Intelligent review system and method and storage medium
By preprocessing and analyzing procurement and response documents through an intelligent review system, and utilizing multiple large language models for collaboration, the problems of high cost and insufficient data volume in manual review are solved, achieving efficient and accurate automated review.
Patent Information
- Application Number
- CN202511107258.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-08
- Publication Date
- 2025-11-21
AI Technical Summary
The existing electronic procurement system has high labor costs for manual review and bid clearing processes, and the general-purpose models cannot meet business requirements and are difficult to train industry-specific models due to data privacy concerns.
An intelligent review system is adopted. After obtaining procurement and response documents and performing preprocessing, the data is sorted out using vectorized models and general large models to extract review clauses and response points. The system is then combined with a re-ranking model for screening and analysis to achieve fully automated review.
It achieves fully automated management of the bid evaluation process, saving labor costs, improving the quality and efficiency of bid evaluation, and providing concise and accurate results that are close to those of manual review, thus avoiding the problem of insufficient data.
Smart Images
Figure CN120996747A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of artificial intelligence, in particular to an intelligent evaluation system and method and a storage medium. BACKGROUND
[0002] The existing electronic procurement system is mainly an information management system for procurement business. In the online evaluation or clearing scene, professional personnel such as project managers and evaluation committees manually sort out and reason data. The clearing and evaluation process takes several hours. In order to ensure the rigor and fairness of the evaluation process, data often needs to be reviewed and verified by multiple people, further increasing the labor cost of the evaluation and clearing process.
[0003] With the outbreak of new generation natural language processing technologies such as large models, user systems can introduce large model technology in the procurement field to achieve higher intelligence and automation, save labor costs and improve procurement quality and efficiency. However, the existing large model technology is mainly a general large model, which does not understand the professional knowledge and business rules of the bidding field deeply. Directly handing over business problems to the large model cannot achieve the results that meet the business requirements. Moreover, the data in the procurement field is mostly private and confidential, and it is also difficult to sort out enough data for industry vertical model training. SUMMARY
[0004] The purpose of the present application is to overcome the defects of the prior art, provide an intelligent evaluation system, method and storage medium, solve the problem of high labor cost in the existing artificial evaluation and clearing process, and solve the problem that the existing general large model cannot achieve the results that meet the business requirements and it is difficult to sort out enough data for industry vertical model training due to data privacy.
[0005] The technical solution to achieve the above-mentioned purpose is:
[0006] The present application provides an intelligent evaluation method, comprising the following steps:
[0007] Obtain procurement files and response files;
[0008] Preprocess the obtained procurement files to obtain vectorized data of the procurement files;
[0009] Preprocess the obtained response files to obtain vectorized data of the response files;
[0010] Use a vectorization model to recall data related to evaluation clauses from the vectorized data of the procurement files, and use a general large model to sort out the recalled data related to evaluation clauses to obtain structured evaluation clause data;
[0011] recall relevant response points from vectorized data of the response file according to the structured data of the review clauses;
[0012] perform response condition analysis on each response point of the structured data of the review clauses by using the general large model to obtain a response condition analysis result;
[0013] give an intelligent review conclusion for the response condition of the response file by using the general large model in combination with the review requirements in the structured data of the review clauses.
[0014] The intelligent review method further improves that after obtaining the structured data of the review clauses, the following steps are further included:
[0015] interpret the clause content in the structured data of the review clauses by using the general large model, and determine whether the clause content has information reference, if yes, recall the corresponding reference content from the vectorized data of the procurement file to complete the information of the clause content, and then determine whether there is information reference in the next round, if no, extract the review elements from the clause content.
[0016] The intelligent review method further improves that recalling relevant response points from vectorized data of the response file according to the structured data of the review clauses includes the following steps:
[0017] recall the response file fragments according to the structured data of the review clauses;
[0018] sort the recalled response file fragments by using a reordering model, and select a predetermined number of response file fragments;
[0019] integrate the selected response file fragments to form a response point preliminary screening result.
[0020] The intelligent review method further improves that performing response condition analysis on each response point of the structured data of the review clauses by using the general large model includes the following steps:
[0021] based on the response point preliminary screening result, filter the content of each response point by using the general large model, determine whether the review elements are responded to, if no, eliminate the corresponding response point, if yes, summarize the response condition of the response point, and record the response condition together with the response point information in the refined screening result;
[0022] after the response point content filtering is completed, reorganize and merge the refined screening result to form a response condition analysis result.
[0023] The intelligent review method further improves that pre-processing the obtained response file to obtain vectorized data of the response file includes the following steps:
[0024] The document is cut according to the file directory index, and the chapter content is read page by page;
[0025] The table type data in the response file needs to be standardized into markdown format for easy reading during the reading process;
[0026] For picture type data in the response file, a multi-modal large model and a license library data are used to locate and extract the picture type and key information during the reading process;
[0027] During the process of reading the chapter content page by page, the page number and the coordinate position information of the content in the page are recorded synchronously;
[0028] After the chapter content of the response file is read, a general large model is used to extract key information and summarize the chapter content;
[0029] The summary of the chapter content of the response file is vectorized and stored, the chapter original text, chapter coordinate information, file type and original file address are vectorized and stored in the metadata, and the project identification and supplier identification are added in the metadata.
[0030] The application also provides a storage medium, wherein the storage medium stores a program of the intelligent evaluation method, and the program of the intelligent evaluation method is executed by a processor to realize the steps of the intelligent evaluation method.
[0031] The application further provides an intelligent evaluation system, comprising:
[0032] An acquisition unit is configured to acquire a procurement file and a response file;
[0033] A preprocessing unit is connected to the acquisition unit and configured to preprocess the acquired procurement file to obtain vectorized data of the procurement file, and preprocess the acquired response file to obtain vectorized data of the response file;
[0034] An evaluation clause extraction unit is connected to the preprocessing unit and configured to use a vectorization model to recall data related to an evaluation clause from the vectorized data of the procurement file, and use a general large model to sort the recalled data related to the evaluation clause to obtain structured data of the evaluation clause;
[0035] A response recall unit is connected to the evaluation clause extraction unit and the preprocessing unit and configured to recall relevant response points from the vectorized data of the response file according to the structured data of the evaluation clause;
[0036] The response analysis unit, connected with the review clause extraction unit and the response recall unit, is configured to perform response situation analysis on each response point of the review clause structured data by using a general large model to obtain a response situation analysis result.
[0037] The response review unit, connected with the review clause extraction unit and the response analysis unit, is configured to give an intelligent review conclusion on the response situation of the response file by using a general large model in combination with the review requirements in the review clause structured data.
[0038] The intelligent review system is further improved in that the system further comprises a review element extraction unit connected with the review clause extraction unit and the preprocessing unit.
[0039] The review element extraction unit is configured to interpret the clause content in the review clause structured data by using a general large model, and determine whether the clause content contains information references.
[0040] The intelligent review system is further improved in that the response recall unit is further configured to recall response file segments according to the review clause structured data.
[0041] The response recall unit is further configured to sort the recalled response file segments by using a reordering model, and select a predetermined number of response file segments.
[0042] The response recall unit is further configured to integrate the selected response file segments to form a response point preliminary screening result.
[0043] The intelligent review system is further improved in that the response analysis unit is further configured to filter the content of each response point by using a general large model based on the response point preliminary screening result, determine whether the review elements are responded to, and if not, eliminate the corresponding response point, and if so, summarize the response situation, and record the response situation together with the response point information in a refined screening result.
[0044] After the content of the response point is filtered, the refined screening result is further sorted and merged to form a response situation analysis result.
[0045] The intelligent review system, method and storage medium have the following advantages:
[0046] The intelligent evaluation system and method of the present application are realized based on multiple large language models, realize full-automatic hosting of the clearing link, save the labor cost of manual evaluation index table carding, accelerate the development of clearing work, and the large model searches the full text of the bid document, avoids the risk of manual staggered missing, and improves the evaluation quality and efficiency.
[0047] The intelligent evaluation system and method of the present application realize the process carding of the behavior and thought of the expert evaluation process through in-depth understanding of the evaluation link of the electronic procurement process, use workflow and task flow to organize general large models, multi-modal large models, reordering models, vectorization models and other artificial intelligence technologies for cooperation, and realize intelligent evaluation, a highly professional business application scenario.
[0048] The intelligent evaluation system and method of the present application are not strongly coupled to models of various scales and application scenarios, can be realized based on various open source / closed source general models on the market, and is a general large model technology applied to intelligent evaluation in the strong professional industry vertical field, which avoids the difficulty of insufficient professional industry field data to train industry vertical field models.
[0049] The intelligent evaluation system and method of the present application deeply understand and simulate the thinking process and behavior of professional expert judges in the evaluation process, the result is more concise and accurate compared with the traditional scheme of directly handing over to the large model, the difference with the evaluation and review behavior of the judges is small, data and results are output in each link, the expert is easier to understand the processing process, has lower unexplainability, and the result is closer to the conclusion of the manual evaluation of the judges. BRIEF DESCRIPTION OF DRAWINGS
[0050] Figure 1 It is a total flow chart of the preprocessing stage of the intelligent evaluation system and method of the present application.
[0051] Figure 2 It is a preprocessing flow chart of the procurement file of the intelligent evaluation system and method of the present application.
[0052] Figure 3 It is a preprocessing flow chart of the response file of the intelligent evaluation system and method of the present application.
[0053] Figure 4 It is a flow chart of the evaluation clause extraction of the intelligent evaluation system and method of the present application.
[0054] Figure 5 It is a flow chart of the evaluation element extraction of the intelligent evaluation system and method of the present application.
[0055] Figure 6 It is a total flow chart of the intelligent evaluation and evaluation calculation of the intelligent evaluation system and method of the present application.
[0056] Figure 7 Flow chart for response point identification in the intelligent review system and method of the present application.
[0057] Figure 8 Flow chart for response situation analysis in the intelligent review system and method of the present application.
[0058] Figure 9 Flow chart for review conclusion reasoning in the intelligent review system and method of the present application.
[0059] Figure 10 Flow chart for the intelligent review method of the present application. DETAILED DESCRIPTION
[0060] The present application will be further described below in conjunction with the accompanying drawings and specific embodiments.
[0061] Referring to Figure 1 , the present application provides an intelligent review system, method and storage medium. By deeply understanding the review business, the review business process is split and combed to construct an intelligent review workflow. Suitable general large language models are selected for integration according to different natural language technology application targets. Traditional program logic, vector database and other technologies are used for assistance. Finally, an intelligent review agent with three core capabilities of intelligent clause interpretation, response intelligent retrieval and review conclusion reasoning is realized. The intelligent review system, method and storage medium of the present application will be described below in conjunction with the accompanying drawings.
[0062] Referring to Figure 1 , the overall flow chart of the preprocessing stage in the intelligent review system and method of the present application is shown. Referring to Figure 6 , the intelligent review-review calculation overall flow chart in the intelligent review system and method of the present application is shown. The intelligent review system of the present application will be described below in conjunction with Figure 1 and Figure 6 .
[0063] As shown in Figure 1 and Figure 6 , the intelligent review system of the present application includes an acquisition unit, a preprocessing unit, a review clause extraction unit, a response recall unit, a response analysis unit and a response review unit. The acquisition unit is connected with the preprocessing unit. The preprocessing unit is connected with the review clause extraction unit and the response recall unit. The review clause extraction unit is connected with the response recall unit, the response analysis unit and the response review unit. The response analysis unit is connected with the response review unit.
[0064] The acquisition unit is configured to acquire the procurement document and the response document; the preprocessing unit is configured to preprocess the acquired procurement document to obtain vectorized data of the procurement document; the preprocessing unit is also configured to preprocess the acquired response document to obtain vectorized data of the response document; the review clause extraction unit is configured to recall data related to the review clause from the vectorized data of the procurement document by using a vectorization model, and to sort the recalled data related to the review clause by using a general large model to obtain structured data of the review clause; the response recall unit is configured to recall relevant response points from the vectorized data of the response document according to the structured data of the review clause; the response analysis unit is configured to analyze the response situation of each response point of the structured data of the review clause by using the general large model to obtain a response situation analysis result; and the response review unit is configured to give an intelligent review conclusion for the response document according to the response situation by using the general large model in combination with the review requirements in the structured data of the review clause.
[0065] The intelligent review system of the present application divides the intelligent review processing process into two stages of document preprocessing and review calculation by understanding and analyzing the electronic procurement business review link, and compresses the time consumption of the intelligent review calculation process into an acceptable time range of the business by means of advance analysis and parallel processing.
[0066] The main goal of the review preprocessing stage of the intelligent review system of the present application is to perform advance analysis and data processing on the procurement document and the response document, and to perform special extraction and analysis process on the review clause. As shown in the figure, the overall flow of the preprocessing stage includes the steps of vectorization of the procurement document, acquisition of all response documents, vectorization of the response documents, storage of the vectorized data, extraction of the review clause, and extraction and understanding of the review elements. Figure 1
[0067] In one specific embodiment of the present application, the acquisition unit of the intelligent review system of the present application can be connected with an online bidding procurement system, receives the response documents sent by each supplier received by the online bidding procurement system, and the procurement document can be uploaded to the bidding procurement system by the person in charge of the bidding procurement project, and then sent to the acquisition unit of the intelligent review system by the bidding procurement system. For example, the bidding procurement system is provided with an online bidding execution time, after the bidding is completed, the received response documents of each supplier are sent to the intelligent review system together with the corresponding procurement document for review.
[0068] In one specific embodiment of the present application, the acquisition unit of the intelligent review system of the present application can also receive the procurement document and the response document uploaded by the user, such as for the offline bidding procurement project, the project manager creates an intelligent bidding project in the intelligent review system after the bidding, and uploads the procurement document and the response document.
[0069] In one specific embodiment of the present application, the preprocessing unit of the present application is used for preprocessing the procurement document and the response document, and the preprocessing flow of the procurement document is as shown in the figure Figure 2 The preprocessing unit is used for reading the procurement document by chapter, and the text type data, table type data and picture type data in the procurement document are read respectively in the process of reading the content by chapter. The table type data in the procurement document needs to be standardized into markdown format for easy reading in the process of reading. The picture type data in the procurement document is understood and refined by using a multi-modal large model in the process of reading, to form a picture description and extract picture content information. After the chapter content of the procurement document is read, a general large model is used to extract key information and summarize the chapter content, the summary of the chapter content is vectorized and stored, and the chapter original text, project information and data information such as bid evaluation method are stored in the metadata together with the summary and stored in the vector database.
[0070] The intelligent review system of the present application vectorizes the key information and summary of the chapter content of the procurement document in preprocessing and stores it in the vector database, which can improve the accuracy of retrieval. The preprocessing unit of the present application is connected with the vectorization model, which vectorizes the summary and processes the chapter original text as metadata to form vectorized data. The preprocessing unit then vectorizes the vectorized data and stores it in the vector database.
[0071] The preprocessing unit of the present application considers the context processing problem of the large model and the vectorization model, and the accuracy of storage and recall of the vector database.
[0072] Further, the preprocessing flow of the response document is as shown in the figure Figure 3As shown, the preprocessing unit is used to split the document according to the file directory index and read the chapter content page by page for the response file, so as to facilitate the positioning of the response content; in the process of reading the content according to the directory, the reading is carried out in the mode of merging chapters after page-by-page reading. The table type data in the response file needs to be standardized into markdown format for easy reading during reading; the picture type data in the response file is positioned and extracted by using a multi-modal large model and a license library data during reading; in the process of reading the chapter content page by page, the page number where the content is located and the coordinate position information in the page are recorded synchronously; after the chapter content of the response file is read, the page information is merged and integrated according to the chapter start and end page number and the content coordinates of the directory index, to obtain detailed chapter content information, and then the general large model is used to extract key information and abstract the chapter content; the abstract of the chapter content of the response file is vectorized and stored, the chapter original text, chapter coordinate information, file type and original file address are vectorized and stored in the metadata, and the project identification and supplier identification are added in the metadata for data isolation control.
[0073] The preprocessing unit vectorizes the abstract of the chapter content of the response file, and the metadata of the chapter original text, chapter coordinate information, file type and original file address is vectorized data processing by using a vectorization model, and the preprocessing unit further stores the vectorized data in a vector database.
[0074] In one specific embodiment of the present application, as shown in Figure 4 As shown, the review clause extraction unit of the present application further analyzes the vectorized data after preprocessing of the procurement file, recalls the content related to the review clause from the vectorized data of the procurement file through a vectorization model, and reselects the content through a reordering model to improve the data recall quality. Then the review clause extraction unit extracts the review clause from the recall result by using a general large model, and guides the general large model to sort out the review clause by stage and node grouping through technical means such as prompt words and professional knowledge RAG supplement, and forms a relatively standard structured data.
[0075] For the online project that has been running in the electronic procurement system, if the structured review clause has been obtained, the extraction of the clause does not need to be re-performed, and the subsequent logic can be directly based on the structured review clause.
[0076] The review clause extraction unit of the present application stores the obtained structured review clause data, so that the structured review clause data can be directly read for the same procurement file in the subsequent process.
[0077] In one specific embodiment of the present application, as shown in Figure 5As shown, the intelligent review system of the present application further comprises a review element extraction unit connected with the review clause extraction unit and the preprocessing unit; the review element extraction unit is used to interpret the clause content in the review clause structured data by using a general large model, and to judge whether the clause content contains information reference, if yes, the corresponding reference content is recalled from the vectorized data of the procurement document to complete the information of the clause content, and then the next round of judgment of whether there is information reference is carried out; if not, the review elements are extracted from the clause content.
[0078] Since the review clause content is mainly aimed at professional judges, there are content references, omissions and the like, which are not friendly to the understanding and reasoning of the model, therefore, the review element extraction unit understands the clause content in the review element extraction process, and also carries out multi-round recall and completion processing on the references.
[0079] Firstly, the general model attempts to interpret the clause content, judges whether the content information is perfect, and for the cases of content reference, omission and the like, the contents of other chapters of the procurement document need to be referred to, and the multi-round recall and information completion link is entered. According to the information to be completed, the chapter content possibly containing information is recalled from the vectorization result of the procurement document, and the target information is extracted from it by using a large model. After the target information is extracted, the original clause content is completed, and it is judged whether it needs to continue information completion, and this link is repeated until the whole information required for review is obtained. Based on the review clause information after completion, the general large model is used to attempt to understand and summarize the requirements of the clause, and the review requirements are summarized into review elements one by one, which are used for subsequent review calculation and reasoning process.
[0080] In one specific embodiment of the present application, as shown in Figure 6 The intelligent review system of the present application further comprises an intelligent review calculation and analysis stage, the main target of which is to carry out the clear bidding and review calculation process of each supplier and each clause based on the results of the preprocessing stage. Considering that the review process may exist in the re-evaluation scene and may be manually intervened, the review calculation and analysis process is designed as an independent process and can be executed multiple times. The total flow of the review calculation includes the steps of response point identification / primary screening, response point situation analysis / precision screening and review result reasoning.
[0081] In one specific embodiment of the present application, as shown in Figure 7 The response recall unit of the present application is also used to recall the response document fragments according to the review clause structured data; is also used to sort the recalled response document fragments by using a reordering model, and select the first set number of response document fragments; is also used to integrate the selected response document fragments to form the response point primary screening result.
[0082] The response calling unit is used to identify response points, call relevant response fragment contents from each supplier response file according to the review elements based on the review element interpretation results of the preprocessing stage. In order to further improve the accuracy of the response calling result and save the calculation resources of the next response condition analysis, a reordering model is introduced to screen the calling result, and the response content with higher relevance is retained top-k, k is a set number, such as 3, 5, etc. Finally, the response points screened out are integrated, and the continuous page coordinate intervals are merged and integrated to form the response point preliminary screening result.
[0083] In one specific embodiment of the present application, as shown in Figure 8 The response analysis unit is also used to screen each response point content based on the response point preliminary screening result using a general large model, judge whether the review element is responded to, if not, the corresponding response point is eliminated, if yes, the response condition of the response point is summarized, and the response condition is recorded in the fine screening result together with the response point information;
[0084] After the response point content screening is completed, the fine screening result is combed and merged again to form the response condition analysis result.
[0085] Specifically, in the response condition analysis stage, the general large model is used to analyze each response point of the review element based on the preliminary screening result. The model is used to screen each response point content, judge whether the review element is responded to, if not, the response point is eliminated; if the response is made, the response condition of the response point to the element is summarized, and the response condition is recorded in the fine screening result together with the response point information.
[0086] After the response point analysis and screening are completed, the response result after fine screening is combed and merged again to form the response condition analysis result, which is used for clear label result display and export.
[0087] In one specific embodiment of the present application, as shown in Figure 9 The response review unit uses a general large model to infer the response condition of each supplier based on the response condition summary of each supplier in the review conclusion reasoning stage, combines the requirements of the review clauses and review elements, and infers the response condition of each supplier to give an intelligent review conclusion whether to pass or a score number according to different clause types.
[0088] The intelligent review system of the application further comprises an operation interface to facilitate the user to log in and perform the operation of bid review or bid clearing by means of the operation interface. The intelligent review system of the application is applied to the bid clearing link of the electronic procurement system or the bid clearing link before bid review, and is used by the procurement managers, expert judges and other users. The procurement managers or expert judges can upload the procurement files and response files after logging in, and then entrust the bid clearing work to the system. The system will automatically analyze the clauses, procurement files and response files, identify the response points, sort out the supplier response and give the reference opinions. After the intelligent review is completed, in the online bid review activity, the experts can refer to the intelligent review results to carry out the bid review work. In the offline business process, the procurement managers can also export the bid clearing results as the bid review index table for reference in the offline bid review activity of the judges.
[0089] The intelligent bid review function of the application is realized based on a plurality of large language models, realizes the full-automatic entrustment of the bid clearing link, saves the labor cost of manually sorting out the bid review index table, accelerates the development of the bid clearing work, and the large model searches the full text of the bid document, avoids the risk of manual misreading or missing, and improves the quality and efficiency of the bid review.
[0090] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.
[0091] The application further provides an intelligent review method, which will be described below.
[0092] As shown in Figure 10 The intelligent review method of the application comprises the following steps:
[0093] Step S11 is performed to obtain the procurement files and the response files; then step S12 is performed;
[0094] Step S12 is performed to pre-process the obtained procurement files to obtain the vectorization data of the procurement files; then step S13 is performed;
[0095] Step S13 is performed to pre-process the obtained response files to obtain the vectorization data of the response files; then step S14 is performed;
[0096] Step S14 is performed to recall the data related to the review clauses from the vectorization data of the procurement files by means of the vectorization model, and sort out the recalled data related to the review clauses by means of the general large model to obtain the structured data of the review clauses; then step S15 is performed;
[0097] Step S15 is performed to recall the related response points from the vectorization data of the response files according to the structured data of the review clauses; then step S16 is performed;
[0098] A step S16 is performed to analyze the response situation of each response point of the review clause structured data by using the general large model to obtain a response situation analysis result; and then a step S17 is performed.
[0099] A step S17 is performed to give an intelligent review conclusion for the response situation of the response file by using the general large model in combination with the review requirements in the review clause structured data.
[0100] In one specific embodiment of the present application, after obtaining the review clause structured data, the following steps are further included:
[0101] The general large model is used to interpret the clause content in the review clause structured data, and it is judged whether the clause content contains information references, if yes, the corresponding reference content is recalled from the vectorized data of the procurement file to complete the information of the clause content, and then the next round of judgment of whether there is information reference is performed; if no, the review elements are extracted from the clause content.
[0102] In one specific embodiment of the present application, for the review clause structured data, the following steps are included to recall the relevant response points from the vectorized data of the response file:
[0103] The response file fragments are recalled according to the review clause structured data;
[0104] The recalled response file fragments are sorted by using a reordering model, and the first set number of response file fragments are selected;
[0105] The selected response file fragments are integrated to form a response point preliminary screening result.
[0106] In one specific embodiment of the present application, the general large model is used to analyze the response situation of each response point of the review clause structured data, including the following steps:
[0107] Based on the response point preliminary screening result, the general large model is used to screen the content of each response point to judge whether the review elements are responded to, if no, the corresponding response point is excluded, if yes, the response situation of the response point is summarized, and the response situation is recorded in the refined screening result together with the response point information;
[0108] After the content screening of the response point is completed, the refined screening result is combed and merged again to form a response situation analysis result.
[0109] In one specific embodiment of the present application, the following steps are included to pre-process the obtained response file to obtain the vectorized data of the response file:
[0110] The document is cut according to the file directory index, and the chapter content is read page by page;
[0111] The table type data in the response file needs to be standardized into markdown format for easy reading during the reading process;
[0112] For picture type data in the response file, a multi-modal large model and license library data are used to locate and extract key information during the reading process;
[0113] During the page-by-page reading of the chapter content, the page number and the coordinate position information of the content in the page are recorded synchronously;
[0114] After the chapter content reading of the response file is completed, a general large model is used to extract key information and summarize the chapter content;
[0115] The summary of the chapter content is vectorized and stored, and the chapter original text, chapter coordinate information, file type, and original file address are vectorized and stored in the metadata, and the project identification and supplier identification are added to the metadata.
[0116] In one specific embodiment of the present application, as shown in Figure 2 The pre-processing of the obtained procurement file to obtain the vectorized data of the procurement file includes the following steps:
[0117] The procurement file is divided and read by chapter, and the text type data, table type data, and picture type data in the procurement file are read respectively during the reading process; the table type data in the procurement file needs to be standardized into markdown format for easy reading during the reading process; for the picture type data in the procurement file, a multi-modal large model is used to understand and refine the picture content, form a picture description, and extract picture content information during the reading process. After the chapter content reading of the procurement file is completed, a general large model is used to extract key information and summarize the chapter content, and the summary of the chapter content is vectorized and stored, and the chapter original text, project information, and bid evaluation method data information are stored in the metadata together with the summary and stored in the vector database.
[0118] For online bidding procurement projects, after the opening and decryption are completed, the system will automatically start the intelligent evaluation preprocessing process, analyze and process the procurement file and the response file, and interpret the bid evaluation clauses. After the preprocessing is completed, the evaluation analysis calculation will be automatically started, the response points are identified, the supplier response situation is summarized, and the evaluation suggestions are given. The procurement manager and experts can directly view the response situation online during the bid evaluation, and the expert judges can evaluate and score according to the response situation description. The intelligent evaluation suggestions can be opened to the procurement manager, experts, and other users for viewing and reference as needed.
[0119] For the offline bidding procurement project, the project manager can create an intelligent clearing project after the opening of the bidding and upload the procurement file and the response file, the system will automatically identify and extract the evaluation clauses from the procurement file, and perform intelligent clearing work such as response point identification, response summary, evaluation suggestion reasoning based on the evaluation clauses. Finally, the procurement manager can view the clearing result from the system and support exporting the result as a clearing result index table, which can be used as a reference for the subsequent evaluation process.
[0120] The application further provides a storage medium, wherein the storage medium stores a program of the intelligent evaluation method, and the program of the intelligent evaluation method is executed by a processor to implement the steps of the intelligent evaluation method.
[0121] The application is described in detail above in combination with the embodiments of the drawings, and those skilled in the art can make various changes to the application according to the above description. Therefore, some details in the embodiments should not constitute a limitation on the application, and the scope of protection of the application is defined by the appended claims.
Claims
1. An intelligent review method, characterized by, The method comprises the following steps: Obtaining procurement documents and response documents; Preprocessing the obtained procurement documents to obtain vectorized data of the procurement documents; Preprocessing the obtained response documents to obtain vectorized data of the response documents; Using a vectorization model to recall data related to the evaluation clauses from the vectorized data of the procurement documents, using a general large model to sort the recalled data related to the evaluation clauses to obtain evaluation clause structured data; Recalling relevant response points from the vectorized data of the response documents for the evaluation clause structured data; Using a general large model to analyze the response situation for each response point of the evaluation clause structured data to obtain a response situation analysis result; Using a general large model to give an intelligent evaluation conclusion for the response situation of the response documents in combination with the evaluation requirements in the evaluation clause structured data.
2. The intelligent review method of claim 1, wherein, After obtaining the evaluation clause structured data, the following steps are further included: Using a general large model to interpret the clause content in the evaluation clause structured data, and judging whether the clause content contains information references, if yes, recalling the corresponding reference content from the vectorized data of the procurement documents to complete the information of the clause content, and then judging whether there is information reference in the next round; if not, extract the evaluation elements from the clause content.
3. The intelligent review method of claim 1 or 2, wherein, Recalling relevant response points from the vectorized data of the response documents for the evaluation clause structured data comprises the following steps: Recalling response document fragments according to the evaluation clause structured data; Using a reordering model to sort the recalled response document fragments, and selecting a predetermined number of response document fragments; Integrating the selected response document fragments to form a response point preliminary screening result.
4. The intelligent review method of claim 3, wherein, Using a general large model to analyze the response situation for each response point of the evaluation clause structured data comprises the following steps: Based on the response point preliminary screening result, using a general large model to screen the content of each response point, judging whether the evaluation elements are responded to, if not, the corresponding response point is excluded, if yes, the response situation of the response point is summarized, and the response situation is recorded together with the response point information into a refined screening result; After the content screening of the response point is completed, the refined screening result is further sorted and merged to form a response situation analysis result.
5. The intelligent review method of claim 1, wherein, The preprocessing of the obtained response documents to obtain vectorized data of the response documents comprises the following steps: Dividing the documents according to the file directory index and reading the chapter content page by page for the response documents; The table type data in the response documents needs to be standardized to markdown format for easy interpretation during reading; For the picture type data in the response documents, a multi-modal large model and a license library data are used to locate and extract the key information of the pictures; During the process of reading the chapter content page by page, the page number and the coordinate position information of the content in the page are recorded synchronously; After the chapter content of the response document is read, a general large model is used to extract key information and summarize the chapter content; The chapter content of the response file is vectorized, and the chapter original text, chapter coordinate information, file type and original file address are vectorized and stored in the metadata, and the project identification and supplier identification are added in the metadata.
6. A storage medium, characterized by The storage medium stores a program of the intelligent review method, and the program of the intelligent review method is executed by the processor to implement the steps of the intelligent review method in any one of claims 1-5.
7. An intelligent review system, characterized by, Comprise: An acquisition unit is configured to acquire a procurement file and a response file; A preprocessing unit connected with the acquisition unit is configured to preprocess the acquired procurement file to obtain vectorized data of the procurement file; Also used for preprocessing the acquired response file to obtain vectorized data of the response file; A review clause extraction unit connected with the preprocessing unit is configured to use a vectorization model to recall data related to review clauses from the vectorized data of the procurement file, and use a general large model to sort the recalled data related to review clauses to obtain structured data of review clauses; A response recall unit connected with the review clause extraction unit and the preprocessing unit is configured to recall relevant response points from the vectorized data of the response file for the structured data of review clauses; A response analysis unit connected with the review clause extraction unit and the response recall unit is configured to use a general large model to analyze the response situation for each response point of the structured data of review clauses to obtain a response situation analysis result; A response review unit connected with the review clause extraction unit and the response analysis unit is configured to use a general large model to give an intelligent review conclusion for the response file in combination with the review requirements in the structured data of review clauses.
8. The intelligent review system of claim 7, wherein, Also comprising an evaluation element extraction unit connected with the review clause extraction unit and the preprocessing unit; The evaluation element extraction unit is configured to use a general large model to interpret the clause content in the structured data of review clauses, and determine whether the clause content contains information references, if yes, recall the corresponding reference content from the vectorized data of the procurement file to complete the information of the clause content, and then determine whether there is information reference in the next round; if not, extract the evaluation elements from the clause content.
9. The intelligent review system of claim 7, wherein, The response recall unit is also configured to recall response file fragments according to the structured data of review clauses; Also used for using a reordering model to sort the recalled response file fragments, and selecting a predetermined number of response file fragments; Also used for data integration of the selected response file fragments to form a response point preliminary screening result.
10. The intelligent review system of claim 9, wherein, The response analysis unit is also configured to use a general large model to filter each response point content based on the response point preliminary screening result, determine whether the response point content responds to the evaluation elements, if not, eliminate the corresponding response point, if yes, summarize the response situation of the response point, and record the response situation together with the response point information in a refined screening result; After the response point content filtering is completed, the refined screening result is sorted and merged again to form the response situation analysis result. The storage medium stores a program of the intelligent review method, and the program of the intelligent review method is executed by the processor to implement the steps of the intelligent review method in any one of claims 1-5.