An AI-based enterprise management software SaaS data processing method and system

Abstract

The application discloses a kind of based on AI enterprise management software SaaS data processing method and system, is related to enterprise management software SaaS data processing technical field, including the following steps: S1, response service request, in tenant running space and delimit enterprise data domain;S2, to the data record in enterprise data domain execution fact element and gather, and be connected as object fact chain;S3, assemble enterprise fact graph, access service and access boundary path Path;S4, transcription request landing point and embed enterprise fact graph, form request anchoring fact graph;S5, input improvement GraphRAG model, form closed evidence subgraph and access fracture shadow band;S6, generate enterprise processing conclusion and split, anti-paste conclusion fragment;S7, shield fracture or not anti-paste fragment, solidify enterprise SaaS data processing result.The application can improve the traceability of enterprise SaaS data processing and access boundary consistency.

Description

Technical Field

[0001] This invention relates to the field of data processing technology for enterprise management software (SaaS), and in particular to a data processing method and system based on AI-powered enterprise management software (SaaS). Background Technology

[0002] As enterprise management software (SaaS) is increasingly used in sales, procurement, inventory, finance, production, cost, approval, supplier management, and customer management, enterprise business data is typically scattered across different functional modules and tenant runtime environments. Existing AI data processing methods often directly retrieve and generate data from business records, retrieval fragments, or graph nodes, making it difficult to simultaneously maintain consistency in business relationships, tenant data boundaries, and conclusion sources. While ordinary GraphRAG models can enhance retrieval performance using graph structures, in enterprise SaaS scenarios, directly deleting invisible graph fragments at access boundaries can easily break evidence paths; retaining cross-boundary fragments may allow invisible business fact elements to enter the prompt context, creating access violation risks. Furthermore, generated results are usually only associated with the overall retrieval context, lacking the constraint of back-attachment between conclusion fragments and evidence positions within the graph, leading to incomplete evidence chains, difficulty in tracing sources, and unfounded generation of conclusions.

[0003] Therefore, how to provide a data processing method and system based on AI-powered enterprise management software (SaaS) is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a data processing method and system based on AI-powered enterprise management software (SaaS). This invention, through the hierarchical construction of enterprise data domains, object fact chains, and enterprise fact graphs, organizes data records scattered across different business functions into an in-graph evidence structure with a business succession order. This allows business requests to enter the corresponding business succession path and access boundary path from the request landing point, avoiding AI processing relying solely on scattered records or isolated retrieval fragments. The improved GraphRAG model forms a closed evidence subgraph through evidence path extension units and embeds a boundary shadow compression structure in the access breakpoint shadow band unit. When an evidence path crosses an access boundary path, the cross-boundary path segment is folded into an access breakpoint shadow band, preserving the preceding and following connection positions, succession direction, and edge order span of the cross-boundary path segment, while simultaneously shielding the business fact element content within the cross-boundary path segment, reducing the risk of invisible data entering the generation evidence. The conclusion generation and back-pasting unit further breaks down the enterprise's processing conclusions into conclusion fragments, and back-pasts them to closed evidence subgraphs or access breakpoint shadow bands according to object pointing, action pointing, and caliber pointing. It also blocks conclusion fragments that touch the access breakpoint shadow band or are not back-pasted, thereby improving the evidence traceability, access boundary consistency, and output reliability of the enterprise's SaaS data processing results.

[0005] A data processing method based on AI-powered enterprise management software (SaaS) according to an embodiment of the present invention includes the following steps: S1. Respond to business requests in enterprise management software (SaaS) and define the enterprise data domain that is consistent with the business request within the tenant's runtime space. S2. Perform fact element aggregation on data records scattered across different business functions within the enterprise data domain, and connect them into object fact chains according to the business object affiliation and business succession order between the data records. S3. Assemble the enterprise fact graph around the object fact chain, connect the business acceptance path within the enterprise fact graph, and connect the access boundary path according to the data visibility range within the tenant's operating space. S4. Rewrite the business request into the request landing point, and embed the request landing point into the corresponding business acceptance path and access boundary path in the enterprise fact graph to form a request anchoring fact graph. S5. Input the request anchoring fact graph into the improved GraphRAG model. The improved GraphRAG model includes evidence path extension unit, access breakpoint shadow band unit and conclusion generation reverse unit. The access breakpoint shadow band unit embeds a boundary shadow compression structure. The evidence path extension unit converges to form a closed evidence subgraph. The access breakpoint shadow band unit folds the evidence path that crosses the access boundary path into an access breakpoint shadow band. S6. The conclusion generation and back-pasting unit uses the closed evidence sub-graph as the generation basis and the access breakpoint shadow band as the shielding boundary to generate the enterprise processing conclusion. The enterprise processing conclusion is then split into conclusion fragments and back-pasted to the closed evidence sub-graph or the access breakpoint shadow band. S7. Block the reverse pasting to the access port shadow band or the conclusion fragment that has not been reverse pasted, and solidify the enterprise SaaS data processing results according to the retained conclusion fragment.

[0006] Optionally, S1 specifically includes: S11. Read the tenant ownership bit and business entry bit carried by the business request in the tenant's runtime space, and map the business entry bit to the business function boundary in the enterprise management software SaaS. S12. Locate the data partition that matches the tenant's affiliation along the business function boundary, and classify the business records that fall into the data partition into the tenant candidate domain. S13. Verify the connection between the business records in the tenant candidate domain and the business entry point, and delete business records that are outside the scope of the business entry point or cross the tenant's ownership point. S14. Merge and retain business records according to the direction of business entry points to form an enterprise data domain consistent with the business request's origin.

[0007] Optionally, S2 specifically includes: S21. Divide the record segments according to the business function to which the data records in the enterprise data domain belong, and gather the record segments that carry changes in business status, changes in object relationships, or results of business actions into business fact elements; S22. Perform object attribution verification on business fact elements and classify business fact elements belonging to the same business object into the same object fact slot. S23. Within the same object fact slot, arrange the business fact elements according to the generation order and business action succession relationship, and merge duplicate business fact elements that point to the same business action result into co-position fact elements. S24. Connect the preceding and following relationships between the corresponding fact elements and the adjacent business fact elements to form the object fact chain of the corresponding business object. S25. Write an object start marker at the beginning of the object fact chain, write an object end marker at the end of the object fact chain, and write an object continuation marker at the cross-object continuation position.

[0008] Optionally, S3 specifically includes: S31. According to the order of the object fact chain, map the business fact element to the fact node in the enterprise fact graph one by one, and configure the in-chain order for the fact node in the same object fact chain; S32. Connect the fact nodes of adjacent chains within the same object's fact chain to form connecting edges within the object; S33. At the position corresponding to the object continuation mark, connect the tail fact node of the previous object fact chain with the head fact node of the next object fact chain to form an inter-object continuation edge. S34. Arrange the intra-object connecting edges and inter-object connecting edges according to the connecting order of fact nodes to form the business connecting paths within the enterprise fact diagram; S35. According to the data visibility range within the tenant's operating space, check the fact nodes and receiving edges on the business receiving path, arrange the continuously visible fact nodes and receiving edges into visible path segments, and mark the connection position between the visible path segments and the invisible graph segments as the boundary break position. S36. Connect visible path segments and boundary disconnect points in the order of business acceptance paths to form access boundary paths within the enterprise fact diagram.

[0009] Optionally, S4 specifically includes: S41. Perform semantic segmentation on business requests to determine the business object, processing action, and scope of the business request; S42. Locate the fact node corresponding to the business object in the enterprise fact diagram, and extract the target path segment corresponding to the scope along the business acceptance path where the fact node is located. S43. Add a new request landing point in the target path segment and connect the request landing point to the fact node corresponding to the business object through the request association edge; S44. Write the processing action to the request landing point, and write the path association edge between the request landing point and the connecting edge in the target path segment. S45. Verify the access boundary path corresponding to the target path segment, and connect the request landing point to the visible path segment or boundary break point in the access boundary path to form a request anchoring fact graph.

[0010] Optionally, the convergence of the evidence path extension unit to form a closed evidence subgraph specifically includes: The evidence path extension unit starts from the request landing point in the request anchoring fact graph, locates the fact node that is directly connected to the request landing point along the request association edge, and extracts the business undertaking path where the fact node is located as the target path segment. Using the fact node directly connected to the request landing point as the starting point of the path, expand the adjacent fact nodes and receiving edges one by one according to the edge order in the target path segment. Merge the fact nodes and receiving edges located in the target path segment into the candidate evidence path, and write the receiving edge to be collapsed mark for the path that crosses the access boundary. For the connecting edges in the candidate evidence path, check the in-chain order of the starting fact node and the ending fact node, and classify the connecting edges with adjacent in-chain order into the in-chain evidence path, and classify the connecting edges on both sides of the object continuation mark into the cross-object evidence path. Perform original graph edge verification on the break points in the intra-chain evidence path and cross-object evidence path. When the fact nodes at both ends of the break point have original receiving edges in the enterprise fact graph and the original receiving edges do not cross the access boundary path, the original receiving edges are returned to the corresponding evidence path. Merge the in-chain evidence paths and cross-object evidence paths under the same request landing point according to the business acceptance direction, and delete the fact nodes and acceptance edges that are separated from the target path segment to obtain the candidate closed evidence subgraph. For each evidence path in the candidate closed evidence subgraph, a path closure score is calculated. The path closure score is obtained by accumulating the number of edges connected to the request landing point, the number of consecutive connecting edges in the chain, and the number of hits of the object continuation mark. The score is deducted according to the number of connecting edges of the path that crosses the access boundary. Evidence paths are truncated from high to low according to the path closure score, and the truncated evidence paths are merged into the closed evidence subgraph.

[0011] Optionally, the access breakpoint shadow band unit folds the evidence path that crosses the access boundary path into an access breakpoint shadow band, specifically including: The access break shadow zone unit is scanned edge by edge along the evidence path in the closed evidence subgraph to locate the position in the graph where the evidence path intersects with the access boundary path; The access break shadow zone unit is defined by the location in the diagram where the evidence path intersects with the access boundary path, and the portion of the evidence path located on both sides of the access boundary path that is still within the visible range is extracted. The boundary shadow compression structure cuts out the cross-boundary path segment from the evidence path with the folding mark, and extracts the visible fact node at the front end of the cross-boundary path segment, the visible fact node at the end of the cross-boundary path segment, and the direction of the cross-boundary path segment in the original evidence path. The boundary shadow compression structure counts the number of internal fact nodes and internal connecting edges of the cross-boundary path segment, calculates the edge order span of the cross-boundary path segment in the original evidence path, and pushes the number of internal fact nodes, the number of internal connecting edges, the edge order span, the position of the visible fact node at the front end, and the position of the visible fact node at the end into the access breakpoint shadow band. The boundary shadow compression structure backfills the access breakpoint shadow band to the original position of the cross-boundary path segment in the original evidence path, and writes content masking marks and reverse placeholder marks in the access breakpoint shadow band; The access break shadow band unit deletes the prompt context index corresponding to the access break shadow band according to the content masking mark, and retains the reverse pasting position of the conclusion fragment corresponding to the access break shadow band according to the reverse pasting placeholder mark.

[0012] Optionally, S6 specifically includes: S61. The conclusion generates the reverse unit, which arranges visible business fact elements along the evidence path in the closed evidence subgraph, and retains the corresponding position of the access breakpoint shadow band as a non-generable position. S62. Using visible business fact elements as generated content and non-generable positions as generated shielding boundaries, organize the prompt context and generate enterprise processing conclusions. S63. Segment the conclusion text according to the business judgment in the enterprise's processing conclusion, and extract the object pointer, action pointer and caliber pointer in each text segment. Keep the text segments with object pointer, action pointer or caliber pointer as conclusion segments. S64. Match the object pointer in the conclusion fragment with the fact node in the closed evidence subgraph using text identifier matching, match the action pointer with the action type of the connecting edge in the closed evidence subgraph using action type matching, match the statement pointer with the request landing point or target path segment using range matching, and obtain the object hit value, action hit value and statement hit value respectively. S65. Accumulate the object hit value, action hit value, and caliber hit value to obtain the back-paste matching value of the conclusion fragment; when any hit position falls into the corresponding position of the access break shadow belt, back-paste the conclusion fragment to the access break shadow belt; for conclusion fragments that do not fall into the access break shadow belt, when the back-paste matching value reaches the back-paste judgment condition and the hit position falls into the same evidence path, the conclusion fragment is back-pasteted to the closed evidence sub-graph. S66. Mark the conclusion fragments that do not match the closed evidence subgraph and do not match the access break shadow band as incomplete re-pasting, and encapsulate the conclusion fragment, re-pasting position and incomplete re-pasting mark into the conclusion fragment re-pasting result.

[0013] Optionally, S7 specifically includes: S71. Read the result of the reverse pasting of the conclusion fragment, write the conclusion fragment whose reverse pasting position falls into the shadow band of the access break point into the masked fragment set, and write the conclusion fragment with the incomplete reverse pasting mark into the abnormal fragment set. S72. Perform evidence path verification on conclusion fragments that are not written into the masked fragment set and abnormal fragment set, and retain conclusion fragments that are pasted back to the closed evidence subgraph and whose corresponding evidence paths are continuous. S73. Delete the conclusion fragments from the masked fragment set and the abnormal fragment set from the enterprise processing conclusions, and compress the arrangement interval of the conclusion fragments on both sides of the deletion position; S74. Bind the retained conclusion fragments to the corresponding closed evidence subgraph positions and solidify them into enterprise SaaS data processing results.

[0014] A data processing system based on AI enterprise management software (SaaS) according to an embodiment of the present invention includes the following modules: The Enterprise Data Domain Delineation Module is used to respond to business requests in enterprise management software (SaaS) and delineate the enterprise data domain that is consistent with the business request within the tenant's runtime space. The fact meta aggregation module is used to aggregate the fact meta records that are scattered across different business functions within the enterprise data domain, and to connect them into an object fact chain according to the business object affiliation and business succession order between the data records. The Enterprise Fact Graph Assembly Module is used to assemble an enterprise fact graph around the object fact chain, connect business acceptance paths within the enterprise fact graph, and connect access boundary paths according to the data visibility range within the tenant's runtime space. The request anchoring module is used to rewrite business requests into request landing points and embed the request landing points into the corresponding business acceptance path and access boundary path in the enterprise fact graph to form a request anchoring fact graph. An improved GraphRAG processing module is used to input the request anchored fact graph into the improved GraphRAG model. The improved GraphRAG model includes an evidence path expansion unit, an access breakpoint shadow band unit, and a conclusion generation reverse unit. The access breakpoint shadow band unit embeds a boundary shadow compression structure. The evidence path expansion unit converges to form a closed evidence subgraph, and the access breakpoint shadow band unit folds the evidence path that crosses the access boundary path into an access breakpoint shadow band. The conclusion generation and reposting module is used to generate enterprise processing conclusions by using the closed evidence sub-graph as the generation basis and the access breakpoint shadow band as the shielding boundary through the conclusion generation and reposting unit. The enterprise processing conclusions are then split into conclusion fragments and reposted to the closed evidence sub-graph or the access breakpoint shadow band. The results solidification module is used to block the shadow band of the access break or the incomplete reverse pasting of the conclusion fragments, and solidify the enterprise SaaS data processing results according to the retained conclusion fragments.

[0015] The beneficial effects of this invention are: 1. This invention, through the hierarchical construction of enterprise data domains, object fact chains, and enterprise fact graphs, organizes data records scattered across different business functions into an in-graph structure with a business sequence, enabling AI processing to no longer rely on isolated business records or fragmented retrieval pieces. The business sequence path preserves the sequential flow relationships between business objects, and the access boundary path synchronously defines the data visibility range within the tenant's operating space, thereby improving the integrity of evidence paths and in-graph traceability during enterprise SaaS data processing.

[0016] 2. This invention sets up access breakpoint shadow band units in the improved GraphRAG model and embeds a boundary shadow compression structure within these units. When an evidence path crosses an access boundary path, the cross-boundary path segment is not directly deleted but is clipped, compressed, and backfilled as an access breakpoint shadow band. The access breakpoint shadow band retains the preceding and following connection positions, continuation directions, and edge order spans of the cross-boundary path segment. Simultaneously, it deletes the corresponding context index using content masking markers, thereby maintaining the identifiability of the evidence chain breakpoint location while reducing the risk of invisible business fact elements entering the generation basis.

[0017] 3. This invention uses a conclusion generation and back-pasting unit to perform fragmented back-pasting on enterprise processing conclusions. It matches the object pointer, action pointer, and caliber pointer in the conclusion fragment to fact nodes, connecting edges, request landing points, or target path segments, respectively, and determines the in-graph landing point of the conclusion fragment based on the back-pasting matching value. Conclusion fragments back-pasted to access breakpoint shadow bands or those not yet back-pasted are masked. The retained conclusion fragments are bound and fixed to the position of the closed evidence subgraph, thereby reducing unfounded and unauthorized generation, and improving the evidence consistency, access boundary consistency, and output reliability of enterprise SaaS data processing results. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a data processing method based on AI-powered enterprise management software (SaaS) proposed in this invention; Figure 2 This is a schematic diagram of the improved GraphRAG model processing structure for a data processing method based on AI enterprise management software SaaS proposed in this invention; Figure 3 This is a structural diagram of a data processing system based on AI enterprise management software (SaaS) proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 A data processing method based on AI-powered enterprise management software (SaaS) includes the following steps: S1. Respond to business requests in enterprise management software (SaaS) and define the enterprise data domain that is consistent with the business request within the tenant's runtime space. S2. Perform fact element aggregation on data records scattered across different business functions within the enterprise data domain, and connect them into object fact chains according to the business object affiliation and business succession order between the data records. S3. Assemble the enterprise fact graph around the object fact chain, connect the business acceptance path within the enterprise fact graph, and connect the access boundary path according to the data visibility range within the tenant's operating space. S4. Rewrite the business request into the request landing point, and embed the request landing point into the corresponding business acceptance path and access boundary path in the enterprise fact graph to form a request anchoring fact graph. S5. Input the request anchoring fact graph into the improved GraphRAG model. The improved GraphRAG model includes evidence path extension unit, access breakpoint shadow band unit and conclusion generation reverse unit. The access breakpoint shadow band unit embeds a boundary shadow compression structure. The evidence path extension unit converges to form a closed evidence subgraph. The access breakpoint shadow band unit folds the evidence path that crosses the access boundary path into an access breakpoint shadow band. S6. The conclusion generation and back-pasting unit uses the closed evidence sub-graph as the generation basis and the access breakpoint shadow band as the shielding boundary to generate the enterprise processing conclusion. The enterprise processing conclusion is then split into conclusion fragments and back-pasted to the closed evidence sub-graph or the access breakpoint shadow band. S7. Block the reverse pasting to the access port shadow band or the conclusion fragment that has not been reverse pasted, and solidify the enterprise SaaS data processing results according to the retained conclusion fragment.

[0021] In this embodiment, S1 specifically includes: S11. Read the tenant ownership bit and business entry bit carried by the business request in the tenant's runtime space, and map the business entry bit to the business function boundary in the enterprise management software SaaS. S12. Locate the data partition that matches the tenant's affiliation along the business function boundary, and classify the business records that fall into the data partition into the tenant candidate domain. S13. Verify the connection between the business records in the tenant candidate domain and the business entry point, and delete business records that are outside the scope of the business entry point or cross the tenant's ownership point. S14. Merge and retain business records according to the direction of business entry points to form an enterprise data domain consistent with the business request's origin.

[0022] In the specific implementation process, after receiving a business request, the enterprise management software (SaaS) reads the tenant ownership bit and business entry bit from the corresponding session token, interface call context, or business request fields. The tenant ownership bit can be the tenant number, enterprise space number, or SaaS instance number, and the business entry bit can be a business menu entry, interface address, process entry, or page function code. The business entry bit corresponds to the business function boundaries in the enterprise management software (SaaS). For example, the sales payment entry corresponds to the business function boundaries formed by sales orders, invoices, payments, and customer transaction records; the procurement approval entry corresponds to the business function boundaries formed by procurement applications, approval workflows, purchase orders, supplier files, supplier arrival, arrival acceptance, warehousing records, procurement reconciliation records, and accounts payable records; the production cost accounting entry corresponds to the business function boundaries formed by production orders, production work orders, material requisition, process reporting, equipment hours, labor hours, manufacturing overhead collection, completion warehousing, and cost transfer records; and the financial accounting entry corresponds to the business function boundaries formed by accounts receivable, accounts payable, expense reimbursement, accounting vouchers, payment and receipt records, cost transfer records, and financial period closing records. By using the joint location of the tenant's home location and the business entry point, the data partition and business processing scope corresponding to the business request within the tenant's operating space are first determined.

[0023] When locating data partitions along business function boundaries, the data tables, business documents, and process records associated with the business function boundaries are first read, and then the tenant affiliation bits carried by these records are verified. Business records whose tenant affiliation bits match the business request are included in the tenant candidate domain, while business records with inconsistent tenant affiliation bits are directly excluded. The tenant candidate domain only covers the range of tenant data that the current business request can reach; the data visibility range within a tenant is reserved until the enterprise fact diagram assembly stage to access the boundary path.

[0024] After the tenant candidate domain is formed, the connection between the business records within the candidate domain and the business entry points is verified. During the verification, the document reference relationships, process flow relationships, and business action relationships between records are checked according to the business flow corresponding to the business entry point. For example, under the sales payment entry point, contract records, order records, invoice records, and payment records are retained in the order of contract to order, order to invoice, and invoice to payment; after-sales records, administrative approval records, and historical records without reference relationships that are related to the same customer but are out of the current payment chain are deleted from the tenant candidate domain. For business records that cross the tenant's ownership position, even if the record name, customer name, or business object name is the same as the current business request, they are also deleted from the tenant candidate domain. For production cost accounting entry points, production orders, production work orders, material requisition records, process reporting records, equipment time records, labor time records, manufacturing overhead collection records, completion and warehousing records, and cost transfer records are retained in the following order: from production order to production execution, from production execution to resource consumption, from resource consumption to expense collection, from expense collection to completion and warehousing, and from completion and warehousing to cost transfer. Historical material requisition records that are detached from the current production order, product batch, or cost object, expense records without corresponding work orders, and transfer records spanning cost periods are deleted from the tenant candidate domain.

[0025] After verifying the business relationship, the retained business records are merged according to the business direction corresponding to the business entry point. The business direction can be determined by business process configuration, record generation order, document reference fields, or business action flow relationship. Taking the sales processing chain as an example, retained records can be arranged according to the direction of contract, order, shipment, invoicing, and payment collection; taking the procurement processing chain as an example, retained records can be arranged according to the direction of purchase request, approval record, purchase order, and inventory receipt record. The arranged business records form an enterprise data domain consistent with the business request's origin.

[0026] In this embodiment, S2 specifically includes: S21. Divide the record segments according to the business function to which the data records in the enterprise data domain belong, and gather the record segments that carry changes in business status, changes in object relationships, or results of business actions into business fact elements; S22. Perform object attribution verification on business fact elements and classify business fact elements belonging to the same business object into the same object fact slot. S23. Within the same object fact slot, arrange the business fact elements according to the generation order and business action succession relationship, and merge duplicate business fact elements that point to the same business action result into co-position fact elements. S24. Connect the preceding and following relationships between the corresponding fact elements and the adjacent business fact elements to form the object fact chain of the corresponding business object. S25. Write an object start marker at the beginning of the object fact chain, write an object end marker at the end of the object fact chain, and write an object continuation marker at the cross-object continuation position.

[0027] In the specific implementation process, after the enterprise data domain is formed, the record fragments are first divided according to the business function to which the data record belongs. During the division, the content of business status changes, object relationship changes, and business action results in the data record are read, and the corresponding record fragments are grouped into business fact elements; page display records, cached logs, and synchronization traces without business continuity are not included in business fact elements. When grouping production cost accounting data records into fact elements, production start-up, material requisition, process completion, working hour confirmation, manufacturing cost collection, completion and warehousing, cost transfer, and cost difference confirmation are respectively organized into business fact elements; cost summary views used only for page display, temporary cached records, and statistical results not included in production orders or cost objects are not included in business fact elements.

[0028] After a business fact element is formed, an object attribution check is performed on it. During the check, the object number, document reference relationship, process instance relationship, or upstream and downstream reference relationship in the business fact element are read, and business fact elements belonging to the same business object are assigned to the same object fact slot. When a business fact element is associated with multiple business objects, it is assigned to the corresponding object fact slot according to the primary reference relationship of the business fact element in the data record.

[0029] Within the same object fact slot, business fact elements are arranged according to their generation order and business action succession relationship. The generation order is determined by the record generation time, process flow order, document version order, or business action number; the business action succession relationship is determined by document reference relationship and process flow relationship. When multiple business fact elements point to the same business action result, the master fact element is retained, and duplicate business fact elements are merged into the same position as the master fact element to form corresponding fact elements.

[0030] Once the corresponding fact elements are determined, the sequential relationships between these fact elements and adjacent business fact elements are linked to form an object fact chain for the corresponding business object. During the linking process, adjacent business fact elements are arranged according to the chronological order of business actions; sequential relationships are established between adjacent business fact elements with clear upstream and downstream references; positions lacking intermediate records are left disconnected, and business fact elements with unclear adjacent relationships are not forcibly linked.

[0031] After an object fact chain is formed, an object start marker is written at the beginning of the chain, and an object end marker is written at the end. When there is a document reference, workflow transition, or business action succession relationship between two object fact chains, an object succession marker is written at the corresponding connection position. The object start marker, object end marker, and object succession marker are retained with the object fact chain.

[0032] In this embodiment, S3 specifically includes: S31. According to the order of the object fact chain, map the business fact element to the fact node in the enterprise fact graph one by one, and configure the in-chain order for the fact node in the same object fact chain; S32. Connect the fact nodes of adjacent chains within the same object's fact chain to form connecting edges within the object; S33. At the position corresponding to the object continuation mark, connect the tail fact node of the previous object fact chain with the head fact node of the next object fact chain to form an inter-object continuation edge. S34. Arrange the intra-object connecting edges and inter-object connecting edges according to the connecting order of fact nodes to form the business connecting paths within the enterprise fact diagram; S35. According to the data visibility range within the tenant's operating space, check the fact nodes and receiving edges on the business receiving path, arrange the continuously visible fact nodes and receiving edges into visible path segments, and mark the connection position between the visible path segments and the invisible graph segments as the boundary break position. S36. Connect visible path segments and boundary disconnect points in the order of business acceptance paths to form access boundary paths within the enterprise fact diagram.

[0033] In the specific implementation process, after the object fact chain is formed, the business fact elements are first expanded according to the arrangement order of the object fact chain. Each business fact element is mapped to a fact node in the enterprise fact diagram. The fact node retains the object fact chain number and the order within the chain corresponding to the business fact element. The order within the chain is arranged according to the sequential order of the business fact elements in the object fact chain. For example, the first business fact element in the same object fact chain is configured as the order 1 within the chain, and the next business fact element is configured as the order 2 within the chain, so that business changes within the same object can maintain the original order in the enterprise fact diagram.

[0034] After completing the fact node mapping, edges are connected between adjacent fact nodes within the same object's fact chain. The connection starts with the fact node in the preceding chain and ends with the fact node in the following chain, forming an intra-object connection edge. These intra-object connection edges are sequentially connected along the object's fact chain, recording state changes, action flows, or connection relationships within the same business object. For positions in the object's fact chain that remain disconnected, connection edges are not forcibly added; only existing fact nodes and existing connection relationships are retained.

[0035] At the position corresponding to the object continuation marker, read the tail fact node of the previous object fact chain and the head fact node of the next object fact chain, and establish an inter-object connection edge between the two fact nodes. Inter-object connection edges are used to connect different object fact chains with a clear business continuation relationship. For example, the tail fact node of the contract fact chain connects to the head fact node of the order fact chain; the tail fact node of the order fact chain connects to the head fact node of the invoicing fact chain; the material requisition fact node of the production work order fact chain connects to the inventory outbound fact node; the completed and put into storage fact node connects to the inventory inbound fact node; and the cost aggregation fact node connects to the financial cost transfer fact node. Inter-object connection edges are only established at positions where the object continuation marker has already been written to avoid incorrectly connecting object fact chains without a business continuation relationship.

[0036] After the intra-object and inter-object connection edges are established, all connection edges are arranged according to the connection order of the fact nodes to form the business connection path within the enterprise fact graph. The business connection path consists of consecutive fact nodes and connection edges, and their arrangement order is consistent with the internal order of the object fact chain and the cross-object connection order. For multiple object fact chains that may be involved in the same business request, the business connection path is arranged according to the connection direction corresponding to the business entry point.

[0037] Next, the fact nodes and connecting edges on the business acceptance path are checked according to the data visibility range within the tenant's operating space. The data visibility range is matched item by item with the business records corresponding to the fact nodes and the business relationships corresponding to the connecting edges; fact nodes and connecting edges that fall continuously within the data visibility range are arranged as visible path segments. The connection position between the visible path segment and the invisible segment in the graph is marked as the boundary break point. The boundary break point is retained in the original order of the business acceptance path, and adjacent visible fact nodes and visible connecting edges are not deleted.

[0038] Finally, visible path segments and boundary breakpoints are concatenated according to the order of the business acceptance paths to form the access boundary paths within the enterprise fact diagram. The access boundary paths and business acceptance paths maintain the same diagram arrangement benchmark: business acceptance paths record the acceptance relationships between business facts, while access boundary paths record visible segments and breakpoints on the same acceptance relationship.

[0039] In this embodiment, S4 specifically includes: S41. Perform semantic segmentation on business requests to determine the business object, processing action, and scope of the business request; S42. Locate the fact node corresponding to the business object in the enterprise fact diagram, and extract the target path segment corresponding to the scope along the business acceptance path where the fact node is located. S43. Add a new request landing point in the target path segment and connect the request landing point to the fact node corresponding to the business object through the request association edge; S44. Write the processing action to the request landing point, and write the path association edge between the request landing point and the connecting edge in the target path segment. S45. Verify the access boundary path corresponding to the target path segment, and connect the request landing point to the visible path segment or boundary break point in the access boundary path to form a request anchoring fact graph.

[0040] In practice, after a business request enters the request anchoring process, semantic segmentation is first performed on the request. Business requests can originate from natural language queries, page query conditions, API call parameters, or process triggering instructions. During segmentation, three parts are extracted: business object, processing action, and scope. The business object corresponds to the contract, order, customer, project, inventory batch, approval process, etc., that the request points to; the processing action corresponds to actions such as querying, summarizing, analyzing, verifying, predicting, and generating suggestions; the scope corresponds to the time range, organizational range, business status range, indicator range, or process stage range. If multiple business objects appear simultaneously in a business request, the primary business object is determined according to the business entry point and request subject, while other business objects are retained as associated objects.

[0041] After identifying the business object, locate the corresponding fact node in the enterprise fact diagram. During the search, first verify the object fact chain to which the fact node belongs, then compare the business object number, object name, document reference relationship, or process instance relationship retained in the fact node. Successfully matched fact nodes are designated as target fact nodes. Once the target fact node is determined, extract the target path segment corresponding to the scope along the business connection path where the target fact node is located, both forward and backward. For example, if the scope is limited to the current quarter, only extract the fact nodes and connection edges in the business connection path whose occurrence time falls within the current quarter; if the scope is limited to the stage of invoiced but not yet paid for, only extract the path segment between the invoicing fact node and the payment fact node; if the scope is limited to the production cost of a certain product batch, only extract the path segment between production orders, material requisition, process reporting, expense collection, completion and warehousing, and cost transfer.

[0042] Once the target path segment is determined, a new request landing point is added to the Enterprise Fact Graph. This request landing point serves as the anchor point for the business request within the Enterprise Fact Graph, and includes the business request number, the business object it points to, the processing action, and the scope of the request. The request landing point is connected to the target fact node corresponding to the business object via request association edges. When a business request involves related objects, the request landing point is connected to both the main business object fact node and the related object fact nodes, preserving the priority of the main business object fact node. Through request association edges, the business request transitions from external request text or interface parameters into the Enterprise Fact Graph.

[0043] Subsequently, the processing action is written to the request endpoint, and a path association edge is written between the request endpoint and the connecting edge in the target path segment. The path association edge connects the request endpoint to a specific connecting position in the target path segment. For query-type business requests, the path association edge connects to the connecting edge in the target path segment directly related to the query object; for analysis-type business requests, the path association edge connects to the continuous connecting edges in the target path segment covering changes in business status; for verification-type business requests, the path association edge connects to the connecting edges in the target path segment involving comparisons between before and after states. After processing with the path association edge, the request endpoint is simultaneously connected to the fact node and the business connecting path.

[0044] Finally, verify the access boundary path corresponding to the target path segment. If the target path segment enters a continuous visible path segment, the request landing point is connected to the corresponding visible path segment; if the target path segment passes through a boundary break point, the request landing point is connected to the boundary break point, and the connection relationship between the request landing point and the boundary break point is preserved. After the above processing, a request anchoring fact graph is formed in the enterprise fact graph, which includes the connection relationship between the request landing point, request associated edges, path associated edges, target path segment, and access boundary path.

[0045] In this embodiment, the process of converging the evidence path extension unit to form a closed evidence subgraph specifically includes: The evidence path extension unit starts from the request landing point in the request anchoring fact graph, locates the fact node that is directly connected to the request landing point along the request association edge, and extracts the business undertaking path where the fact node is located as the target path segment. Using the fact node directly connected to the request landing point as the starting point of the path, expand the adjacent fact nodes and receiving edges one by one according to the edge order in the target path segment. Merge the fact nodes and receiving edges located in the target path segment into the candidate evidence path, and write the receiving edge to be collapsed mark for the path that crosses the access boundary. For the connecting edges in the candidate evidence path, check the in-chain order of the starting fact node and the ending fact node, and classify the connecting edges with adjacent in-chain order into the in-chain evidence path, and classify the connecting edges on both sides of the object continuation mark into the cross-object evidence path. Perform original graph edge verification on the break points in the intra-chain evidence path and cross-object evidence path. When the fact nodes at both ends of the break point have original receiving edges in the enterprise fact graph and the original receiving edges do not cross the access boundary path, the original receiving edges are returned to the corresponding evidence path. Merge the in-chain evidence paths and cross-object evidence paths under the same request landing point according to the business acceptance direction, and delete the fact nodes and acceptance edges that are separated from the target path segment to obtain the candidate closed evidence subgraph. For each evidence path in the candidate closed evidence subgraph, a path closure score is calculated. The path closure score is obtained by summing the number of edges connected to the request landing point, the number of consecutive connecting edges in the chain, and the number of hits of the object continuation mark. The score is deducted according to the number of connecting edges of the path that crosses the access boundary. Evidence paths are truncated from high to low according to the path closure score, and the truncated evidence paths are merged into the closed evidence subgraph.

[0046] In the specific implementation process, after the request anchored fact graph enters the improved GraphRAG model, the evidence path expansion unit first locates the starting point of the evidence path from the request landing point. The request landing point is already connected to the fact node corresponding to the business object through the request association edge. Therefore, the evidence path expansion unit finds the fact node directly connected to the request landing point along the request association edge and reads the business acceptance path where the fact node is located. The part of the business acceptance path that falls within the scope of the application is truncated as the target path segment. For example, when the business request is limited to a certain time interval, a certain business stage, or a certain processing scope, only the fact nodes and acceptance edges within that scope are retained.

[0047] Once the target path segment is determined, the evidence path expansion unit uses the fact node directly connected to the request landing point as the path starting point and reads adjacent fact nodes and connecting edges one by one according to the edge order in the target path segment. For each connecting edge read, it checks whether the connecting edge is still within the target path segment. Fact nodes and connecting edges located within the target path segment are merged into the candidate evidence path; when a connecting edge crosses an access boundary path, the connecting edge remains in its original position on the candidate evidence path, but a "to be collapsed" mark is written on the connecting edge. The "to be collapsed" mark does not carry invisible business content; it only records that the connecting edge needs to be processed by the access breakpoint shadow band unit.

[0048] After candidate evidence paths are formed, the evidence path expansion unit performs a chain-based verification of the connecting edges within the candidate evidence paths. For connecting edges within the same object's fact chain, the evidence path expansion unit reads the chain-based order of the fact nodes at both ends of the connecting edge; connecting edges with adjacent chain-based order are assigned to the intra-chain evidence path. For connecting edges located on both sides of the object continuation marker, the evidence path expansion unit checks whether the connecting edge connects the tail fact node of the previous object's fact chain and the head fact node of the next object's fact chain; connecting edges that pass the check are assigned to the cross-object evidence path.

[0049] When there are breaks in the evidence path within the chain or across objects, the evidence path extension unit returns to the enterprise fact graph to verify the original connecting edge. If the fact nodes at both ends of the break have original connecting edges in the enterprise fact graph, and the original connecting edges do not cross the access boundary path, then the original connecting edges are returned to the corresponding evidence path; if there are no original connecting edges in the enterprise fact graph, or if the original connecting edges cross the access boundary path, then the break is retained, and no new connecting edge is added.

[0050] After completing the chain-level verification, cross-object verification, and original graph edge verification, the evidence path expansion unit merges the intra-chain evidence paths and cross-object evidence paths under the same request landing point according to the business direction. During the merging, fact nodes and connecting edges that are detached from the target path segment are deleted, and only intra-graph segments with a direct or continuous connecting relationship to the request landing point are retained. The retained intra-graph segments form a candidate closed evidence subgraph. The fact nodes in the closed evidence subgraph come from the target path segment, the connecting edges come from the original connecting relationships in the enterprise fact graph, and the cross-object positions are supported by object continuation markers. Therefore, the closed evidence subgraph can serve as the intra-graph evidence range for the enterprise's processing conclusion.

[0051] To reduce the number of low-relevance paths entering the closed evidence subgraph, the evidence path expansion unit can also calculate a path closure score. During calculation, the number of request landing point edges, the number of consecutive receiving edges within the chain, and the number of object continuation marker hits in each candidate evidence path are first counted, and these three quantities are summed. Then, the number of receiving edges with pending collapse markers in the candidate evidence path is counted, and this number is subtracted from the summed result. After obtaining the path closure score, the evidence path expansion unit arranges the candidate evidence paths from highest to lowest score, and incorporates evidence paths whose scores meet the truncation criteria into the closed evidence subgraph. The truncation criteria can be determined based on the average path closure score of candidate evidence paths under the same request landing point; candidate evidence paths with a path closure score not lower than the average score are incorporated into the closed evidence subgraph. When the number of candidate evidence paths exceeds a set number, the top N paths are truncated from highest to lowest path closure score, where N can be 3 to 10.

[0052] In this embodiment, the access fracture shadow band unit folds the evidence path that crosses the access boundary path into an access fracture shadow band, specifically including: The access break shadow zone unit is scanned edge by edge along the evidence path in the closed evidence subgraph to locate the position in the graph where the evidence path intersects with the access boundary path; The access break shadow zone unit is defined by the location in the diagram where the evidence path intersects with the access boundary path, and the portion of the evidence path located on both sides of the access boundary path that is still within the visible range is extracted. The boundary shadow compression structure cuts out the cross-boundary path segment from the evidence path with the folding mark, and extracts the visible fact node at the front end of the cross-boundary path segment, the visible fact node at the end of the cross-boundary path segment, and the direction of the cross-boundary path segment in the original evidence path. The boundary shadow compression structure counts the number of internal fact nodes and internal connecting edges of the cross-boundary path segment, calculates the edge order span of the cross-boundary path segment in the original evidence path, and pushes the number of internal fact nodes, the number of internal connecting edges, the edge order span, the position of the visible fact node at the front end, and the position of the visible fact node at the end into the access breakpoint shadow band. The boundary shadow compression structure backfills the access breakpoint shadow band to the original position of the cross-boundary path segment in the original evidence path, and writes content masking marks and reverse placeholder marks in the access breakpoint shadow band; The access break shadow band unit deletes the prompt context index corresponding to the access break shadow band according to the content masking mark, and retains the reverse pasting position of the conclusion fragment corresponding to the access break shadow band according to the reverse pasting placeholder mark.

[0053] In the specific implementation process, after receiving the closed evidence subgraph, the access breakpoint shadow strip unit first scans the evidence path in the closed evidence subgraph edge by edge. The scanning objects are the receiving edges and the fact nodes at both ends of the receiving edges in the evidence path, and the scanning order is executed according to the edge sequence in the evidence path. When a receiving edge already has a folding mark, or when the fact nodes at both ends of the receiving edge fall on opposite sides of the access boundary path, the access breakpoint shadow strip unit registers the position of the receiving edge as the in-graph position where the evidence path and the access boundary path intersect. This in-graph position retains the edge sequence in the original evidence path to avoid the cross-boundary position being lost in the evidence path.

[0054] The access boundary shadow zone unit is defined by the point of intersection within the graph. It searches forward for fact nodes and connecting edges still within the visible range, extracting the evidence path portion before the intersection point. Simultaneously, it searches backward for fact nodes and connecting edges that return to the visible range, extracting the evidence path portion after the intersection point. Both the forward and backward evidence path portions are retained in the closed evidence subgraph, while the portion crossing the access boundary is not directly written into the prompt context. Through this processing, the closed evidence subgraph retains the visible path segments, while the portion crossing the boundary is processed using the boundary shadow compression structure.

[0055] The boundary shadow compression structure extracts cross-boundary path segments from the evidence path marked with folding indicators. During extraction, the last visible fact node before the intersection is taken as the front visible fact node of the cross-boundary path segment, and the first visible fact node after the intersection is taken as the end visible fact node. The continuation direction of the cross-boundary path segment in the original evidence path is then read. The continuation direction can be determined according to the increasing edge order, the object fact chain arrangement direction, or the object continuation mark direction. The fact node content and continuation edge content within the cross-boundary path segment are not included in the reading process; only the number of internal fact nodes, the number of internal continuation edges, and the edge order span are retained. The number of internal fact nodes and the number of internal continuation edges are obtained by statistically analyzing the node number sequence and continuation edge number sequence in the cross-boundary path segment. The statistical process only reads the number of numbers and the edge order, without reading the business field content, business description text, and continuation edge content within the fact nodes.

[0056] Subsequently, the boundary shadow compression structure performs compression processing on the cross-boundary path segments. The compression process first counts the number of fact nodes within the cross-boundary path segment, then counts the number of internal connecting edges, and calculates the edge order span between the front-end visible fact node and the end-end visible fact node. The positions of the front-end and end-end visible fact nodes, the connecting direction, the number of internal fact nodes, the number of internal connecting edges, and the edge order span are compressed into the access breakpoint shadow band. The access breakpoint shadow band thus retains the graph structure outline of the cross-boundary path segment but does not carry the internal business fact element content of the cross-boundary path segment.

[0057] After the access breakpoint shadow band is formed, the boundary shadow compression structure backfills the access breakpoint shadow band to its original position in the original evidence path for the cross-boundary path segment. During backfilling, the access breakpoint shadow band occupies the corresponding edge sequence interval in the original evidence path, with its front end aligned with the visible fact nodes before the intersection position and its end aligned with the visible fact nodes after the intersection position. The boundary shadow compression structure writes content masking markers and reverse placeholder markers into the access breakpoint shadow band. The content masking markers correspond to the business fact element content within the cross-boundary path segment; the reverse placeholder markers correspond to the breakpoint positions that the conclusion fragment may touch.

[0058] After backfilling is completed, the access breakpoint shadow band unit deletes the corresponding prompt context index according to the content masking mark. After the prompt context index is deleted, the enterprise processing conclusion generation stage can only read the visible fact nodes, the direction of connection, and the compressed path structure information at both ends of the access breakpoint shadow band, and cannot read the business fact element content inside the cross-boundary path segment. The access breakpoint shadow band unit also retains the back-paste position of the conclusion fragment corresponding to the access breakpoint shadow band according to the back-paste placeholder mark. When the conclusion fragment hits the back-paste position during the back-paste process, the conclusion fragment is marked as falling into the access breakpoint shadow band.

[0059] In this embodiment, S6 specifically includes: S61. The conclusion generates the reverse unit, which arranges visible business fact elements along the evidence path in the closed evidence subgraph, and retains the corresponding position of the access breakpoint shadow band as a non-generable position. S62. Using visible business fact elements as generated content and non-generable positions as generated shielding boundaries, organize the prompt context and generate enterprise processing conclusions. S63. Segment the conclusion text according to the business judgment in the enterprise's processing conclusion, and extract the object pointer, action pointer and caliber pointer in each text segment. Keep the text segments with object pointer, action pointer or caliber pointer as conclusion segments. S64. Match the object pointer in the conclusion fragment with the fact node in the closed evidence subgraph using text identifier matching, match the action pointer with the action type of the connecting edge in the closed evidence subgraph using action type matching, match the statement pointer with the request landing point or target path segment using range matching, and obtain the object hit value, action hit value and statement hit value respectively. S65. Accumulate the object hit value, action hit value, and caliber hit value to obtain the back-paste matching value of the conclusion fragment; when any hit position falls into the corresponding position of the access break shadow belt, back-paste the conclusion fragment to the access break shadow belt; for conclusion fragments that do not fall into the access break shadow belt, when the back-paste matching value reaches the back-paste judgment condition and the hit position falls into the same evidence path, the conclusion fragment is back-pasteted to the closed evidence sub-graph. S66. Mark the conclusion fragments that do not match the closed evidence subgraph and do not match the access break shadow band as incomplete re-pasting, and encapsulate the conclusion fragment, re-pasting position and incomplete re-pasting mark into the conclusion fragment re-pasting result.

[0060] In the specific implementation process, after receiving the closed evidence subgraph and the access breakpoint shadow band, the conclusion generation and reposting unit first arranges the visible business fact elements along the evidence path in the closed evidence subgraph. During arrangement, the fact nodes and connecting edges are read according to the edge order in the evidence path, and business fact elements without content masking markers are written into the visible fact sequence; the path position corresponding to the access breakpoint shadow band is reserved as an empty position and marked as an ungenerateable position. The ungenerateable position only retains the front visible fact node, the end visible fact node, the connecting direction, the edge order span, and the reposting placeholder marker, and does not write the business fact element content inside the cross-boundary path segment.

[0061] When organizing the prompt context, the conclusion generation and reposting unit uses business fact elements from the visible fact sequence as the generated content, and writes non-generable locations as masked boundaries. The prompt context is arranged in the order of request landing point, target path segment, visible fact sequence, and masked boundaries, ensuring that the generated content of the enterprise processing conclusion can only come from visible business fact elements in the closed evidence subgraph. For the location corresponding to the access breakpoint shadow band, the prompt context only retains the structural prompt "an access breakpoint exists at this location," without carrying the business content inside the breakpoint. The conclusion generation and reposting unit generates the enterprise processing conclusion accordingly.

[0062] After the enterprise processing conclusion is generated, the conclusion generation and reposting unit splits the conclusion text according to the business judgment. During splitting, periods, semicolons, line breaks, business judgment connectors, and indicator judgment phrases serve as sentence boundaries, extracting each text segment that expresses an independent business judgment. For each text segment, object pointers, action pointers, and scope pointers are extracted. Object pointers correspond to business object names, object numbers, document names, or process objects; action pointers correspond to business actions or status changes such as query, exception, delay, increase, decrease, pass, rejection, and incomplete; scope pointers correspond to the processing scope recorded in the time range, organizational scope, indicator range, business stage, or request endpoint. Text segments with object pointers, action pointers, or scope pointers are retained as conclusion segments; connecting texts without business judgment pointers are not included in the conclusion segment reposting process.

[0063] After the conclusion fragment is formed, the conclusion generation and back-posting unit performs three types of matching. First, the object pointer is matched with the fact nodes in the closed evidence subgraph using text identifier matching. The matching objects include the business object name, object number, and document identifier in the fact node. Second, the action pointer is matched with the receiving edge type in the closed evidence subgraph using action type matching. The matching objects include state transition type, document reference type, and business action type. Third, the scope pointer is matched with the request landing point or target path segment using range matching. The matching objects include the processing scope in the request landing point, the start and end edge sequence of the target path segment, and the time or stage range covered by the target path segment. A successful match is recorded as a valid hit, yielding the object hit value, action hit value, and scope hit value, respectively. Unsuccessful matches are recorded as invalid hits. The object hit value, action hit value, and scope hit value can be either 0 or 1. A value of 1 is used when the corresponding pointer in the conclusion fragment is completely consistent with the object in the graph or remains consistent after synonym normalization; otherwise, a value of 0 is used. The anti-patch matching value is the sum of the three types of hit values. The anti-patch judgment condition can be set to an anti-patch matching value of not less than 2 and an object hit value of 1.

[0064] The conclusion generation and back-paste unit accumulates the object hit value, action hit value, and caliber hit value to obtain the back-paste matching value of the conclusion fragment. The back-paste determination condition can be set to a valid object hit value and at least one of the action hit value or caliber hit value, or it can be set to a cumulative total of two valid hits across the three types of hit values. When the back-paste matching value meets the back-paste determination condition, and the hit positions fall within the same evidence path, the conclusion fragment is back-pasteped to the corresponding fact node, connecting edge, or target path segment position in the closed evidence subgraph. If any hit position of the object pointer, action pointer, or caliber pointer falls within the corresponding position of the access breakpoint shadow zone, the conclusion fragment is back-pasteped to the access breakpoint shadow zone.

[0065] For conclusion fragments that do not match a closed evidence subgraph and do not match an access breakpoint shadow band, the conclusion generation and re-attaching unit writes an incomplete re-attaching mark. The conclusion generation and re-attaching unit encapsulates the conclusion fragment, re-attaching position, re-attaching match value, hit type, and incomplete re-attaching mark into a conclusion fragment re-attaching result. The conclusion fragment re-attaching result enters the result solidification stage, where the result solidification stage performs masking processing on the conclusion fragments re-attached to the access breakpoint shadow band and the incomplete re-attached conclusion fragments, while retaining the conclusion fragments re-attached to a closed evidence subgraph with a continuous evidence path.

[0066] Typical GraphRAG models usually recall in-graph content along entity relationships, semantically similar nodes, or community summaries, and organize the recalled nodes, relationship texts, and summary content into a generation context. For multi-tenant business data in enterprise management software (SaaS), typical GraphRAG is prone to two types of processing defects when encountering access boundaries: if invisible in-graph segments are directly deleted, the business connection path will break at the boundary, and the generation model can only see incomplete evidence; if cross-boundary path segments are retained, the business fact meta-content within the cross-boundary path segments may enter the prompt context, creating the risk of access exceeding the boundary. Simultaneously, the results generated by typical GraphRAG usually only establish a reference relationship with the overall context, making it difficult to determine whether each business conclusion falls within visible evidence, access breakpoints, or areas without evidence.

[0067] The improved GraphRAG model in this implementation first uses the evidence path expansion unit to converge the fact nodes and connecting edges around the request landing point into a closed evidence subgraph, so that the retrieval result is no longer a discrete set of nodes, but a set of evidence paths with a business connection order. When the evidence path crosses the access boundary path, the access breakpoint shadow band unit calls the boundary shadow compression structure to perform clipping, statistics, compression, and backfilling processing on the cross-boundary path segment: the front visible fact nodes, the back visible fact nodes, the connection direction, the number of internal fact nodes, the number of internal connecting edges, and the edge order span of the cross-boundary path segment are compressed into the access breakpoint shadow band, and the prompt context index corresponding to the business fact element content inside the cross-boundary path segment is deleted. Thus, the access breakpoint shadow band retains the structural position and connection relationship of the cross-boundary path segment in the original evidence path, but does not carry invisible business content.

[0068] During the conclusion generation phase, the conclusion generation back-paste unit uses visible business fact elements in the closed evidence subgraph as the generated content and the access breakpoint shadow band as the shielding boundary to generate the enterprise processing conclusion. After the enterprise processing conclusion is generated, the conclusion generation back-paste unit further extracts the object pointer, action pointer, and caliber pointer in the conclusion fragment and matches them with fact nodes, connecting edges, request landing points, or target path segments respectively. Conclusion fragments that can fall into the same evidence path are back-pastemed to the closed evidence subgraph, conclusion fragments that touch the access breakpoint shadow band are back-pastemed to the access breakpoint shadow band, and conclusion fragments that cannot fall into the position in the graph are marked as incomplete back-paste.

[0069] Through the above modifications, the improved GraphRAG model forms a dual structure of "closed evidence subgraph + access breakpoint shadow zone" in the graph retrieval stage and a constraint path of "visible fact generation + conclusion fragment back-pasting" in the generation stage. Compared with ordinary GraphRAG, this model transforms access boundaries into a localizable shadow structure within the graph and decomposes the generated results and back-pasting them to the evidence positions within the graph. This preserves the structural continuity of cross-boundary evidence paths, avoids invisible business fact elements from entering the generation basis, and can identify conclusions without evidence and conclusion fragments that touch access breaks, thereby improving the evidence traceability, access boundary consistency, and output reliability of enterprise SaaS data processing results.

[0070] In this embodiment, S7 specifically includes: S71. Read the result of the reverse pasting of the conclusion fragment, write the conclusion fragment whose reverse pasting position falls into the shadow band of the access break point into the masked fragment set, and write the conclusion fragment with the incomplete reverse pasting mark into the abnormal fragment set. S72. Perform evidence path verification on conclusion fragments that are not written into the masked fragment set and abnormal fragment set, and retain conclusion fragments that are pasted back to the closed evidence subgraph and whose corresponding evidence paths are continuous. S73. Delete the conclusion fragments from the masked fragment set and the abnormal fragment set from the enterprise processing conclusions, and compress the arrangement interval of the conclusion fragments on both sides of the deletion position; S74. Bind the retained conclusion fragments to the corresponding closed evidence subgraph positions and solidify them into enterprise SaaS data processing results.

[0071] In the specific implementation process, the result consolidation step first reads the conclusion fragment back-painting results. The conclusion fragment back-painting results include the conclusion fragment, back-painting position, back-painting matching value, and an incomplete back-painting mark. Each conclusion fragment is first categorized according to its back-painting position: conclusion fragments whose back-painting positions fall within the access breakpoint shadow zone are written into the shielded fragment set; conclusion fragments with an incomplete back-painting mark are written into the abnormal fragment set. The shielded fragment set corresponds to conclusions that touch the access boundary breakpoint, while the abnormal fragment set corresponds to conclusions that do not fall into the closed evidence subgraph or the access breakpoint shadow zone.

[0072] After classification, evidence path verification is performed on conclusion fragments not written into the masked fragment set and the abnormal fragment set. During verification, the corresponding reversed position of the conclusion fragment is read first, and then the closed evidence subgraph containing the reversed position is searched. If the object pointer, action pointer, and caliber pointer corresponding to the conclusion fragment all fall within the same evidence path, and the fact nodes and connecting edges in the evidence path are arranged continuously, then the conclusion fragment is retained. If the reversed position falls into the closed evidence subgraph, but the corresponding evidence path has a break, lacks original graph connecting edge support, or deviates from the target path segment, then the conclusion fragment is added to the abnormal fragment set.

[0073] Subsequently, conclusion fragments from the masked fragment set and the abnormal fragment set are deleted from the enterprise processing conclusions. During deletion, the original order of the conclusion fragments in the enterprise processing conclusions is preserved, and only the intervals between fragments on either side of the deletion position are compressed. For example, if there is a deleted masked fragment between two retained conclusion fragments, the two retained conclusion fragments are arranged directly adjacent to each other; if the deletion position is at the beginning or end of a sentence, the corresponding conjunctions, transition words, or meaningless residual fragments are deleted simultaneously to avoid broken expressions in the finalized results.

[0074] Finally, the retained conclusion fragments are bound to the corresponding closed evidence subgraph positions. The binding objects include the conclusion fragment number, the backed-up fact node, the backed-up connecting edge, the request landing point, and the target path segment. After binding, the retained conclusion fragments are written into the enterprise SaaS data processing result in compressed order, and the closed evidence subgraph positions, backed-up matching values, the set of masked fragments, and the set of abnormal fragments are also fixed. The resulting enterprise SaaS data processing result only retains conclusion fragments that can return to the closed evidence subgraph; fragments touching the access breakpoint shadow zone and fragments lacking in-graph evidence support are not included in the final output.

[0075] refer to Figure 3 A data processing system based on AI-powered enterprise management software (SaaS) includes the following modules: The Enterprise Data Domain Delineation Module is used to respond to business requests in enterprise management software (SaaS) and delineate the enterprise data domain that is consistent with the business request within the tenant's runtime space. The fact meta aggregation module is used to aggregate the fact meta records that are scattered across different business functions within the enterprise data domain, and to connect them into an object fact chain according to the business object affiliation and business succession order between the data records. The Enterprise Fact Graph Assembly Module is used to assemble an enterprise fact graph around the object fact chain, connect business acceptance paths within the enterprise fact graph, and connect access boundary paths according to the data visibility range within the tenant's runtime space. The request anchoring module is used to rewrite business requests into request landing points and embed the request landing points into the corresponding business acceptance path and access boundary path in the enterprise fact graph to form a request anchoring fact graph. An improved GraphRAG processing module is used to input the request anchored fact graph into the improved GraphRAG model. The improved GraphRAG model includes an evidence path expansion unit, an access breakpoint shadow band unit, and a conclusion generation reverse unit. The access breakpoint shadow band unit embeds a boundary shadow compression structure. The evidence path expansion unit converges to form a closed evidence subgraph, and the access breakpoint shadow band unit folds the evidence path that crosses the access boundary path into an access breakpoint shadow band. The conclusion generation and reposting module is used to generate enterprise processing conclusions by using the closed evidence sub-graph as the generation basis and the access breakpoint shadow band as the shielding boundary through the conclusion generation and reposting unit. The enterprise processing conclusions are then split into conclusion fragments and reposted to the closed evidence sub-graph or the access breakpoint shadow band. The results solidification module is used to block the shadow band of the access break or the incomplete reverse pasting of the conclusion fragments, and solidify the enterprise SaaS data processing results according to the retained conclusion fragments.

[0076] Example 1: To verify the feasibility of this invention in practice, it was applied to a SaaS platform for enterprise management software used by a manufacturing company. This platform includes nine business functions: sales management, procurement management, inventory management, production management, cost accounting, financial management, supplier management, approval management, and customer management. The platform is configured with 18 tenant runtime spaces, and the test tenants include 126 accounts, nine types of role permissions, and approximately 526,000 business records. This example selects six types of business requests for testing: sales revenue analysis, procurement approval verification, inventory anomaly tracking, production cost accounting, supplier performance verification, and financial cost carry-over. A total of 200 business requests were constructed, of which 120 involve a single business object, and 80 involve cross-business object relationships.

[0077] During implementation, after receiving a business request, the enterprise management software (SaaS) first reads the tenant affiliation and business entry point from the request. It then maps the business entry point to the boundaries of functions such as sales, purchasing, inventory, production, cost, finance, supplier, approval, or customer management. Next, it delineates data partitions consistent with the tenant affiliation along these functional boundaries. For sales revenue analysis requests, the enterprise data domain retains records of contracts, orders, invoices, payments received, and customer transactions. For purchasing approval verification requests, the enterprise data domain retains purchase requests, approval processes, purchase orders, and warehousing records. Business records that cross the tenant affiliation or fall outside the scope of the business entry point are excluded.

[0078] Subsequently, fact elements are consolidated for data records in the enterprise data domain. Taking a sales payment analysis request as an example, contract creation, order generation, shipment registration, invoicing record, and payment confirmation are each organized into business fact elements and assigned to the corresponding object fact slots according to their business object affiliation. Interface synchronization records and manual confirmation records that repeatedly point to the same order status change are merged into co-occurring fact elements. Each business fact element is chained together into an object fact chain according to the business sequence of contract to order, order to shipment, shipment to invoicing, and invoicing to payment, with corresponding markers written at the beginning, end, and cross-object continuation positions of the chain.

[0079] During the enterprise fact graph assembly phase, each business fact element is mapped to a fact node. Intra-object connection edges are established between adjacent fact nodes within the same object's fact chain, and inter-object connection edges are established at locations with cross-object continuation markers. Intra-object and inter-object connection edges are arranged according to the connection order of the fact nodes to form business connection paths. Then, the fact nodes and connection edges on the business connection paths are checked against the data visibility range within the tenant's runtime space. Continuously visible portions are arranged into visible path segments, and the connection positions between visible path segments and invisible graph segments are marked as boundary breakpoints, forming access boundary paths.

[0080] After a business request is transcribed into a request endpoint, the request endpoint is connected to the corresponding fact node via request-related edges and to the receiving edge in the target path segment via path-related edges. The improved GraphRAG model receives the request and anchors it to the fact graph. The evidence path expansion unit expands the evidence path outward along the request endpoint, incorporating evidence paths that are continuous within the chain, clearly connected across objects, and whose path closure scores meet the truncation criteria into the closed evidence subgraph. Receiving edges that cross access boundary paths are marked as to be folded, and the access breakpoint shadow band unit calls the boundary shadow compression structure for processing. The boundary shadow compression structure cuts out the cross-boundary path segment, counts the number of fact nodes, the number of receiving edges, and the edge order span within the cross-boundary path segment, and pushes the positions of the visible fact nodes at the front and end of the segment, along with the path compression information, into the access breakpoint shadow band, and deletes the corresponding hint context index of the access breakpoint shadow band.

[0081] The conclusion generation and back-pasting unit uses visible business fact elements in the closed evidence subgraph as the generated content and the access breakpoint shadow band as the shielding boundary to generate the enterprise processing conclusion. After generation, the enterprise processing conclusion is split into conclusion fragments, and the object pointer, action pointer, and caliber pointer in the conclusion fragments are extracted. The object pointer is matched with the fact node by performing text identifier matching, the action pointer is matched with the receiving edge type by performing action type matching, and the caliber pointer is matched with the request landing point or target path segment by performing range matching. Conclusion fragments whose back-pasting matching value meets the back-pasting judgment condition and whose hit position falls within the same evidence path are back-pasted to the closed evidence subgraph; conclusion fragments that hit the access breakpoint shadow band are written to the shielded fragment set; conclusion fragments that are not back-pasted are written to the abnormal fragment set. Finally, the conclusion fragments that are back-pasted to the closed evidence subgraph and whose evidence paths are continuous are retained and solidified as the enterprise SaaS data processing result.

[0082] In this embodiment, the evidence path integrity rate is the proportion of test requests that can be traced from the request landing point to a continuous evidence path out of the total number of test requests; the out-of-bounds content hit rate is the proportion of the number of conclusion fragments in the enterprise processing conclusions that touch on invisible business fact elements out of the total number of conclusion fragments; the conclusion re-paste pass rate is the proportion of the number of conclusion fragments that complete the re-paste of the closed evidence subgraph out of the total number of conclusion fragments; the manual review quantity is the number of conclusion fragments that are written into the masked fragment set or the abnormal fragment set and require manual review; and the average response time is the average time from when the business request enters the enterprise management software SaaS to when the enterprise SaaS data processing result is solidified.

[0083] Table 1. Comparison of different data processing methods in a company's SaaS environment

[0084] As shown in Table 1, the ordinary keyword retrieval method has a fast response speed, but it can only hit partial business records, resulting in a low evidence path completeness rate. Business judgments in the generated results are difficult to reconcile into a complete evidence path. The ordinary GraphRAG method can improve evidence coverage by utilizing nodes and relationships in the enterprise fact graph, but it lacks breakpoint compression processing at access boundary paths, posing a risk that cross-boundary business fact elements may enter the prompt context. The permission-filtered GraphRAG method directly deletes invisible fragments within the graph, preventing out-of-bounds content from entering the generated evidence; however, the business connection path is broken at the boundary, leading to a decrease in the conclusion reconciliation pass rate.

[0085] This invention preserves visible evidence paths by closing evidence subgraphs, retains the structural positions of paths crossing access boundaries by access breakpoint shadow bands, and filters out conclusion fragments that touch access breakpoint shadow bands or have no landing points within the graph by back-pasting conclusion fragments. This keeps the out-of-bounds content hit rate at 0%, while increasing the evidence path integrity rate to 88.5% and the conclusion back-pasting pass rate to 82.0%. These results demonstrate that this invention can balance the integrity of business evidence chains, the consistency of access boundaries, and the traceability of AI processing conclusions in enterprise SaaS multi-tenant and multi-role data environments.

[0086] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A data processing method based on AI-powered enterprise management software (SaaS), characterized in that, Includes the following steps: S1. Respond to business requests in enterprise management software (SaaS) and define the enterprise data domain that is consistent with the business request within the tenant's runtime space. S2. Perform fact element aggregation on data records scattered across different business functions within the enterprise data domain, and connect them into object fact chains according to the business object affiliation and business succession order between the data records. S3. Assemble the enterprise fact graph around the object fact chain, connect the business acceptance path within the enterprise fact graph, and connect the access boundary path according to the data visibility range within the tenant's operating space. S4. Rewrite the business request into the request landing point, and embed the request landing point into the corresponding business acceptance path and access boundary path in the enterprise fact graph to form a request anchoring fact graph. S5. Input the request anchoring fact graph into the improved GraphRAG model. The improved GraphRAG model includes evidence path extension unit, access breakpoint shadow band unit and conclusion generation reverse unit. The access breakpoint shadow band unit embeds a boundary shadow compression structure. The evidence path extension unit converges to form a closed evidence subgraph. The access breakpoint shadow band unit folds the evidence path that crosses the access boundary path into an access breakpoint shadow band. S6. The conclusion generation and back-pasting unit uses the closed evidence sub-graph as the generation basis and the access breakpoint shadow band as the shielding boundary to generate the enterprise processing conclusion. The enterprise processing conclusion is then split into conclusion fragments and back-pasted to the closed evidence sub-graph or the access breakpoint shadow band. S7. Block the reverse pasting to the access port shadow band or the conclusion fragment that has not been reverse pasted, and solidify the enterprise SaaS data processing results according to the retained conclusion fragment; The access breakpoint shadow band unit folds the evidence path that crosses the access boundary path into an access breakpoint shadow band, specifically including: The access break shadow zone unit is scanned edge by edge along the evidence path in the closed evidence subgraph to locate the position in the graph where the evidence path intersects with the access boundary path; The access break shadow zone unit is defined by the location in the diagram where the evidence path intersects with the access boundary path, and the portion of the evidence path located on both sides of the access boundary path that is still within the visible range is extracted. The boundary shadow compression structure cuts out the cross-boundary path segment from the evidence path with the folding mark, and extracts the visible fact node at the front end of the cross-boundary path segment, the visible fact node at the end of the cross-boundary path segment, and the direction of the cross-boundary path segment in the original evidence path. The boundary shadow compression structure counts the number of internal fact nodes and internal connecting edges of the cross-boundary path segment, calculates the edge order span of the cross-boundary path segment in the original evidence path, and pushes the number of internal fact nodes, the number of internal connecting edges, the edge order span, the position of the visible fact node at the front end, and the position of the visible fact node at the end into the access breakpoint shadow band. The boundary shadow compression structure backfills the access breakpoint shadow band to the original position of the cross-boundary path segment in the original evidence path, and writes content masking marks and reverse placeholder marks in the access breakpoint shadow band; The access break shadow band unit deletes the prompt context index corresponding to the access break shadow band according to the content masking mark, and retains the reverse pasting position of the conclusion fragment corresponding to the access break shadow band according to the reverse pasting placeholder mark.

2. The data processing method based on AI enterprise management software SaaS according to claim 1, characterized in that, S1 specifically includes: S11. Read the tenant ownership bit and business entry bit carried by the business request in the tenant's runtime space, and map the business entry bit to the business function boundary in the enterprise management software SaaS. S12. Locate the data partition that matches the tenant's affiliation along the business function boundary, and classify the business records that fall into the data partition into the tenant candidate domain. S13. Verify the connection between the business records in the tenant candidate domain and the business entry point, and delete business records that are outside the scope of the business entry point or cross the tenant's ownership point. S14. Merge and retain business records according to the direction of business entry points to form an enterprise data domain consistent with the business request's origin.

3. The data processing method based on AI enterprise management software SaaS according to claim 1, characterized in that, S2 specifically includes: S21. Divide the record segments according to the business function to which the data records in the enterprise data domain belong, and gather the record segments that carry changes in business status, changes in object relationships, or results of business actions into business fact elements; S22. Perform object attribution verification on business fact elements and classify business fact elements belonging to the same business object into the same object fact slot. S23. Within the same object fact slot, arrange the business fact elements according to the generation order and business action succession relationship, and merge duplicate business fact elements that point to the same business action result into co-position fact elements. S24. Connect the preceding and following relationships between the corresponding fact elements and the adjacent business fact elements to form the object fact chain of the corresponding business object. S25. Write an object start marker at the beginning of the object fact chain, write an object end marker at the end of the object fact chain, and write an object continuation marker at the cross-object continuation position.

4. The data processing method based on AI enterprise management software SaaS according to claim 1, characterized in that, S3 specifically includes: S31. According to the order of the object fact chain, map the business fact element to the fact node in the enterprise fact graph one by one, and configure the in-chain order for the fact node in the same object fact chain; S32. Connect the fact nodes of adjacent chains within the same object's fact chain to form connecting edges within the object; S33. At the position corresponding to the object continuation mark, connect the tail fact node of the previous object fact chain with the head fact node of the next object fact chain to form an inter-object continuation edge. S34. Arrange the intra-object connecting edges and inter-object connecting edges according to the connecting order of fact nodes to form the business connecting paths within the enterprise fact diagram; S35. According to the data visibility range within the tenant's operating space, check the fact nodes and receiving edges on the business receiving path, arrange the continuously visible fact nodes and receiving edges into visible path segments, and mark the connection position between the visible path segments and the invisible graph segments as the boundary break position. S36. Connect visible path segments and boundary disconnect points in the order of business acceptance paths to form access boundary paths within the enterprise fact diagram.

5. The data processing method based on AI enterprise management software SaaS according to claim 1, characterized in that, S4 specifically includes: S41. Perform semantic segmentation on business requests to determine the business object, processing action, and scope of the business request; S42. Locate the fact node corresponding to the business object in the enterprise fact diagram, and extract the target path segment corresponding to the scope along the business acceptance path where the fact node is located. S43. Add a new request landing point in the target path segment and connect the request landing point to the fact node corresponding to the business object through the request association edge; S44. Write the processing action to the request landing point, and write the path association edge between the request landing point and the connecting edge in the target path segment. S45. Verify the access boundary path corresponding to the target path segment, and connect the request landing point to the visible path segment or boundary break point in the access boundary path to form a request anchoring fact graph.

6. The data processing method based on AI enterprise management software SaaS according to claim 1, characterized in that, The convergence of the evidence path extension unit to form a closed evidence subgraph specifically includes: The evidence path extension unit starts from the request landing point in the request anchoring fact graph, locates the fact node that is directly connected to the request landing point along the request association edge, and extracts the business undertaking path where the fact node is located as the target path segment. Using the fact node directly connected to the request landing point as the starting point of the path, expand the adjacent fact nodes and receiving edges one by one according to the edge order in the target path segment. Merge the fact nodes and receiving edges located in the target path segment into the candidate evidence path, and write the receiving edge to be collapsed mark for the path that crosses the access boundary. For the connecting edges in the candidate evidence path, check the in-chain order of the starting fact node and the ending fact node, and classify the connecting edges with adjacent in-chain order into the in-chain evidence path, and classify the connecting edges on both sides of the object continuation mark into the cross-object evidence path. Perform original graph edge verification on the break points in the intra-chain evidence path and cross-object evidence path. When the fact nodes at both ends of the break point have original receiving edges in the enterprise fact graph and the original receiving edges do not cross the access boundary path, the original receiving edges are returned to the corresponding evidence path. Merge the in-chain evidence paths and cross-object evidence paths under the same request landing point according to the business acceptance direction, and delete the fact nodes and acceptance edges that are separated from the target path segment to obtain the candidate closed evidence subgraph. For each evidence path in the candidate closed evidence subgraph, a path closure score is calculated. The path closure score is obtained by summing the number of edges connected to the request landing point, the number of consecutive connecting edges in the chain, and the number of hits of the object continuation mark. The score is deducted according to the number of connecting edges of the path that crosses the access boundary. Evidence paths are truncated from high to low according to the path closure score, and the truncated evidence paths are merged into the closed evidence subgraph.

7. The data processing method based on AI enterprise management software SaaS according to claim 1, characterized in that, S6 specifically includes: S61. The conclusion generates the reverse unit, which arranges visible business fact elements along the evidence path in the closed evidence subgraph, and retains the corresponding position of the access breakpoint shadow band as a non-generable position. S62. Using visible business fact elements as generated content and non-generable positions as generated shielding boundaries, organize the prompt context and generate enterprise processing conclusions. S63. Segment the conclusion text according to the business judgment in the enterprise's processing conclusion, and extract the object pointer, action pointer and caliber pointer in each text segment. Keep the text segments with object pointer, action pointer or caliber pointer as conclusion segments. S64. Match the object pointer in the conclusion fragment with the fact node in the closed evidence subgraph using text identifier matching, match the action pointer with the action type of the connecting edge in the closed evidence subgraph using action type matching, match the statement pointer with the request landing point or target path segment using range matching, and obtain the object hit value, action hit value and statement hit value respectively. S65. Accumulate the object hit value, action hit value, and caliber hit value to obtain the back-paste matching value of the conclusion fragment; when any hit position falls into the corresponding position of the access break shadow belt, back-paste the conclusion fragment to the access break shadow belt; for conclusion fragments that do not fall into the access break shadow belt, when the back-paste matching value reaches the back-paste judgment condition and the hit position falls into the same evidence path, the conclusion fragment is back-pasteted to the closed evidence sub-graph. S66. Mark the conclusion fragments that do not match the closed evidence subgraph and do not match the access break shadow band as incomplete re-pasting, and encapsulate the conclusion fragment, re-pasting position and incomplete re-pasting mark into the conclusion fragment re-pasting result.

8. The data processing method based on AI enterprise management software SaaS according to claim 1, characterized in that, Specifically, S7 includes: S71. Read the result of the reverse pasting of the conclusion fragment, write the conclusion fragment whose reverse pasting position falls into the shadow band of the access break point into the masked fragment set, and write the conclusion fragment with the incomplete reverse pasting mark into the abnormal fragment set. S72. Perform evidence path verification on conclusion fragments that are not written into the masked fragment set and abnormal fragment set, and retain conclusion fragments that are pasted back to the closed evidence subgraph and whose corresponding evidence paths are continuous. S73. Delete the conclusion fragments from the masked fragment set and the abnormal fragment set from the enterprise processing conclusions, and compress the arrangement interval of the conclusion fragments on both sides of the deletion position; S74. Bind the retained conclusion fragments to the corresponding closed evidence subgraph positions and solidify them into enterprise SaaS data processing results.

9. A data processing system based on AI enterprise management software (SaaS), comprising executing the data processing method based on AI enterprise management software (SaaS) as described in any one of claims 1 to 8, characterized in that, Includes the following modules: The Enterprise Data Domain Delineation Module is used to respond to business requests in enterprise management software (SaaS) and delineate the enterprise data domain that is consistent with the business request within the tenant's runtime space. The fact meta aggregation module is used to aggregate the fact meta records that are scattered across different business functions within the enterprise data domain, and to connect them into an object fact chain according to the business object affiliation and business succession order between the data records. The Enterprise Fact Graph Assembly Module is used to assemble an enterprise fact graph around the object fact chain, connect business acceptance paths within the enterprise fact graph, and connect access boundary paths according to the data visibility range within the tenant's runtime space. The request anchoring module is used to rewrite business requests into request landing points and embed the request landing points into the corresponding business acceptance path and access boundary path in the enterprise fact graph to form a request anchoring fact graph. An improved GraphRAG processing module is used to input the request anchored fact graph into the improved GraphRAG model. The improved GraphRAG model includes an evidence path expansion unit, an access breakpoint shadow band unit, and a conclusion generation reverse unit. The access breakpoint shadow band unit embeds a boundary shadow compression structure. The evidence path expansion unit converges to form a closed evidence subgraph, and the access breakpoint shadow band unit folds the evidence path that crosses the access boundary path into an access breakpoint shadow band. The conclusion generation and reposting module is used to generate enterprise processing conclusions by using the closed evidence sub-graph as the generation basis and the access breakpoint shadow band as the shielding boundary through the conclusion generation and reposting unit. The enterprise processing conclusions are then split into conclusion fragments and reposted to the closed evidence sub-graph or the access breakpoint shadow band. The results solidification module is used to block the shadow band of the access break or the incomplete reverse pasting of the conclusion fragments, and solidify the enterprise SaaS data processing results according to the retained conclusion fragments.

Images