A link backtracking-based trusted large model enterprise scenario contract review system and method
By assigning unique identifiers to contract document fragments and building a backtracking mechanism, the problems of difficulty in tracing and insufficient credibility in the existing contract review system are solved, realizing accurate tracing and transparency of contract review results, and improving the credibility and efficiency of the system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DONGFANG ELECTRIC GROUP DIGITAL TECHNOLOGY CO LTD
- Filing Date
- 2025-12-29
- Publication Date
- 2026-07-03
AI Technical Summary
Existing contract review systems suffer from problems such as difficulty in tracing the source of review results, lack of explicit vector-text correlation, and insufficient credibility of large-scale model reviews, resulting in low efficiency and low credibility.
By assigning unique identifiers (slice IDs) to contract document fragments, an explicit association between fragments and vector library module representations is achieved, and a full-process backtracking mechanism for large model review is constructed, enabling review results to be directly traced back to the original clauses and making the review process transparent.
It significantly improves the credibility and acceptability of the contract review system, reduces the time cost of manual searches, and enhances the closed-loop efficiency of the review process and the credibility of the system.
Smart Images
Figure FT_1
Abstract
Description
Technical Field
[0001] This invention belongs to the field of contract review technology, specifically a trusted large-scale model enterprise scenario contract review system and method based on link backtracking. Background Technology
[0002] In business operations, contract review is a crucial step in risk control, directly impacting a company's compliance and economic interests. Traditional manual contract review relies on professional experience, resulting in inefficiency, high labor costs, and susceptibility to overlooking risks due to human error. With the development of artificial intelligence technology, contract review systems based on Natural Language Processing (NLP) and large-scale models are increasingly being applied in enterprise scenarios to improve review efficiency and accuracy.
[0003] From a general perspective in the technology field, AI-driven contract review systems typically include functions such as text parsing, risk point identification, and intelligent suggestions. Related technologies are frequently seen in patent literature and academic journals; the former proposes review schemes based on rules and machine learning models, while the latter focuses on semantic understanding of contract documents and risk point extraction.
[0004] Existing contract review systems have shortcomings in areas such as tracing the link between review results and the original text, accurately associating contract document fragments with vector representations, and verifying the credibility of large-scale model reviews.
[0005] The difficulty in tracing review results: Most existing systems only output a list of risk points and suggested modifications, failing to directly link and locate the risk points to specific clauses in the original contract. This forces reviewers to manually search for the corresponding locations in the contract document, increasing the workload for subsequent modifications and verification, and impacting the efficiency of the closed-loop review process.
[0006] The vector-text association is not explicit: After the contract document segments are vectorized and stored in the vector library module, the correspondence between the contract document segments and the original text lacks explicit identification (such as bookmarks or identifiers that can be directly located). When the large model recalls similar segments for review based on vectors, users cannot quickly confirm the specific location of the segment in the original text, which affects the verifiability of the review results.
[0007] The credibility of large-scale model reviews is insufficient: When large-scale models perform review tasks, the review process is in a "black box" state. The input-output derivation process lacks the ability to trace back the link. Users cannot clearly know which contract fragments the review results are based on or what logic is used to derive them. This non-traceable review mechanism leads to insufficient basis for the review results. In compliance scenarios of enterprises that focus on compliance and risk management, this seriously reduces the credibility and acceptability of the system. Summary of the Invention
[0008] The purpose of this invention is to address the shortcomings of existing technologies by providing a reliable large-scale model enterprise contract review system based on link back to the original clauses. This system establishes a precise link between contract document fragments and review results, enabling direct tracing of review results back to the original clauses. It assigns unique identifiers (such as slice IDs) to contract document fragments, establishing an explicit association between fragments and vector library modules, ensuring that fragments retrieved based on vector retrieval can be accurately traced back to their original positions. Furthermore, it constructs a full-process link backtracking mechanism for large-scale model review, making the derivation basis and logical process of review conclusions transparent, allowing users to trace the derivation process of review results, and significantly improving the credibility and acceptability of the system in enterprise compliance scenarios.
[0009] The technical objective of this invention is achieved through the following technical solution: A trusted large-scale model-based contract review system for enterprise scenarios, based on link backtracking, includes: The module includes a document editor, a backend engineering module, a vectorization tool module, a vector library module, an audit checklist management system module, and a workflow module. The backend engineering module is communicatively connected to the document editor module, vectorization tool module, review checklist management system module, and workflow module, respectively. The vectorization tool module is communicatively connected to the vector library module and the workflow module, respectively. The backend engineering module is used to receive the contract document uploaded by the user, and to divide the contract document into multiple independent contract clause fragments according to the preset segmentation logic through its contract document processing unit, and to send the contract clause fragments to the vectorization tool module. The vectorization tool module is used to perform semantic vectorization processing on each received contract clause fragment, generate a unique slice ID for each contract clause fragment, and construct an associated data structure containing document ID, slice ID, original text of contract clause fragment, and vector representation. The vector representation and its corresponding slice ID are stored in the vector library module, and the mapping information containing the correspondence between slice ID and original text of contract clause fragment is returned to the backend engineering module.
[0010] The backend engineering module is also used to send the received mapping information to the document editor module; The document editor module is used to create a bookmark for each contract clause fragment in the displayed contract document according to the mapping information. The name of each bookmark corresponds one-to-one with a slice ID, thereby realizing the explicit binding between the original contract clause fragment and the slice ID. The review list management system module is used to store and manage multiple review lists, each review list containing at least one review item, and each review item includes at least the review point name and review logic. The workflow module is used to retrieve the target review list selected by the user from the review list management system module when a review task is triggered; for each review item in the target review list, the vectorization tool module is called to vectorize the review logic text of the review item to obtain a review item vector; based on the review item vector, the top K contract clause fragments with the highest similarity in the vector library module are retrieved as target contract clause fragments, and their slice IDs are obtained, where K is a preset positive integer; the review items and the target contract clause fragments containing the slice IDs and corresponding original texts are input into the large model unit for review analysis to obtain the review results; The backend engineering module also includes a review result processing unit, which receives the review result returned by the workflow module, associates the review result with the corresponding slice ID and contract clause fragment, and pushes it to the document editor module. The document editor module is also used to activate the corresponding bookmark based on the slice ID contained in the review result after receiving the review result, so as to locate the target contract clause fragment in the contract document and realize the link back from the review result to the original contract clause.
[0011] Preferably, the vectorization tool module is specifically used for: The contract clause fragments are semantically encoded using a pre-trained language model based on the Transformer architecture to generate corresponding vector representations; Establish and maintain associated records containing document ID, slice ID, original text of contract clause fragments, and vector representations.
[0012] Preferably, the workflow module further includes a data input unit; The data input unit is communicatively connected to the review list management system, the vectorization tool, and the large model unit, respectively. The data input unit is used to obtain the target review list from the review list management system, call the vectorization tool for each review item to perform vectorization processing to obtain the review item vector, and retrieve the contract clause fragments with the top K similarity rankings from the vector library based on the review item vectors as the target contract clause fragments, where K is a preset positive integer, and input the review item and the retrieved target contract clause fragments into the large model unit. The large model unit is used to perform semantic analysis and risk identification on the target contract clause fragment based on the review items, generate a review result containing the slice ID, and send the review result to the backend engineering module.
[0013] Preferably, the preset segmentation logic includes: segmenting according to contract clause number, or segmenting according to paragraph semantic integrity.
[0014] Preferably, the review results include at least the review point name, risk warning, modification suggestions, and the slice ID of the target contract clause segment.
[0015] Preferably, the review item further includes a risk level; correspondingly, the review result also includes the risk level.
[0016] Preferably, the document editor module is a web page editor, and the bookmark is a hypertext link that supports clicking to jump or an internal anchor point in the document.
[0017] Preferably, the vector library module is a Milvus vector database or a Pinecone vector database.
[0018] A contract review method for enterprise scenarios based on a trusted large-scale model with link backtracking includes the following steps: S1. Receive the contract document to be reviewed and divide it into multiple independent contract clause fragments according to preset rules; S2. Perform semantic vectorization processing on each of the contract clause fragments to generate a corresponding vector representation and a unique slice ID. Store the vector representation and its corresponding slice ID in the vector library and establish a mapping relationship between the slice ID and the contract clause fragment. S3. Based on the mapping relationship, establish a positioning tag in the contract document that is bound to the corresponding slice ID for each contract clause fragment; S4. Obtain the review items in the review list; perform vectorization processing on the review items to obtain the review item vector; based on the review item vector, retrieve the contract clause fragments with the top K similarity rankings in the vector library as target contract clause fragments, and obtain their slice IDs, where K is a preset positive integer; S5. Input the review item and the target contract clause fragment containing the slice ID and the corresponding original text into the large model for review and analysis, and generate review results; the review results shall at least contain the slice ID of the target contract clause fragment; S6. Based on the slice ID in the review result, locate and highlight the target contract clause fragment in the contract document using the corresponding positioning mark, thereby achieving a backtracking from the review result to the original contract text.
[0019] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention establishes a precise link between contract document fragments and review results, enabling direct tracing of review results back to the original clauses; by assigning unique identifiers (such as slice IDs) to contract document fragments, it achieves an explicit association between fragments and vector library module representations, ensuring that fragments retrieved based on vector retrieval can be accurately traced back to their original positions; and by constructing a full-process link tracing mechanism for large-scale model review, it makes the derivation basis and logical process of review conclusions transparent, allowing users to trace the derivation process of review results, significantly improving the credibility and acceptability of the system in enterprise compliance scenarios.
[0020] 2. The system of the present invention adopts a modular design. The review checklist management system module can customize exclusive review checklists according to different contract types of enterprises, and the workflow module can be adapted to the access of large models in different industries. It has good scalability and scenario adaptability, and can meet the diverse contract review needs of enterprises. Attached Figure Description
[0021] Figure 1 This is an architecture diagram of the Trusted Large Model Enterprise Scenario Contract Review System based on Link Backtracking of the present invention; Attached image labels: 1—Document editor module; 2—Backend engineering module; 3—Vectorization tool module; 4—Vector library module; 5—Review checklist management system module; 6—Workflow module. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0023] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0024] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.
[0025] like Figure 1As shown, a trustworthy large-scale model enterprise scenario contract review system based on link backtracking includes: document editor module 1, backend engineering module 2, vectorization tool module 3, vector library module 4, review list management system module 5, and workflow module 6.
[0026] Backend engineering module 2 communicates with document editor module 1, vectorization tool module 3, review checklist management system module 5 and workflow module 6 respectively; vectorization tool module 3 communicates with vector library module 4 and workflow module 6 respectively.
[0027] Backend engineering module 2 is used to receive contract documents uploaded by users, and use its contract document processing unit to divide the contract document into multiple independent contract clause fragments according to the preset segmentation logic, and send the contract clause fragments to vectorization tool module 3.
[0028] In this embodiment, the backend engineering module 2 includes a contract document processing unit and a review result processing unit.
[0029] The preset segmentation logic includes: segmenting by contract clause number or segmenting by paragraph semantic integrity, etc.
[0030] In this embodiment, the backend engineering module 2 adopts a microservice architecture and includes a contract document processing unit and a review result processing unit. The backend engineering module 2 is deployed on an enterprise private cloud or public cloud server and is responsible for receiving, processing, and forwarding data.
[0031] The vectorization tool module 3 is used to perform semantic vectorization processing on each received contract clause fragment, generate a unique slice ID for each contract clause fragment, and construct an associated data structure containing document ID, slice ID, original text of contract clause fragment and vector representation. The vector representation and its corresponding slice ID are stored in the vector library module 4, and the mapping information containing the correspondence between slice ID and original text of contract clause fragment is returned to the backend engineering module 2.
[0032] In this embodiment, the model used by the vectorization tool module 3 for semantic vectorization processing is a pre-trained language model based on the Transformer architecture, including BERT or GPT series models. Vectorization tool module 3 is developed based on an open-source vector processing framework and deployed on a GPU server to ensure vectorization efficiency.
[0033] Backend engineering module 2 is also used to send the received mapping information to document editor module 1. Document editor module 1 is used to create a bookmark for each contract clause fragment in the displayed contract document according to the mapping information. The name of each bookmark corresponds one-to-one with a slice ID, realizing an explicit binding between the original contract clause fragments and slice IDs. Review list management system module 5 is used to store and manage multiple review lists. Each review list contains at least one review item. The review item includes at least the review point name and review logic.
[0034] Workflow module 6 is used to retrieve the target review list selected by the user from the review list management system module 5 when the review task is triggered; for each review item in the target review list, the vectorization tool module 3 is called to vectorize the review logic text of the review item to obtain the review item vector; based on the review item vector, the top K contract clause fragments with the highest similarity are retrieved from the vector library module 4 as target contract clause fragments, and their slice IDs are obtained, where K is a preset positive integer; the review items and the target contract clause fragments containing the slice IDs and corresponding original texts are input into the large model unit for review analysis to obtain the review results.
[0035] In practical use, the vector library module 4 employs specialized vector databases for efficient storage and retrieval of contract fragment vectors. Examples include the Milvus vector database and the Pinecone vector database.
[0036] like Figure 1 As shown, workflow module 6 also includes a data input unit; the data input unit is communicatively connected to the review checklist management system module 5, the vectorization tool module 3, and the large model unit; the data input unit is used to obtain the target review checklist from the review checklist management system, call the vectorization tool for each review item to perform vectorization processing to obtain the review item vector, and retrieve the contract clause fragments with the top K similarity rankings from the vector library based on the review item vector as the target contract clause fragments, where K is a preset positive integer, and input the review item and the retrieved target contract clause fragments into the large model unit.
[0037] The large model unit is used to perform semantic analysis and risk identification on target contract clause fragments based on review items, generate review results containing slice IDs, and send the review results to backend engineering module 2.
[0038] The workflow engine is developed based on open-source frameworks such as Dify. Large models can be connected to private models trained internally by enterprises or large model APIs of mainstream cloud service providers (such as Azure OpenAI, Alibaba Cloud Tongyi Qianwen, etc.).
[0039] In this embodiment, the review items also include risk level; correspondingly, the review results also include risk level.
[0040] The backend engineering module 2 also includes a review result processing unit, which receives the review results returned by the workflow module 6, associates the review results with the corresponding slice ID and contract clause fragments, and then pushes them to the document editor module 1.
[0041] In this embodiment, the review results include at least the review point name, risk warning, modification suggestions, and the slice ID of the target contract clause segment.
[0042] The document editor module 1 is also used to activate the corresponding bookmark based on the slice ID contained in the review result after receiving the review result, so as to locate the target contract clause fragment in the contract document and realize the link back from the review result to the original contract clause.
[0043] In this embodiment, the document editor is developed based on Web technology, supports access from mainstream browsers, and has functions such as bookmark management, text display, and annotation of review results. When document editor module 1 is a Web page editor, bookmarks are hypertext links or internal document anchors that support clickable navigation.
[0044] By using a "bookmark-slice ID-review result" association mechanism, each review result can directly locate the specific clause fragment in the original contract text. After receiving the review result, users can immediately jump to the corresponding location through the document editor's bookmarks to make modifications or verifications, significantly reducing the time cost of manual searching and improving the closed-loop efficiency of contract review and rectification.
[0045] Through a complete chain of "review item - recall slice - large model reasoning - review result - original document signature", users can clearly trace the derivation process of each review result: that is, based on which contract fragments and according to what logic of the review item, the large model draws its conclusion. This chain of traceability is crucial in enterprise compliance scenarios, significantly improving the system's credibility and persuasiveness, enabling enterprise users to confidently rely on the system for contract risk management.
[0046] A contract review method for enterprise scenarios based on a trusted large-scale model with link backtracking includes the following steps: S1. Contract Document Segmentation: Receives contract documents to be reviewed and segments them into multiple independent contract clause fragments according to preset rules.
[0047] In this embodiment, the user uploads a contract document and initiates a review. Specifically, the user uploads the contract document to be reviewed through the system interface, selects the corresponding review checklist (such as selecting "2025 Enterprise Sales Contract Review Checklist" from the review checklist management system), and initiates the review task.
[0048] In this embodiment, after a user initiates a contract review task and uploads a contract document, the contract document first enters the contract document processing unit of the backend engineering module 2. The contract document processing unit divides the contract document into n independent contract clause fragments (denoted as contract clause No. 001 to No. xxx) according to a preset segmentation logic (such as by contract clause number, paragraph semantics, etc.).
[0049] For example, when numbering contract clauses, the default segmentation logic is to use "Article X" as an identifier to segment the contract clauses.
[0050] S2. Vectorization and storage steps: Perform semantic vectorization processing on each contract clause fragment to generate a corresponding vector representation and a unique slice ID. Store the vector representation and its corresponding slice ID in the vector library and establish a mapping relationship between the slice ID and the contract clause fragment. In this embodiment, the backend engineering module 2 passes the contract clause fragments from step S1 to the vectorization tool module 3. The vectorization tool module 3 performs semantic vectorization processing on each contract clause fragment (e.g., using the encoding capabilities of pre-trained models such as BERT and GPT to generate vectors), and simultaneously generates a unique slice ID for each contract clause fragment (e.g., slice_id.1 to slice_id.n), and constructs a relational data structure of "document id - slice ID - contract clause fragment - vector". Here, "document id" identifies different contract documents, and "slice ID" uniquely identifies different contract clause fragments within the same contract.
[0051] Vectorization tool module 3 stores the generated vectors and corresponding slice IDs into vector library module 4, and at the same time sends the mapping relationship of "slice ID - contract clause fragment" back to backend engineering module 2.
[0052] S3. Explicit Association: Based on the mapping relationship, establish a positioning tag in the contract document that is bound to the corresponding slice ID for each contract clause fragment; In this embodiment, the location marker is a bookmark. Specifically, after receiving the slice ID from the vectorization tool module 3, the backend engineering module 2 passes it to the document editor module 1. The document editor module 1 creates a bookmark for each contract clause fragment in the current contract document, with each bookmark name corresponding one-to-one with a slice ID (e.g., bookmark No. 001 is associated with slice _id.1, corresponding to contract clause No. 001; bookmark No. 002 corresponds to slice _id.2, corresponding to contract clause No. 002; and so on). In this way, an explicit and precise binding is achieved between the original contract clause fragments and the slice IDs in the vector library module 4, laying the foundation for subsequent link tracing.
[0053] S4. Intelligent Search: Obtain the review items in the review list; perform vectorization processing on the review items to obtain the review item vector; based on the review item vector, search the vector library for the contract clause fragments with the highest similarity ranking among the top K as the target contract clause fragments, and obtain their slice IDs, where K is a preset positive integer; In this embodiment, when a user initiates a review, they need to select the corresponding review list (such as "General Review List for Enterprise Procurement Contracts") from module 5 of the review list management system.
[0054] Backend engineering module 2 retrieves the review list selected by the user (such as review items No.01 to No.xx) from review list management system module 5 and passes it to workflow module 6.
[0055] In this embodiment, the target contract clause fragment is a recalled slice. The vectorization tool module 3 is sequentially invoked to vectorize the text such as "review logic" of the review item. Then, the vector library module 4 retrieves and recalls the top k slices (i.e., the vectors of the contract clause fragments) with the highest similarity to the review item. The vector library module 4 returns the recalled slices carrying the corresponding slice ID, which can then be associated with the specific clauses of the original contract.
[0056] By using the slice IDs generated by the vectorization tool module 3 and the bookmark bindings from the document editor module 1, an explicit, one-to-one correspondence between contract document fragments and vectors in the vector library module 4 is achieved. When the large model unit recalls similar fragments based on vectors, users can quickly confirm the location of the fragment in the original text through the slice ID, verify the rationality of the recall, and enhance the interpretability and verifiability of the review process.
[0057] S5. Large Model Review: Input the review items and the target contract clause fragment containing the slice ID and the corresponding original text into the large model for review and analysis, and generate review results; the review results shall at least contain the slice ID of the target contract clause fragment; In this embodiment, workflow module 6 inputs the "review items (including review point name, risk level, review logic, risk warning, and modification suggestions)" and the "recalled slices (including slice ID and corresponding contract clause fragments)" into the large model unit. Based on the requirements of the review items, the large model unit performs semantic analysis and risk identification on the recalled contract slices, generating review results containing information such as "review point name, risk level, risk warning, modification suggestions, slice ID.xxx, and contract clause No.m".
[0058] S6. Backtracking of review results: Based on the slice ID in the review results, the target contract clause fragment is located in the contract document through the corresponding positioning mark, thereby achieving backtracking from the review results to the original contract text.
[0059] In this embodiment, the review results output by the large model unit are transmitted to the review result processing unit of the backend project. The review result processing unit associates the review results with the corresponding slice ID and contract clause fragment, and then pushes them to the document editor module 1. When users view the review results in the document editor module 1, they can directly jump to the corresponding bookmark position through the slice ID or contract clause identifier carried in the review results, realizing a link back from the review results to the original contract clauses.
[0060] The technical solutions provided by the embodiments of the present invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the embodiments of the present invention. The descriptions of the embodiments above are only for helping to understand the principles of the embodiments of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the embodiments of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A trustworthy large-scale model-based contract review system for enterprise scenarios based on link backtracking, characterized in that, include: The module includes a document editor, a backend engineering module, a vectorization tool module, a vector library module, an audit checklist management system module, and a workflow module. The backend engineering module is communicatively connected to the document editor module, vectorization tool module, review checklist management system module, and workflow module, respectively. The vectorization tool module is communicatively connected to the vector library module and the workflow module, respectively. The backend engineering module is used to receive the contract document uploaded by the user, and to divide the contract document into multiple independent contract clause fragments according to the preset segmentation logic through its contract document processing unit, and to send the contract clause fragments to the vectorization tool module. The vectorization tool module is used to perform semantic vectorization processing on each received contract clause fragment, generate a unique slice ID for each contract clause fragment, and construct an associated data structure containing document ID, slice ID, original text of contract clause fragment and vector representation. The vector representation and its corresponding slice ID are stored in the vector library module, and the mapping information containing the correspondence between slice ID and original text of contract clause fragment is returned to the backend engineering module. The backend engineering module is also used to send the received mapping information to the document editor module; The document editor module is used to create a bookmark for each contract clause fragment in the displayed contract document according to the mapping information. The name of each bookmark corresponds one-to-one with a slice ID, thereby realizing the explicit binding between the original contract clause fragment and the slice ID. The review list management system module is used to store and manage multiple review lists, each review list containing at least one review item, and each review item includes at least the review point name and review logic. The workflow module is used to retrieve the target review list selected by the user from the review list management system module when a review task is triggered; for each review item in the target review list, the vectorization tool module is called to vectorize the review logic text of the review item to obtain a review item vector; based on the review item vector, the top K contract clause fragments with the highest similarity in the vector library module are retrieved as target contract clause fragments, and their slice IDs are obtained, where K is a preset positive integer; the review items and the target contract clause fragments containing the slice IDs and corresponding original texts are input into the large model unit for review analysis to obtain the review results; The backend engineering module also includes a review result processing unit, which is used to receive the review results returned by the workflow module, associate the review results with the corresponding slice ID and contract clause fragments, and push them to the document editor module; The document editor module is also used to activate the corresponding bookmark based on the slice ID contained in the review result after receiving the review result, so as to locate the target contract clause fragment in the contract document and realize the link back from the review result to the original contract clause.
2. The enterprise scenario contract review system based on link backtracking according to claim 1, characterized in that, The vectorization tool module is specifically used for: The contract clause fragments are semantically encoded using a pre-trained language model based on the Transformer architecture to generate corresponding vector representations; Establish and maintain associated records containing document ID, slice ID, original text of contract clause fragments, and vector representations.
3. The enterprise scenario contract review system based on link backtracking according to claim 1, characterized in that, The workflow module also includes a data input unit; The data input unit is communicatively connected to the review list management system, the vectorization tool, and the large model unit, respectively. The data input unit is used to obtain the target review list from the review list management system, call the vectorization tool for each review item to perform vectorization processing to obtain the review item vector, and retrieve the contract clause fragments with the top K similarity rankings from the vector library based on the review item vectors as the target contract clause fragments, where K is a preset positive integer, and input the review item and the retrieved target contract clause fragments into the large model unit. The large model unit is used to perform semantic analysis and risk identification on the target contract clause fragment based on the review items, generate a review result containing the slice ID, and send the review result to the backend engineering module.
4. The enterprise scenario contract review system based on link backtracking trustworthy large model as described in claim 1, characterized in that, The preset segmentation logic includes: segmenting by contract clause number or segmenting by paragraph semantic integrity.
5. The enterprise scenario contract review system based on link backtracking according to claim 1, characterized in that, The review results shall include at least the review point name, risk warning, modification suggestions, and the slice ID of the target contract clause segment.
6. The enterprise scenario contract review system based on link backtracking according to claim 1, characterized in that, The review items also include risk levels; correspondingly, the review results also include the risk levels.
7. The enterprise scenario contract review system based on link backtracking according to claim 1, characterized in that, The document editor module is a web page editor, and the bookmarks are hypertext links or internal document anchors that support click-through.
8. The enterprise scenario contract review system based on link backtracking according to claim 1, characterized in that, The vector library module is either the Milvus vector database or the Pinecone vector database.
9. A contract review method for enterprise scenarios based on a trusted large-scale model with link backtracking, characterized in that, Includes the following steps: S1. Receive the contract document to be reviewed and divide it into multiple independent contract clause fragments according to preset rules; S2. Perform semantic vectorization processing on each of the contract clause fragments to generate a corresponding vector representation and a unique slice ID. Store the vector representation and its corresponding slice ID in the vector library and establish a mapping relationship between the slice ID and the contract clause fragment. S3. Based on the mapping relationship, establish a positioning tag in the contract document that is bound to the corresponding slice ID for each contract clause fragment; S4. Obtain the review items in the review list; perform vectorization processing on the review items to obtain the review item vector; based on the review item vector, retrieve the contract clause fragments with the top K similarity rankings in the vector library as target contract clause fragments, and obtain their slice IDs, where K is a preset positive integer; S5. Input the review item and the target contract clause fragment containing the slice ID and the corresponding original text into the large model for review and analysis, and generate review results; the review results shall at least contain the slice ID of the target contract clause fragment; S6. Based on the slice ID in the review result, locate and highlight the target contract clause fragment in the contract document using the corresponding positioning mark, thereby achieving a backtracking from the review result to the original contract text.