splicing operation introduces a semantic span change or a sudden change in the interpretation direction. For connection points that are judged to be "semantic transition interruptions", a transition interpretation node is inserted. The inserted node is retrieved from the structured semantic template library, and the semantic overlap with the previous and next nodes is required to be greater than 0.7. After that, the semantic dependency relationship between the splicing segment and the adjacent nodes is re-established, and the "interpretation prefix mark bit" and "causal predecessor number" between the path nodes are updated to complete the reconstruction of the semantic relationship chain. The reconstructed path node sequence is reorganized into a semantic structure graph to form path semantic relationship reconstruction data. The data format includes: a reconstructed node directed edge table, a semantic label update log, an interpretation function conduction path, and a splicing operation record list.

[0079] Step S345: performing path structure simulation correction processing based on the path semantic relationship reconstruction data to obtain path structure simulation correction data.

[0080] In an embodiment of the present invention, after completing semantic relationship reconstruction, a path structure simulation and correction process 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 compared with the original structured interpretable path data to identify structurally evolved sections and unchanged sections. Subsequently, with the evolved sections as the center, the path simulation engine module is called to simulate the response of the modified path structure under typical input semantic query conditions. The simulation process adopts a three-step model of semantic query-interpretation expansion-path backtracking, and records indicators such as the interpretation path length, number of node activations, interpretation redundancy rate, and terminal arrival rate in each round of simulation. If the simulation results meet the semantic coherence and structural convergence conditions, the path is marked as "corrected successfully". Otherwise, the correction strategy rollback is continued. Finally, the path structure simulation correction data is generated, including the path number, the final corrected structure diagram, the semantic interpretation path node sequence, the simulation verification indicator table, and the structure replacement success flag, which is used for subsequent feedback extraction and strategy optimization phase calls.

[0081] Preferably, step S4 includes the following steps: Step S41: performing path simulation correction feedback extraction based on the path structure simulation correction data, thereby obtaining path simulation correction feedback data; In an embodiment of the present invention, a path simulation correction feedback extraction operation is implemented based on the path structure simulation correction data. The specific operation process includes: obtaining the path structure simulation correction data, which includes the corrected path node sequence, the adjustment of the connection relationship between nodes, and the corresponding semantic label change records. By comparing the original structured interpretable path data and the simulated corrected data, the path structure change indicators are calculated, such as the number of node additions and deletions, the frequency of connection edge changes, and the adjustment amplitude of the semantic labels. Furthermore, the node topology analysis tool is used to identify the key change points in the corrected path, including new semantic jumps, missing connections, and sequence adjustments. Based on the above analysis results, the path simulation correction feedback data is extracted. The data format includes the correction point coordinates, change type, change amplitude, and corresponding timestamp, which serves as the input basis for the rationality verification of the subsequent correction strategy, and realizes quantitative feedback on the correction effect.

[0082] Step S42: verifying the rationality of the correction strategy based on the feedback data of the path simulation correction to obtain rationality data of the path correction strategy; In an embodiment of the present invention, the rationality verification of the correction strategy is implemented for the path simulation correction feedback data. The specific operation process is: extract key correction indicators from the path simulation correction feedback data, such as the correction point distribution density, correction amplitude and correction frequency. Use a structural integrity detection algorithm to perform connectivity detection and circular dependency identification on the corrected path structure to determine whether the correction causes a break in the path structure or an abnormal loop. Combined with semantic coherence detection, the semantic consistency of the correction node is evaluated to ensure the semantic fluency of the corrected node. Utilize a multi-index scoring mechanism to calculate the rationality score of the correction strategy based on the rationality of the distribution of the correction points, the integrity of the path structure and the semantic coherence. Output the scoring results in the form of path correction strategy rationality data. The data content includes the rationality score value, the abnormal correction area identifier and the relevant correction strategy parameters to support subsequent strategy optimization processing.

[0083] 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; In an embodiment of the present invention, an interpretable retrieval path strategy is optimized based on the path correction strategy rationality data. The specific operational process includes: receiving the rationality score and anomaly identification in the path correction strategy rationality data, and adjusting the path correction strategy parameters, such as node priority weights, path connection constraints, and semantic continuity thresholds, based on the scoring results. An optimization algorithm (such as a heuristic search or iterative optimization algorithm) is used to adjust the path search space and node connection rules to enhance the semantic coherence and structural integrity of the path. By constructing a multi-scheme comparison framework, the impact of different strategy parameter combinations on the path structure and semantic coherence is evaluated, and the optimal strategy parameter combination is selected to generate interpretable path retrieval strategy optimization data, which includes the optimized strategy parameter settings, adjustment range, and strategy effect evaluation indicators for subsequent path updates.

[0084] 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.

[0085] In an embodiment of the present invention, based on the interpretable path retrieval strategy optimization data, the agent interpretable retrieval path update processing is implemented. The specific operation is: using the optimized strategy parameters, the node weights, connection relationships and semantic labels in the structured interpretable path data are readjusted. Through the path update module, the node priority is adjusted according to the optimization strategy, the path nodes are added or reduced, and the connections between the nodes are rebuilt to ensure that the path reaches a new balance in semantic coherence and structural integrity. The path update process relies on the path topology construction tool to maintain the causal logic chain between nodes in real time and re-label the semantic structure of the path nodes. After the update is completed, the agent interpretable retrieval path update data is output. The data structure contains the updated path node list, connection topology relationship, semantic labeling results and update strategy parameters to ensure the traceability and interpretability of the path update and support the subsequent retrieval and verification operations of the agent.

[0086] The present invention also provides an agent-interpretable retrieval path generation system for executing the agent-interpretable retrieval path verification method described above. The agent-interpretable retrieval path generation system includes: The associated knowledge retrieval processing module is used to obtain multimodal input data of the intelligent agent; perform associated knowledge retrieval processing based on the multimodal input data of the intelligent agent to obtain intelligent agent associated knowledge retrieval data; The interpretable path determination module is used to construct node multi-path initial structure data based on the agent semantic association knowledge retrieval data; determine path node causal reasoning chain data based on the node multi-path initial structure data; extract path node semantic interpretation information based on 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 interpretation information; A path structure simulation and correction module is used to detect interpretable path deviation identification data based on structured interpretable path data; perform path structure simulation and correction processing on the interpretable path deviation identification data to obtain path structure simulation and correction data; The retrieval path update module is used to perform interpretable retrieval path strategy optimization processing on the path structure simulation correction data to obtain interpretable path retrieval strategy optimization data; and perform intelligent agent interpretable retrieval path update processing on the structured interpretable path data according to the interpretable path retrieval strategy optimization data to obtain intelligent agent interpretable retrieval path update data.

[0087] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A method for verifying an agent-interpretable retrieval path, characterized in that: The following steps are involved: Step S1: Acquire multimodal input data of the intelligent agent; perform associated knowledge retrieval processing based on the multimodal input data of the intelligent agent to obtain intelligent agent associated knowledge retrieval data; Step S2: constructing node multi-path initial structure data based on the agent semantic association knowledge retrieval data; determining path node causal reasoning chain data based on the node multi-path initial structure data; extracting path node semantic interpretation information based on the multi-path node initial path structure data; Determine structured interpretable path data based on path node causal reasoning chain data and path node semantic interpretation information; 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; Step S4: performing interpretable retrieval path strategy optimization processing based on the path structure simulation correction data to obtain interpretable path retrieval strategy optimization data; According to the interpretable path retrieval strategy optimization data, the structured interpretable path data is processed by the intelligent agent interpretable retrieval path update to obtain the intelligent agent interpretable retrieval path update data.

2. The agent-interpretable retrieval path verification method according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: Obtaining agent multimodal input data; Step S12: performing cross-modal attention adjustment processing based on the agent multimodal input data, thereby obtaining agent multimodal weighted fusion data; Step S13: performing semantic representation fusion processing on the agent multimodal weighted fusion data, thereby obtaining input fusion semantic representation data; Step S14: performing associated knowledge retrieval processing based on the input fusion semantic representation data, thereby obtaining agent associated knowledge retrieval data.

3. The agent-interpretable retrieval path verification method according to claim 2, characterized in that: Step S14 includes the following steps: Step S141: extracting semantic query representation based on the input fusion semantic representation data to obtain input semantic query representation data; Step S142: performing semantic similarity calculation based on the input semantic query representation data to obtain input semantic similarity data; Step S143: performing keyword search processing based on the input semantic similarity data and the input semantic query representation data to obtain input semantic keyword search data; Step S144: determining input semantic candidate knowledge segment data based on the input semantic keyword retrieval data and the input semantic similarity data; Step S145: performing context sentence consistency evaluation on the input semantic candidate knowledge fragment data, thereby obtaining semantic context sentence consistency data; Step S146: performing knowledge fragment screening based on the semantic context sentence consistency data to obtain input semantic knowledge fragment data; Step S147: performing knowledge traceability processing based on the input semantic knowledge fragment data to obtain knowledge traceability data; Step S148: Perform associated knowledge retrieval processing based on the knowledge traceability data and the input semantic knowledge fragment data, thereby obtaining intelligent agent semantic associated knowledge retrieval data.

4. The agent-interpretable retrieval path verification method according to claim 1, characterized in that: In step S2, constructing node multi-path initial structure data based on agent semantic association knowledge retrieval data includes: Extract semantically pointing feature data based on the agent's semantically related knowledge retrieval data; Perform semantic node abstraction processing based on semantic pointing feature data to obtain path candidate node data; Perform semantic connection relationship recognition processing based on the path candidate node data to obtain node semantic connection relationship data; Perform path constraint screening on node semantic connection relationship data to obtain path effective connection structure data; Constructing a multi-path topology network based on the effective path connection structure data, thereby obtaining multi-path node topology structure data; Performing weight initialization processing on multi-path node topology structure data to obtain initial weight data of node topology structure; Node multi-path initial structure data is constructed based on node topology structure initial weight data and path effective connection structure data.

5. The agent-interpretable retrieval path verification method according to claim 1, characterized in that: Determining the path node causal reasoning chain data based on the node multi-path initial structure data in step S2 includes: Parsing the node control output field according to the node multi-path initial structure data to obtain the node control output field data; Parsing the node response input field according to the node multi-path initial structure data to obtain the node response input field data; determining path node control dependency data based on node response input field data and node control output field data; Determine the node path behavior evolution law based on the path node control dependency data; Analyze the inter-node trigger dependency relationship based on the node path behavior evolution law to obtain the inter-node dependency trigger data; Identify the causal basic unit data of path nodes based on the trigger data of the dependency relationship between nodes; The path node causal reasoning chain data is determined based on the path node causal basic unit data.

6. The agent-interpretable retrieval path verification method according to claim 1, characterized in that: In step S2, determining structured interpretable path data based on the path node causal reasoning chain data and the path node semantic interpretation information includes: Perform chain sequence reconstruction processing based on the path node causal reasoning chain data to obtain path node chain sequence reconstruction data; Determine the path node time sequence positioning data based on the path node chain sequence reconstruction data; Perform field alignment processing on the path node temporal positioning data and the path node semantic interpretation information data to obtain semantic field alignment data; Perform path node semantic structure annotation processing on the semantic field alignment data to obtain path node semantic annotation structure data; Perform semantic continuity analysis based on the semantic annotation structure data of the path nodes to obtain path semantic continuity data; Based on the path semantic continuity evaluation data, semantically credible segments are identified to obtain path semantically credible segment structure data; Merge the path semantic trusted segment structure data with the path node semantic annotation structure data to generate structured path expression template data; Based on the structured path expression template data, the path structure is uniformly output and processed to generate structured interpretable path data.

7. The agent-interpretable retrieval path verification method according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: detecting interpretable path deviation identification data based on the structured interpretable path data; Step S32: performing deviation cause analysis based on the interpretable path deviation identification data to obtain interpretable path deviation cause feature data; Step S33: generating an explainable path deviation correction strategy based on the explainable path deviation cause feature data to obtain an explainable 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.

8. The agent-interpretable retrieval path verification method according to claim 7, characterized in that: Step S31 includes the following steps: Step S311: Analyze node semantic jumps based on the structured interpretable path data to obtain path node semantic jump data; Step S312: performing a comparison process based on the path node semantic interpretation information and the path node semantic jump data to obtain path structure logic consistency deviation data; Step S313: performing path semantic interruption location clustering processing based on the path structure logic consistency deviation data to obtain path semantic interruption clustering data; Step S314: performing a path interpretation strength attenuation evaluation based on the path structure logic consistency deviation data to obtain path interpretation strength attenuation data; Step S315: determining the path semantic structure offset based on the path interpretation strength attenuation data to obtain path semantic structure offset data; Step S316: performing deviation weight normalization processing based on the path semantic structure offset data to obtain path semantic deviation comprehensive data; Step S317: Detect interpretable path deviation identification data based on the path semantic deviation comprehensive data and the path semantic structure offset data.

9. The agent-interpretable retrieval path verification method according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: performing path simulation correction feedback extraction based on the path structure simulation correction data, thereby obtaining path simulation correction feedback data; Step S42: verifying the rationality of the correction strategy based on the feedback data of the path simulation correction to obtain rationality data of the path correction strategy; 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.

10. An agent-interpretable retrieval path generation system, characterized in that: For executing the agent-interpretable search path verification method according to claim 1, the agent-interpretable search path generation system comprises: The associated knowledge retrieval processing module is used to obtain multimodal input data of the intelligent agent; perform associated knowledge retrieval processing based on the multimodal input data of the intelligent agent to obtain intelligent agent associated knowledge retrieval data; The interpretable path determination module is used to construct node multi-path initial structure data based on the agent semantic association knowledge retrieval data; determine path node causal reasoning chain data based on the node multi-path initial structure data; extract path node semantic interpretation information based on 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 interpretation information; A path structure simulation and correction module is used to detect interpretable path deviation identification data based on structured interpretable path data; perform path structure simulation and correction processing on the interpretable path deviation identification data to obtain path structure simulation and correction data; The retrieval path update module is used to perform interpretable retrieval path strategy optimization processing on the path structure simulation correction data to obtain interpretable path retrieval strategy optimization data; and perform intelligent agent interpretable retrieval path update processing on the structured interpretable path data according to the interpretable path retrieval strategy optimization data to obtain intelligent agent interpretable retrieval path update data.

Images