Skill package-based contract intelligent double-track verification method and device

Abstract

The embodiment of the application provides a contract intelligent double-track verification method and device based on a skill package, the method is that a first LLM model uses each to-be-verified item injected in a compliance skill package to determine whether each to-be-verified item exists in a to-be-compliance-verified text, extracts original text description corresponding to the to-be-verified item that exists to generate a determination result item, a first rule engine combines comprehensive compliance rules to judge each determination result item to generate a compliance verification result; an intelligent agent injects each to-be-extracted field obtained by analyzing consistency verification requirements and a to-be-consistency-verified text set into a consistency skill package, extracts an extraction result set corresponding to each to-be-extracted field from the to-be-consistency-verified text set based on the consistency skill package, a text priority rule and an extraction and merging rule, selects a comparison strategy corresponding to the field type of each to-be-extracted field to verify the extraction result set, generates a consistency verification result, and determines the verification result based on the compliance verification result and the consistency verification result.

Description

Technical Field

[0001] This application relates to the field of text verification technology, specifically to a smart dual-track verification method and device for contracts based on skill packs. Background Technology

[0002] In enterprise OA (Office Automation) systems, contract approval is a high-frequency and high-risk business process. With the continuous advancement of enterprise informatization, digital contract management systems have been widely applied in the daily operations of various enterprises. However, the core verification work of contract approval still faces significant challenges. Contract approval verification mainly falls into two categories: consistency verification, used to check whether the content of contract attachments is consistent with the data in the approval form and quotation, including comparison of key information such as amount, details, and terms; and compliance verification, checking whether the contract terms comply with the company's management system, including checking for prohibited clauses, verifying payment method requirements, and checking warranty period restrictions. Existing technologies mostly rely on rule engines for the approval of structured content, such as deterministic content like amounts and details, while semi-structured and unstructured content relies on manual review, resulting in low automation levels and poor review efficiency and accuracy in contract approval. Summary of the Invention

[0003] To address the problems in the existing technology, this application provides a skill pack-based intelligent dual-track verification method and device for contracts, which can effectively improve the automation level and accuracy of contract approval.

[0004] To solve at least one of the above problems, this application provides the following technical solution: Firstly, this application provides a dual-track intelligent contract verification method based on a skill package, comprising: obtaining compliance verification requirements and consistency verification requirements of a document to be verified; an intelligent agent parsing the compliance verification requirements to obtain multiple items to be verified and comprehensive compliance rules; injecting the compliance verification text of the document to be verified, the multiple items to be verified, and the comprehensive compliance rules into the configuration text of a compliance skill package; using a first LLM model in the compliance skill package to determine whether each item to be verified exists in the compliance verification text, extracting the original text description corresponding to the existing items to be verified to generate a judgment result item; injecting each judgment result item back into the configuration text; and utilizing a first rule engine in the compliance skill package. The comprehensive compliance rules are combined to make a comprehensive judgment on each of the judgment result items, generating a compliance verification result. The intelligent agent decomposes the consistency verification requirements to obtain multiple fields to be extracted. Each field to be extracted and the text set to be consistent in the document to be verified are injected into the consistency skill package. Based on the consistency skill package, text priority rules, and extraction and merging rules, the extraction result set corresponding to each field to be extracted is extracted from the text set to be consistent in the document to be verified. The extraction result set is then verified by selecting a comparison strategy corresponding to the field type of each field to be extracted, generating a consistency verification result. Based on the compliance verification result and the consistency verification result, the verification result of the document to be verified is generated.

[0005] In some embodiments, the agent parses the compliance verification requirements to obtain multiple items to be verified and comprehensive compliance rules, and injects the compliance verification text of the document to be verified, the multiple items to be verified, and the comprehensive compliance rules into the configuration text of the compliance skill package. This includes: the agent parses the compliance verification requirements to obtain multiple compliance sub-verification requirements; performs declarative definition processing on each compliance sub-verification requirement to obtain items to be verified with the verification content corresponding to the compliance sub-verification requirement as the field name; constructs comprehensive compliance rules based on each verification content and whether each verification content is allowed to exist; injects the text area information corresponding to the compliance verification text in the document to be verified into the verification attachment characters in the configuration text of the compliance skill package; injects the multiple items to be verified into the verification characters in the configuration text according to a preset format; and injects the comprehensive compliance rules into the rule characters in the configuration text.

[0006] In some embodiments, the method of using the first LLM model in the compliance skill package to determine whether each of the items to be verified exists in the text to be verified, and extracting the original text description corresponding to the existing items to be verified to generate a judgment result item, includes: after the agent modifies the configuration text, based on the first compliance verification command in the compliance verification command sequence of the compliance skill package, calling the first LLM model to perform semantic analysis on the text to be verified in combination with each of the items to be verified, indicating the existence of the verification result identifier of the first item to be verified in the text to be verified, and extracting the original text description corresponding to the first item to be verified from the text to be verified; indicating the non-existence of the verification result identifier of the second item to be verified in the text to be verified, and using an empty character as the original text description corresponding to the second item to be verified, and using each item to be verified, each verification result identifier, and each original text description as a judgment result item corresponding to each item to be verified in a format recognizable by the first rule engine.

[0007] In some embodiments, each of the judgment result items is injected back into the configuration text, and the first rule engine in the compliance skill package, combined with the comprehensive compliance rules, is used to comprehensively judge each of the judgment result items to generate a compliance verification result. This includes: injecting each of the judgment result items into the semantic recognition result characters in the configuration text in a set form, and marking the completion of the first compliance verification command; after the agent recognizes the completion of the first compliance verification command, it calls the first rule engine to obtain each of the judgment result items from the configuration text based on the second compliance verification command in the order of the compliance verification commands; if the number of judgment result items is equal to the number of judgment result items in the comprehensive compliance rules, the first rule engine extracts and verifies the judgments sequentially according to the order of the judgment result items in the comprehensive compliance rules. The target verification result identifier in the target judgment result item corresponding to the item is used to determine the sub-compliance verification result based on the verification judgment item and the target verification result identifier. If the sub-compliance verification result is not passed, the compliance verification result is determined to be not passed and the first reason for not passing is determined. The verification extraction based on the verification judgment item sequence is repeated, and the target verification result identifier in the target judgment result item corresponding to the next verification judgment item is determined based on the next verification judgment item and the target verification result, until the next verification judgment item is the last verification judgment item in the verification judgment sequence. If multiple sub-compliance verification results are all passed, the compliance verification result is determined to be passed. If there is a sub-compliance verification result that is uncertain, the compliance verification result is determined to be unknown, and an unknown reason is generated.

[0008] In some embodiments, the step of extracting the extraction result set corresponding to each field to be extracted from the set of texts to be validated for consistency verification based on the consistency skill package, text priority rules, and extraction and merging rules includes: after the agent injects each field to be extracted and the set of texts to be validated for consistency verification of the document to be validated into the consistency skill package, based on the first consistency verification command in the consistency verification command sequence of the consistency skill package, calling the second LLM model in conjunction with the text priority rules to select at least two target validation texts from the set of texts to be validated for consistency verification; for identical texts to be validated for consistency verification in the set of texts to be validated for consistency verification, the text to be validated for consistency verification with higher text version priority is taken as the target validation text of the identical texts to be validated for consistency verification. The different texts to be validated in the text set are used as the target validation texts; the second LLM model extracts the field content corresponding to each of the fields to be extracted from at least two target validation texts, and uses the target validation text corresponding to the field content as the field source of the field content. If there are multiple identical target validation texts with the same text version priority, the multiple field contents corresponding to the field to be extracted are merged according to the field type of the field to be extracted and combined with the extraction and merging rules to obtain the field content corresponding to the field to be extracted. Each field content, each field source, and each field to be extracted are used as each comparison item, and the multiple comparison items corresponding to the same field to be extracted are used as the extraction result set corresponding to the field to be extracted.

[0009] In some embodiments, the step of selecting a comparison strategy corresponding to the field type of each of the fields to be extracted to verify the extraction result set and generate a consistency verification result includes: for each field to be extracted, determining a judgment engine and a comparison strategy based on the field type of the field to be extracted; if the field type is numeric, boolean, or structured list, then determining to use a second rule engine combined with a comparison strategy corresponding to the field type to perform consistency judgment on the extraction result set corresponding to the field to be extracted, generating a sub-consistency result; if the field type is a one-to-many detail list, then determining to use the second rule engine combined with a detail list comparison strategy to perform consistency judgment on the extraction result set corresponding to the field to be extracted. A sub-consistency result is generated. If the field type is a many-to-many detail list, the second LLM model combined with a semantic splitting comparison strategy is used to determine the consistency of the extraction result set corresponding to the field to be extracted, and a sub-consistency result is generated. If the field type is time-based, the second LLM model combined with a time period comparison strategy is used to determine the consistency of the extraction result set corresponding to the field to be extracted, and a sub-consistency result is generated. If there is a sub-consistency result that fails, the consistency verification result is failed, and a second reason for failure is generated. If all the sub-consistency results pass, the consistency verification result is passed. If there is a sub-consistency result that is uncertain, the consistency verification result is unknown, and an unknown reason is generated.

[0010] In some embodiments, generating the verification result of the document to be verified based on the compliance verification result and the consistency verification result includes: the agent reviews the compliance verification result and the consistency verification result based on a two-dimensional cross-decision matrix to generate the verification result, the verification result including agreement, disagreement, and manual confirmation; the verification result is passed if and only if the compliance verification result is passed and the consistency verification result is passed; if the verification result is manually confirmed, if a first unknown reason is obtained from the compliance verification result, the anomaly judgment result item of the first unknown reason and the comprehensive compliance rule are sent to relevant personnel, so that the relevant personnel can determine whether the document to be verified is compliant based on the original description in the anomaly judgment result, the document to be verified, and the comprehensive compliance rule; if a second unknown reason is obtained from the consistency verification result, multiple comparison fragments corresponding to the anomaly extraction field of the second unknown reason are sent to the relevant personnel, so that the relevant personnel can perform consistency verification on the multiple comparison fragments and determine whether the set of documents to be verified is consistent.

[0011] Secondly, this application provides a contract intelligent dual-track verification device based on a skill package, comprising: an acquisition unit, used to acquire compliance verification requirements and consistency verification requirements of a document to be verified; an intelligent agent parses the compliance verification requirements to obtain multiple items to be verified and comprehensive compliance rules; and injects the compliance verification text of the document to be verified, the multiple items to be verified, and the comprehensive compliance rules into the configuration text of a compliance skill package; and a compliance verification unit, used to determine whether each item to be verified exists in the compliance verification text using a first LLM model in the compliance skill package, and extract the original text description corresponding to the existing items to be verified to generate a judgment result item; injecting each judgment result item back into the configuration text; and using the first rule in the compliance skill package... The engine then combines the comprehensive compliance rules to make a comprehensive judgment on each of the judgment result items and generates a compliance verification result. The consistency verification unit is used by the agent to decompose the consistency verification requirements to obtain multiple fields to be extracted. Each field to be extracted and the text set to be consistent verified of the document to be verified are injected into the consistency skill package. Based on the consistency skill package, text priority rules and extraction and merging rules, the extraction result set corresponding to each field to be extracted is extracted from the text set to be consistent verified. The extraction result set is then verified by selecting a comparison strategy corresponding to the field type of each field to be extracted, generating a consistency verification result. Based on the compliance verification result and the consistency verification result, the verification result of the document to be verified is generated.

[0012] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the skill pack-based intelligent dual-track verification method for contracts.

[0013] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the skill pack-based intelligent dual-track verification method for contracts.

[0014] Fifthly, this application provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the skill pack-based intelligent dual-track verification method for contracts.

[0015] As can be seen from the above technical solution, this application provides a contract intelligent dual-track verification method and device based on skill packages. The verification of the document to be verified is split into two parallel tracks: compliance verification and consistency verification. Compliance verification is handled by a compliance skill package, and consistency verification is handled by a consistency skill package. Utilizing independent dual-track verification (consistency verification package verification and compliance verification package verification) allows compliance verification and consistency verification to be performed in parallel, improving verification efficiency. The intelligent agent parses the compliance verification requirements to obtain multiple items to be verified and comprehensive compliance rules. The text to be verified, the multiple items to be verified, and the comprehensive compliance rules are injected into the configuration text of the compliance skill package. This organizes the originally unstructured text to be verified into a structured input that can be processed by the first LLM model and the first rule engine, directly improving the executability and accuracy of the compliance verification. Specifically, the first LLM model is used to determine the existence of the item to be verified and extract the original text description. The first rule engine combines the comprehensive compliance rules for a comprehensive judgment, organically combining the semantic understanding capability of the LLM model with the deterministic judgment capability of the rule engine. For the compliance verification of unstructured text to be verified, it retains the flexibility of semantic recognition while ensuring the controllability and interpretability of the comprehensive judgment. On the consistency track, text priority rules and extraction and merging rules are introduced, and corresponding comparison strategies are selected according to field type. This allows texts to be verified from different versions and attachments to be optimized, merged, and then compared, thus adapting to the complex scenarios of multiple versions and attachments in contract practice. Finally, the verification result is generated based on the compliance verification result and the consistency verification result. It can be seen that this solution proposes a fully automated verification process based on intelligent agent automatic injection of configuration files - dual-track verification - automatic decision-making of dual-track verification results, which greatly improves the efficiency, accuracy, and automation of text verification. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the intelligent dual-track verification method for contracts based on skill packs in an embodiment of this application. Figure 2 This is a schematic diagram of the two-dimensional cross-decision matrix in an embodiment of this application; Figure 3 This is a structural diagram of the smart dual-track verification device for contracts based on skill packs in an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0020] To address the problems existing in the prior art, this application provides a dual-track intelligent contract verification method and apparatus based on skill packages. The verification of the document to be verified is split into two parallel tracks: compliance verification and consistency verification. Compliance verification is handled by a compliance skill package, and consistency verification by a consistency skill package. Utilizing independent dual-track verification (consistency verification package verification and compliance verification package verification) allows compliance and consistency verification to be performed in parallel, improving verification efficiency. An intelligent agent parses the compliance verification requirements to obtain multiple verification items and comprehensive compliance rules. The text to be verified, the multiple verification items, and the comprehensive compliance rules are injected into the configuration text of the compliance skill package. This organizes the originally unstructured text into a structured input that can be processed by the first LLM model and the first rule engine, directly improving the executability and accuracy of the compliance verification. Specifically, the first LLM model is used to determine the existence of the item to be verified and extract the original text description. The first rule engine combines the comprehensive compliance rules for a comprehensive judgment, organically combining the semantic understanding capability of the LLM model with the deterministic judgment capability of the rule engine. For the compliance verification of unstructured text to be verified, it retains the flexibility of semantic recognition while ensuring the controllability and interpretability of the comprehensive judgment. On the consistency track, text priority rules and extraction and merging rules are introduced, and corresponding comparison strategies are selected according to field type. This allows texts to be verified from different versions and attachments to be optimized, merged, and then compared, thus adapting to the complex scenarios of multiple versions and attachments in contract practice. Finally, the verification result is generated based on the compliance verification result and the consistency verification result. It can be seen that this solution proposes a fully automated verification process based on intelligent agent automatic injection of configuration files - dual-track verification - automatic decision-making of dual-track verification results, which greatly improves the efficiency, accuracy, and automation of text verification.

[0021] To improve the accuracy of contract review, this application provides an embodiment of a skill-based intelligent dual-track contract verification method, see [link to relevant documentation]. Figure 1 The intelligent dual-track verification method for contracts based on skill packs specifically includes the following: Step S110: Obtain the compliance verification requirements and consistency verification requirements of the document to be verified. The intelligent agent parses the compliance verification requirements to obtain multiple items to be verified and comprehensive compliance rules. The compliance verification text of the document to be verified, the multiple items to be verified, and the comprehensive compliance rules are injected into the configuration text of the compliance skill package.

[0022] The document to be verified refers to a contract document that needs to be reviewed and verified. This document may contain multiple versions and attachments. Attachments can include approval forms, quotations, technical attachment price lists, technical specifications, performance plans, acceptance standard documents, and other materials related to the contract text. Compliance verification requirements refer to the verification requirements for the document to be verified in terms of meeting the compliance requirements of the contract's clauses and elements. These requirements determine whether the content of the contract text in the document to be verified complies with regulations. Consistency verification requirements refer to the verification requirements that the contract text and attachments in the document to be verified must maintain consistency. The intelligent agent parses the verification requirements (including compliance and consistency verification requirements) and, based on the parsing results, calls the corresponding skill packages (compliance verification requirements call the compliance skill package, consistency verification requirements call the consistency skill package). This activates the LLM model and rule engine within the skill package to intelligently and automatically execute the verification tasks corresponding to the verification requirements (consistency verification tasks or compliance verification tasks). Combining the verification results from multiple skill packages, the intelligent agent automatically determines the verification result of the document to be verified. The intelligent agent is the main entity that automates the intelligent dual-track verification of contract consistency and compliance. "Items to be verified" refers to specific verification objects obtained after parsing the compliance verification requirements and needing to be determined to exist in the text to be verified. "Comprehensive compliance rules" refers to rules constructed based on each item to be verified and its permissible constraints, used to determine the compliance verification results. "Text to be verified" refers to the contract text in the document to be verified used for compliance verification. For example, if the compliance verification requirement is to check whether there are non-compliant clauses such as "lifetime free upgrade guarantee" or "no concurrency limit" in text 1 to be verified, then "lifetime free upgrade clause" and "no concurrency limit clause" are two items to be verified. The comprehensive compliance rule states that if either "lifetime free upgrade clause" or "no concurrency limit clause" exists in the text, the compliance verification is considered to have failed. "Compliance skill package" refers to a functional unit that encapsulates the configuration text, verification command sequence, LLM model call command, and rule engine call command required for compliance verification. "Configuration text" refers to the text carrier in the compliance skill package used to carry the text to be verified, items to be verified, comprehensive compliance rules, and subsequent judgment result annotation content.

[0023] Specifically, in step S110, the agent parses the compliance verification requirements to obtain multiple items to be verified and comprehensive compliance rules. The agent then injects the compliance verification text of the document to be verified, the multiple items to be verified, and the comprehensive compliance rules into the configuration text of the compliance skill package. This includes: the agent parses the compliance verification requirements to obtain multiple compliance sub-verification requirements; declaratively defines each compliance sub-verification requirement to obtain items to be verified with the verification content corresponding to the compliance sub-verification requirement as the field name; constructs comprehensive compliance rules based on each verification content and whether each verification content is allowed to exist; injects the text area information corresponding to the compliance verification text in the document to be verified into the verification attachment characters in the configuration text of the compliance skill package; injects the multiple items to be verified into the verification characters in the configuration text according to a preset format; and injects the comprehensive compliance rules into the rule characters in the configuration text.

[0024] Among them, compliance sub-verification requirements refer to indivisible individual verification requirements obtained after parsing compliance verification requirements. For example, if the compliance verification requirement is to check whether there are non-compliant clauses such as a lifetime free upgrade guarantee or no concurrency limit in the text to be verified (1), then the first compliance sub-verification requirement is to check whether there are non-compliant clauses such as a lifetime free upgrade guarantee in the text to be verified, and the second compliance sub-verification requirement is to check whether there are non-compliant clauses such as no concurrency limit in the text to be verified. Declarative definition processing refers to the processing method of using the verification content corresponding to each compliance sub-verification requirement as the definition field name, so that the field name representing the item to be verified contains the verification content of the compliance sub-verification requirement, thereby expressing the compliance sub-verification requirement in a declarative form. For example, a compliance sub-verification requirement might refer to checking whether a non-compliant clause guaranteeing lifetime free upgrades exists in the text to be verified. The item to be verified would be "Lifetime Free Upgrade Clause." The field name "Lifetime Free Upgrade Clause" identifies the verification content required for this item: identifying whether the text contains content semantically identical to "Lifetime Free Upgrade Clause." If the text to be verified is about unlimited free upgrades for A, then "unlimited free upgrades" is semantically identical to item 1 of the "Lifetime Free Upgrade Clause," thus confirming the presence of item 1 in the text. The verification content refers to the specific audit content targeted by the compliance sub-verification requirement, and this specific audit content is used as the field name for the item to be verified. Text area information refers to the text name corresponding to the text to be verified in the document to be verified, and its specific location in the contract text corresponding to that text name. This specific location can be indicated by paragraph number, page number, etc., or by the beginning and end text information of the text area. For example, in contract text 1 of the text to be verified, the text area information begins with "Clause 1" and ends with "The above clauses have all been confirmed." Verification characters refer to the character positions in the configuration text used to carry multiple verification items. Rule characters refer to the character positions in the configuration text used to carry comprehensive compliance rules. Verification attachment characters refer to the character positions in the configuration text used to carry the text area information corresponding to the text to be verified. Preset format refers to the pre-agreed format followed when injecting multiple verification items into the configuration text.For example, if the character to be verified is represented as "concepts", and multiple items to be verified are injected into the configuration text, it is represented as: `concepts{item to be verified 1{"exists": bool, "original description": string}, item to be verified 2{"exists": bool, "original description": string}}`. Here, "exists" indicates whether the corresponding item to be verified exists in the contract text. `bool` is a data type; if the corresponding item to be verified exists in the contract text, `bool` is "yes". "Original description" refers to the description of the corresponding item to be verified in the contract text. For example, if item to be verified 1 is "lifetime free upgrade clause", and the contract text contains the description "unlimited free upgrade", then "unlimited free upgrade" is the "original description" of item to be verified 1. The expression format of item to be verified 1{"exists": bool, "original description": string} is a preset format.

[0025] Therefore, by parsing compliance verification requirements through an intelligent agent, multiple sub-compliance verification requirements are obtained. Each sub-compliance verification requirement is then declaratively defined to produce verification items named after the verification content. Simultaneously, the compliance judgment of the verification content is extracted from the verification items and incorporated into the comprehensive compliance rules. This transforms the verification requirements from business-oriented compliance judgments into standardized, semantic verification items recognizable by the first LLM model, and deterministic comprehensive compliance rules recognizable by the first rule engine. This achieves the separation of the responsibilities of "semantic recognition" and "compliance judgment." Furthermore, based on each verification item and whether its existence is permitted, comprehensive compliance rules are constructed. This explicitly solidifies the semantic compliance constraints of "what content should be present and what content is not allowed" into comprehensive compliance rules with deterministic judgment items as their carriers. This allows the first rule engine to make deterministic judgments based on the recognition results of the first LLM model, thus fully integrating the semantic recognition capabilities of the first LLM model with the deterministic judgment capabilities of the first rule engine. This solution declaratively defines compliance sub-verification requirements as fields to be verified, with the verification content as the field name. The same set of declarations is used to automatically map and generate the recognition target of the first LLM model and the comprehensive compliance rules of the first rule engine. This ensures that "what the LLM extracts" and "what the rules judge" originate from the same declarative definition and naturally maintain a one-to-one correspondence. This fundamentally eliminates the hidden danger of mismatch between the extraction target and the rule conditions, laying the foundation for the deep integration of LLM model semantic recognition and rule engine compliance judgment.

[0026] Furthermore, since the item to be verified only bears the semantic recognition responsibility of "whether the content exists," while the judgment of "whether it exists or not is compliant" is uniformly handled by the comprehensive compliance rules, the subsequent first LLM model does not need to understand complex compliance logic, but only needs to complete the targeted semantic existence recognition, reducing the dependence of compliance judgment on the stability of the output of the first LLM model. The comprehensive compliance rules exist independently in the form of defined judgment items, which indirectly makes the compliance judgment standards independently maintainable, reusable and auditable. When compliance requirements are adjusted, only the corresponding judgment items in the comprehensive compliance rules need to be modified, without changing the definition of the item to be verified or the calling logic of the first LLM model, thereby enhancing the configurability and maintainability of compliance verification in the scenario of changing requirements. Furthermore, since the items to be verified are defined declaratively, business personnel can automatically map new items to be verified and new comprehensive compliance rules simply by adding or deleting the corresponding compliance sub-verification requirements when expanding or modifying verification needs. This is without having to modify the semantic recognition logic of the first LLM model or the underlying code of the first rule engine. This enhances the scalability and configurability of compliance verification rules and items, enabling compliance verification to quickly adapt to the differentiated compliance requirements of different contract types and business scenarios in a low-cost and low-modification manner. It avoids the problems of inconsistent compliance configuration standards, low efficiency, and easy errors caused by relying on the experience of approval personnel to set compliance verification items and rules.

[0027] Step S120: Use the first LLM model in the compliance skill package to determine whether each item to be verified exists in the text to be verified, extract the original text description corresponding to the existing item to be verified to generate a judgment result item, inject each judgment result item back into the configuration text, and use the first rule engine in the compliance skill package in combination with the comprehensive compliance rules to make a comprehensive judgment on each judgment result item to generate a compliance verification result.

[0028] The first LLM model refers to a large language model used for semantic analysis of the text to be verified, determining the existence of the items to be verified, and extracting the original text description. The judgment result item refers to a structured result unit generated by the first LLM model for each item to be verified, containing its existence or non-existence and the original text description. For example, if item 1 is "lifetime free upgrade clause," and the contract text contains the description "unlimited free upgrades," the judgment result item for item 1 would be: Item 1 {"Exists": Yes, "Original Description": "Unlimited free upgrades"}. The first rule engine refers to a rule processing engine that combines comprehensive compliance rules to comprehensively judge each judgment result item to generate a compliance verification result. The compliance verification result refers to the judgment result made on whether the text to be verified is compliant.

[0029] Specifically, in step S120, the first LLM model in the compliance skill package is used to determine whether each of the items to be verified exists in the text to be verified, and the original text descriptions corresponding to the existing items to be verified are extracted to generate a judgment result item. This includes: after the agent modifies the configuration text, based on the first compliance verification command in the compliance verification command sequence of the compliance skill package, the first LLM model is called to perform semantic analysis on the text to be verified in combination with each of the items to be verified. The verification result identifier of the first item to be verified in the text to be verified is marked as existing, and the original text description corresponding to the first item to be verified is extracted from the text to be verified. The verification result identifier of the second item to be verified that does not exist in the text to be verified is marked as non-existent, and an empty character is used as the original text description corresponding to the second item to be verified. Each item to be verified, each verification result identifier, and each original text description are used as a judgment result item corresponding to each item to be verified in a format recognizable by the first rule engine.

[0030] The compliance verification command sequence refers to the order in which the compliance verification commands are executed sequentially as agreed upon in the compliance skill package. The first compliance verification command is the first command executed in the compliance verification command sequence, used to invoke the first LLM model for semantic analysis. Semantic analysis refers to the semantic understanding and judgment performed by the first LLM model on the compliance verification text based on each item to be verified. The first item to be verified refers to the item to be verified that is determined to exist in the compliance verification text after semantic analysis. The verification result identifier is used to indicate whether the item to be verified exists or not in the compliance verification text. The original text description refers to the original text content extracted from the compliance verification text that corresponds to the existing item to be verified. The second item to be verified refers to the item to be verified that is determined not to exist in the compliance verification text after semantic analysis. The empty character refers to the null value content in the original text description corresponding to the non-existent second item to be verified. The first rule engine-recognizable format refers to organizing each item to be verified, each verification result identifier, and each original text description into a format that the first rule engine can read and process, such as JSON format.

[0031] Therefore, after modifying the configuration text, the agent solidifies the verification execution logic based on the compliance verification command sequence preset by the compliance skill package. It prioritizes driving the first LLM model through the first compliance verification command to complete the accurate semantic analysis of the text to be verified, achieving standardized, batch compliance identification and judgment of all verification items. For existing first verification items, the corresponding original text description is accurately extracted; for non-existent second verification items, empty characters are uniformly matched as the matching result, achieving differentiated and standardized processing of different verification results and avoiding the problem of lack of result traceability in traditional text verification. Simultaneously, each verification item, each verification result identifier, and each original text description is organized into judgment result items according to a format recognizable by the first rule engine. The semantic recognition results output by the LLM model are transformed into structured, uniformly formatted data, indirectly providing directly readable input for the subsequent comprehensive judgment of each judgment result item by the first rule engine in conjunction with comprehensive compliance rules, achieving a deep connection between non-deterministic semantic output and deterministic rule processing. Because the existing items to be verified retain the corresponding original text descriptions, the judgment results are indirectly made traceable and interpretable. When manual confirmation is required later, the specific expressions in the text to be verified can be traced back to the original text descriptions, which enhances the auditability of the compliance verification.

[0032] Specifically, in step S120, each of the aforementioned judgment result items is injected back into the configuration text. The first rule engine in the compliance skill package, combined with the comprehensive compliance rules, performs a comprehensive judgment on each of the judgment result items to generate a compliance verification result. This includes: injecting each of the judgment result items into the semantic recognition result characters in the configuration text in a set format, and marking the completion of the first compliance verification command; after recognizing the completion of the first compliance verification command, the agent, based on the second compliance verification command in the compliance verification command sequence, calls the first rule engine to retrieve each of the judgment result items from the configuration text; if the number of judgment result items is equal to the number of judgment result items in the comprehensive compliance rules, the first rule engine extracts and verifies them sequentially according to the order of the judgment result items in the comprehensive compliance rules. The target verification result identifier in the target judgment result item corresponding to the verification judgment item is used to determine the sub-compliance verification result. If the sub-compliance verification result is not passed, the compliance verification result is determined to be not passed and a first reason for not passing is determined. The verification extraction based on the verification judgment item sequence is repeated, and the target verification result identifier in the target judgment result item corresponding to the next verification judgment item is extracted. The sub-compliance verification result is determined based on the next verification judgment item and the target verification result, until the next verification judgment item is the last verification judgment item in the verification judgment sequence. If multiple sub-compliance verification results are all passed, the compliance verification result is determined to be passed. If there is a sub-compliance verification result that is uncertain, the compliance verification result is determined to be unknown, and an unknown reason is generated.

[0033] The semantic recognition result characters refer to the characters in the configuration text used to carry the judgment result items in a set-based back-annotation format. The second compliance verification command refers to the compliance verification command executed after the first compliance verification command in the compliance verification command sequence, used to call the first rule engine. The verification judgment item order refers to the order in which each verification judgment item in the compliance verification result is processed. The target judgment result item refers to the judgment result item corresponding to the current verification judgment item. The target verification result identifier refers to the verification result identifier extracted from the target judgment result item for judgment. The sub-compliance verification result refers to the local compliance judgment result determined based on a single verification judgment item and the target verification result identifier. The sub-compliance verification result is either pass, fail, or uncertain. The first reason for failure refers to the reason for the failure of the compliance verification result when the sub-compliance verification result is fail. Uncertainty indicates that no target judgment result item matching the judgment result item was found. Unknown reason refers to the reason why the generated compliance verification result is unknown when there is an uncertain sub-compliance verification result. If the number of the judgment result items is not equal to the number of the judgment result items in the comprehensive compliance rule, and the first rule engine is the last item in the verification judgment item, the first rule engine will give an unknown compliance verification result and give the reason that no matching comprehensive compliance rule was found.

[0034] Therefore, after the first compliance verification command completes LLM semantic recognition, the second compliance verification command triggers the rule engine verification. The judgment result items are uniformly injected back into the dedicated semantic recognition result characters in the configuration text in a set form, and the command execution status is marked. This achieves decoupling and state synchronization between the semantic recognition and rule judgment stages, ensuring that the two-level verification process is orderly, controllable, and traceable. Simultaneously, this solution adds a multi-layered anomaly fallback mechanism. For anomalies such as uncertain sub-verification results, mismatch between the number of judgment result items and the comprehensive compliance rules, or the existence of unverified targets, corresponding unknown reasons are generated and unknown verification results are output. This improves the judgment dimensions of compliance verification and breaks through the limitation of traditional contract compliance verification which only supports a binary judgment of pass or fail. It ensures the accuracy and comprehensiveness of compliance verification results in standard scenarios, effectively identifies special scenarios such as data matching anomalies and rule adaptation anomalies, and significantly improves the completeness, fault tolerance, and scenario adaptability of contract compliance verification by providing accurate and detailed anomaly tracing evidence for subsequent manual review. Furthermore, unlike traditional rule engine solutions that can only automatically judge structured data such as form fields by configuring fixed rules, but cannot understand unstructured text content such as contract attachments (for example, they cannot determine whether the main body of the contract contains prohibitive clauses such as "lifetime free upgrades" which only exist in free text and are not reflected in form fields; their capabilities are limited by the degree of data structuring and cannot perform semantic analysis on free text), this solution uses the semantic recognition capabilities of the first LLM model to identify unstructured text content in contract attachments as standardized items to be verified. This enables the first rule engine to perform deterministic compliance judgments on unstructured text that it could not previously handle, breaking through the limitation of traditional rule engines' reliance on the degree of data structuring.

[0035] Step S130: The agent decomposes the consistency verification requirement to obtain multiple fields to be extracted. It injects each field to be extracted and the text set to be verified of the document to be verified into the consistency skill package. Based on the consistency skill package, text priority rules, and extraction and merging rules, it extracts the extraction result set corresponding to each field to be extracted from the text set to be verified. It then selects a comparison strategy corresponding to the field type of each field to be extracted to verify the extraction result set and generates a consistency verification result. Based on the compliance verification result and the consistency verification result, it generates the verification result of the document to be verified.

[0036] Here, "fields to be extracted" refers to the fields whose content needs to be extracted from the text set to be verified, obtained after decomposing the consistency verification requirements. "Text set to be verified" refers to the text set in the document to be verified, containing multiple texts (e.g., multiple versions, multiple attachments) used for consistency verification. "Consistency skill set" refers to the functional unit that encapsulates the command sequence, model call commands, and comparison strategy scheduling logic required for consistency verification. "Text priority rule" refers to the rule used to select identical texts from the text set to be verified according to version priority. "Extraction and merging rule" refers to the rule used to merge the content of multiple text versions corresponding to the same field to be extracted. "Extraction result set" refers to the set of content extracted from the text set to be verified, corresponding to each field to be extracted. "Field type" refers to the data category (e.g., numeric, boolean, or time-based) and data format category (e.g., structured list, one-to-many detail list, many-to-many detail list) of the field to be extracted, used to determine the corresponding comparison strategy. "Comparison strategy" refers to the method used to perform consistency verification on the extraction result set, corresponding to the field type of the field to be extracted. The consistency verification result refers to the judgment made on whether the set of texts to be verified is consistent. The verification result refers to the final audit result of the document to be verified, which is generated based on the compliance verification result and the consistency verification result.

[0037] Specifically, step S130, which involves extracting the set of extraction results corresponding to each field to be extracted from the set of texts to be validated based on the consistency skill package, text priority rules, and extraction and merging rules, includes: after the agent injects each field to be extracted and the set of texts to be validated for consistency verification into the consistency skill package, it calls the second LLM model based on the first consistency verification command in the consistency verification command sequence of the consistency skill package, combined with the text priority rules, to select at least two target verification texts from the set of texts to be validated for consistency verification. For identical texts to be validated for consistency verification in the set of texts to be validated for consistency verification, the text with the higher text version priority is taken as the target verification text of the identical texts to be validated for consistency verification. The different verification texts in the verification text set are taken as the target verification texts; the second LLM model extracts the field content corresponding to each of the fields to be extracted from at least two target verification texts, and takes the target verification text corresponding to the field content as the field source of the field content. If there are multiple identical target verification texts with the same text version priority, the multiple field contents corresponding to the field to be extracted are merged according to the field type of the field to be extracted and combined with the extraction and merging rules to obtain the field content corresponding to the field to be extracted. Each field content, each field source, and each field to be extracted are taken as each comparison item, and the multiple comparison items corresponding to the same field to be extracted are taken as the extraction result set corresponding to the field to be extracted.

[0038] The consistency verification command sequence refers to the order in which the consistency verification commands are executed, as agreed upon in the consistency skill package. The first consistency verification command is the first command executed in the consistency verification command sequence, used to call the second LLM model to select the target verification text. The second LLM model is the large language model used to select the target verification text from the set of texts to be consistent and extract field content. The target verification text is the text selected from the set of texts to be consistent according to text priority rules, used to extract field content for consistency verification, such as contract text or attachments. Text version priority refers to the version number of the text to be consistent, used to determine which version of the text to use for consistency verification among identical texts. Field content refers to the specific content extracted by the second LLM model from the target verification text that corresponds to the field to be extracted. Field source refers to the target verification text corresponding to the field content, used to indicate the source of the field content. The comparison item refers to the comparison unit composed of the field content, field source, and the field to be extracted.

[0039] For example, when the set of texts to be validated contains three texts to be validated (Text 1 and Text 2), namely Text 1 - Initial Version, Text 1 - Revised Version 1, and Text 1 - Revised Version 2, the text priority rule is followed: first the final version, then the revised version, and finally the initial version. Text 1 - Revised Version 1 and Text 1 - Revised Version 2 are selected as two target validation texts, and Text 2 is also included as a target validation text. This set of texts to be validated is used to verify whether the fields to be extracted in Text 1 and Text 2 are consistent. The second LLM model extracts the field content corresponding to each field to be extracted from the three target validation texts and labels which text to be validated corresponds to each field content. After extracting the content of each field to be extracted, for a single field to be extracted containing multiple field contents from different versions of the same text to be validated, an extraction and merging rule matching the field type of the field to be extracted is selected to merge the multiple field contents. For example, if the field type of the field to be extracted is Boolean, then OR merging (or concatenation) is used. OR merging is used to concatenate multiple field contents using the OR operator. For example, the field content corresponding to field 1 to be extracted is "field content 1" OR "field content 2"; if the field type of the field to be extracted is... If the field type is a scalar type, the first non-empty value is selected. That is, the field content that comes first when sorted in ascending order of scalar values ​​(numerical, time, or other scalar values) is selected. For example, if the expected deadline is field 2 and the corresponding field content is 2025.02.03 and 2025.01.09, then 2025.01.09 will be selected as the field content of field 2. If the field type of the field to be extracted is a list type, then all the field content corresponding to the field to be extracted will be concatenated into a non-empty list.

[0040] Therefore, after injecting the fields to be extracted and the set of texts to be consistent, the agent, based on the first consistency verification command, calls the second LLM model to select at least two target verification texts from the set of texts to be consistent. For identical texts to be consistent, the text with the higher version priority is used as the target verification text; for different texts to be consistent, the text with the higher version priority is used as the target verification text. This directly achieves redundancy removal and selection for multiple versions and attachments of text, avoiding repeated comparisons of different versions of the same content that could interfere with consistency judgment, thus improving the accuracy and efficiency of consistency verification. The second LLM model extracts the field content corresponding to each field to be extracted from the at least two target verification texts and uses the corresponding target verification text as the field source, directly ensuring that each field content carries a clear field source, improving the traceability of consistency verification results. In addition, when there are multiple identical target verification texts with the same text version priority, this solution creatively introduces an adaptive merging strategy based on field type (Boolean, scalar, list), which automatically completes the merging and unification of information with standardized logic (such as OR merging, taking the first non-empty value, list concatenation), eliminating ambiguity in content selection among texts of the same level. Finally, the merged results carrying the source of the fields are constructed into a set of comparison items for subsequent verification. This enables the system to not only determine the result when it detects inconsistencies, but also to accurately trace back to the specific source text that caused the difference. This greatly enhances the interpretability of the entire consistency verification process and the ability to locate problems.

[0041] Specifically, step S130, which involves selecting a comparison strategy corresponding to the field type of each field to be extracted to verify the extraction result set and generate a consistency verification result, includes: for each field to be extracted, determining a judgment engine and a comparison strategy based on the field type of the field to be extracted; if the field type is numeric, boolean, or structured list, then determining to use a second rule engine combined with a comparison strategy corresponding to the field type to perform consistency judgment on the extraction result set corresponding to the field to be extracted, generating a sub-consistency result; if the field type is a one-to-many detail list, then determining to use the second rule engine combined with a detail list comparison strategy to perform consistency verification on the extraction result set corresponding to the field to be extracted. The process involves determining whether to generate a sub-consistency result. If the field type is a many-to-many detail list, then the second LLM model combined with a semantic splitting comparison strategy is used to determine the consistency of the extraction result set corresponding to the field to be extracted, and a sub-consistency result is generated. If the field type is time-based, then the second LLM model combined with a time period comparison strategy is used to determine the consistency of the extraction result set corresponding to the field to be extracted, and a sub-consistency result is generated. If there is a sub-consistency result that fails, then the consistency verification result is failed, and a second reason for failure is generated. If all the sub-consistency results pass, then the consistency verification result is passed. If there is a sub-consistency result that is uncertain, then the consistency verification result is unknown, and an unknown reason is generated.

[0042] The decision engine refers to the engine used to perform consistency determination based on the field type of the field to be extracted, including the second rule engine or the second LLM model. Numeric type refers to a field type where the field content is a numerical value. Boolean type refers to a field type where the field content is a binary value such as yes / no or true / false. Structured list type refers to a field type where the field content is a structured list. The second rule engine refers to the rule processing engine used to perform consistency determination on the extracted result set of field types such as numeric, Boolean, structured list, or one-to-many detail list. A sub-consistency result refers to the local result generated by performing consistency determination on the extracted result set of a single field to be extracted. A one-to-many detail list refers to a list-type field type where one subject corresponds to multiple details. The detailed list comparison strategy refers to the comparison strategy corresponding to the one-to-many detailed list, used to determine the consistency of its extracted result set. Specifically, the detailed list comparison strategy refers to comparing field by field according to the matching key. For example, for product quantity and production qualification rate in the detailed list, the comparison uses product quantity as the matching key to compare whether the product quantity in multiple field sources is consistent, and at the same time, it uses production qualification rate as the matching key to compare whether the product quantity in multiple field sources is consistent. A many-to-many detailed list refers to a list-type field type where multiple subjects and multiple details have a many-to-many relationship. The semantic splitting comparison strategy refers to the comparison strategy corresponding to the many-to-many detailed list, executed by the second LLM model, and used to determine consistency through semantic splitting. Time-type refers to a field type where the field content is time or period. The period comparison strategy refers to the comparison strategy corresponding to the time-type, executed by the second LLM model, used to determine the consistency of the period. Specifically, the time period comparison strategy refers to normalizing the time unit and using the normalized field content for consistency comparison. For example, "five years" and "60 months" are uniformly converted to months, so "5 years" is also converted to "60 months". The second reason for failure refers to the reason why the generated consistency check result fails when there are sub-consistency results that fail.

[0043] Specifically, the semantic segmentation and comparison strategy refers to pre-injecting domain knowledge into the second LLM model, enabling it to understand the segmentation relationships, conversion rules, and cross-line calculation capabilities within the text. This allows the second LLM model to perform correct consistency comparisons of complex lists such as many-to-many detail lists. Segmentation relationships refer to how a single line of general description in a contract text is broken down into multiple lines of specific details in a quotation. Conversion rules refer to the second LLM model's ability to understand the mathematical relationships between different ways of expressing the same thing in different texts. Execution calculation capability refers to the second LLM model's ability to automatically perform mathematical summations based on segmentation relationships and conversion rules. For example, the text to be compared includes text A (the main contract): "Purchase 'A8-N Multi-Organization Edition', with a license size of 1500 concurrent connections." The second LLM model extracts the following field content for field 1: {Product: "A8-N Multi-Organization Edition", Size: "1500 concurrent connections"} - the field source is text A. The text to be compared also includes text B (quotation): which lists the specific charges in two lines: A8-N multi-organization version basic concurrent license: 1 set (80 concurrent); A8-N multi-organization version additional concurrent license: 1420 concurrent. The content of field 1 to be extracted from the second LLM model is a list containing two lines of details, and the field source is text B. At this point, the second LLM model, based on domain knowledge, understands that there is a splitting relationship for field 1 to be extracted from text A. From the second LLM model's perspective, this splitting relationship can be understood as: "I (the second LLM model) know that the product 'A8-N Multi-Organization Edition' can be sold through a combination of 'Basic License' and 'Add-on Licenses.' Therefore, the two lines of details in text B are likely splits of the total amount in text A, and I need to treat them as a whole for comparison." Simultaneously, from the second LLM model's perspective, the conversion rule it understands is: "In the first line of text B, I (the second LLM model) see that the unit of 'Basic Concurrent License' is 'Set,' and the scale is '80 Concurrent.' This defines a conversion relationship: 1 Set of Basic License = 80 Concurrent. At this point, I calculate the number of concurrent connections provided by the Basic License: 1..." 80 = 80. Finally, the second LLM model, based on the previously understood splitting relationship and conversion rules, performs cross-line calculations, adding the concurrency provided by the basic license (80) and the concurrency provided by the additional license (1420), representing it as "80 + 1420 = 1500 concurrency". This 1500 concurrency is then used as the field content of the field to be extracted, 1, for text B. The 1500 concurrency corresponding to the field to be extracted, 1, for text B is compared with the 1500 concurrency corresponding to the field to be extracted, 1, for text A, and the sub-consistency result of the field to be extracted, 1, is determined to be successful.

[0044] For example, if the field type is numeric, the comparison strategy is a numeric comparison strategy. This strategy first removes commas and spaces, converts the numerical units of the text description into numerical values ​​(e.g., converting "10,000" to "10,000"), and then converts percentages to decimals (e.g., converting 20% ​​to 0.2). Consistency comparison is then performed based on the converted field content. If the field type is Boolean, a two-value comparison strategy is used, mapping "compliant," "checkmark," and "yes" to "yes," and mapping "discompliant," "cross," and "no" to "no." Consistency comparison is then performed based on the mapped field content. If the field type is a structured list, a payment agreement comparison strategy is used, performing a node-by-node comparison according to each row or column of the structured list.

[0045] Therefore, by constructing a field-type-driven adaptive hybrid engine comparison architecture, the solution systematically solves the technical challenge of single comparison technologies failing to balance the judgment efficiency of simple fields with the semantic understanding depth of complex fields, achieving accurate, efficient, and in-depth coverage of consistency verification for multiple types of contract fields. Its core benefit lies in the organic combination of dynamic routing and strategy specialization: for fields with clearly defined rules, such as numeric, boolean, and structured lists, the solution automatically routes them to the second rule engine. Through preset precise conversion and matching logic (such as numeric unit normalization, boolean semantic mapping, and list node-by-node comparison), efficient and deterministic comparison is performed, ensuring zero illusion and high reliability for basic comparison tasks. For complex comparison scenarios rich in business logic and deep semantics, such as one-to-many detail lists, many-to-many detail lists, and time-type fields, the solution intelligently schedules the second LLM model and configures it with dedicated advanced comparison strategies. In particular, by injecting domain knowledge into the comparison of many-to-many detail lists, the second LLM model evolved to understand product splitting relationships, master quantity conversion rules (such as "1 set = 80 concurrent transactions"), and perform cross-line mathematical summation. This truly achieved a leap from "superficial textual similarity" to "substantive content consistency," overcoming the persistent problem of traditional methods struggling to handle heterogeneous detail comparisons. Simultaneously, for the time-based field period comparison strategy, unit normalization eliminated misjudgments caused by differences in expression such as "five years" and "60 months." Utilizing a hybrid scheduling mechanism of "simple field rule-based approach and complex field semantic approach," the efficiency of the rule engine and the intelligence of the LLM were maximized while ensuring overall verification accuracy, ultimately forming a comprehensive consistency verification system that combines breadth, accuracy, and depth. Furthermore, since the many-to-many detail list adopts the second LLM model combined with semantic splitting and comparison strategy, and the time type adopts the second LLM model combined with the period comparison strategy, the consistency judgment scenarios that are difficult to exhaustively enumerate with fixed rules, such as cross-line splitting and summation of quotation details and different period expression forms, can be handled with the semantic understanding capability of the second LLM model, making up for the shortcomings of the second rule engine in handling unstructured content that requires semantic conversion.

[0046] Specifically, in step S130 above, generating the verification result of the document to be verified based on the compliance verification result and the consistency verification result includes: the agent reviews the compliance verification result and the consistency verification result based on a two-dimensional cross-decision matrix to generate the verification result, the verification result including agreement, disagreement, and manual confirmation; the verification result is passed if and only if the compliance verification result is passed and the consistency verification result is passed; if the verification result is manually confirmed, if a first unknown reason is obtained from the compliance verification result, the anomaly judgment result item of the first unknown reason and the comprehensive compliance rule are sent to relevant personnel, so that the relevant personnel can determine whether the text to be verified is compliant based on the original description in the anomaly judgment result, the item to be verified, and the comprehensive compliance rule; if a second unknown reason is obtained from the consistency verification result, multiple comparison segments corresponding to the anomaly extraction field of the second unknown reason are sent to the relevant personnel, so that the relevant personnel can perform consistency verification on the multiple comparison segments and determine whether the set of texts to be verified is consistent.

[0047] The two-dimensional cross-decision matrix refers to a decision matrix that generates verification results by cross-checking compliance verification results and consistency verification results as two dimensions. Manual confirmation refers to a state of the verification result, indicating a situation requiring confirmation by relevant personnel. The first unknown reason refers to an unknown reason obtained from the compliance verification result that necessitates manual compliance confirmation, and the agent cannot provide a clear answer regarding compliance approval or disapproval. The anomaly judgment result item refers to the judgment result item corresponding to the first unknown reason that needs to be sent to relevant personnel. Relevant personnel refers to the personnel who receive the anomaly judgment result item or comparison fragment and perform manual confirmation. The second unknown reason refers to an unknown reason obtained from the consistency verification result that necessitates personnel to make a consistency judgment, and the agent cannot provide a clear answer regarding consistency approval or disapproval. The anomaly to be extracted field refers to the field to be extracted corresponding to the second unknown reason. The comparison fragment refers to the content and source of multiple fields corresponding to the anomaly to be extracted field that need to be sent to relevant personnel for consistency verification.

[0048] For example, such as Figure 2As shown, the compliance verification results of the compliance skill package and the consistency verification results of the consistency skill package are sent to the agent's two-dimensional cross-decision matrix. The two-dimensional cross-decision matrix utilizes the principle that "when the consistency verification result fails, regardless of whether the compliance verification result is pass, fail, or unknown, the verification result is disagreement." In other words, the two-dimensional cross-decision matrix prioritizes content correctness over policy compliance. When the content is inconsistent, it indicates that the document to be verified contains errors. In this case, regardless of whether the content complies with the policy, the document to be verified cannot pass the review and needs to be revised. When the content is consistent or its consistency is unknown, the verification result is determined by combining the content compliance of the document to be verified. Specifically, for example... Figure 3 As shown, when both the consistency check and compliance check are passed, the check result is "agree". When both the consistency check and compliance check are failed, the check result is "disagree". When both the consistency check and compliance check are unknown, the check result is "manual confirmation". When both the consistency check and compliance check are unknown, the check result is "manual confirmation" regardless of the result.

[0049] Therefore, the intelligent agent reviews the compliance and consistency verification results based on a two-dimensional cross-decision matrix and generates verification results including pass, fail, and manual confirmation. A verification result is considered passable only if both the compliance and consistency verification results are passable. This directly achieves unified cross-decision for the dual-track verification results and uses a strict "double pass" condition as the threshold for automatic release, ensuring that the document to be verified is only judged as passable when both compliance and consistency requirements are met, avoiding the risk of missed judgments caused by passing only one dimension. Simultaneously, when the verification result is manually confirmed and a first unknown reason is obtained from the compliance verification result, the anomaly judgment result item for the first unknown reason and the comprehensive compliance rules are sent to relevant personnel. This allows relevant personnel to judge whether the text to be verified is compliant based on the original description, the item to be verified, and the comprehensive compliance rules. Indirectly, this allows compliance difficulties that the system cannot automatically determine to be pushed to relevant personnel, while simultaneously providing complete judgment basis such as the original description, the item to be verified, and the comprehensive compliance rules. This reduces the information retrieval burden during manual confirmation and improves the efficiency and accuracy of manual confirmation. When a second unknown cause is obtained from the consistency verification result, multiple comparison segments corresponding to the abnormal fields to be extracted for the second unknown cause are sent to relevant personnel so that they can perform consistency verification on the multiple comparison segments. Indirectly, even consistency-related difficult cases that the system cannot automatically determine are also pushed to relevant personnel along with specific comparison segments, thus forming a human-machine collaborative verification closed loop. That is, most cases that can be automatically determined are handled automatically by the system, and only a small number of unknown cases are handed over to relevant personnel for confirmation. This ensures the reliability of the verification results while indirectly reducing the overall manual input.

[0050] In summary, this solution splits the verification of documents to be verified into two parallel tracks: compliance verification and consistency verification. The compliance skill package handles compliance verification, and the consistency skill package handles consistency verification, respectively. Utilizing dual-track independent verification (consistency verification package verification and compliance verification package verification) allows compliance and consistency verification to proceed in parallel, improving verification efficiency. An intelligent agent parses the compliance verification requirements to obtain multiple verification items and comprehensive compliance rules. These, along with the text to be verified, the multiple verification items, and the comprehensive compliance rules, are injected into the configuration text of the compliance skill package. This organizes the originally unstructured text into a structured input that can be processed by the first LLM model and the first rule engine, directly improving the executability and accuracy of compliance verification. Specifically, the first LLM model is responsible for determining the existence of verification items and extracting their original descriptions. The first rule engine combines these with the comprehensive compliance rules for a comprehensive judgment. This organically combines the semantic understanding capability of the LLM model with the deterministic judgment capability of the rule engine, preserving the flexibility of semantic recognition while ensuring the controllability and interpretability of the comprehensive judgment for compliance verification of unstructured text. On the consistency track, text priority rules and extraction / merging rules are introduced, and corresponding comparison strategies are selected according to field type. This allows texts from different versions and attachments to be compared after being selected and merged, thus adapting to the complex scenarios of multiple versions and attachments coexisting in contract practice. Finally, a verification result is generated based on the compliance verification result and the consistency verification result. Therefore, this solution proposes a fully automated verification process based on intelligent agent automatic configuration file injection, dual-track verification, and automatic decision-making based on dual-track verification results, which greatly improves the efficiency, accuracy, and automation of contract verification.

[0051] To improve the efficiency and accuracy of contract verification, this application provides an embodiment of a skill-pack-based intelligent dual-track contract verification device for implementing all or part of the skill-pack-based intelligent dual-track contract verification method. See [link to embodiment]. Figure 3 The intelligent dual-track verification device for contracts based on skill packs specifically includes the following components: The acquisition unit 10 is used to acquire the compliance verification requirements and consistency verification requirements of the document to be verified. The intelligent agent parses the compliance verification requirements to obtain multiple items to be verified and comprehensive compliance rules, and injects the compliance verification text of the document to be verified, the multiple items to be verified, and the comprehensive compliance rules into the configuration text of the compliance skill package.

[0052] The compliance verification unit 20 is used to determine whether each item to be verified exists in the text to be verified using the first LLM model in the compliance skill package, extract the original text description corresponding to the existing item to be verified to generate a judgment result item, inject each judgment result item back into the configuration text, and use the first rule engine in the compliance skill package in combination with the comprehensive compliance rules to make a comprehensive judgment on each judgment result item to generate a compliance verification result.

[0053] The consistency verification unit 30 is used by the agent to decompose the consistency verification requirements to obtain multiple fields to be extracted. It injects each field to be extracted and the text set to be consistent in the document to be verified into the consistency skill package. Based on the consistency skill package, text priority rules and extraction and merging rules, it extracts the extraction result set corresponding to each field to be extracted from the text set to be consistent in the document to be verified. It then selects a comparison strategy corresponding to the field type of each field to be extracted to verify the extraction result set, generates a consistency verification result, and generates the verification result of the document to be verified based on the compliance verification result and the consistency verification result.

[0054] As described above, the intelligent dual-track contract verification device based on skill packages provided in this application splits the verification of the document to be verified into two parallel tracks: compliance verification and consistency verification. The compliance skill package handles compliance verification, and the consistency skill package handles consistency verification, respectively. Utilizing independent dual-track verification (consistency verification package verification and compliance verification package verification) allows compliance verification and consistency verification to be performed in parallel, improving verification efficiency. The intelligent agent parses the compliance verification requirements to obtain multiple items to be verified and comprehensive compliance rules. The text to be verified, the multiple items to be verified, and the comprehensive compliance rules are injected into the configuration text of the compliance skill package. This organizes the originally unstructured text to be verified into a structured input that can be processed by the first LLM model and the first rule engine, directly improving the executability and accuracy of the compliance verification. Specifically, the first LLM model is used to determine the existence of the item to be verified and extract the original text description. The first rule engine combines the comprehensive compliance rules for a comprehensive judgment, organically combining the semantic understanding capability of the LLM model with the deterministic judgment capability of the rule engine. For the compliance verification of unstructured text to be verified, it retains the flexibility of semantic recognition while ensuring the controllability and interpretability of the comprehensive judgment. On the consistency track, text priority rules and extraction and merging rules are introduced, and corresponding comparison strategies are selected according to field type. This allows texts to be verified from different versions and attachments to be optimized, merged, and then compared, thus adapting to the complex scenarios of multiple versions and attachments in contract practice. Finally, the verification result is generated based on the compliance verification result and the consistency verification result. It can be seen that this solution proposes a fully automated verification process based on intelligent agent automatic injection of configuration files - dual-track verification - automatic decision-making of dual-track verification results, which greatly improves the efficiency, accuracy, and automation of text verification.

[0055] To further illustrate this solution, this application also provides a specific application example of using the above-mentioned skill-pack-based intelligent dual-track verification device to implement the skill-pack-based intelligent dual-track verification method for contracts, which specifically includes the following: In some embodiments, the agent parses the compliance verification requirements to obtain multiple items to be verified and comprehensive compliance rules, and injects the compliance verification text of the document to be verified, the multiple items to be verified, and the comprehensive compliance rules into the configuration text of the compliance skill package. This includes: the agent parses the compliance verification requirements to obtain multiple compliance sub-verification requirements; performs declarative definition processing on each compliance sub-verification requirement to obtain items to be verified with the verification content corresponding to the compliance sub-verification requirement as the field name; constructs comprehensive compliance rules based on each verification content and whether each verification content is allowed to exist; injects the text area information corresponding to the compliance verification text in the document to be verified into the verification attachment characters in the configuration text of the compliance skill package; injects the multiple items to be verified into the verification characters in the configuration text according to a preset format; and injects the comprehensive compliance rules into the rule characters in the configuration text.

[0056] In some embodiments, the method of using the first LLM model in the compliance skill package to determine whether each of the items to be verified exists in the text to be verified, and extracting the original text description corresponding to the existing items to be verified to generate a judgment result item, includes: after the agent modifies the configuration text, based on the first compliance verification command in the compliance verification command sequence of the compliance skill package, calling the first LLM model to perform semantic analysis on the text to be verified in combination with each of the items to be verified, indicating the existence of the verification result identifier of the first item to be verified in the text to be verified, and extracting the original text description corresponding to the first item to be verified from the text to be verified; indicating the non-existence of the verification result identifier of the second item to be verified in the text to be verified, and using an empty character as the original text description corresponding to the second item to be verified, and using each item to be verified, each verification result identifier, and each original text description as a judgment result item corresponding to each item to be verified in a format recognizable by the first rule engine.

[0057] In some embodiments, each of the judgment result items is injected back into the configuration text, and the first rule engine in the compliance skill package, combined with the comprehensive compliance rules, is used to comprehensively judge each of the judgment result items to generate a compliance verification result. This includes: injecting each of the judgment result items into the semantic recognition result characters in the configuration text in a set form, and marking the completion of the first compliance verification command; after the agent recognizes the completion of the first compliance verification command, it calls the first rule engine to obtain each of the judgment result items from the configuration text based on the second compliance verification command in the order of the compliance verification commands; if the number of judgment result items is equal to the number of judgment result items in the comprehensive compliance rules, the first rule engine extracts and verifies the judgments sequentially according to the order of the judgment result items in the comprehensive compliance rules. The target verification result identifier in the target judgment result item corresponding to the item is used to determine the sub-compliance verification result based on the verification judgment item and the target verification result identifier. If the sub-compliance verification result is not passed, the compliance verification result is determined to be not passed and the first reason for not passing is determined. The verification extraction based on the verification judgment item sequence is repeated, and the target verification result identifier in the target judgment result item corresponding to the next verification judgment item is determined based on the next verification judgment item and the target verification result, until the next verification judgment item is the last verification judgment item in the verification judgment sequence. If multiple sub-compliance verification results are all passed, the compliance verification result is determined to be passed. If there is a sub-compliance verification result that is uncertain, the compliance verification result is determined to be unknown, and an unknown reason is generated.

[0058] In some embodiments, the step of extracting the extraction result set corresponding to each field to be extracted from the set of texts to be validated for consistency verification based on the consistency skill package, text priority rules, and extraction and merging rules includes: after the agent injects each field to be extracted and the set of texts to be validated for consistency verification of the document to be validated into the consistency skill package, based on the first consistency verification command in the consistency verification command sequence of the consistency skill package, calling the second LLM model in conjunction with the text priority rules to select at least two target validation texts from the set of texts to be validated for consistency verification; for identical texts to be validated for consistency verification in the set of texts to be validated for consistency verification, the text to be validated for consistency verification with higher text version priority is taken as the target validation text of the identical texts to be validated for consistency verification. The different texts to be validated in the text set are used as the target validation texts; the second LLM model extracts the field content corresponding to each of the fields to be extracted from at least two target validation texts, and uses the target validation text corresponding to the field content as the field source of the field content. If there are multiple identical target validation texts with the same text version priority, the multiple field contents corresponding to the field to be extracted are merged according to the field type of the field to be extracted and combined with the extraction and merging rules to obtain the field content corresponding to the field to be extracted. Each field content, each field source, and each field to be extracted are used as each comparison item, and the multiple comparison items corresponding to the same field to be extracted are used as the extraction result set corresponding to the field to be extracted.

[0059] In some embodiments, the step of selecting a comparison strategy corresponding to the field type of each of the fields to be extracted to verify the extraction result set and generate a consistency verification result includes: for each field to be extracted, determining a judgment engine and a comparison strategy based on the field type of the field to be extracted; if the field type is numeric, boolean, or structured list, then determining to use a second rule engine combined with a comparison strategy corresponding to the field type to perform consistency judgment on the extraction result set corresponding to the field to be extracted, generating a sub-consistency result; if the field type is a one-to-many detail list, then determining to use the second rule engine combined with a detail list comparison strategy to perform consistency judgment on the extraction result set corresponding to the field to be extracted. A sub-consistency result is generated. If the field type is a many-to-many detail list, the second LLM model combined with a semantic splitting comparison strategy is used to determine the consistency of the extraction result set corresponding to the field to be extracted, and a sub-consistency result is generated. If the field type is time-based, the second LLM model combined with a time period comparison strategy is used to determine the consistency of the extraction result set corresponding to the field to be extracted, and a sub-consistency result is generated. If there is a sub-consistency result that fails, the consistency verification result is failed, and a second reason for failure is generated. If all the sub-consistency results pass, the consistency verification result is passed. If there is a sub-consistency result that is uncertain, the consistency verification result is unknown, and an unknown reason is generated.

[0060] In some embodiments, generating the verification result of the document to be verified based on the compliance verification result and the consistency verification result includes: the agent reviews the compliance verification result and the consistency verification result based on a two-dimensional cross-decision matrix to generate the verification result, the verification result including agreement, disagreement, and manual confirmation; the verification result is passed if and only if the compliance verification result is passed and the consistency verification result is passed; if the verification result is manually confirmed, if a first unknown reason is obtained from the compliance verification result, the anomaly judgment result item of the first unknown reason and the comprehensive compliance rule are sent to relevant personnel, so that the relevant personnel can determine whether the document to be verified is compliant based on the original description in the anomaly judgment result, the document to be verified, and the comprehensive compliance rule; if a second unknown reason is obtained from the consistency verification result, multiple comparison fragments corresponding to the anomaly extraction field of the second unknown reason are sent to the relevant personnel, so that the relevant personnel can perform consistency verification on the multiple comparison fragments and determine whether the set of documents to be verified is consistent.

[0061] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the skill pack-based smart dual-track verification method for contracts.

[0062] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described intelligent dual-track verification method for contracts based on skill packs.

[0063] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described intelligent dual-track verification method for contracts based on skill packages.

[0064] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0065] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0066] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0067] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0068] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart dual-track verification method for contracts based on skill packages, characterized in that, The method includes: Obtain the compliance verification requirements and consistency verification requirements of the document to be verified. The intelligent agent parses the compliance verification requirements to obtain multiple items to be verified and comprehensive compliance rules. The compliance verification text of the document to be verified, the multiple items to be verified, and the comprehensive compliance rules are injected into the configuration text of the compliance skill package. The first LLM model in the compliance skill package is used to determine whether each item to be verified exists in the text to be verified for compliance, and the original text description corresponding to the existing item to be verified is extracted to generate a judgment result item. Each judgment result item is injected back into the configuration text. The first rule engine in the compliance skill package is used in combination with the comprehensive compliance rules to make a comprehensive judgment on each judgment result item to generate a compliance verification result. The intelligent agent decomposes the consistency verification requirement to obtain multiple fields to be extracted. It injects each field to be extracted and the text set to be consistent in the document to be verified into a consistency skill package. Based on the consistency skill package, text priority rules, and extraction and merging rules, it extracts the extraction result set corresponding to each field to be extracted from the text set to be consistent in the document. It then selects a comparison strategy corresponding to the field type of each field to be extracted to verify the extraction result set, generates a consistency verification result, and generates the verification result of the document to be verified based on the compliance verification result and the consistency verification result.

2. The intelligent dual-track verification method for contracts based on skill packages according to claim 1, characterized in that, The intelligent agent parses the compliance verification requirements to obtain multiple items to be verified and comprehensive compliance rules. It then injects the compliance verification text of the document to be verified, the multiple items to be verified, and the comprehensive compliance rules into the configuration text of the compliance skill package, including: The intelligent agent parses the compliance verification requirements to obtain multiple compliance sub-verification requirements. Each compliance sub-verification requirement is defined declaratively to obtain a field to be verified with the verification content corresponding to the compliance sub-verification requirement as the field name. A comprehensive compliance rule is constructed based on each verification content and whether each verification content is allowed to exist. The text area information corresponding to the compliance verification text in the document to be verified is injected into the verification attachment characters in the configuration text of the compliance skill package. Multiple items to be verified are injected into the verification characters in the configuration text according to a preset format. The comprehensive compliance rules are injected into the rule characters in the configuration text.

3. The intelligent dual-track verification method for contracts based on skill packages according to claim 1, characterized in that, The first LLM model in the compliance skills package is used to determine whether each of the items to be verified exists in the text to be verified, and the original text descriptions corresponding to the existing items to be verified are extracted to generate a judgment result item, including: After modifying the configuration text, the intelligent agent calls the first LLM model to perform semantic analysis on the text to be verified based on the first compliance verification command in the compliance verification command sequence of the compliance skill package, and combines each of the items to be verified to identify the existence of the first item to be verified in the text to be verified. The agent also extracts the original text description corresponding to the first item to be verified from the text to be verified. The verification result identifier of the second item to be verified that does not exist in the text to be verified is indicated as not existing, and an empty character is used as the original text description corresponding to the second item to be verified. Each item to be verified, each verification result identifier, and each original text description are used as the judgment result item corresponding to each item to be verified in a format that can be recognized by the first rule engine.

4. The intelligent dual-track verification method for contracts based on skill packages according to claim 3, characterized in that, Each of the aforementioned judgment result items is fed back into the configuration text. The first rule engine in the compliance skill package, combined with the comprehensive compliance rules, is used to perform a comprehensive judgment on each of the aforementioned judgment result items, generating a compliance verification result, including: Each of the aforementioned judgment result items is fed back into the semantic recognition result characters in the configuration text in a set form, and the execution of the first compliance verification command is marked as completed. After the agent recognizes that the first compliance verification command has been executed, it calls the first rule engine to obtain each of the aforementioned judgment result items from the configuration text based on the second compliance verification command in the sequence of compliance verification commands. If the number of the judgment result items is equal to the number of the judgment result items in the comprehensive compliance rule, the first rule engine extracts the target verification result identifier from the target judgment result item corresponding to the verification judgment item in the order of the judgment result items in the comprehensive compliance rule, and determines the sub-compliance verification result based on the verification judgment item and the target verification result identifier. If the sub-compliance verification result is not passed, the compliance verification result is determined to be not passed and the first reason for not passing is determined. Repeatedly perform the sequential verification extraction based on the verification decision item, extract the target verification result identifier in the target decision result item corresponding to the next verification decision item, determine the sub-compliance verification result based on the next verification decision item and the target verification result, until the next verification decision item is the last verification decision item in the verification decision sequence. If multiple sub-compliance verification results are all passed, then the compliance verification result is determined to be passed. If there is a sub-compliance verification result that is uncertain, then the compliance verification result is determined to be unknown, and an unknown reason is generated.

5. The intelligent dual-track verification method for contracts based on skill packages according to claim 1, characterized in that, The step of extracting the set of extraction results corresponding to each field to be extracted from the set of texts to be validated for consistency based on the consistency skill package, text priority rules, and extraction and merging rules includes: After the agent injects the fields to be extracted and the set of texts to be validated of the document to be validated into the consistency skill package, it calls the second LLM model and combines the text priority rules to select at least two target validation texts from the set of texts to be validated. For the same texts to be validated in the set of texts to be validated, the text with higher text version priority is taken as the target validation text of the same texts to be validated. The different texts to be validated in the set of texts to be validated are taken as the target validation texts. The second LLM model extracts the field content corresponding to each of the fields to be extracted from at least two target verification texts, and uses the target verification text corresponding to the field content as the field source of the field content. If there are multiple identical target verification texts with the same text version priority, the multiple field contents corresponding to the field to be extracted are merged according to the field type of the field to be extracted and combined with the extraction and merging rules to obtain the field content corresponding to the field to be extracted. Each field content, each field source, and each field to be extracted are used as comparison items, and multiple comparison items corresponding to the same field to be extracted are used as the extraction result set corresponding to the field to be extracted.

6. The intelligent dual-track verification method for contracts based on skill packages according to claim 1, characterized in that, The step of selecting a comparison strategy corresponding to the field type of each of the fields to be extracted to verify the extraction result set and generate a consistency verification result includes: For each of the fields to be extracted, a judgment engine and comparison strategy are determined based on the field type of the field to be extracted. If the field type is numeric, boolean, or structured list, then the second rule engine is used in combination with the comparison strategy corresponding to the field type to perform consistency judgment on the extraction result set corresponding to the field to be extracted, and a sub-consistency result is generated. If the field type is a one-to-many detail list, then the second rule engine combined with the detail list comparison strategy is used to determine the consistency of the extraction result set corresponding to the field to be extracted, and a sub-consistency result is generated. If the field type is a many-to-many detail list, then the second LLM model combined with the semantic splitting comparison strategy is used to determine the consistency of the extraction result set corresponding to the field to be extracted, and a sub-consistency result is generated. If the field type is time-based, then the second LLM model combined with the period comparison strategy is used to determine the consistency of the extraction result set corresponding to the field to be extracted, and a sub-consistency result is generated. If there is a sub-consistency result that fails, the consistency verification result is failed, and a second reason for failure is generated. If all the sub-consistency results pass, the consistency verification result is passed. If there is a sub-consistency result that is uncertain, the consistency verification result is unknown, and an unknown reason is generated.

7. The intelligent dual-track verification method for contracts based on skill packages according to claim 1, characterized in that, The step of generating the verification result of the document to be verified based on the compliance verification result and the consistency verification result includes: The intelligent agent reviews the compliance verification result and the consistency verification result based on a two-dimensional cross-decision matrix, and generates the verification result, which includes agreement, disagreement, and manual confirmation. The verification result is passed if and only if the compliance verification result is passed and the consistency verification result is passed. If the verification result is manually confirmed, and a first unknown reason is obtained from the compliance verification result, the anomaly judgment result item of the first unknown reason and the comprehensive compliance rule are sent to relevant personnel so that the relevant personnel can determine whether the text to be verified is compliant based on the original description in the anomaly judgment result, the item to be verified, and the comprehensive compliance rule. If a second unknown cause is obtained from the consistency verification result, then multiple comparison segments corresponding to the abnormal fields to be extracted for the second unknown cause are sent to the relevant personnel so that the relevant personnel can perform consistency verification on the multiple comparison segments and determine whether the text set to be consistent is consistent.

8. A smart dual-track verification device for contracts based on skill packages, characterized in that, The device includes: The acquisition unit is used to acquire the compliance verification requirements and consistency verification requirements of the document to be verified. The intelligent agent parses the compliance verification requirements to obtain multiple items to be verified and comprehensive compliance rules, and injects the compliance verification text of the document to be verified, the multiple items to be verified, and the comprehensive compliance rules into the configuration text of the compliance skill package. The compliance verification unit is used to determine whether each item to be verified exists in the text to be verified using the first LLM model in the compliance skill package, extract the original text description corresponding to the existing item to be verified to generate a judgment result item, inject each judgment result item back into the configuration text, and use the first rule engine in the compliance skill package in combination with the comprehensive compliance rules to make a comprehensive judgment on each judgment result item to generate a compliance verification result. A consistency verification unit is used by the agent to decompose the consistency verification requirement into multiple fields to be extracted. Each field to be extracted and the text set to be verified of the document to be verified are injected into a consistency skill package. Based on the consistency skill package, text priority rules, and extraction and merging rules, an extraction result set corresponding to each field to be extracted is extracted from the text set to be verified. A comparison strategy corresponding to the field type of each field to be extracted is selected to verify the extraction result set, generating a consistency verification result. Based on the compliance verification result and the consistency verification result, a verification result for the document to be verified is generated.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the smart dual-track verification method for contracts based on skill packs as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the skill pack-based intelligent dual-track verification method for contracts as described in any one of claims 1 to 7.

Images