An agent intelligent auditing method suitable for multi-source heterogeneous enterprise-level expense invoices

Abstract

The application relates to the technical field of intelligent auditing of bills, and particularly discloses an Agent intelligent auditing method suitable for multi-source heterogeneous enterprise-level expense bills. The method comprises the following steps: extracting an auditing rule feature to determine an exclusive constraint set; uniformly standardizing an output graph by rendering an attachment flow, and directionally extracting a normalized element set; constructing a weighted correlation matrix by analyzing multi-dimensional feature similarity, and performing exclusive mapping to obtain a calibrated element pair; establishing a judgment graph according to a trigger condition, and generating an ordered verification chain through sorting and compiling; inputting the verification chain into an Agent analysis, combining a state snapshot to obtain a single judgment conclusion, and performing a hash operation on the aggregated judgment conclusion and a traceable identifier to obtain a tamper-proof evidence chain. The application realizes control of a large model reasoning state and irreversible solidification of bottom-layer auditing data, and is beneficial to improving the accuracy of enterprise-level complex expense accounting and the absolute reliability of auditing traceability.

Description

Technical Field

[0001] This invention relates to the field of intelligent document review technology, specifically to an Agent-based intelligent review method adapted to multi-source heterogeneous enterprise-level expense documents. Background Technology

[0002] As large corporations continue to expand their operations, the number of subsidiary units and the volume of cross-channel expense reimbursement are growing rapidly, leading to a corresponding increase in the complexity of reimbursement scenarios, document types, and supporting materials. Traditional expense review relies on a dual-stage review process involving initial and secondary manual checks, depending on the experience of finance personnel. This approach is prone to omissions and errors when dealing with a large volume of diverse documents, and suffers from low standardization. Currently, the industry generally hopes to automate the entire expense review process through artificial intelligence technology.

[0003] Existing intelligent expense review solutions, such as standardized SaaS expense review tools and document recognition solutions based on single-point large language models, adopt globally unified review rules. They lack a layered knowledge base that adapts to multiple entities within the group and cannot match the differentiated systems of different member companies (such as control standards for different years). Applying these solutions indiscriminately can lead to a large number of compliance misjudgments.

[0004] Meanwhile, the existing solution relies on a single bill of lading system for standard documents and does not have a mechanism for normalizing elements of multi-source heterogeneous forms, resulting in the inability to integrate different fields and layouts between parallel systems. The existing technology has very weak attachment compatibility, lacks a unified multi-format image rendering mechanism and intelligent pagination fallback strategy, and cannot handle complex mixed attachments containing multiple formats and non-standard layouts, resulting in unstructured attachments still requiring manual verification.

[0005] Therefore, this invention provides an Agent-based intelligent auditing method that adapts to multi-source heterogeneous enterprise-level expense documents. Summary of the Invention

[0006] The purpose of this invention is to provide an Agent-based intelligent auditing method that adapts to multi-source heterogeneous enterprise-level expense documents, in order to solve the aforementioned background problems.

[0007] The objective of this invention can be achieved through the following technical solutions: An agent-based intelligent auditing method adapted to multi-source heterogeneous enterprise-level expense documents includes the following steps: Extract the heterogeneous attachment stream bound to the documents to be reviewed, and perform format routing and proportional rendering to output a unified standard diagram; obtain the constraint hint words containing field validation requirements in the exclusive constraint set, and perform targeted extraction and normalization processing on the unified standard diagram based on the constraint hint words to obtain a normalized element set; Extract the monetary features contained in the normalized feature set to generate the retrieval primary key. When there is a conflict of the same amount in the matching candidate attachments, parse the multi-dimensional feature similarity between the form and the attachment to construct a weighted association matrix. Based on the weighted association matrix, perform exclusive mapping and missing item completion on the candidate attachments to obtain the calibration feature pairs. The control logic of the exclusive constraint set is parsed and the pre-trigger conditions are extracted. A directed decision graph is built based on the pre-trigger conditions and topological sorting is performed to obtain the sorted verification nodes. The calibration element pairs and verification nodes are mapped and bound to the context blocks, and an ordered verification chain is compiled. The ordered verification chain is input into the Agent to perform chain-like logical reasoning to obtain a single judgment conclusion; the traceability identifier mapped by the single judgment conclusion in the unified specification graph and the exclusive constraint set is parsed, and the single judgment conclusion and the traceability identifier are aggregated and hashed to generate a hash pointer, thus obtaining a tamper-proof evidence chain.

[0008] Furthermore, the method for determining the specific constraint set is as follows: Extract the hierarchical identifiers and control lists from the multi-source audit rule documents to construct a multidimensional index tree. Extract the business attributes of the documents to be audited and perform hierarchical recall within the multidimensional index tree to obtain the initial rule set. Perform intersection calculation on the boundary constraints of the initial rule set to determine the exclusive constraint set.

[0009] Furthermore, the process of performing the intersection calculation is as follows: Extract constraint extrema by analyzing the effective period and extreme value range contained in the initial rule set, establish control boundaries based on constraint extrema, and generate boundary constraint matrix; The control boundaries of the boundary constraint matrix are subjected to extreme value comparison. Based on the comparison results, conflict states are identified, and overlapping and redundant terms are eliminated according to the conflict states to determine the exclusive constraint set.

[0010] Furthermore, the process of performing the targeted extraction and normalization is as follows: Iterate through the exclusive constraint set to extract the field validation requirements of the target business fields, concatenate the field validation requirements with the preset instruction template execution text, and configure the generation of constraint prompt words; Constraint prompts and a unified specification graph are simultaneously input into a multimodal network. The cross-attention mechanism of the multimodal network is used to locate image feature regions and perform targeted extraction to obtain initial key-value pairs. The initial key-value pairs are cleaned and normalized, and the text is aligned with the original form fields based on the mapping dictionary to obtain a normalized feature set.

[0011] Furthermore, the process of outputting the unified specification diagram is as follows: Parse the file encoding attributes of the heterogeneous attachment stream, perform format routing and dispatch, call the conversion component corresponding to the file encoding attributes to perform pixel decoding, and generate the basic image stream; Extract the boundary dimensions of the base image stream and perform a proportional scaling operation with a preset standard long side threshold to output a unified standard image that removes layout differences.

[0012] Furthermore, the process of performing the exclusive mapping and missing term completion is as follows: Traverse the generated weighted correlation matrix to determine the minimum value; Extract the row and column indices of the minimum value, and force-bind the corresponding form detail points with the candidate attachment nodes to establish a one-to-one exclusive mapping relationship; Using the invoice data in the structured form as the absolute judgment benchmark, error correction verification is performed on the feature elements of the candidate attachments; Based on error correction and verification, complete the missing item completion operation and bidirectional coverage operation for the corresponding candidate attachments; For the remaining valid invoice data that exists in the form but is not matched in the candidate attachment subset, automatically create standardized blank invoice attachment entries and map and backfill the form elements to fill in the missing information; After completing the exclusive mapping, missing item completion, and two-way overlay operations, the form elements and attachment elements are combined and locked to obtain the calibration element pair.

[0013] Furthermore, the process of constructing the weighted correlation matrix is ​​as follows: Extract the monetary features contained in the normalized feature set and generate the retrieval primary key. Based on the retrieval primary key, initiate a matching traversal of candidate attachments. During the traversal, identify and extract the subset of candidate attachments that are in a state of equal amount conflict. Extract multidimensional association features from the candidate attachment subset and the original form sub-segments, calculate the temporal and textual differences of the multidimensional association features respectively, and then combine the preset bias weights with the temporal and textual differences to construct a weighted association matrix.

[0014] Furthermore, the process of binding the node mapping with the context block is as follows: Extract the required field attributes of the scheduled verification nodes, map the required field attributes to the calibration features and load them into the data slots to obtain the loaded node set; The total character length of the load node set is calculated, and the load node set is divided into blocks based on the context capacity scale to obtain task blocks. The task blocks are then compiled to generate an ordered verification chain.

[0015] Furthermore, the process of extracting the scheduled verification nodes is as follows: Parse the syntactic structure of each rule item within the exclusive constraint set, separate the pre-triggered conditions and post-validation actions, and establish a directed decision graph of the causal relationship of business validation. Extract the conditional parent nodes of the directed decision graph, substitute the calibration feature pairs into the conditional parent nodes to perform Boolean pre-computation, cut invalid dependency connection edges based on the pre-computation results, and generate a pruned decision graph. Perform a topological sort on the pruning decision graph to obtain the sorted check nodes.

[0016] Furthermore, the process of obtaining the tamper-proof evidence chain is as follows: The ordered verification chain is input into the Agent to perform logical reasoning at each node, and a state snapshot is constructed based on the probability matrix output by the reasoning. Extract the expected data format preset by the current verification node and the actual data format of the initial inference text. Perform format matching verification between the expected data format and the actual data format. When the format matching verification is consistent, trigger state inheritance. When the format matching verification is different, call the state snapshot to overwrite the memory working area to perform state rollback and obtain a single judgment conclusion. The target verification elements contained in the single judgment conclusion are analyzed, and the reverse lookup is performed on the unified specification diagram and the exclusive constraint set using the target verification elements to obtain the traceability identifier. The single judgment conclusion and the traceability identifier are converted into the underlying byte stream and then concatenated and aggregated. A secure hash algorithm is called to perform bitwise operations on the concatenated and aggregated result, generating a hash pointer and obtaining a tamper-proof evidence chain.

[0017] The beneficial effects of this invention are as follows: 1. Constructing a multi-dimensional index tree and performing hierarchical recall establishes rule associations encompassing multiple dimensions such as cost, channel, and year, which helps mitigate the shortcomings of inconsistent management standards across multiple business entities in large groups. Based on the calculation results, overlapping and redundant items are eliminated, and conflict states are handled, thereby generating a dedicated constraint set. This set can automatically identify and adjust the logical inversion of boundaries between upper and lower-level management dimensions, ensuring that the rule set issued to the system backend highly matches the actual verification needs of the current documents. Constraint prompts are configured in conjunction with the dedicated constraint set and synchronously input into a multimodal network with a unified specification diagram. A cross-attention mechanism is used to locate feature regions and generate key-value pairs, which are then cleaned and normalized to output a unified element set. This helps reduce interference from redundant background pixels and improves the reliability of reading core business data from multiple sources in heterogeneous business environments.

[0018] 2. A weighted association matrix is ​​constructed. Exclusive mapping is performed on the minimum value of the association matrix. After establishing the exclusive mapping, invoice data from structured forms is used as the judgment benchmark to perform bidirectional coverage and missing item completion on matching features to generate calibration feature pairs. This bidirectional cross-correction and data gap-filling fusion mechanism helps reduce feature bias caused by optical analysis or manual data entry by business personnel. A large model context capacity scale is introduced to perform block isolation and encapsulation of data sequences, generating an ordered verification chain. This helps control the tensor dimension of a single input model and prevents model attention decay caused by merging massive amounts of extremely long text inputs.

[0019] 3. By leveraging the Agent to perform progressive logical reasoning based on an ordered verification chain, a workspace state snapshot containing a probability matrix is ​​simultaneously constructed. The actual text data format of the reasoning output is compared in real time. When a format collapse or mutation is detected, the state snapshot is directly called to overwrite the current memory to execute process rollback. This enables the generative large model to have a task self-repair and forced data correction mechanism in the verification process, improving the stability of the output structure of multi-step chain reasoning. For the target verification elements in the final output conclusion, a reverse lookup of spatial coordinates and knowledge base clauses is performed. Physical coordinate data is extracted and combined with the clause number to generate a traceability identifier, completing the irreversible physical solidification and high-strength auditable record of the underlying AI logical reasoning process. Attached Figure Description

[0020] The invention will now be further described with reference to the accompanying drawings.

[0021] Figure 1 This is a system architecture and module data flow diagram of an Agent-based intelligent auditing method adapted to multi-source heterogeneous enterprise-level expense documents provided in this embodiment of the invention; Figure 2 This is a flowchart of an Agent-based intelligent auditing method adapted to multi-source heterogeneous enterprise-level expense documents according to the present invention; Figure 3 This is a flowchart illustrating how the calibration element pairs are obtained according to the present invention. Detailed Implementation

[0022] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0023] like Figure 1 As shown, this system architecture fully demonstrates the processing chain of multi-source heterogeneous data from the input end to the underlying Agent central hub. The group's multi-source heterogeneous data is divided into structured forms and unstructured attachments, which are respectively processed through a dual-line system of audit points and audit elements; In response to the key points of the review, the dedicated constraint set construction module extracts feature words by separating text and matching identifiers, and then establishes a multi-dimensional index tree to determine the tree topology; the constraint boundary precise adjustment module generates a boundary constraint matrix by unifying the extreme value dimensions, and destroys overlapping and redundant terms by comparing extreme values ​​and judging logical inversion, and determines the rule control boundary after adjustment convergence. For the review elements, the attachment intelligent rendering and extraction module performs physical splitting or matrix splicing on the heterogeneous attachment stream, and outputs a unified standard map after proportional pixel rendering; combined with constraint prompts, it performs cross-attention addressing, and retryes based on the tensor mean append instruction to separate and obtain the normalized elements; The element-equal-amount conflict calibration module constructs a weighted association matrix based on the primary key of the amount. By extracting the minimum value of the matrix and bidirectional coverage with the benchmark, it completes the exclusive mapping. The two data paths converge downwards to the ordered verification chain compilation module. The system establishes a directed decision graph with bound control identifiers, substitutes the calibration values, performs Boolean solving and cuts off invalid connected edges. After topological sorting and isolation compilation, an ordered verification chain is generated.

[0024] The orderly verification chain is input into the Agent intelligent audit center, and the logic deduction is performed in conjunction with the state snapshot anti-collapse rollback mechanism to obtain the audit result with a detailed list and the underlying anti-tampering hash.

[0025] Example 1:

[0026] like Figure 2 and Figure 3 As shown, an Agent-based intelligent auditing method for adapting to multi-source heterogeneous enterprise-level expense documents includes the following steps: S10. Extract the hierarchical identifiers and control lists from the multi-source audit rule documents to construct a multi-dimensional index tree. Extract the business attributes of the documents to be audited and perform hierarchical recall within the multi-dimensional index tree to obtain the initial rule set. Perform intersection calculation on the boundary constraints of the initial rule set to determine the exclusive constraint set. The process of extracting hierarchical identifiers and control lists from multi-source audit rule documents to construct a multidimensional index tree, extracting business attributes of documents to be audited, and performing hierarchical recall within the multidimensional index tree to obtain the initial rule set is as follows: S101. Extract the hierarchical identifiers and control lists from the source documents of the rules to generate feature words, and construct a multidimensional index tree based on the associated attributes of the feature words; Specifically, the method for generating feature terms and constructing a multidimensional index tree is as follows: The preset text parsing component is invoked to perform paragraph structure segmentation on the rule source documents from multiple sources, separating the title text and body content of the rule source documents; Using a pre-defined fee thesaurus and system dictionary, the fee type and hierarchical identifiers of different bill of lading channels are matched from the title text; Simultaneously apply regular expression matching logic to identify the effective year of the system, the applicable control list, and the corresponding control level weight from the main text content, and encapsulate the expense type, level identifier, effective year of the system, control list, and control level weight together as feature words; Specify the expense type as the root node, the hierarchy identifier as the second-level branch, the effective year of the system as the third-level branch, and the control list as the end node; According to the preset tree-like hierarchical rules, all feature words are mapped to the corresponding branches and node positions in sequence to build a multi-dimensional index tree containing hierarchical relationships; S102. Extract the business attributes of the documents to be reviewed to establish search conditions, and use the search conditions to perform hierarchical recall in the multidimensional index tree to obtain the initial rule set. Specifically, the method for establishing search criteria and performing hierarchical recall to obtain the initial rule set is as follows: Parse the incoming documents to be reviewed, extract the fee types, hierarchical identifiers, policy effective year, and control lists built into the structured form, and combine them into search criteria; Input the search criteria into the multidimensional index tree and initiate a depth-first search from the root node; If the search criteria match the attribute value of the current node, continue searching down to the next level node along the matching path. If the search criteria do not match the attribute values ​​of the current node, the node backtracking mechanism is triggered, and the comparison is transferred to an adjacent node at the same level. After the search reaches the end node of the multidimensional index tree, all the subdivided control rules attached to the successfully matched path are extracted, and all the subdivided control rules are aggregated and processed to obtain the initial rule set; The process of calculating the intersection of the boundary constraints of the initial rule set to determine the exclusive constraint set is as follows: S103. Extract constraint extreme values ​​from the effective period and extreme value range contained in the initial rule set, establish control boundaries based on constraint extreme values, and generate boundary constraint matrix; Specifically, the method for extracting constraint extrema, establishing control boundaries based on constraint extrema, and generating boundary constraint matrices is as follows: Semantic word segmentation is performed on the rule text in the initial rule set to separate the effective period of the time attribute, the extreme value range of the quota attribute, and the control level weight; Perform timestamp format conversion on the effective period to separate the lower time limit value and the upper time limit value; Perform uniform dimension processing on the extreme value range to extract the lower limit and upper limit of the amount; The lower time limit value and the upper time limit value are combined to determine the time control boundary; The lower limit and upper limit of the amount are combined to determine the limit control boundary; Traverse all control boundaries in the initial rule set, and arrange the time control boundary, quota control boundary and corresponding control level weights according to the corresponding rule number execution matrix to generate a structured boundary constraint matrix; S104. Perform extreme value comparison on the control boundary of the boundary constraint matrix, identify conflict states based on the comparison results, remove overlapping and redundant terms according to the conflict states, and determine the exclusive constraint set. Specifically, the method for determining the exclusive constraint set is as follows: For multiple limit control boundaries within the boundary constraint matrix, filter all lower limit values ​​and extract the maximum lower limit value; Simultaneously, filter all upper limit values ​​to extract the minimum upper limit value, and combine the maximum lower limit value and the minimum upper limit value to form the intersection boundary of the amounts; The minimum upper limit of the amount at the boundary of the amount intersection is compared and determined by the logic; If the minimum upper limit of the extracted amount is greater than or equal to the maximum lower limit of the amount, then the intersection boundary of the amounts is determined to be converged, and it is confirmed to be a conflict-free state. If the minimum upper limit of the extracted amount is less than the maximum lower limit of the amount, it is determined that there is a logical inversion at the boundary of the intersection of amounts, and it is confirmed as a significant conflict state. When a significant conflict is identified, the control level weight associated with the control boundary is extracted, and the limit control boundaries corresponding to the lower-level subsidiaries are selected and retained based on the control level weight. The remaining limit control boundaries are marked as overlapping and redundant items and memory destruction operations are performed. For example, if the limit control boundary of the group's head office is 400-500 yuan, the corresponding management level weight is Level 1, while the limit control boundary of the lower-level subsidiary is 200-300 yuan, the corresponding management level weight is Level 2; after extreme value comparison, the minimum value of the upper limit is calculated to be 300 yuan, which is less than the maximum value of the lower limit of 400 yuan, thus triggering a significant conflict state; at this time, according to the setting rule of the management standard tilting towards the lower-level business end, the system selects and retains the subsidiary limit control boundary corresponding to the Level 2 management level weight, and marks the head office limit control boundary corresponding to the Level 1 management level weight as an overlapping redundant item and performs a memory destruction operation; When the state is confirmed to be conflict-free, the converged limit control boundary is directly retained, and the memory destruction operation of overlapping redundant items is skipped. For time control boundaries, the same extreme value comparison and state determination logic as that used for quota control boundaries is employed to perform filtering. Aggregate the full set of valid time control boundaries and limit control boundaries that have been filtered and retained, and repackage them to output a dedicated constraint set.

[0027] S20. Perform format routing and proportional rendering on the heterogeneous attachment stream bound to the documents to be reviewed, and output a unified standard diagram; parse the constraint hints in the exclusive constraint set that contain field validation requirements, and perform targeted extraction and normalization processing on the unified standard diagram based on the constraint hints to obtain a normalized element set. The method for performing format routing and proportional rendering on heterogeneous attachment streams bound to documents awaiting review, and outputting a unified standard diagram, is as follows: S201. Parse the file encoding attributes of the heterogeneous attachment stream, perform format routing and dispatch, call the conversion component corresponding to the file encoding attributes to perform pixel decoding, and generate the basic image stream; Specifically, the method for generating the basic image stream is as follows: Read the heterogeneous attachment stream bound to the document to be reviewed, and extract the underlying file identifier header and memory layout characteristics of the heterogeneous attachment stream; The file encoding type is determined based on the underlying file identifier header; If the encoding type is portable document format, the heterogeneous attachment stream is routed and distributed to the document rendering branch. At this time, it is determined whether the current attachment stream is a multi-page aggregated layout based on the memory layout features. If so, the pixel contour density of each page is calculated using the edge detection operator, and the high-density feature pages are physically and independently split. Then, the document conversion engine is called to extract the independently split page view and convert it into raster pixels to generate a single-page basic image stream. If the encoding type is Extensible Markup Language format, the heterogeneous attachment stream is routed to the structured parsing branch, the tag parsing component is called to separate the text nodes and embedded images, and coordinate rendering is performed according to the preset layout rules to generate the basic image stream; If the encoding type is a standard image compression format, the heterogeneous attachment stream is routed and distributed to the direct-read branch. At this time, it is determined whether the current attachment stream is a discrete image arrangement based on the memory layout characteristics. If so, the image decoding engine is directly called to decode and restore the byte stream of the standard image compression format, generating multiple independent discrete pixel matrices. The system timestamps corresponding to the multiple discrete pixel matrices are extracted, and the absolute deviation between adjacent system timestamps is calculated. When the absolute deviation is less than the preset time aggregation threshold, the physical coordinates of the multiple independent discrete pixel matrices are stitched together to generate a fused basic image stream. For example, the portable document format is a PDF file commonly used in the reimbursement process, the extensible markup language format is an XML file of a digital invoice, the standard image compression format is a JPG file or a PNG file containing a restaurant receipt, and the time aggregation threshold is preset to 3 seconds. S202. Extract the boundary dimensions of the basic image stream and perform a proportional scaling operation with the preset standard long side threshold to output a unified standard image with layout differences removed. The process of outputting the unified standard diagram is as follows: Read the extreme values ​​of pixel coordinates of the base image stream to determine the original pixel width and original pixel height of the base image stream; The system's pre-set standardized parameter library is called to extract the standard long-side threshold for multimodal recognition; For example, the standard long side threshold is preset to 2048 pixels; Compare the original pixel width and the original pixel height, and select the maximum value of the two as the current long side size; Calculate the quotient of the standard long side threshold and the current long side size to obtain the proportional scaling factor; The original pixel width and original pixel height are simultaneously multiplied by a proportional scaling factor to calculate the target pixel width and target pixel height. The bicubic interpolation algorithm is called to smoothly map the pixel matrix of the base image stream to the grid space defined by the target pixel width and the target pixel height, and output a uniform standard image with layout differences removed. The process of parsing the constraint hints in the exclusive constraint set, which contain field validation requirements, and then performing targeted extraction and normalization on the unified specification diagram based on these constraint hints to obtain the normalized feature set is as follows: S203. Traverse the exclusive constraint set to extract the field validation requirements of the target business fields, concatenate the field validation requirements with the preset instruction template execution text, and configure the generation of constraint prompt words; Specifically, the process of configuring and generating constraint prompts is as follows: Traverse the exclusive constraint set determined in S10 and extract the target business fields that each verification node in the exclusive constraint set depends on. For example, the target business fields include the invoice time field and the total amount including tax field in the invoice recognition scenario, and the merchant name field and the number of people traveling with the business entertainment scenario. Extract the field validation requirements attached to the target business fields, and use a semantic analyzer to convert the field validation requirements into structured constraint statements; Read the multimodal instruction template loaded into memory; Perform string concatenation operation between the structured constraint statement and the multimodal instruction template, and append key-value pair identifiers to the end of the concatenation to qualify the data type, and configure the generation of constraint prompt words; S204. Simultaneously input the constraint prompts and the unified specification diagram into the multimodal network, and use the cross-attention mechanism of the multimodal network to locate the image feature regions and perform directional extraction to obtain the initial key-value pairs; Specifically, the process of obtaining the initial key-value pairs is as follows: Activate the pre-deployed multimodal network, input the unified standard graph into the visual encoding layer of the multimodal network, and use the convolution operation of the visual encoding layer to extract the visual feature tensor representing the local physical texture; The constraint prompts are simultaneously input into the text encoding layer of the multimodal network, and the text feature tensor representing the semantic instructions is extracted using the self-attention network of the text encoding layer. The visual feature tensor and the text feature tensor are input together into the cross-attention layer; Within the cross-attention layer, the similarity of the inner product between the text feature tensor and the visual feature tensor is calculated, and attention weights are assigned accordingly. Based on the extreme values ​​of attention weights, cluster centers are used to locate image feature regions corresponding to the unified normalization map; The decoding layer of the multimodal network performs character reconstruction parsing on the pixel clusters contained in the image feature region, outputs initial key-value pairs, and simultaneously extracts the probability tensor representing the parsing accuracy. Calculate the tensor mean of the probability tensor and determine whether the tensor mean is greater than or equal to a preset precision lower limit threshold. If the tensor mean is greater than or equal to the lower precision threshold, then the initial key-value pair of the current output is determined directly. If the tensor mean is less than the precision lower limit threshold, control instructions representing contrast enhancement and local coordinate clipping are dynamically added to the constraint cue words. This drives the multimodal network to recalculate the inner product similarity and attention weights using the constraint cue words after the added control instructions, until the tensor mean corresponding to the re-output result is greater than or equal to the precision lower limit threshold, thus determining the final initial key-value pair. For example, the preset lower limit of accuracy is set to 0.85; S205. Clean and normalize the numerical format of the initial key-value pairs, and perform text alignment with the original form fields based on the mapping dictionary to obtain a normalized element set; The process of cleaning and normalizing the numerical format of the initial key-value pairs, and aligning the text with the original form fields based on the mapping dictionary to obtain the normalized feature set is as follows: Extract the time and monetary components contained in the initial key-value pairs; Determine the data separator type for the time component; If the time component contains Chinese characters or slash separators, the regular expression matching tree is called to extract the corresponding year, month, and day value segments, and zero padding is performed on odd-numbered months and odd-numbered days to concatenate and generate standard time values. If the time component does not contain Chinese characters or slash separators, it is determined that the current time component conforms to the unified epoch format, and the current time component is directly saved as a standard time value. The preset character filtering regular expression is used to remove currency symbols and thousands separators from the amount, separating the pure numerical string. Then, according to the precision control logic of retaining two decimal places, the pure numerical string is converted to floating point to generate the standard amount value. For example, based on real enterprise-level expense documents, the currency symbol specifically includes the RMB symbol or the USD symbol, and the thousands separator is the English character comma; The system's embedded mapping dictionary is invoked to perform a semantic association comparison between the dictionary keys corresponding to the standard time value and the standard amount value and the native form fields built into the structured form. If the semantic association comparison results are consistent, the standard time value and the standard amount value will be overwritten and written into the data slot corresponding to the original field of the form. If the semantic association comparison results are inconsistent, an extended data node will be dynamically created in the structured form, and the standard time value and standard amount value will be written into the extended data node. The original form fields and extended data nodes after the data writing operation are aggregated and encapsulated to obtain a unified element set with highly consistent data precision and text format.

[0028] S30. Extract the monetary features contained in the normalized feature set to generate the retrieval primary key. When there is a conflict of the same amount in the matching candidate attachments, parse the multi-dimensional feature similarity between the form and the attachment to construct a weighted association matrix. Based on the weighted association matrix, perform exclusive mapping and missing item completion on the candidate attachments to obtain the calibration feature pairs. The process of extracting monetary features from the normalized feature set to generate the retrieval primary key, and constructing a weighted association matrix by parsing the multidimensional feature similarity between the form and the attachment when matching candidate attachments with the same amount conflict, is as follows: S301. Extract the monetary features contained in the normalized element set and generate the retrieval primary key. Based on the retrieval primary key, initiate a matching traversal of candidate attachments. During the traversal, identify and extract the subset of candidate attachments that are in a state of equal amount conflict. Specifically, the process of generating the primary key for retrieval and extracting a subset of candidate attachments is as follows: Iterate through the normalized element set output by S20 and extract the standard amount value corresponding to the structured form; Using the standard amount value as the primary key, data traversal and matching are initiated in the candidate attachment set cached by the system; The number of candidate attachments that match the primary key is counted to obtain the concurrent attachment count. If the number of concurrent attachments is equal to the preset single baseline value, it is determined that the form invoice and the attachment invoice are associated one-to-one. The current candidate attachment is locked and a mapping pointer is established to skip the disambiguation action in the multi-dimensional space. If the number of concurrent attachments exceeds the preset single benchmark value, it is determined that multiple transactions with the same amount are concurrently occurring in the current document, and a conflict of equal amounts is identified. If a conflict is confirmed, extract all candidate attachments that match the primary key of the search and aggregate and encapsulate them into a subset of candidate attachments. If the number of concurrent attachments is equal to the preset empty state value, it is determined that the current form is missing the corresponding attachment material support, triggering the missing item exception flag and terminating the matching traversal of the current node. For example, if there are two taxi fare details for official business with the same amount of 50 yuan in the form, and the candidate attachment set happens to scan and parse two taxi invoices with the same amount of 50 yuan, the system cannot complete the one-to-one matching logic by relying solely on the retrieval primary key of the amount. In this case, the system will include the two taxi invoices in the candidate attachment subset; wherein, the preset single benchmark value is set to 1, and the empty state value is set to 0. S302. Extract the multi-dimensional association features of the candidate attachment subset and concurrent form details, calculate the temporal difference and textual difference of the multi-dimensional association features respectively, and integrate the preset bias weights with the temporal difference and textual difference to construct a weighted association matrix. Specifically, the method for extracting multidimensional association features, calculating temporal and textual differences, and constructing a weighted association matrix is ​​as follows: Extract the explicit details of concurrent forms that have the same amount of conflict to form a concurrent form set; Perform bidirectional nested traversal on the concurrent set of forms and the subset of candidate attachments. For each pair of form details and candidate attachments, extract the corresponding attachment invoicing time and attachment merchant name in the normalized element set, and simultaneously extract the form filling time and form filling merchant in the corresponding form details. Perform a difference operation between the form filling time and the invoice issuance time in the attachment to calculate the absolute time period between the two; Divide the absolute time period by the system's preset time tolerance threshold to calculate the normalized time difference. The character edit distance algorithm is used to perform semantic span verification between the merchant names in the attachments and the merchants filled in on the form, and the number of operation steps for the differences in the string form between the two is calculated. Divide the number of difference operation steps by the total number of characters in the longer string to calculate the text difference degree that reflects the deviation of the characters. Extract the first bias weight and the second bias weight preset by the system, multiply the time difference by the first bias weight to obtain the time penalty component, and multiply the text difference by the second bias weight to obtain the text penalty component. The time penalty component and the text penalty component are summed to obtain the comprehensive deviation value representing the multidimensional comprehensive spatial distance of each pair of nodes. After completing the bidirectional nested traversal, all comprehensive deviation values ​​are arranged into a matrix according to the sequence index of the details in the concurrent set of the form as the row coordinate and the sequence index of the attachments in the candidate attachment subset as the column coordinate, and a weighted correlation matrix is ​​generated. For example, the preset time tolerance judgment limit is set to 24 hours, the first bias weight is set to 0.6, and the second bias weight is set to 0.4; The process of performing exclusive mapping and missing item completion on candidate attachments based on the weighted correlation matrix to obtain calibration element pairs is as follows: S303. Extract the minimum value of the weighted correlation matrix, perform exclusive mapping, and perform bidirectional coverage and missing item completion on the matching features based on the form invoice benchmark to obtain the calibration feature pair; Specifically, the method for performing exclusive mapping and performing bidirectional coverage and missing item completion on the matching features based on the form invoice benchmark to obtain the calibration feature pairs is as follows: Traverse the generated weighted correlation matrix, perform a matrix element scan, and determine the minimum value. Extract the row and column indices of the minimum value, and force-bind the corresponding form detail points with the candidate attachment nodes to establish a one-to-one exclusive mapping relationship; After the exclusive mapping relationship is established, the invoice data of the structured form is used as the absolute judgment benchmark to perform error correction verification on the feature elements of the candidate attachments; If the invoice date identified in the candidate attachment is determined to be empty or garbled, the standard invoice date corresponding to the original field in the form will be automatically filled in to complete the missing item completion operation. If it is determined that the merchant name and other features of the candidate attachments have recognition errors caused by optical character recognition, then the recognition results of the standardized element text coverage anomaly of the original form fields will be uniformly adopted to complete the two-way coverage operation. For the remaining valid invoice data that exists in the form but is not matched in the candidate attachment subset, automatically create standardized blank invoice attachment entries and map and backfill the form elements to fill in the missing information; After completing the exclusive mapping, missing item completion, and bidirectional overlay operations, the form elements and attachment elements are combined and locked to obtain a calibration element pair that removes recognition bias and matching ambiguity.

[0029] Example 2:

[0030] Please see Figure 2 As shown, an Agent-based intelligent auditing method for adapting to multi-source heterogeneous enterprise-level expense documents includes the following steps: S40. Parse the control logic of the exclusive constraint set and extract the pre-trigger conditions. Based on the pre-trigger conditions, establish a directed decision graph and perform topological sorting to obtain the sorted verification nodes. Perform node mapping and context block binding on the calibration element pairs and verification nodes, and compile to generate an ordered verification chain. The process of parsing the control logic of the exclusive constraint set and extracting the pre-trigger conditions, building a directed decision graph based on the pre-trigger conditions and performing topological sorting to obtain the arranged verification nodes is as follows: S401. Parse the syntax structure of each rule item in the exclusive constraint set, separate the pre-trigger conditions and post-validation actions, and establish a directed decision graph of the causal relationship of business validation. Specifically, the method for constructing a directed decision graph is as follows: Iterate through the exclusive constraint set and input each rule item of the exclusive constraint set into the pre-built dependency parser; The dependency parser is used to perform part-of-speech tagging on the rule items, separating the pre-triggered conditions of the premise constraints and the post-validation actions of the execution instructions; Determine if the current rule item has a prerequisite triggering condition; If there is no prerequisite triggering condition, the current rule item is defined as an unconditional node and assigned a default global activation flag; If a pre-triggering condition exists, a graph data structure is constructed in the system memory, defining the pre-triggering condition as the condition parent node in the graph structure, and the post-validation action as the action child node in the graph structure. Synchronously parse the control level weights attached to the rule items, and convert the control level weights into strong control labels or weak control labels; use directed memory pointers to establish dependency connection edges from the condition parent node to the action child node, and bind the corresponding strong control labels or weak control labels to the dependency connection edges; traverse all rule items to complete the insertion operation of nodes and dependency connection edges, and generate a directed decision graph representing the causal relationship of business verification. S402. Extract the conditional parent node of the directed decision graph, substitute the calibration feature pairs into the conditional parent node to perform Boolean pre-calculation, cut invalid dependency connection edges according to the pre-calculation results, and generate a pruning decision graph. Specifically, the process of extracting conditional parent nodes, performing Boolean pre-computation, and generating the pruning decision graph is as follows: Extract the calibration element pairs determined in S30 and traverse all condition parent nodes in the directed decision graph; Identify the logical expression type of the parent node of the condition and determine whether the logical expression type contains only arithmetic relational operators; If the logical expression type contains semantically ambiguous words that cannot be directly calculated, then the state of the dependent connection edge of the parent node of the condition remains unchanged and is handed over to the downstream process for processing; If the logical expression type contains only arithmetic relational operators, then the numerical features in the calibration feature pair are substituted into the conditional parent node to perform Boolean algebra solution; Based on the Boolean algebra solution, if the calculated output is the Boolean true state, then the judgment condition is met and the state of the dependent connection edges remains unchanged. If the calculated output is a Boolean pseudostate, the condition is not met, and the identifier bound to the corresponding dependent connection edge is extracted. If a strong control flag is extracted, the dependent connection edges between the parent node of the condition and the action child node are removed, and the isolated action child node after the edge is broken is released to realize the graph physical pruning of invalid branches and generate a pruning decision graph. If a weak control identifier is extracted, the dependent connection edge connectivity of the directed decision graph is maintained, and a risk warning field is injected into the context state of the corresponding action child node. The corresponding branch is retained and a pruning decision graph is generated, which is then handed over to the downstream process for degradation deduction. For example, if the parent node of the condition is an amount greater than 1,000 yuan and is bound to a strong control identifier, and the logarithm of the calibration element is 500 yuan, resulting in a false state in the Boolean solution, the system will directly destroy the subsequent associated director approval action sub-node to achieve a pre-emptive physical blockade; if it is bound to a weak control identifier, the action sub-node will be retained and a risk warning field will be injected to prompt manual attention. S403. Perform topological sorting on the pruning decision graph to obtain the sorted check nodes; Specifically, the process of performing topological sorting is as follows: After the traversal operation is completed, calculate the in-degree value of each node in the pruning decision graph; Initialize the decision queue in system memory, push nodes with an in-degree value of zero into the decision queue, perform a loop dequeue operation and decrement the in-degree value of adjacent nodes in turn, and generate a scheduled verification node that removes the loop deadlock and filters invalid branches. The process of binding calibration element pairs and verification nodes to execution node mapping and context blocks, and compiling to generate an ordered verification chain, is as follows: S404. Extract the required field attributes of the scheduled verification nodes, map the required field attributes to the calibration features and load them into the data slots to obtain the loaded node set; The method for obtaining the load node set is as follows: Read the arranged verification nodes one by one in the order of arrangement, and extract the necessary field attributes required for the current verification node to execute the verification logic; Read the calibration element pairs determined in S30 and extract the actual element key names built into the calibration element pairs; Perform a string matching comparison between the required field attributes and the actual feature key names; If the matching results are consistent, extract the corresponding numerical features from the calibration feature pair and fill them into the data slot of the current verification node; If the matching results are inconsistent, it is determined that the current expense report is missing the data required for the verification node, and a preset missing mark is written into the data slot of the current verification node. After all the data slots of the fully scheduled verification nodes have been loaded, the scheduled verification nodes carrying real business data will be aggregated and encapsulated to obtain the loaded node set. S405. Calculate the total character length of the load node set, perform block isolation on the load node set based on the context capacity scale to obtain task blocks, and compile the task blocks to generate an ordered verification chain.

[0031] The method for generating an ordered verification chain is as follows: Extract all context data and verification instructions bound within the load node set, and use a character statistics function to calculate the total character length of the load node set; The system's preset large model context capacity scale is used to extract the single attention threshold. Compare the sum of character lengths with the single attention threshold. If the total character length exceeds the single attention threshold, a global delimiter is inserted according to the topological hierarchy boundary within the loaded node set to cut and isolate the long sequence into multiple independent task blocks. The first and last sequences of the multiple task blocks are combined and locked to generate an ordered verification chain. If the total character length does not exceed the single attention threshold, the entire set of loaded nodes is set as a single task block, and sequence combination and locking compilation are directly performed on the single task block to generate an ordered verification chain.

[0032] S50. Input the ordered verification chain into the Agent to perform chain-like logical reasoning and obtain a single judgment conclusion; parse the traceability identifier mapped by the single judgment conclusion in the unified specification diagram and the exclusive constraint set, aggregate the single judgment conclusion and the traceability identifier, perform hash operation to generate a hash pointer, and obtain a tamper-proof evidence chain. The process of inputting an ordered verification chain into the Agent to perform chain-like logical reasoning and obtain a single judgment conclusion is as follows: S501. Input the ordered verification chain into the Agent and perform logical reasoning at each node. Construct a state snapshot based on the probability matrix output by the reasoning. Specifically, the process of performing logical reasoning is as follows: Extract the ordered verification chain generated by S40 compilation and input the ordered verification chain into the preset Agent memory working area; According to the node order of the ordered verification chain, the current verification node is extracted and loaded into the inference engine to drive the Agent to perform intent understanding and logical reasoning, and output the initial inference text; Synchronously extract the probability floating-point value corresponding to each output character when the Agent generates the initial inference text, and arrange all the probability floating-point values ​​according to the character order to generate a probability matrix; The context data, initial inference text, and probability matrix of the current verification node are packaged and compressed to construct a state snapshot of the current node, and the state snapshot is stored in persistent storage medium. S502. Extract the expected data format preset by the current verification node and the actual data format of the initial inference text. Perform format matching verification between the expected data format and the actual data format. Trigger state inheritance when the format matching verification is consistent. Call the state snapshot to overwrite the memory working area to perform state rollback when the format matching verification is different, and obtain a single judgment conclusion. Specifically, the process of obtaining a single-item judgment conclusion is as follows: Extract the expected data format preset by the current verification node, and synchronously call the format parser to extract the actual data format of the initial inference text; Perform a structure matching and verification between the actual data format and the expected data format; If the actual data format is completely consistent with the expected data format, that is, the format matching verification is consistent, then it is determined that there is no illusion deviation in the chain logic reasoning, the state inheritance mechanism is activated, the initial reasoning text is determined as the single judgment conclusion, and the next verification node is loaded in sequence. If the actual data format is inconsistent with the expected data format, i.e. the format matching verification is different, it is determined that the chain logic reasoning has collapsed, the state rollback mechanism is activated, the state snapshot of the previous node is extracted from the persistent storage medium, the state snapshot is called to overwrite the current Agent's memory working area to complete the state rollback, and a forced format constraint prompt word is injected into the current verification node to drive the Agent to re-execute the logic reasoning until the actual data format of the re-output is completely consistent with the expected data format, and a single judgment conclusion without format deviation is output. The process of parsing the traceability identifier mapped to the unified standard diagram and the exclusive constraint set for a single judgment conclusion, aggregating the single judgment conclusion and the traceability identifier, performing a hash operation to generate a hash pointer, and obtaining the tamper-proof evidence chain is as follows: S503. Analyze the target verification elements contained in the single judgment conclusion, and use the target verification elements to perform reverse lookup in the unified specification diagram and the exclusive constraint set respectively, and obtain the traceability identifier. Specifically, the method for parsing and obtaining the traceability identifier is as follows: The output of a single judgment conclusion is decomposed into fields to separate the target verification elements that participate in the current business judgment; Using the target verification element as the retrieval primary key, initiate a reverse lookup of pixel space coordinates in the unified standard map generated in S20; Extract the bounding rectangle boundary of the target verification element in the unified standard map, extract the vertex pixel positions of the bounding rectangle boundary and determine them as the physical coordinates of the image; Simultaneously, the business attributes carried by the target verification elements are input into the exclusive constraint set determined by S10 to initiate a traversal search and match the institutional clause text that supports the determination of the target verification elements. Extract the unique identifier code of the policy clause text from the underlying database table of the enterprise management system and determine it as the clause number; By using commas as separators, the physical coordinates of the image and the clause number are concatenated into a string, and the traceability identifier representing the physical source of the data and the basis of the verification rules is obtained. S504. Convert the single judgment conclusion and the traceability identifier into the underlying byte stream, perform splicing and aggregation, call the secure hash algorithm to perform bitwise operations on the splicing and aggregation result, generate a hash pointer and obtain a tamper-proof evidence chain; Specifically, the method for obtaining a tamper-proof chain of evidence is as follows: The system's unified character encoding set is invoked to convert the individual judgment conclusion and the traceability identifier into the corresponding conclusion byte sequence and traceability byte sequence, respectively. The conclusion byte sequence and the source byte sequence are serialized and concatenated in ascending order of timestamps to generate an aggregated data block. Input the aggregated data block into the system's preset secure hash function; Within the secure hash function, non-linear bitwise XOR operations and circular shift operations are performed on the aggregated data block using a preset initial hash constant. By using bitwise XOR operations and cyclic shift operations, variable-length aggregated data blocks are compressed and mapped into fixed-length hexadecimal feature strings; Extract the fixed-length hexadecimal characteristic string and determine it as a hash pointer; Call the node communication interface of the distributed ledger to synchronously write the single-item judgment conclusion, traceability identifier, and hash pointer into the block body of the underlying blockchain; Receive the block height and transaction receipt returned by the underlying blockchain to confirm the underlying data mapping relationship and obtain a tamper-proof evidence chain.

[0033] The two embodiments of the present invention have been described in detail above, but the content described is only a preferred embodiment of the present invention and should not be considered as limiting the scope of the present invention. All equivalent changes and improvements made in accordance with the scope of the present invention should still fall within the scope of the present invention.

Claims

1. An agent-based intelligent auditing method adapted to multi-source heterogeneous enterprise-level expense documents, characterized in that, Includes the following steps: Extract the heterogeneous attachment stream bound to the documents to be reviewed, and perform format routing and proportional rendering to output a unified standard diagram; obtain the constraint hint words containing field validation requirements in the exclusive constraint set, and perform targeted extraction and normalization processing on the unified standard diagram based on the constraint hint words to obtain a normalized element set; Extract the monetary features contained in the normalized feature set to generate the retrieval primary key. When there is a conflict of the same amount in the matching candidate attachments, parse the multi-dimensional feature similarity between the form and the attachment to construct a weighted association matrix. Based on the weighted association matrix, perform exclusive mapping and missing item completion on the candidate attachments to obtain the calibration feature pairs. The control logic of the exclusive constraint set is analyzed and the pre-trigger conditions are extracted. A directed decision graph is built based on the pre-trigger conditions and a topological sort is performed to obtain the sorted verification nodes. The calibration elements are paired with the verification nodes, and node mapping and context block binding are performed to generate an ordered verification chain. The ordered verification chain is input into the Agent to perform chain-like logical reasoning to obtain a single judgment conclusion; the traceability identifier mapped by the single judgment conclusion in the unified specification graph and the exclusive constraint set is parsed, and the single judgment conclusion and the traceability identifier are aggregated and hashed to generate a hash pointer, thus obtaining a tamper-proof evidence chain.

2. The Agent-based intelligent auditing method for adapting to multi-source heterogeneous enterprise-level expense documents according to claim 1, characterized in that: The method for determining the specific constraint set is as follows: Extract the hierarchical identifiers and control lists from the multi-source audit rule documents to construct a multidimensional index tree. Extract the business attributes of the documents to be audited and perform hierarchical recall within the multidimensional index tree to obtain the initial rule set. Perform intersection calculation on the boundary constraints of the initial rule set to determine the exclusive constraint set.

3. The Agent-based intelligent auditing method for adapting to multi-source heterogeneous enterprise-level expense documents according to claim 2, characterized in that: The process of performing the intersection calculation is as follows: Extract constraint extrema by analyzing the effective period and extreme value range contained in the initial rule set, establish control boundaries based on constraint extrema, and generate boundary constraint matrix; The control boundaries of the boundary constraint matrix are subjected to extreme value comparison. Based on the comparison results, conflict states are identified, and overlapping and redundant terms are eliminated according to the conflict states to determine the exclusive constraint set.

4. The Agent-based intelligent auditing method for adapting to multi-source heterogeneous enterprise-level expense documents according to claim 1, characterized in that: The process of performing the targeted extraction and normalization is as follows: Iterate through the exclusive constraint set to extract the field validation requirements of the target business fields, concatenate the field validation requirements with the preset instruction template execution text, and configure the generation of constraint prompt words; Constraint prompts and a unified specification graph are simultaneously input into a multimodal network. The cross-attention mechanism of the multimodal network is used to locate image feature regions and perform targeted extraction to obtain initial key-value pairs. The initial key-value pairs are cleaned and normalized, and the text is aligned with the original form fields based on the mapping dictionary to obtain a normalized feature set.

5. The Agent-based intelligent auditing method for adapting to multi-source heterogeneous enterprise-level expense documents according to claim 4, characterized in that: The process of outputting the unified specification diagram is as follows: Parse the file encoding attributes of the heterogeneous attachment stream, perform format routing and dispatch, call the conversion component corresponding to the file encoding attributes to perform pixel decoding, and generate the basic image stream; Extract the boundary dimensions of the base image stream and perform a proportional scaling operation with a preset standard long side threshold to output a unified standard image that removes layout differences.

6. The Agent-based intelligent auditing method for adapting to multi-source heterogeneous enterprise-level expense documents according to claim 1, characterized in that: The process of performing the exclusive mapping and missing term completion is as follows: Traverse the generated weighted correlation matrix to determine the minimum value; Extract the row and column indices of the minimum value, and force-bind the corresponding form detail points with the candidate attachment nodes to establish a one-to-one exclusive mapping relationship; Using the invoice data in the structured form as the absolute judgment benchmark, error correction verification is performed on the feature elements of the candidate attachments; Based on error correction and verification, complete the missing item completion operation and bidirectional coverage operation for the corresponding candidate attachments; For the remaining valid invoice data that exists in the form but is not matched in the candidate attachment subset, automatically create standardized blank invoice attachment entries and map and backfill the form elements to fill in the missing information; After completing the exclusive mapping, missing item completion, and two-way overlay operations, the form elements and attachment elements are combined and locked to obtain the calibration element pair.

7. The Agent-based intelligent auditing method for adapting to multi-source heterogeneous enterprise-level expense documents according to claim 6, characterized in that: The process of constructing the weighted correlation matrix is ​​as follows: Extract the monetary features contained in the normalized feature set and generate the retrieval primary key. Based on the retrieval primary key, initiate a matching traversal of candidate attachments. During the traversal, identify and extract the subset of candidate attachments that are in a state of equal amount conflict. Extract multidimensional association features from the candidate attachment subset and the original form sub-segments, calculate the temporal and textual differences of the multidimensional association features respectively, and then combine the preset bias weights with the temporal and textual differences to construct a weighted association matrix.

8. The Agent-based intelligent auditing method for adapting to multi-source heterogeneous enterprise-level expense documents according to claim 1, characterized in that: The process of binding the node mapping with the context block is as follows: Extract the required field attributes of the scheduled verification nodes, map the required field attributes to the calibration features and load them into the data slots to obtain the loaded node set; The total character length of the load node set is calculated, and the load node set is divided into blocks based on the context capacity scale to obtain task blocks. The task blocks are then compiled to generate an ordered verification chain.

9. The Agent-based intelligent auditing method for adapting to multi-source heterogeneous enterprise-level expense documents according to claim 8, characterized in that: The process of extracting the scheduled verification nodes is as follows: Parse the syntactic structure of each rule item within the exclusive constraint set, separate the pre-triggered conditions and post-validation actions, and establish a directed decision graph of the causal relationship of business validation. Extract the conditional parent nodes of the directed decision graph, substitute the calibration feature pairs into the conditional parent nodes to perform Boolean pre-computation, cut invalid dependency connection edges based on the pre-computation results, and generate a pruned decision graph. Perform a topological sort on the pruning decision graph to obtain the sorted check nodes.

10. The Agent-based intelligent auditing method for adapting to multi-source heterogeneous enterprise-level expense documents according to claim 1, characterized in that: The process of obtaining the tamper-proof evidence chain is as follows: The ordered verification chain is input into the Agent to perform logical reasoning at each node, and a state snapshot is constructed based on the probability matrix output by the reasoning. Extract the expected data format preset by the current verification node and the actual data format of the initial inference text. Perform format matching verification between the expected data format and the actual data format. When the format matching verification is consistent, trigger state inheritance. When the format matching verification is different, call the state snapshot to overwrite the memory working area to perform state rollback and obtain a single judgment conclusion. The target verification elements contained in the single judgment conclusion are analyzed, and the reverse lookup is performed on the unified specification diagram and the exclusive constraint set using the target verification elements to obtain the traceability identifier. The single judgment conclusion and the traceability identifier are converted into the underlying byte stream and then concatenated and aggregated. A secure hash algorithm is called to perform bitwise operations on the concatenated and aggregated result, generating a hash pointer and obtaining a tamper-proof evidence chain.

Images