A business node suggestion generation method based on rule engine and large model double-link fusion

Abstract

The present application relates to the technical field of natural language processing, and provides a business node suggestion generation method based on rule engine and large model double-link fusion, comprising: concurrently starting a rule engine link and a large language model link, respectively processing a same contract slice text in parallel, the rule engine link outputting a first node suggestion set, and the large language model link outputting a second node suggestion set; constructing a joint primary key composed of a business stage name and a target node name as an identifier for identifying the uniqueness of a business node; initializing an ordered key-value mapping set, first storing each node object of the first node suggestion set, and then traversing each node object of the second node suggestion set, and covering or appending according to whether the joint primary key already exists in the mapping set; extracting all value objects in the ordered key-value mapping set, and outputting a comprehensive business node suggestion snapshot. The present application can significantly improve the coverage range of node identification while ensuring high accuracy.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method for generating business node suggestions based on the fusion of a rule engine and a large model. Background Technology

[0002] In enterprise-level project management systems, contract documents are the core basis for project initiation and execution. The system typically needs to extract key performance clauses from the signed contract text and automatically map them to work breakdown structure nodes in the project baseline for task allocation, progress tracking, and risk control. The level of automation and intelligence in this process directly impacts the ease of use of the project management system and the efficiency of enterprise contract performance.

[0003] Currently, existing technologies mainly employ the following two types of technical solutions when implementing automatic mapping from contract terms to business nodes.

[0004] The first type is the pure rule engine solution. This solution uses regular expressions or a pre-defined keyword dictionary for text matching. For example, when the system recognizes the phrase "delivery," it automatically generates a "delivery and acceptance" node. Rule engine solutions have the advantages of fast execution speed and 100% controllable results, but their generalization ability is extremely poor. Because natural language expressions in contract texts are highly variable, the same business meaning may appear in different sentence structures, synonyms, or even inverted structures. Rule engines are prone to missing or misidentifying key nodes due to slight variations in expression, resulting in many critical nodes not being correctly extracted, severely impacting the system's usability and coverage.

[0005] The second category is the pure large language model solution. Large language models possess powerful contextual semantic understanding and generalization reasoning capabilities, enabling them to handle non-standard natural language expressions and, to some extent, overcome the generalization limitations of rule engines. However, large language model solutions face the following prominent problems in practical engineering implementation: high response latency, failing to meet the real-time requirements of business scenarios; output results exhibiting a degree of randomness, i.e., the "illusion" phenomenon, which may lead to the model ignoring or misidentifying basic nodes that should be stably extracted; high API call costs, resulting in poor economic efficiency in large-scale contract parsing tasks; and a lack of interpretability in the model's output, making it difficult to meet the needs of enterprise-level auditing and traceability.

[0006] Simply combining the rule engine with the large language model in series or in parallel will lead to two serious problems: First, duplicate node suggestions may occur. For example, if both the rule engine and the large language model suggest a "final verification" node, redundant rows may appear in the front-end WBS tree table. Second, the identification conclusions may be contradictory. Different links may give different node suggestions for the same contract clause, which will seriously trouble the secondary review by business personnel and reduce work efficiency.

[0007] Chinese patent CN121707781A discloses a method for intelligent contract review based on a large language model. This method employs a "large model identification + triage decision" architecture, where the large model handles both identification and triage, while the rule engine only processes the complex review items that have been triaged. The two types of results correspond to different review items (simple and complex), preventing the same review item from being processed simultaneously by both links. However, in actual business scenarios, the same contract clause may be identified by both the large model and the rule engine, leading to different conclusions. Since no conflict detection or adjudication mechanism is provided, there is a risk of contradictory results.

[0008] Therefore, how to provide a solution that can fully mine the deep semantic information in the contract text while ensuring the stable extraction of basic nodes, and effectively eliminate the duplication and conflict between multiple strategies, thus providing a highly reliable and disaster-tolerant solution for intelligent parsing of enterprise-level contracts, has become an urgent technical problem to be solved. Summary of the Invention

[0009] In view of this, in order to overcome the shortcomings of the existing technology, the present invention aims to provide a business node suggestion generation method based on the fusion of rule engine and large model dual links.

[0010] This invention provides a method for generating business node suggestions based on the fusion of a rule engine and a large model, the method comprising:

[0011] Step S1: Simultaneously start the rule engine link and the large language model link, and process the same contract slice text in parallel. The rule engine link outputs the first node suggestion set, and the large language model link outputs the second node suggestion set.

[0012] Step S2: Construct a composite primary key consisting of the business stage name and the target node name, as an identifier to uniquely identify the business node;

[0013] Step S3: Initialize the ordered key-value pair mapping set. First, store each node object of the first node suggestion set, then traverse each node object of the second node suggestion set, and overwrite or append according to whether its composite primary key already exists in the mapping set.

[0014] Step S4: Extract all value objects from the ordered key-value pair mapping set and output a comprehensive business node suggested snapshot.

[0015] Optionally, in the business node suggestion generation method based on the fusion of rule engine and large model of the present invention, in step S1, the rule engine link outputs the first node suggestion set in the following manner:

[0016] Pre-build a mapping dictionary between pre-defined keywords and business stages and target nodes;

[0017] After receiving the contract slice text, high-speed lexical matching is performed in memory to identify preset keywords appearing in the slice text. Based on the matched keywords, the corresponding business stage name and target node name are searched from the mapping dictionary.

[0018] The found business stage name and target node name are assembled into node objects in the first node suggestion set, where each node object contains the base node name.

[0019] Optionally, in the business node suggestion generation method based on the fusion of rule engine and large model of the present invention, in step S1, the large language model link outputs the second node suggestion set in the following manner:

[0020] Combine the contract slice text, the whitelist of allowed nodes in the project, and the current context information into structured prompt words;

[0021] Send the structured prompt to the enterprise-level large language model and force the large language model to output in a preset JSON format. The preset JSON format must include at least the target node name, business stage name, reason for difference, evidence page number, and original evidence reference field.

[0022] After receiving the JSON output returned by the large language model, perform JSON deserialization and validate the output node names according to the node whitelist.

[0023] The verified node objects are assembled into a second set of node suggestions, in which each node object contains a node name, reason for inference, and original evidence information.

[0024] Optionally, in the business node suggestion generation method based on the fusion of rule engine and large model of the present invention, in step S2, a composite primary key is constructed as follows: the business stage name and the target node name are concatenated by a delimiter to form a composite string as the composite primary key, wherein the delimiter is a special character that will not appear in the business node name, and the business stage name and the target node name are subjected to null value verification and standardization processing before concatenation.

[0025] Optionally, in the business node suggestion generation method based on the fusion of rule engine and large model dual links of the present invention, in step S3, the ordered key-value pair mapping set is a chain hash table that preserves the insertion order. When traversing the first node suggestion set and the second node suggestion set, the ordered key-value pair mapping set maintains the storage order of key-value pairs according to the order in which the nodes are first inserted. When finally outputting the comprehensive business node suggestion snapshot, the node order in the ordered key-value pair mapping set is consistent with the timeline of business flow.

[0026] Optionally, in the business node suggestion generation method based on the fusion of rule engine and large model dual links of the present invention, in step S3, when the node object corresponding to the same composite primary key exists in the second node suggestion set, the node object in the second node suggestion set is used to overwrite the node object in the first node suggestion set; if the composite primary key of the node object in the second node suggestion set does not exist, the corresponding node object is appended as a new element to the ordered key-value pair mapping set.

[0027] Optionally, in the business node suggestion generation method based on the fusion of rule engine and large model dual links of the present invention, in step S3, the node objects in the first node suggestion set are overwritten by the node objects in the second node suggestion set in the following manner: the business nodes corresponding to the joint primary key are identified by the rule engine link and the large language model link, and without generating duplicate suggestion rows, the rich node objects in the second node suggestion set that carry the reason for inference and the original text evidence are used to replace the original sparse node objects that only contain the basic node names in the ordered key-value pair mapping set.

[0028] Optionally, in the business node suggestion generation method based on the fusion of rule engine and large model dual links of the present invention, in step S3, when traversing the first node suggestion set, the nodes are stored in the mapping set in the order in which they are identified by the rule engine link; when traversing the second node suggestion set to perform overwrite or append operations, the relative order of the existing key-value pairs in the mapping set is not changed; for the new nodes to be appended, they are appended to the end of the mapping set in the traversal order of the second node suggestion set.

[0029] Optionally, in the business node suggestion generation method based on the fusion of rule engine and large model of the present invention, step S4 outputs a snapshot of comprehensive business node suggestions in the following manner:

[0030] Extract all value objects from the ordered key-value pair mapping set into a list structure, serialize the list structure into JSON format, and write it into the persistent snapshot of the current baseline recognition task;

[0031] The front-end WBS tree table component reads this persistent snapshot. For nodes generated solely by the rule engine chain, it only displays suggested actions. For nodes generated by the large language model or those that have been overwritten, it renders and displays the reasoning, evidence page number, and original evidence citation in the AI ​​suggestion and evidence summary columns for business personnel to trace and confirm.

[0032] Optionally, in the business node suggestion generation method based on the fusion of rule engine and large model of the present invention, in step S4, multiple target nodes in the same business stage are arranged in the order in which they are first identified in the output comprehensive business node suggestion snapshot.

[0033] The business node suggestion generation method based on the dual-link fusion of rule engine and large model of the present invention has the following beneficial technical effects:

[0034] I. Break through the bottleneck of a single technology and achieve complementary advantages

[0035] A dual-link concurrent architecture of a rule engine and a large language model is adopted. The rule engine link ensures the stable extraction of basic nodes, solving the problem of node "forgetting" or misidentification caused by the randomness of the large language model. The large language model link extracts business nodes from long-tail clauses and non-standard expressions through deep semantic understanding, compensating for the weak generalization ability of the rule engine. The two links complement each other, significantly improving the coverage of node recognition while ensuring high accuracy.

[0036] II. Eliminate multi-strategy conflicts and achieve zero duplicate output.

[0037] By constructing a composite primary key consisting of "business stage name + target node name", the node suggestions output by the dual-link system are precisely deduplicated. This completely eliminates duplicate suggestion rows that are prone to occur when the rule engine and the large language model are processed concurrently, ensuring the cleanliness and readability of the front-end WBS tree structure.

[0038] Third, enhance the dimensionality of evidence and improve its interpretability.

[0039] It automatically supplements the original evidence and reasoning logic for the rule-hitting items that were originally without explanation, so that the front-end page can provide business personnel with complete AI suggestions and evidence summary support, which greatly reduces the cognitive threshold of manual review.

[0040] IV. Possesses high disaster recovery capabilities to ensure business continuity.

[0041] Even in extreme cases where the enterprise's private large language model experiences network anomalies, service timeouts, or configuration errors, it can automatically and silently degrade, relying solely on the basic node suggestions output by the rule engine to maintain basic operation, thus preventing business interruption and improving engineering robustness and production environment availability.

[0042] Fifth, maintain the order of business operations and conform to cognitive habits.

[0043] The final output nodes should be arranged in the same order as the natural timeline of the business process, which aligns with the cognitive habits of business users and eliminates the need for additional sorting, thus improving the user experience.

[0044] VI. Reduce overall costs and achieve engineering economics.

[0045] It significantly reduces the number of calls to and reliance on large language models, effectively lowering API call costs and system response latency. Furthermore, it can be quickly deployed without relying on large-scale labeled data, demonstrating strong engineering economics and significant potential for widespread adoption. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a schematic diagram of the process of generating a business node suggestion based on the dual-link fusion of rule engine and large model according to an exemplary embodiment 1 of the present invention, which outputs a first node suggestion set through the rule engine link;

[0048] Figure 2 This is a schematic diagram of the process of generating a business node suggestion based on the dual-link fusion of rule engine and large model according to an exemplary embodiment 1 of the present invention, which outputs a second node suggestion set through the large language model link;

[0049] Figure 3 This is a flowchart illustrating the coverage and append logic of the business node suggestion generation method based on the fusion of rule engine and large model dual links according to exemplary embodiment 1 of the present invention.

[0050] Figure 4 A schematic diagram illustrating the process of sequential maintenance for the business node suggestion generation method based on the fusion of rule engine and large model according to exemplary embodiment 1 of the present invention;

[0051] Figure 5 This is a schematic diagram illustrating the process of outputting a comprehensive business node suggestion snapshot according to the business node suggestion generation method based on the fusion of rule engine and large model dual links in exemplary embodiment 1 of the present invention. Detailed Implementation

[0052] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0053] It should be noted that, in the absence of conflict, the following embodiments and features can be combined with each other; and, based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0054] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0055] Example 1

[0056] Exemplary embodiment 1 of the present invention provides a method for generating business node suggestions based on the dual-link fusion of rule engine and large model. In this embodiment, the method of the present invention is implemented in the following manner:

[0057] Step S1: Simultaneously start the rule engine link and the large language model link, and process the same contract slice text in parallel. The rule engine link outputs the first node suggestion set, and the large language model link outputs the second node suggestion set.

[0058] Figure 1 This is a schematic diagram illustrating the process of a business node suggestion generation method based on the dual-link fusion of rule engine and large model according to an exemplary embodiment 1 of the present invention, which outputs a first node suggestion set through the rule engine link. Figure 1 As shown, in this embodiment, the rule engine chain outputs the first node suggestion set in the following manner:

[0059] A pre-built mapping dictionary is constructed between business-preset keywords and business stages and target nodes. After receiving the contract slice text, high-speed lexical matching is performed in memory to identify the preset keywords appearing in the slice text. Based on the matched keywords, the corresponding business stage name and target node name are searched from the mapping dictionary. The searched business stage name and target node name are assembled into node objects in the first node suggestion set. Each node object in this set contains the base node name.

[0060] Figure 2 This is a schematic diagram illustrating the process of a business node suggestion generation method based on the dual-link fusion of a rule engine and a large model according to an exemplary embodiment 1 of the present invention, which outputs a second node suggestion set through a large language model link. Figure 2 As shown, in this embodiment, the large language model link outputs the second node suggestion set in the following manner:

[0061] The contract slice text, the whitelist of allowed nodes in the project, and the current context information are combined into a structured prompt word. This structured prompt word is sent to the enterprise-level large language model, which is forced to output in a preset JSON format. The preset JSON format includes at least the target node name, business stage name, reason for difference, evidence page number, and original evidence citation fields. After receiving the JSON output returned by the large language model, a JSON deserialization operation is performed, and the output node names are validated for legality based on the node whitelist. The validated node objects are assembled into a second node suggestion set, in which each node object contains a node name, reason for inference, and original evidence information.

[0062] Step S2: Construct a composite primary key consisting of the business stage name and the target node name, as an identifier to uniquely identify the business node.

[0063] In this embodiment, the business stage name and the target node name are concatenated using a delimiter to form a composite string as the composite primary key. The delimiter is a special character that will not appear in the business node name to ensure the uniqueness and resolvability of the composite primary key. Before concatenation, the business stage name and the target node name are checked for null values ​​and standardized to avoid the primary key becoming invalid or ambiguous due to null values.

[0064] In practical applications, a preprocessing step can be performed on the node suggestion set before constructing the composite primary key. For example, the integrity of each node object in the first and second node suggestion sets can be checked to ensure that the business stage name and the target node name are not empty and conform to the business naming convention. For node objects with missing fields or abnormal formats, they can be discarded or filled with default values ​​according to their source link. Only after the node objects have been preprocessed can they enter the subsequent composite primary key construction and mapping set merging process to ensure the correctness of the deduplication and merging logic.

[0065] Step S3: Initialize the ordered key-value pair mapping set. First, store each node object of the first node suggestion set, then traverse each node object of the second node suggestion set, and overwrite or append according to whether its composite primary key already exists in the mapping set.

[0066] In this embodiment, the ordered key-value pair mapping set is a chained hash table that preserves the insertion order. When traversing the first node suggestion set and the second node suggestion set, the ordered key-value pair mapping set maintains the storage order of key-value pairs according to the order in which nodes are first inserted. When finally outputting the comprehensive business node suggestion snapshot, the node order in the ordered key-value pair mapping set is consistent with the timeline of business flow, ensuring that the node order displayed on the front end meets business expectations.

[0067] Figure 3This is a flowchart illustrating the coverage and append logic of the business node suggestion generation method based on the dual-link fusion of rule engine and large model according to exemplary embodiment 1 of the present invention. Figure 3 As shown in this embodiment, when the node object corresponding to the same composite primary key exists in the second node suggestion set, the node object in the second node suggestion set is used to overwrite the node object in the first node suggestion set; if the composite primary key of the node object in the second node suggestion set does not exist, the corresponding node object is appended as a new element to the ordered key-value pair mapping set.

[0068] When it is necessary to overwrite node objects in the first node suggestion set with node objects from the second node suggestion set, the business nodes corresponding to the joint primary key are identified through the rule engine link and the large language model link. Without generating duplicate suggestion rows, the rich node objects from the second node suggestion set, which carry inference reasons and original text evidence, are used to replace the original, sparse node objects in the ordered key-value pair mapping set that only contain basic node names. This achieves an enhanced level of evidence for the dual-link overlapping identification items, providing the front-end page with complete audit traceability information and AI suggestion explanation support.

[0069] Figure 4 This is a schematic diagram illustrating the process of sequential maintenance for the business node suggestion generation method based on the dual-link fusion of rule engine and large model according to exemplary embodiment 1 of the present invention. Figure 4 As shown in this embodiment, when traversing the first node suggestion set, the nodes are stored in the mapping set in the order they are identified by the rule engine link; when traversing the second node suggestion set to perform overwrite or append operations, the relative order of the existing key-value pairs in the mapping set is not changed; for the new nodes to be appended, they are appended to the end of the mapping set in the traversal order of the second node suggestion set.

[0070] Step S4: Extract all value objects from the ordered key-value pair mapping set and output a comprehensive business node suggested snapshot.

[0071] Figure 5 This is a schematic diagram illustrating the process of outputting a comprehensive business node suggestion snapshot according to the business node suggestion generation method based on the dual-link fusion of rule engine and large model according to exemplary embodiment 1 of the present invention. Figure 5 As shown, this embodiment outputs a suggested snapshot of the integrated service node in the following manner:

[0072] All value objects in the ordered key-value pair mapping set are extracted into a list structure, which is then serialized into JSON format and written to the persistent snapshot of the current baseline recognition task. The front-end WBS (Work Breakdown Structure) tree table component reads this persistent snapshot. For nodes generated solely by the rule engine chain, only suggested actions are displayed. For nodes generated by the large language model or those that have been overwritten, the AI ​​suggestion and evidence summary columns render and display the reasoning, evidence page number, and original evidence citation for business personnel to trace and confirm.

[0073] It should be noted that in the comprehensive business node suggestion snapshot output by the method of this embodiment, multiple target nodes within the same business stage are arranged according to the time sequence in which they are first identified, thus fully preserving the natural order of business flow.

[0074] In this embodiment, the rule engine link and the large language model link are architecturally independent and execute concurrently. An anomaly in one link does not affect the normal operation of the other link. When the large language model link returns to a failure state due to network anomaly, service timeout, or configuration error, it silently degrades, skips the traversal and overwrite operation of the second node suggestion set, and only outputs a snapshot of comprehensive business node suggestions based on the first node suggestion set, ensuring that the business can still maintain basic operation in the extreme case that the large language model is unavailable.

[0075] Example 2

[0076] Exemplary embodiment 2 of the present invention provides a business node suggestion generation method based on the dual-link fusion of rule engine and large model. In this embodiment, a dual-track concurrent architecture of rule extraction and large model inference is adopted at the identification layer, and a deduplication and merging mechanism based on composite primary key is introduced at the result fusion layer to achieve business-level complementarity of "rules guaranteeing the lower limit and models raising the upper limit".

[0077] Specifically, in this embodiment, the method of the present invention is implemented in the following manner:

[0078] 1. Rule engine link (link A) identification:

[0079] Extract the contract text slices and load the pre-defined fixed keywords and corresponding stage mapping dictionary. For example, matching "software development" or "implementation services" maps to "requirements analysis → solution design → deployment and launch". Perform high-speed lexical matching in memory and output the first set of highly deterministic node suggestions (Set A). This set has extremely high accuracy but does not explain complex contextual differences; it only outputs basic node names.

[0080] 2. Deep reasoning of large language model links (link B):

[0081] The same contract slice text, the whitelist of allowed nodes for the project, and the current context information are combined into a structured Prompt, which is then sent to the Enterprise Large Language Model (LLM). The model not only outputs node name suggestions but also mandates the output of the "diffReason," "evidencePageNo," and "evidenceText" for each suggestion. After JSON deserialization and whitelist verification, a second set of highly interpretable node suggestions (Set B) is output.

[0082] 3. Dual-link fusion and composite primary key deduplication:

[0083] Construct a composite primary key "phaseName#targetNodeName", which is the phase name plus the target node name, as a unique identifier to determine whether node suggestions are duplicated. Import collections A and B into the merging engine (mergeSuggestionItems), iterate through and compare the two collections, and execute the merge and overwrite logic.

[0084] 4. Implementation of complementary retention strategy:

[0085] If a composite primary key appears only in set A or set B, the union is directly taken and retained to ensure that neither the rule engine's safety net nor the large model's supplementary items are lost.

[0086] If set A and set B have duplicates on the same composite primary key, meaning both links suggested the same node at the same stage, the rule engine's result should be forcibly discarded, and the LLM link's result should be retained first. The LLM result carries detailed original evidence and reasoning; prioritizing the retention of LLM results provides complete interpretability support for the "evidence summary" and "AI suggestions" on the front-end page.

[0087] 5. Structured snapshot output and disk persistence:

[0088] The final set of suggestions after fusion and deduplication is packaged and written into the persistent snapshot of the current baseline recognition task, so that the front-end WBS tree table can highlight and display it for manual confirmation.

[0089] Example 3

[0090] Exemplary embodiment 3 of the present invention provides a method for generating business node suggestions based on the fusion of a rule engine and a large model. In this embodiment, the method of the present invention is implemented according to the following steps:

[0091] Step 1: Trigger the rule engine for recognition

[0092] When a user initiates a contract comparison or identification task, a lightweight rule parsing process is first executed. For example, for goods contracts, if the text contains phrases such as "supply" or "equipment arrival," the rule engine generates a suggestion with "phaseName" set to "implementation phase" and "targetNodeName" set to "delivery and acceptance." This process takes only milliseconds, ensuring that even if subsequent models time out, the most basic node skeleton can still be provided.

[0093] Step 2: Trigger intelligent optimization of the large language model

[0094] The context, which includes file slices, existing WBS nodes, and a node whitelist, is packaged into a Prompt and requested from the LLM. Based on contextual reasoning, the model not only discovers "delivery acceptance" but also extracts the "initial inspection" node through deep semantic understanding, and outputs the corresponding inference basis (diffReason) and original text citation (evidenceText) strictly according to the JSON Schema.

[0095] Step 3: Mapping of Union Primary Keys

[0096] After retrieving the data returned from both links, the backend fusion service (mergeSuggestionItems) initializes a LinkedHashMap with String keys and SuggestionItem values ​​to ensure the stability of the final output order. It iterates through the collection generated by the rule engine, concatenates a composite string using "phaseName#targetNodeName", and stores it as the key in the Map.

[0097] Step 4: Seamless Coverage of High-Value Results

[0098] The system iterates through the collection generated by the large model, concatenating keys using the same rules. When attempting to store data into the Map, if the key already exists (meaning the node extracted by the large model is exactly the same as that extracted by the rule engine), the system leverages the Map's overriding property to directly replace the bare object generated by the rule engine with a rich object from the large model containing the "evidenceText" field, achieving "evidence enhancement" for the same business node. If the key does not exist, it indicates a generalization discovery unique to the large model, and the data is directly appended to the Map.

[0099] Step 5: Front-end rendering and interpretation output

[0100] The Map's Values ​​are converted into a List and passed to the front end. The front-end WBS tree table component reads this list. For nodes generated solely by the rule engine, only suggested actions are displayed. For nodes generated by a large model or after overwriting, the complete reasoning and original page number are rendered in the "AI Suggestion / Evidence Summary" column for business personnel to quickly trace and confirm the source.

[0101] The business node suggestion generation method based on the dual-link fusion of rule engine and large model in this invention has the following beneficial technical effects:

[0102] I. Break through the bottleneck of a single technology and achieve complementary advantages

[0103] A dual-link concurrent architecture of a rule engine and a large language model is adopted. The rule engine link ensures the stable extraction of basic nodes, solving the problem of node "forgetting" or misidentification caused by the randomness of the large language model. The large language model link extracts business nodes from long-tail clauses and non-standard expressions through deep semantic understanding, compensating for the weak generalization ability of the rule engine. The two links complement each other, significantly improving the coverage of node recognition while ensuring high accuracy.

[0104] II. Eliminate multi-strategy conflicts and achieve zero duplicate output.

[0105] By constructing a composite primary key consisting of "business stage name + target node name", the node suggestions output by the dual-link system are precisely deduplicated. During the fusion process, duplicate nodes are automatically identified and overwrite or append operations are performed, completely eliminating duplicate suggestion rows that are easily generated when the rule engine and the large language model are processed concurrently, ensuring the cleanliness and readability of the front-end WBS tree structure.

[0106] Third, enhance the dimensionality of evidence and improve its interpretability.

[0107] When both links simultaneously identify the same business node, the rules engine's output, which only contains the basic node name, is forcibly discarded. The rich results from the large language model, which include reasoning, evidence page numbers, and original text citations, are retained instead. This automatically supplements previously unexplained rule hits with original evidence and reasoning logic, enabling the front-end page to provide business personnel with complete AI suggestions and evidence summaries, significantly reducing the cognitive barrier to manual review.

[0108] IV. Possesses high disaster recovery capabilities to ensure business continuity.

[0109] The rules engine pipeline and the large language model pipeline are architecturally independent and execute concurrently. Even in extreme cases where the enterprise's private large language model experiences network anomalies, service timeouts, or configuration errors, it can automatically and silently degrade, relying solely on the basic node suggestions output by the rules engine pipeline to maintain basic operation, thus preventing business interruption and improving engineering robustness and production environment availability.

[0110] Fifth, maintain the order of business operations and conform to cognitive habits.

[0111] A linked hash table that preserves insertion order is used as the storage structure for the fusion results, and the output order is maintained strictly according to the time sequence in which nodes are first identified. The final output node order is recommended to be consistent with the natural timeline of business processes, conforming to the cognitive habits of business personnel, eliminating the need for additional sorting processing, and improving user experience.

[0112] VI. Reduce overall costs and achieve engineering economics.

[0113] Compared to pure large language model solutions, using a rule engine to handle the identification of a large number of basic nodes significantly reduces the number of calls to and reliance on large language models, effectively lowering API call costs and system response latency. Furthermore, it can be deployed quickly without relying on large-scale labeled data, demonstrating good engineering economics and promotional value.

[0114] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0115] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for generating business node suggestions based on the fusion of a rule engine and a large model, characterized in that, The method includes: Step S1: Simultaneously start the rule engine link and the large language model link, and process the same contract slice text in parallel. The rule engine link outputs the first node suggestion set, and the large language model link outputs the second node suggestion set. Step S2: Construct a composite primary key consisting of the business stage name and the target node name, as an identifier to uniquely identify the business node; Step S3: Initialize the ordered key-value pair mapping set. First, store each node object of the first node suggestion set, then traverse each node object of the second node suggestion set, and overwrite or append according to whether its composite primary key already exists in the mapping set. Step S4: Extract all value objects from the ordered key-value pair mapping set and output a comprehensive business node suggested snapshot.

2. The business node suggestion generation method based on the dual-link fusion of rule engine and large model as described in claim 1, characterized in that, In step S1, the rule engine outputs the first node suggestion set as follows: Pre-build a mapping dictionary between pre-defined keywords and business stages and target nodes; After receiving the contract slice text, high-speed lexical matching is performed in memory to identify preset keywords appearing in the slice text. Based on the matched keywords, the corresponding business stage name and target node name are searched from the mapping dictionary. The found business stage name and target node name are assembled into node objects in the first node suggestion set, where each node object contains the base node name.

3. The business node suggestion generation method based on the dual-link fusion of rule engine and large model as described in claim 1, characterized in that, In step S1, the large language model link outputs the second node suggestion set in the following manner: Combine the contract slice text, the whitelist of allowed nodes in the project, and the current context information into structured prompt words; Send the structured prompt to the enterprise-level large language model and force the large language model to output in a preset JSON format. The preset JSON format must include at least the target node name, business stage name, reason for difference, evidence page number, and original evidence reference field. After receiving the JSON output returned by the large language model, perform JSON deserialization and validate the output node names according to the node whitelist. The verified node objects are assembled into a second set of node suggestions, in which each node object contains a node name, reason for inference, and original evidence information.

4. The business node suggestion generation method based on the dual-link fusion of rule engine and large model as described in claim 1, characterized in that, In step S2, the composite primary key is constructed as follows: the business stage name and the target node name are concatenated using a delimiter to form a composite string as the composite primary key. The delimiter is a special character that will not appear in the business node name. The business stage name and the target node name are checked for null values ​​and standardized before concatenation.

5. The business node suggestion generation method based on the dual-link fusion of rule engine and large model as described in claim 1, characterized in that, In step S3, the ordered key-value pair mapping set is a chained hash table that preserves the insertion order. When traversing the first node suggestion set and the second node suggestion set, the ordered key-value pair mapping set maintains the storage order of key-value pairs according to the order in which the nodes are first inserted. When finally outputting a snapshot of the comprehensive business nodes, the order of nodes in the ordered key-value pair mapping set is kept consistent with the timeline of business flow.

6. The business node suggestion generation method based on the dual-link fusion of rule engine and large model as described in claim 1, characterized in that, In step S3, when the node object corresponding to the same composite primary key exists in the second node suggestion set, the node object in the second node suggestion set is overwritten with the node object in the first node suggestion set; if the composite primary key of the node object in the second node suggestion set does not exist, the corresponding node object is appended as a new element to the ordered key-value pair mapping set.

7. The business node suggestion generation method based on the dual-link fusion of rule engine and large model as described in claim 6, characterized in that, In step S3, the node objects in the first node suggestion set are overwritten with node objects in the second node suggestion set in the following manner: the business nodes corresponding to the joint primary key are identified through the rule engine link and the large language model link. Without generating duplicate suggestion rows, the rich node objects in the second node suggestion set that carry the reason for inference and the original text evidence are used to replace the original sparse node objects that only contain the basic node names in the ordered key-value pair mapping set.

8. The business node suggestion generation method based on the dual-link fusion of rule engine and large model according to claim 1, characterized in that, In step S3, when traversing the first node suggestion set, the nodes are stored in the mapping set in the order they are identified by the rule engine link; when traversing the second node suggestion set to perform overwrite or append operations, the relative order of the existing key-value pairs in the mapping set is not changed; for the new nodes to be appended, they are appended to the end of the mapping set in the traversal order of the second node suggestion set.

9. The business node suggestion generation method based on the dual-link fusion of rule engine and large model according to claim 1, characterized in that, In step S4, output the suggested snapshot of the integrated service node as follows: Extract all value objects from the ordered key-value pair mapping set into a list structure, serialize the list structure into JSON format, and write it into the persistent snapshot of the current baseline recognition task; The front-end WBS tree table component reads this persistent snapshot. For nodes generated solely by the rule engine chain, it only displays suggested actions. For nodes generated by the large language model or those that have been overwritten, it renders and displays the reasoning, evidence page number, and original evidence citation in the AI ​​suggestion and evidence summary columns for business personnel to trace and confirm.

10. The business node suggestion generation method based on the dual-link fusion of rule engine and large model according to claim 1, characterized in that, In step S4, in the output comprehensive business node suggestion snapshot, multiple target nodes within the same business stage are arranged according to the time sequence in which they were first identified.

Images