Intelligent examination method and system for bidding documents

Abstract

The application relates to a bidding document intelligent review method and system. The method obtains an index segment set of a bidding document and a review dimension list, generates a dimension review instruction according to each review rule, and forms one or more review inputs in combination with the context capacity of a target large model; each review input is combined with the corresponding dimension review instruction and input into the target large model to obtain a model review result including review items and result segment indexes; the model review results corresponding to the same review dimension are formed into a dimension review result, the corresponding segment content is obtained according to the result segment index, and a bidding document review conclusion is generated by summarizing. The system includes a task management module, a review module and a result management module. The application can adapt to bidding documents of different lengths, unify the segmented review results, and associate the review items with the original text content.

Description

Technical Field

[0001] This application relates to the field of computer information processing technology, and in particular to a method and system for intelligent review of tender documents. Background Technology

[0002] With the development of electronic bidding and procurement, bidding documents typically contain a large amount of content, including qualification requirements, commercial terms, technical requirements, payment agreements, performance requirements, liability for breach of contract, and review rules. Currently, bidding document review mainly involves reviewers checking each item of the document according to dimensions such as legality and compliance, completeness, reasonableness, consistency, and fair competition. Because bidding documents are lengthy, contain numerous review items, and related content is scattered across different chapters, manual review requires repeated searching and comparison with the original text, resulting in a significant workload. Furthermore, different reviewers may have differing understandings and interpretations of the same review rules.

[0003] Currently, large language models can be used for automatic review of tender documents. This involves organizing the model input according to preset review rules, allowing the large language model to identify content in the document relevant to those rules, and outputting approved, rejected, and corresponding review comments. For shorter tender documents, the entire document can be used as model input for a comprehensive review; for longer tender documents, the content typically needs to be segmented to avoid exceeding the model's contextual capacity.

[0004] However, existing processing methods struggle to organize review content reasonably according to review rules and model context capacity when reviewing lengthy tender documents. When document content corresponding to the same review dimension is divided into multiple parts for separate review, the review results of each part are independent, making it difficult to form a complete dimension review conclusion according to a unified review standard. It also easily leads to confusion between locally undetected relevant content and actual missing content in the tender document. Furthermore, the original text index output by the model may differ from the actual input document fragments, affecting the accurate correspondence between the review conclusion and the original text content. Therefore, existing technologies still suffer from insufficient adaptability to long document review, difficulty in uniformly processing segmented review results, and insufficient traceability of review conclusions. Summary of the Invention

[0005] The purpose of this application is to address the problem of the difficulty in uniformly processing the review results of lengthy tender documents.

[0006] According to one aspect of this application, a method for intelligent review of tender documents is provided, comprising: S10. Obtain the set of index fragments and the list of review dimensions of the tender document. The set of index fragments includes multiple index fragments with fragment indexes and fragment content, and the list of review dimensions includes multiple review dimensions with review rules. S20. Generate dimension review instructions according to each of the review rules, determine the total input length of each dimension review instruction and the index fragment set, and form one or more review inputs corresponding to each review dimension according to the total input length and the context capacity of the target large model, wherein the review inputs retain the corresponding fragment indexes; S30. Combine each of the review inputs with the corresponding dimension review instructions and input them into the target large model to obtain the corresponding model review results. The model review results include review items and result fragment indexes. S40. Determine the single model review result corresponding to the same review dimension as the dimension review result, or merge multiple corresponding model review results into the dimension review result; obtain the corresponding fragment content according to the result fragment index, summarize the dimension review results, and generate the tender document review conclusion.

[0007] Preferably, in step S20, generating dimensional review instructions according to each of the review rules includes: Determine the current review dimension from the list of review dimensions, parse the review rule corresponding to the current review dimension, and obtain the review objective, pass conditions, and fail trigger conditions; Based on the review objective, the pass conditions, and the fail trigger conditions, the review rules are decomposed into at least one review point to obtain a set of review points; Obtain the preset review process and preset output format, and combine the set of review points, the preset review process and the preset output format to obtain the dimension review instruction corresponding to the current review dimension; Associate the dimension review instruction with the current review dimension, so that one or more review inputs corresponding to the current review dimension share the dimension review instruction.

[0008] Preferably, step S20 further includes: Obtain the encoding rules corresponding to the target large model, and determine the maximum token length allowed to be input into the target large model as the context capacity; According to the encoding rules, calculate the instruction token length of the dimension review instruction and the fragment input token length of all fragment indices and fragment contents in the index fragment set; The total input length is determined by the sum of the instruction token length and the fragment input token length. When the total input length does not exceed the context capacity, all the index fragments are combined according to the order in which they are arranged in the tender document to obtain a review input; When the total input length exceeds the context capacity, the instruction token length and the preset reserved token length are subtracted from the context capacity to obtain the allowed token length. The set of index segments is combined into multiple review inputs according to the order of the index segments in the tender document, so that each review input includes at least one index segment and the token length of each review input does not exceed the allowed token length.

[0009] Preferably, in step S30, the corresponding model review result is obtained, including: The current review input is determined from one or more review inputs corresponding to the current review dimension, and the fragment indices in the current review input are summarized to obtain the input index set; The current review input is combined with the dimension review instruction and then input into the target large model to obtain the candidate model review result. The candidate model review result includes candidate review items and candidate result fragment indexes corresponding to the candidate review items. Determine whether each candidate result fragment index exists in the input index set, retain the candidate result fragment index and the corresponding candidate review item that exist in the input index set, and discard the candidate result fragment index and the corresponding candidate review item that do not exist in the input index set; Based on the retained candidate review items and the candidate result fragment index, the model review result corresponding to the current review input is obtained; when there are no retained candidate review items, the model review result with empty review items is obtained; each review input is processed in sequence to obtain one or more model review results corresponding to the current review dimension.

[0010] Preferably, in step S40, determining a single model review result corresponding to the same review dimension as a dimension review result, or merging multiple corresponding model review results into a dimension review result, includes: Obtain one or more model review results corresponding to the current review dimension, where each review item is either a pass or a fail item; When the current review dimension corresponds to a model review result, the passed items and the failed items in the model review result are respectively determined as the dimension passed item set and the dimension failed item set; When the current review dimension corresponds to multiple model review results, the union of the passed items and the failed items in the multiple model review results is taken to obtain the set of passed items and the set of failed items for the dimension. When the set of failed items for a dimension is not empty, the dimension status of the currently reviewed dimension is determined to be failed; when the set of failed items for a dimension is empty and the set of passed items for a dimension is not empty, the dimension status is determined to be passed; when both the set of failed items for a dimension and the set of passed items for a dimension are empty, the dimension status is determined to be not involved. The dimension is combined with the item set, the dimension failure item set, and the dimension status to obtain the dimension review result corresponding to the current review dimension.

[0011] Preferably, step S40 further includes: Obtain the result fragment index corresponding to each review item in the set of items that passed the dimension and the set of items that failed the dimension, and summarize the result fragment indexes to obtain the dimension-related index set; When the dimension status is not involved and the set of indexes related to the dimension is empty, determine whether the failure trigger condition includes missing content related to the current review dimension; When the failure trigger condition includes missing content related to the current review dimension, a content missing failure item is generated, the content missing failure item is added to the dimension failure item set, and the dimension status is updated to failure; When the failure trigger condition does not include missing content related to the current review dimension, the dimension status remains as not involved. The dimension review result is updated based on the processed set of items that passed the dimension, the set of items that failed the dimension, and the dimension status.

[0012] Preferably, when the set of dimension-related indexes is not empty and the set of dimension-failed items includes multiple failed items, step S40 further includes: Each item that failed includes a problem conclusion, suggested modifications, review criteria, and at least one index of the result fragment; The multiple failed items are grouped according to the problem conclusion of each of the failed items to obtain at least one failed item group, and the multiple failed items in each failed item group have the same problem conclusion; For each group of failed items that includes multiple failed items, the multiple failed items are merged into a target failed item, the target failed item includes the problem conclusion, a set of modification suggestions consisting of multiple modification suggestions, a set of review criteria consisting of multiple review criteria, and a set of result fragment indexes consisting of multiple result fragment indexes; Replace the corresponding multiple failed items in the set of failed items for each of the aforementioned targets with the failed items for each dimension, and update the dimension review result based on the replaced set of failed items for each dimension.

[0013] Preferably, step S30 includes: The dimension review instruction corresponding to each review dimension and one or more review inputs are combined to obtain the corresponding dimension review task; Obtain a preset concurrency level, and execute multiple dimension review tasks in batches according to the preset concurrency level until each dimension review task is completed or terminated abnormally. In each of the dimensional review tasks, the review inputs are combined with the dimensional review instructions and then input into the target large model to obtain the model review results corresponding to each of the review inputs. The review results of each model are associated with the corresponding review dimensions.

[0014] Preferably, multiple dimensional review tasks are executed in batches and in parallel, including: Obtain the task execution status of each of the aforementioned dimension review tasks, and determine the abnormal dimension review tasks that are experiencing execution errors based on the task execution status; Terminate the abnormal dimension review task and do not retry the abnormal dimension review task. Associate the task execution status corresponding to the abnormal dimension review task with the corresponding review dimension to obtain an abnormal task record. Continue executing the other dimension review tasks besides the abnormal dimension review task, until the other dimension review tasks are completed or terminated abnormally; Associate the model review result corresponding to the completed dimension review task with the corresponding review dimension, and output the abnormal task record.

[0015] This application also provides an intelligent tender document review system, which, when running, executes the intelligent tender document review method as described in any of the above claims, including: The task management module is used to receive tender documents and review configurations, generate review task data, and output the review task data to the review module. The review configurations include multiple review dimensions. The review module is connected to the task management module and is used to receive the review task data, call the target large model to perform the tender document review, generate the model review result, and output the model review result to the result management module. The result management module is connected to the review module and is used to receive the model review results, form the dimension review results corresponding to each of the review dimensions, and summarize the review results of each of the dimensions to generate the tender document review conclusion.

[0016] This application offers the following advantages: It generates dimensional review instructions based on the review rules corresponding to each review dimension, and, combined with the contextual capacity of the target large model, forms one or more review inputs for the set of index fragments in the tender document, ensuring that tender documents of varying lengths can adapt to the input range of the target large model. For the same review dimension, each review input is processed separately using the corresponding dimensional review instructions, and the resulting model review results are directly determined or merged into the dimensional review results. This ensures that the segmented review items are still aggregated according to a unified review standard, reducing the problem of isolated local results. Simultaneously, it retrieves the corresponding fragment content based on the result fragment index in the model review results, and summarizes the review results of each dimension to generate the tender document review conclusion. This establishes a clear connection between the review items and the original text of the tender document, improving the completeness, consistency, and traceability of the review conclusion. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a logic block diagram of the intelligent review method for tender documents described in one embodiment of this application. Detailed Implementation

[0019] To facilitate understanding of this application, a more complete description will be provided below with reference to the accompanying drawings. Preferred embodiments of this application are shown in the drawings. However, this application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of this application.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0021] Please refer to Figure 1 Yes, the previous embodiment should also retain the complete technical content of each step, and then explain them separately. It can be directly replaced with the following content: One embodiment of this application provides a method for intelligent review of tender documents, including: S10. Obtain the set of index fragments and the list of review dimensions of the tender documents. The set of index fragments includes multiple index fragments with fragment indexes and fragment content, and the list of review dimensions includes multiple review dimensions with review rules.

[0022] In this step, it should be noted that the tender document can be a text file, a portable document file, or an electronic document obtained through format conversion. The index fragment set is an ordered set formed after organizing the content of the tender document, with each index fragment corresponding to a portion of the content in the tender document. The fragment index is used to uniquely identify the corresponding index fragment, and the fragment content is the text content within that index fragment that is used for subsequent review and processing. The index fragments are arranged according to the chronological order of their corresponding content in the tender document, so that the index fragment set can represent the content order of the tender document.

[0023] Review dimensions define the scope of the review, while review rules describe the review requirements and judgment criteria for each dimension. Different review dimensions correspond to their own review rules, ensuring that subsequent processing focuses on a single review matter and avoids mixing review requirements of different natures in the same judgment process. Review dimensions can be configured according to actual review needs, and multiple review dimensions together constitute the scope of this tender document review.

[0024] S20. Generate dimension review instructions according to each review rule, determine the total input length of each dimension review instructions and index fragment set, and form one or more review inputs corresponding to each review dimension based on the total input length and the context capacity of the target large model. The review inputs retain the corresponding fragment indexes.

[0025] In this step, it's important to note that the dimensional review instruction is model control information generated based on the corresponding review rules. It is used to define the review content, judgment scope, and result organization method adopted by the target large model for the current review dimension. Different review dimensions generate corresponding dimensional review instructions, enabling each review dimension to be processed independently according to its own review rules.

[0026] The total input length is the overall length of the corresponding dimension review instructions and index fragment set organized according to the model input format. The context capacity of the target large model is the range of input content that the target large model can receive in a single call. For each review dimension, the corresponding total input length is compared with the context capacity of the target large model, and the number of review inputs corresponding to that review dimension is determined accordingly. When the corresponding content can be input in one model call, the index fragment set is formed into one review input; when the corresponding content cannot be input in one model call, the index fragment set is formed into multiple review inputs.

[0027] Each review input consists of at least one index fragment, and retains the fragment index corresponding to each index fragment. Thus, even after the tender document is organized into multiple review inputs, the content of each fragment remains clearly identified. Multiple review inputs corresponding to the same review dimension are used in conjunction with the dimension review instructions for that dimension, adjusting only the range of fragment content for each input without changing the corresponding review rules.

[0028] S30. Combine each review input with the corresponding dimension review instruction and input them into the target large model to obtain the corresponding model review results. The model review results include review items and result fragment indexes.

[0029] In this step, it should be noted that for each review dimension, the corresponding review inputs are combined with the dimension review instructions, so that the target large model, while receiving the tender document fragments, simultaneously obtains the review requirements applied to those fragments. One review input corresponds to one model processing step, resulting in a corresponding model review result; when the same review dimension corresponds to multiple review inputs, multiple model review results corresponding to each review input are obtained.

[0030] The model review result is a structured result generated by the target large model according to the dimensional review instructions. Review items record the judgments made by the target large model regarding the current review dimension and the current review input. Result fragment indexes indicate the index fragments related to that review item. A model review result may include one or more review items, and each review item may correspond to one or more result fragment indices. The target large model establishes the association between review items and the content of the tender document by outputting result fragment indices, without having to repeatedly output the complete fragment content in the model review result.

[0031] S40. Determine the review result of a single model corresponding to the same review dimension as the dimension review result, or merge the review results of multiple corresponding models into the dimension review result. Obtain the corresponding fragment content according to the result fragment index, summarize the review results of each dimension, and generate the tender document review conclusion.

[0032] In this step, it's important to note that the dimensional review results are organized according to individual review dimensions. When the same review dimension corresponds to only one review input, the single model review result corresponding to that input is determined as the dimensional review result for that review dimension. When the same review dimension corresponds to multiple review inputs, the model review results corresponding to each input are merged according to their common review dimension, thus aggregating review items scattered across different review inputs into a single dimensional review result. Model review results corresponding to different review dimensions are processed separately and are not mixed when forming the dimensional review results.

[0033] For review items with result fragment indexes in the model review results, the corresponding fragment content is searched in the index fragment set according to the result fragment index, and the found fragment content is associated with the corresponding review item. After processing each review dimension, the review results of each dimension are summarized according to the review dimension list to generate the tender document review conclusion. The tender document review conclusion can record the corresponding review items and their fragment content according to different review dimensions, and can be used for result display, review report generation, or manual review.

[0034] The technical solution implemented in this embodiment can generate dimensional review instructions for different review dimensions, and, combined with the context capacity of the target large model, adapt the tender document into one or more review inputs, enabling tender documents of different lengths to complete the review along the same processing link. When the tender document is organized into multiple review inputs, multiple review inputs corresponding to the same review dimension use the same dimensional review instruction, and after model processing, the model review results are merged according to the review dimensions, so that the review items obtained from decentralized processing can form a complete dimensional review result. By passing the fragment index between the index fragment, the review input, and the model review result, the corresponding fragment content can be obtained according to the result fragment index, establishing a clear association between the review items and the original text of the tender document, and forming a tender document review conclusion organized according to the review dimensions.

[0035] In one specific embodiment, step S20 involves generating dimensional review instructions based on each review rule, including: Determine the current review dimension from the list of review dimensions, parse the review rules corresponding to the current review dimension, and obtain the review objectives, approval conditions, and rejection trigger conditions.

[0036] In this step, it should be noted that the current review dimension is the review dimension currently pending processing in the review dimension list. The review objective is used to define the review object and scope targeted by the current review dimension; the pass conditions describe the conditions that must be met for the relevant content in the tender document to comply with the review rules; the fail trigger conditions describe the conditions that must be met when the relevant content in the tender document has omissions, conflicts, violations, or other non-compliance with the review rules. When parsing the review rules, the content in the review rules used to describe the object to be verified, compliance requirements, and abnormal situations is identified and semantically categorized into review objectives, pass conditions, and fail trigger conditions. For review rules that only specify fail situations, the review objective and fail trigger conditions can be determined based on the corresponding fail situation; for review rules that simultaneously specify positive requirements and prohibitive requirements, pass conditions and fail trigger conditions are formed respectively.

[0037] Based on the review objectives, approval conditions, and disapproval trigger conditions, the review rules are decomposed into at least one review point, resulting in a set of review points.

[0038] In this step, it should be noted that the review points are review items that can be independently verified and judged. When a review rule involves multiple objects to be verified, multiple judgment conditions, or multiple abnormal situations, the review rule is decomposed into multiple review points according to different objects to be verified and their corresponding judgment conditions. When a review rule involves only one object to be verified, that object to be verified is identified as one review point. Each review point includes the content to be verified and the pass or fail trigger conditions corresponding to that content. All review points together cover the review rules corresponding to the current review dimension. For multiple review points that are logically related, the sequential judgment relationship or common satisfaction relationship between the review points can also be recorded, so that the target large model can verify the relevant content in the bidding documents item by item according to the corresponding relationship.

[0039] Obtain the preset review process and preset output format, combine the set of review points, preset review process and preset output format to obtain the dimension review instruction corresponding to the current review dimension.

[0040] In this step, it's important to note that the preset review process defines the processing order and judgment method for each review point in the target large model. The preset review process may include identifying content fragments related to the review points, determining whether the relevant content fragments meet the pass or fail trigger conditions, and generating corresponding review items. If no relevant content fragments are identified, the preset review process may further define whether content loss constitutes a failure based on the corresponding fail trigger conditions. The preset output format defines the data organization of the model review results and may include fields such as review item category, problem conclusion, modification suggestions, review basis, and result fragment index. Combining the review point set, preset review process, and preset output format according to the preset instruction template ensures that the resulting dimension review instruction simultaneously includes the content to be reviewed, judgment conditions, processing flow, and result format, corresponding to the current review dimension.

[0041] Associate the dimension review instruction with the current review dimension, so that one or more review inputs corresponding to the current review dimension can share the dimension review instruction.

[0042] In this step, it's important to establish a correspondence between the current review dimension and the dimension review instructions, and determine the applicable review inputs for each dimension review instruction based on this correspondence. When the current review dimension corresponds to one review input, the dimension review instruction is combined with that review input and then input into the target large model. When the current review dimension corresponds to multiple review inputs, the same dimension review instruction is combined with each review input separately, without changing the review points, pass conditions, fail trigger conditions, preset review processes, and preset output formats regardless of the different fragment content contained in each review input. Different review dimensions are associated with their respective dimension review instructions, enabling each review input to invoke the corresponding dimension review instruction based on its assigned review dimension.

[0043] The technical solution implemented in this embodiment can convert review rules presented to reviewers into dimensional review instructions consisting of review objectives, judgment conditions, and review points. This allows the target large model to process the content of the tender document item by item, focusing on review matters that can be independently judged. By uniformly configuring review points, review processes, and output formats in the dimensional review instructions, and by allowing one or more review inputs corresponding to the same review dimension to share the same dimensional review instructions, the multi-model processing can adopt the same review scope, judgment conditions, and result structure, providing a unified data foundation for the subsequent merging of review results from multiple models under the same review dimension.

[0044] In one specific embodiment, step S20 further includes: Obtain the encoding rules corresponding to the target large model, and determine the maximum token length allowed for input to the target large model as the context capacity.

[0045] In this step, it's important to note that different target models can employ different text encoding and token partitioning methods, and the same text may correspond to different token lengths under different encoding rules. Therefore, we first obtain the encoding rule corresponding to the current target model to uniformly calculate the token length for subsequent input content. Context capacity characterizes the maximum token range that the target model can receive in a single processing iteration, and it can be determined based on the target model's model type, version, or calling parameters. When the target model is switched, the corresponding new encoding rule and maximum token length are obtained.

[0046] According to the encoding rules, calculate the instruction token length of the dimension review instruction and the fragment input token length of all fragment indices and fragment contents in the index fragment set.

[0047] In this step, it should be noted that the instruction token length is the number of tokens obtained after encoding the dimension review instruction according to the encoding rules. The fragment input token length is the number of tokens obtained after encoding all fragment indices and fragment contents of the target large model to be input into the index fragment set according to the same encoding rules. When calculating the fragment input token length, the fragment index and the corresponding fragment content are counted as the input content of the same index fragment, so that the fragment index can be retained when the review input is formed later, and the fragment index is not excluded from the input length calculation.

[0048] The total input length is determined by the sum of the instruction token length and the fragment input token length.

[0049] In this step, it should be noted that the total input length corresponds to the token length used when the dimension review instructions and all index fragments are input into the target large model under the current review dimension. Since the dimension review instructions corresponding to different review dimensions may have different content and lengths, the corresponding total input length is calculated separately for each review dimension, and the overall input method or the segmented input method is determined based on the total input length.

[0050] When the total input length does not exceed the context capacity, all index fragments are combined according to their order of arrangement in the tender document to obtain a review input.

[0051] In this step, it's important to note that when the dimensional review instruction and all index fragments fall within the single input range of the target large model, the index fragment set is not segmented. Instead, all index fragments are combined into a single review input according to the order in which their corresponding content appears in the tender document. During the combination process, each index fragment retains its corresponding fragment index and content, enabling the target large model to obtain all the content to be reviewed in the tender document in a single processing step, and to identify different fragment contents based on their fragment indexes.

[0052] When the total input length exceeds the context capacity, the instruction token length and the preset reserved token length are subtracted from the context capacity to obtain the allowed token length. The set of index fragments is then combined into multiple review inputs according to the order of each index fragment in the tender document, so that each review input includes at least one index fragment and the token length of each review input does not exceed the allowed token length.

[0053] In this step, it should be noted that the preset reserved token length is the token length in the context capacity that is not allocated to the index fragment, used to reserve a preset length margin when forming each review input. The allowed token length is used to limit the maximum token length that an index fragment can occupy in a single review input, and it varies with the instruction token length of the current dimension review instruction.

[0054] When generating multiple review inputs, the index fragments are combined sequentially according to their order in the tender document. Once the token length of the current combination reaches the allowed token length range, subsequent index fragments are combined into the next review input. Each index fragment retains its fragment index and content as a whole, and the correspondence between the fragment index and content remains unchanged despite segmentation processing. All review inputs collectively contain all index fragments in the index fragment set, and are then combined with the dimension review instructions corresponding to the current review dimension before being input into the target large model.

[0055] The technical solution implemented in this embodiment can accurately determine the token length of the dimension review instructions and the set of index fragments based on the encoding rules corresponding to the target large model, and adaptively select the overall input or segmented input method based on the context capacity of the target large model. When the tender document can be input in a single model call, the overall processing of all index fragments is maintained; when the tender document exceeds the range of a single input, the set of index fragments is organized into multiple review inputs that meet the allowed token length according to the original text order. Thus, without changing the order of index fragments and the index correspondence, tender documents of different lengths can be adapted to the context capacity of the target large model, and an input basis is provided for multiple review inputs under the same review dimension to adopt a unified dimension review instruction.

[0056] In one specific embodiment, step S30 obtains the corresponding model review result, including: Determine the current review input from one or more review inputs corresponding to the current review dimension, and summarize the fragment indices in the current review input to obtain the input index set.

[0057] In this step, it's important to note that the current review input refers to the review input of the target large model to be input under the current review dimension. When the current review dimension corresponds to only one review input, that review input is determined as the current review input; when the current review dimension corresponds to multiple review inputs, the current review inputs are determined sequentially according to the order of the index fragments contained in each review input in the tender document. The input index set consists of the fragment indices of each index fragment in the current review input, used to characterize the range of fragments actually received by the current model call. Since each fragment index is used to uniquely identify the corresponding index fragment, the range of indices that the target large model can reference in this processing can be determined based on the input index set.

[0058] The current review input is combined with the dimension review instruction and then input into the target large model to obtain the candidate model review results. The candidate model review results include candidate review items and the candidate result fragment index corresponding to the candidate review items.

[0059] In this step, it should be noted that the current review input provides the target large model with the content of the segment to be reviewed and its index, while the dimensional review instruction provides the target large model with the review points, judgment conditions, review process, and output format corresponding to the current review dimension. The two are combined and input into the target large model, which then processes the segment content in the current review input according to the dimensional review instruction and generates the candidate model review results according to the preset output format.

[0060] Candidate review items are the initial judgments formed by the target model based on the current review input. They may include at least one of the following: review item category, issue conclusion, modification suggestions, or review basis. Candidate result fragment indexes indicate the index fragments associated with the corresponding candidate review item. One candidate review item can correspond to one or more candidate result fragment indices, and the same candidate result fragment index can be associated with multiple candidate review items. At this stage, the candidate model review results have not yet undergone index validity verification and are not directly used as the model review results corresponding to the current review input.

[0061] Determine whether each candidate result fragment index exists in the input index set. If it exists, retain the candidate result fragment index and its corresponding candidate review item. If it does not exist, discard the candidate result fragment index and its corresponding candidate review item.

[0062] In this step, it's important to note that each candidate result fragment index in the candidate model review results is matched against the fragment indices in the input index set. If a candidate result fragment index exists in the input index set, it indicates that the index points to the index fragment actually included in the current review input; therefore, the candidate result fragment index and its corresponding candidate review item are retained. If a candidate result fragment index does not exist in the input index set, it indicates that the index does not point to the index fragment actually received by the current model call; therefore, the candidate result fragment index and its corresponding candidate review item are discarded.

[0063] When a candidate review item corresponds to multiple candidate result fragment indices, each candidate result fragment index can be validated separately, and those that exist in the input index set can be retained. If none of the candidate result fragment indices corresponding to the candidate review item exist in the input index set, then the candidate review item is discarded. By limiting the validation scope to the input index set corresponding to the current review input, it is possible to avoid incorrectly associating fragment indices from other review inputs or invalid indices generated by the target large model with the current candidate review item.

[0064] Based on the retained candidate review items and candidate result fragment indices, the model review result corresponding to the current review input is obtained. When there are no retained candidate review items, a model review result with empty review items is obtained. Each review input is processed sequentially to obtain one or more model review results corresponding to the current review dimension.

[0065] In this step, it's important to note that candidate review items that pass the index validity check are designated as review items, and the retained candidate result fragment indices are designated as the corresponding result fragment indices, thus forming the model review result corresponding to the current review input. If all candidate review items are discarded due to invalid candidate result fragment indices, or if the target model does not generate any candidate review items, a model review result corresponding to the current review input is still generated, and the review items in this model review result are set to empty. This situation is not directly identified as a model call anomaly or a failure of the current review dimension.

[0066] After processing the current review input, a new current review input is determined from the remaining review inputs corresponding to the current review dimension. The processes of generating the input index set, calling the model, validating the index, and generating the model review result are repeated until all review inputs corresponding to the current review dimension have been processed. When the current review dimension corresponds to only one review input, one model review result is obtained; when the current review dimension corresponds to multiple review inputs, multiple model review results are obtained, each corresponding to a separate review input, and each model review result maintains its association with its corresponding review input.

[0067] The technical solution implemented in this embodiment can use the actual fragment indexes contained in the current review input as the verification scope to determine the validity of the candidate result fragment indexes output by the target large model, and discard candidate review items that do not point to the indexed fragments in the current review input, thereby reducing erroneous associations caused by inconsistencies between the model output indexes and the actual input content. By retaining the verified review items and their result fragment indexes, each model review result can correspond to a clear review input and original text fragment, and when no valid review item is obtained, a model review result with empty review items is formed, providing a data foundation for subsequent differentiation of content not covered, content missing, and model execution anomalies.

[0068] In one specific embodiment, step S40, determining a single model review result corresponding to the same review dimension as a dimension review result, or merging multiple corresponding model review results into a dimension review result, includes: Retrieve one or more model review results corresponding to the current review dimension. Each review item is either a pass or a fail item.

[0069] In this step, it should be noted that the current review dimension is the review dimension from which the review result is to be formed. Based on the correlation between the current review dimension and the model review result, the model review result generated by each review input corresponding to the current review dimension is obtained. When the current review dimension corresponds to one review input, one model review result is obtained; when the current review dimension corresponds to multiple review inputs, multiple model review results are obtained. Passed items record review items in the tender document that meet the corresponding review rules, while failed items record review items in the tender document that meet the failure trigger conditions. Each model review result may include only passed items, only failed items, both passed and failed items, or no review items.

[0070] When the current review dimension corresponds to a model review result, the passed items and failed items in the model review result are respectively determined as the set of passed items and the set of failed items for the dimension.

[0071] In this step, it's important to note that when the set of index fragments from the tender document forms a review input under the current review dimension, the model review result corresponding to that input covers the content to be reviewed under this current review dimension. Therefore, according to the category of the review item, items that pass the model review result are categorized into the dimension's pass item set, and items that fail are categorized into the dimension's fail item set. When there are no pass items in the model review result, the dimension's pass item set is empty; when there are no fail items in the model review result, the dimension's fail item set is empty.

[0072] When the current review dimension corresponds to multiple model review results, the union of the passed and failed items in the multiple model review results is taken to obtain the set of passed items and the set of failed items for the dimension.

[0073] In this step, it's important to note that when the current review dimension corresponds to multiple review inputs, each model's review result reflects the review status of its respective input. Passing items from multiple model review results are grouped into the same category and their union is taken to obtain a set of passing items covering all review inputs. Similarly, failing items from multiple model review results are grouped into the same category and their union is taken to obtain a set of failing items covering all review inputs. When taking the union, identical review items are retained only once, while different review items from different review inputs are all retained. This ensures that review items scattered across multiple model review results are grouped according to the current review dimension without changing the passing or failing category of each review item.

[0074] When the set of failed items for a dimension is not empty, the dimension status of the currently reviewed dimension is determined to be failed. When the set of failed items for a dimension is empty and the set of passed items for a dimension is not empty, the dimension status is determined to be passed. When both the set of failed items for a dimension and the set of passed items for a dimension are empty, the dimension status is determined to be not involved.

[0075] In this step, it's important to note that the dimension status is used to characterize the overall review status of the current review dimension. When at least one item in the dimension's failed item set is failed, the dimension status is determined to be failed, regardless of whether the dimension's passed item set is empty. This ensures that any failed items identified in any review input are reflected in the overall status of the current review dimension. When the dimension's failed item set is empty and the dimension's passed item set is not empty, it indicates that all review items identified under the current review dimension are passed, and the dimension status is determined to be passed. When both sets are empty, it indicates that none of the model review results have resulted in any review items that can be categorized as passed or failed, and the dimension status is determined to be uninvolved. The uninvolved status only indicates that there are no passed or failed items in the current model review results; it does not directly indicate whether the tender document complies with or does not comply with the corresponding review rules.

[0076] The dimension review result is obtained by combining the item set, the item set of the dimension that failed, and the dimension status.

[0077] In this step, it's important to note that the set of approved and rejected dimension items, categorized by review item type, are linked to the dimension status determined by the contents of these two sets. These sets are then combined according to a pre-defined dimension result structure to form the dimension review result for the current review dimension. The dimension review result simultaneously records the specific review items and the overall dimension status under the current review dimension, allowing the dimension status to be interpreted by the corresponding set of approved and rejected dimension items. After processing the current review dimension, other review dimensions can be processed sequentially to obtain the dimension review results for each dimension.

[0078] The technical solution implemented in this embodiment can, based on the number of model review results corresponding to the same review dimension, adopt either a single-result direct aggregation or a multi-result classification and merging processing method to uniformly classify the passed and failed items scattered across different review inputs into the corresponding dimension review results. By determining the dimension status of failed, passed, or not involved according to the content of the dimension's failed item set and passed item set, the failed items appearing in the local review input can be reflected in the overall status of the current review dimension, and differentiating between the existence of passed content, the existence of failed content, and the absence of review items, thereby forming dimension review results corresponding to the content and status.

[0079] In one specific embodiment, step S40 further includes: Obtain the result fragment index corresponding to each review item in the set of approved and unapproved items of the dimension, and summarize the result fragment indexes to obtain the set of dimension-related indexes.

[0080] In this step, it's important to note that the dimension-related index set records the range of index fragments for review items that have been formed under the current review dimension. The result fragment indices corresponding to each review item in both the dimension's passed and failed item sets are extracted, and then summarized. When multiple review items correspond to the same result fragment index, one such index can be retained in the dimension-related index set; when a review item corresponds to multiple result fragment indices, each result fragment index is added to the dimension-related index set separately. Therefore, the dimension-related index set can be used to determine whether there are any review items in the current dimension's review results that are related to the content of the tender document fragments.

[0081] When the dimension status is "not involved" and the dimension-related index set is empty, determine whether the failure trigger condition includes missing content related to the current review dimension.

[0082] In this step, it's important to note that a dimension status of "not involved" and an empty set of related indexes indicates that none of the model review results corresponding to the current review dimension have resulted in either a pass or fail item, nor are there any result fragment indexes associated with the review item. At this point, the current review dimension cannot be directly determined as pass or fail. It's necessary to consider the failure trigger conditions corresponding to the current review dimension to determine whether the absence of relevant content in the tender document falls under the failure circumstances stipulated in the review rules. Specifically, it's possible to identify whether the failure trigger conditions include missing materials, incomplete clauses, lack of necessary content, or other conditions based on the absence of relevant content.

[0083] When the failure trigger condition includes missing content related to the current review dimension, a content missing failure item is generated, the content missing failure item is added to the dimension failure item set, and the dimension status is updated to failure.

[0084] In this step, it should be noted that when the trigger condition defines missing content as a failure, even if a corresponding result fragment index cannot be obtained from the tender document, a content missing failure item needs to be generated based on the failure trigger condition. The content missing failure item records the review conclusion that necessary content required for verification in the current review dimension is not present in the tender document. It can include missing content, corresponding review requirements, and supplementary suggestions. Since the content missing failure item is generated based on the absence of relevant content, it does not need to be configured with a result fragment index. After adding the content missing failure item to the dimension failure item set, the dimension failure item set is updated from an empty set to a non-empty set, and the dimension status of the current review dimension is updated from not involved to failed.

[0085] If the triggering conditions do not include missing content related to the current review dimension, the dimension status remains "not involved".

[0086] In this step, it's important to note that if the failure trigger condition does not explicitly define missing content as a failure circumstance, the absence of relevant content does not directly indicate that the tender documents do not comply with the review rules. Therefore, no content-missing failure items are generated, the set of passed and failed items for each dimension remains empty, and the dimension status of the current review dimension remains "not involved." "Not involved" indicates that no identifiable content corresponding to the current review dimension was obtained from the tender documents during this review; it does not equate to the current review dimension passing or failing.

[0087] Update the dimension review results based on the processed set of approved items, the set of unapproved items, and the dimension status.

[0088] In this step, it's important to note that the set of approved items, the set of rejected items, and the dimension status, after content missing assessment, are recombined to replace the original dimension review result for the current dimension. When a content missing item is generated, the updated dimension review result includes both the rejected item and the rejected dimension status. When no content missing item is generated, the updated dimension review result keeps both the set of approved and rejected items empty and records the unaddressed dimension status. This allows the dimension review result to distinguish between two scenarios: no relevant content was found but its presence is not required, and no relevant content was found where the absence itself constitutes a rejection.

[0089] The technical solution implemented in this embodiment can distinguish between "not involved" and "content missing" when the model review results do not generate a pass or fail item, based on the failure trigger conditions corresponding to the current review dimension. When the relevant content missing is a failure condition, a content missing failure item is generated and the dimension status is updated; when the relevant content missing is not a failure condition, the "not involved" status is maintained, avoiding a direct pass or fail conclusion simply because the relevant fragment was not detected, thereby ensuring that the dimension review results match the content completeness requirements in the corresponding review rules.

[0090] In one specific embodiment, when the set of dimension-related indexes is not empty and the set of dimension-failed items includes multiple failed items, step S40 further includes: Each failed item includes a problem conclusion, suggested modifications, review basis, and at least one result fragment index.

[0091] In this step, it should be noted that the "Issue Conclusion" describes the specific issues in the tender document that do not comply with the review rules corresponding to the current review dimension; the "Modification Suggestion" records the handling methods for this issue; the "Review Basis" records the review rules or judgment conditions on which the "Issue Conclusion" is based; and the "Result Segment Index" points to the index segments related to the "Issue Conclusion." Each rejected item is associated with the corresponding segment content in the tender document through at least one result segment index. When an issue involves multiple index segments, a rejected item may include multiple result segment indices to collectively indicate the document content on which the "Issue Conclusion" is based.

[0092] Based on the conclusion of each failed item, multiple failed items are grouped to obtain at least one failed item group. Multiple failed items in each failed item group have the same conclusion.

[0093] In this step, it's important to note that the problem conclusions of each failed item in the set of failed items for each dimension are read sequentially, and the problem conclusions of different failed items are compared to see if they are the same. Failed items with the same problem conclusion are grouped into the same failed item group, while failed items with different problem conclusions are grouped into different failed item groups. Each failed item group corresponds to one problem conclusion and may include one or more failed items. When the same problem appears in different review inputs or different index segments, multiple failed items formed by the corresponding model review results are grouped into the same failed item group; failed items that do not have the same problem conclusion as other failed items form separate failed item groups.

[0094] For each group of failed items, the multiple failed items are merged into a target failed item. The target failed item includes the problem conclusion, a set of modification suggestions consisting of multiple modification suggestions, a set of review basis consisting of multiple review basis, and a set of result fragment indexes consisting of multiple result fragment indexes.

[0095] In this step, it should be noted that only groups of failed items containing multiple failed items are merged; groups of failed items containing only one failed item are not merged. For the failed item groups to be merged, the common problem conclusions of all failed items are extracted, and these problem conclusions are identified as the problem conclusions of the target failed item; modification suggestions for each failed item are summarized to obtain a set of modification suggestions; the review basis for each failed item is summarized to obtain a set of review basis; and the result fragment indices corresponding to each failed item are summarized to obtain a set of result fragment indices.

[0096] When summarizing the modification suggestions, review basis, and result fragment index, duplicate content can be deleted, while the differences between different failed items can be retained. The target failed items are simultaneously associated with multiple occurrences of the same issue in the tender document through the result fragment index set, and the processing information and judgment basis corresponding to each failed item before merging are retained through the modification suggestion set and review basis set.

[0097] Replace multiple failed items in the set of failed items for each dimension with the failed items for each target, and update the dimension review results based on the replaced set of failed items for each dimension.

[0098] In this step, it's important to note that in the set of failed dimension items, multiple failed items that have already been merged are deleted, and a target failed item formed by these multiple failed items is added to the corresponding position. Failed items that haven't been merged are retained in the set of failed dimension items. After processing each group of failed items, a replaced set of failed dimension items is obtained. This replaced set of failed dimension items is then recombined with the set of passed dimension items and the dimension status to update the dimension review result corresponding to the currently reviewed dimension. Since the replaced set of failed dimension items still includes at least one failed item, the dimension status of the currently reviewed dimension remains failed.

[0099] The technical solution implemented in this embodiment can group multiple failed items with the same problem conclusion under the same review dimension into target failed items, and centrally retain the corresponding modification suggestions, review basis, and result fragment index in the target failed items. When the same problem appears in different review inputs or different positions in the tender documents, it can reduce the duplicate recording of problem conclusions in the dimension review results, while retaining multiple original text positions and review information corresponding to the same problem, so that the dimension review results are centrally organized according to problem type.

[0100] In one specific embodiment, step S30 includes: The corresponding dimension review task is obtained by combining the dimension review instructions and one or more review inputs for each review dimension.

[0101] In this step, it's important to note that dimensional review tasks are model processing tasks organized according to review dimensions. Each dimensional review task corresponds to a review dimension and includes the corresponding dimensional review instruction and one or more review inputs. The dimensional review instruction defines the review content, judgment conditions, processing flow, and output format of the current dimensional review task. The review input provides the fragment index and fragment content to be reviewed. When a review dimension corresponds to multiple review inputs, all review inputs are grouped into the dimensional review task corresponding to that dimension, allowing each review input to share the corresponding dimensional review instruction within the same task. Different review dimensions form separate dimensional review tasks to maintain the independence of the review rules and model processing results for each review dimension.

[0102] Obtain the preset concurrency level, and execute multiple dimension review tasks in batches according to the preset concurrency level until the review tasks of each dimension are completed or terminated abnormally.

[0103] In this step, it's important to note that the preset concurrency level limits the number of dimensional review tasks that can be executed simultaneously. Based on the preset concurrency level, the current batch of dimensional review tasks is determined from the remaining unexecuted dimensional review tasks, and these tasks are executed in parallel. Once a dimensional review task in the current batch completes or terminates abnormally, the next batch is determined from the remaining unexecuted dimensional review tasks, and so on, until all dimensional review tasks have completed or terminated abnormally. When the number of dimensional review tasks does not exceed the preset concurrency level, all dimensional review tasks can be executed in parallel as a single batch; when the number of dimensional review tasks exceeds the preset concurrency level, they are divided into multiple batches according to the preset concurrency level. Each dimensional review task has an independent execution process; if a single dimensional review task terminates abnormally, it does not change the model input content or review rules of other dimensional review tasks.

[0104] In each dimension review task, each review input is combined with the dimension review instruction and then input into the target large model to obtain the model review result corresponding to each review input.

[0105] In this step, it should be noted that for the currently executed dimensional review task, each review input in that dimensional review task is obtained sequentially, and each review input is combined with the dimensional review instruction in that dimensional review task to form the corresponding model input. The target large model reviews the fragment content in the current review input according to the dimensional review instruction and outputs the model review result corresponding to that review input.

[0106] When a dimensional review task includes only one review input, model processing is performed once, resulting in one model review result. When a dimensional review task includes multiple review inputs, model processing is performed on each review input separately, resulting in multiple model review results corresponding to each review input. All review inputs within the same dimensional review task use the same dimensional review instructions, and the review points, judgment conditions, and output format remain unchanged regardless of the content or scope of the review input segments.

[0107] Associate the review results of each model with the corresponding review dimensions.

[0108] In this step, it's important to note that when generating model review results, the dimension review task that produced the result is recorded. Based on the correspondence between dimension review tasks and review dimensions, the model review result is associated with the corresponding review dimension. When a review dimension corresponds to only one model review result, a correspondence is established between that review dimension and the model review result; when a review dimension corresponds to multiple model review results, the multiple model review results are associated with the same review dimension.

[0109] The associated model review results can retain the corresponding task identifier, review dimension identifier, or other information that can represent the corresponding relationship, so that one or more model review results can be obtained according to the review dimension later. The model review results corresponding to different review dimensions are associated separately, and the model review results generated by one review dimension are not included in other review dimensions.

[0110] The technical solution implemented in this embodiment can organize dimension review instructions and review inputs on a unit basis, and execute multiple dimension review tasks in batches according to a preset concurrency level, enabling different review dimensions to independently call the target large model. When the same review dimension corresponds to multiple review inputs, each review input shares the same dimension review instruction and generates its own model review result, ensuring consistency in the review rules and result format used during the segmented input process. By associating each model review result with its corresponding review dimension, a clear data relationship can be provided for subsequent aggregation and merging of model review results according to review dimensions.

[0111] In one specific embodiment, multiple dimensions of review tasks are executed in batches and in parallel, including: Obtain the task execution status of each dimension of the review task, and determine the abnormal dimension review task based on the task execution status.

[0112] In this step, it's important to note that the task execution status records the current state of the dimensional review task during execution, which can include pending execution, in-process execution, completed execution, and execution exception states. After the dimensional review task begins execution, its model invocation process undergoes status checks. If the target model fails to return results, the model invocation is interrupted, execution times out, the returned data does not conform to the preset format, or other situations occur that prevent further execution, the corresponding task execution status is determined to be an execution exception state, and the dimensional review task is identified as an abnormal dimensional review task. The task execution status of different dimensional review tasks is recorded separately, allowing for the individual identification of execution exceptions for each dimensional review task.

[0113] Terminate the abnormal dimension review task without retrying it, associate the task execution status of the abnormal dimension review task with the corresponding review dimension, and obtain the abnormal task record.

[0114] In this step, it's important to note that after identifying the abnormal dimension review task, the processing of any incomplete tasks related to that task is halted, and the target model is not re-executed for the abnormal dimension review task. Based on the correspondence between abnormal dimension review tasks and review dimensions, the execution anomaly status is associated with the corresponding review dimension, forming an abnormal task record. This record can include the corresponding review dimension, task execution status, and anomaly information to characterize the review dimension that failed to complete the review normally. The abnormal task record and the model review results obtained from normal execution are saved separately to avoid directly using task execution anomalies as the conclusion of whether the tender document content passed or failed.

[0115] Continue executing all review tasks except for the abnormal dimension review task, until the other dimension review tasks are completed or terminated abnormally.

[0116] In this step, it's important to note that terminating an abnormal dimension review task does not terminate other ongoing dimension review tasks within the same batch, nor does it cancel unexecuted dimension review tasks in subsequent batches. For other ongoing dimension review tasks within the same batch, the corresponding model calls and result retrieval continue. After the current batch finishes processing, dimension review tasks in subsequent batches continue to execute according to the preset concurrency level. If other dimension review tasks complete normally, their execution status is updated to "execution completed." If other dimension review tasks also encounter execution anomalies, they are identified as new abnormal dimension review tasks and terminated and recorded in the same manner until all dimension review tasks have either completed or terminated abnormally.

[0117] Associate the model review results corresponding to the completed dimension review task with the corresponding review dimension, and output the abnormal task record.

[0118] In this step, it's important to note that for dimensional review tasks with a completed execution status, one or more model review results generated by that task are retrieved. Based on the correspondence between dimensional review tasks and review dimensions, each model review result is associated with its corresponding review dimension for subsequent dimensional review results. For abnormal dimensional review tasks, their execution status is not mixed into the normal model review results; instead, the corresponding abnormal task record is output. Therefore, the final output can separately represent review dimensions that have completed review normally and their model review results, as well as review dimensions that have not completed review normally and their abnormal task records.

[0119] The technical solution implemented in this embodiment can detect and record the execution status of review tasks for each dimension separately, and terminate the task without retrying when an anomaly occurs in a single dimension review task, while continuing to execute other dimension review tasks. Therefore, an anomaly in model invocation for a single review dimension will not interrupt the processing of other review dimensions, and the anomaly task record can distinguish between task execution anomalies and the review results of the tender document content, allowing normally completed model review results to continue subsequent processing according to the corresponding review dimension.

[0120] In this embodiment, it should be noted that the task management module, review module, and result management module can be software functional units deployed on the same server or different computing nodes, and the modules exchange data through interfaces. The task management module, as the system's input, receives the tender documents and review configuration uploaded by the user. The review configuration defines the review dimensions included in this review and the corresponding review requirements for each dimension. The task management module generates review task data based on the tender documents and review configuration. This review task data may include a task identifier, a review dimension identifier, dimension review instructions, and one or more review inputs corresponding to the content of the tender documents, and then sends the review task data to the review module.

[0121] After receiving the review task data, the review module organizes model calls according to the review dimensions, inputs each review input and its corresponding dimension review instruction into the target large model, and obtains the model review results returned by the target large model. The target large model can be a large language model that supports structured text input and output. The model review results can include review items, result fragment indexes, and judgment content corresponding to the review items. For cases where the same review dimension includes multiple review inputs, the review module obtains the model review results corresponding to each review input separately and retains the correlation between the model review results and the review dimension.

[0122] In a specific example, the review module calls the target large model according to a preset concurrency level, which can be set to 4, enabling review tasks corresponding to different review dimensions to be processed in parallel. Multiple review inputs corresponding to the same review dimension use the same dimension review instructions. When a single review task encounters a call exception, the review module records the corresponding task execution status and continues to process other review tasks, without directly using the call exception as the review conclusion of the tender document content.

[0123] The results management module receives the model review results and organizes them according to the review dimensions. When a review dimension corresponds to only one model review result, that model review result is formed into the corresponding dimension review result; when a review dimension corresponds to multiple model review results, the multiple model review results are aggregated into the same dimension review result. The results management module can also obtain the corresponding tender document content based on the result fragment index in the model review results, associate the review items with the corresponding original text content, and summarize the review results of each dimension according to the review dimensions in the review configuration to generate the tender document review conclusion. The tender document review conclusion may include the dimension status of each review dimension, specific review items, corresponding original text content, and abnormal task records, and is output to the results display interface or the review report generation unit.

[0124] The technical solution implemented in this embodiment forms a continuous data processing link through the task management module, review module, and result management module, enabling the tender documents and review configurations to be sequentially converted into review task data, model review results, dimension review results, and tender document review conclusions. By organizing model calls and result processing by review dimension, different review dimensions can adopt corresponding review requirements, and multiple model review results under the same review dimension can be uniformly aggregated. By associating the result fragment index in the model review results with the content of the tender documents, the corresponding original text evidence can be retained while outputting the review conclusions, and task execution anomalies and review results of the tender document content can be recorded separately.

[0125] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for intelligent review of tender documents, characterized in that, include: S10. Obtain the set of index fragments and the list of review dimensions of the tender document. The set of index fragments includes multiple index fragments with fragment indexes and fragment content, and the list of review dimensions includes multiple review dimensions with review rules. S20. Generate dimension review instructions according to each of the review rules, determine the total input length of each dimension review instruction and the index fragment set, and form one or more review inputs corresponding to each review dimension according to the total input length and the context capacity of the target large model, wherein the review inputs retain the corresponding fragment indexes; S30. Combine each of the review inputs with the corresponding dimension review instructions and input them into the target large model to obtain the corresponding model review results. The model review results include review items and result fragment indexes. S40. Determine the single model review result corresponding to the same review dimension as the dimension review result, or merge multiple corresponding model review results into the dimension review result; obtain the corresponding fragment content according to the result fragment index, summarize the dimension review results, and generate the tender document review conclusion.

2. The intelligent review method for tender documents according to claim 1, characterized in that, In step S20, dimensional review instructions are generated according to each of the review rules, including: Determine the current review dimension from the list of review dimensions, parse the review rule corresponding to the current review dimension, and obtain the review objective, pass conditions, and fail trigger conditions; Based on the review objective, the pass conditions, and the fail trigger conditions, the review rules are decomposed into at least one review point to obtain a set of review points; Obtain the preset review process and preset output format, and combine the set of review points, the preset review process and the preset output format to obtain the dimension review instruction corresponding to the current review dimension; Associate the dimension review instruction with the current review dimension, so that one or more review inputs corresponding to the current review dimension share the dimension review instruction.

3. The intelligent review method for tender documents according to claim 2, characterized in that, Step S20 also includes: Obtain the encoding rules corresponding to the target large model, and determine the maximum token length allowed to be input into the target large model as the context capacity; According to the encoding rules, calculate the instruction token length of the dimension review instruction and the fragment input token length of all fragment indices and fragment contents in the index fragment set; The total input length is determined by the sum of the instruction token length and the fragment input token length. When the total input length does not exceed the context capacity, all the index fragments are combined according to the order in which they are arranged in the tender document to obtain a review input; When the total input length exceeds the context capacity, the instruction token length and the preset reserved token length are subtracted from the context capacity to obtain the allowed token length. The set of index segments is combined into multiple review inputs according to the order of the index segments in the tender document, so that each review input includes at least one index segment and the token length of each review input does not exceed the allowed token length.

4. The intelligent review method for tender documents according to claim 3, characterized in that, In step S30, the corresponding model review results are obtained, including: The current review input is determined from one or more review inputs corresponding to the current review dimension, and the fragment indices in the current review input are summarized to obtain the input index set; The current review input is combined with the dimension review instruction and then input into the target large model to obtain the candidate model review result. The candidate model review result includes candidate review items and candidate result fragment indexes corresponding to the candidate review items. Determine whether each candidate result fragment index exists in the input index set, retain the candidate result fragment index and the corresponding candidate review item that exist in the input index set, and discard the candidate result fragment index and the corresponding candidate review item that do not exist in the input index set; Based on the retained candidate review items and the candidate result fragment index, the model review result corresponding to the current review input is obtained; when there are no retained candidate review items, the model review result with empty review items is obtained; each review input is processed in sequence to obtain one or more model review results corresponding to the current review dimension.

5. The intelligent review method for tender documents according to claim 4, characterized in that, In step S40, determining a single model review result corresponding to the same review dimension as a dimension review result, or merging multiple corresponding model review results into a dimension review result, includes: Obtain one or more model review results corresponding to the current review dimension, where each review item is either a pass or a fail item; When the current review dimension corresponds to a model review result, the passed items and the failed items in the model review result are respectively determined as the dimension passed item set and the dimension failed item set; When the current review dimension corresponds to multiple model review results, the union of the passed items and the failed items in the multiple model review results is taken to obtain the set of passed items and the set of failed items for the dimension. When the set of failed items for a dimension is not empty, the dimension status of the currently reviewed dimension is determined to be failed; when the set of failed items for a dimension is empty and the set of passed items for a dimension is not empty, the dimension status is determined to be passed; when both the set of failed items for a dimension and the set of passed items for a dimension are empty, the dimension status is determined to be not involved. The dimension is combined with the item set, the dimension failure item set, and the dimension status to obtain the dimension review result corresponding to the current review dimension.

6. The intelligent review method for tender documents according to claim 5, characterized in that, Step S40 also includes: Obtain the result fragment index corresponding to each review item in the set of items that passed the dimension and the set of items that failed the dimension, and summarize the result fragment indexes to obtain the dimension-related index set; When the dimension status is not involved and the set of indexes related to the dimension is empty, determine whether the failure trigger condition includes missing content related to the current review dimension; When the failure trigger condition includes missing content related to the current review dimension, a content missing failure item is generated, the content missing failure item is added to the dimension failure item set, and the dimension status is updated to failure; When the failure trigger condition does not include missing content related to the current review dimension, the dimension status remains as not involved. The dimension review result is updated based on the processed set of items that passed the dimension, the set of items that failed the dimension, and the dimension status.

7. The intelligent review method for tender documents according to claim 6, characterized in that, When the set of dimension-related indexes is not empty and the set of dimension-failed items includes multiple failed items, step S40 further includes: Each of the aforementioned failed items includes a problem conclusion, suggested modifications, review criteria, and at least one index of the aforementioned result fragment; The multiple failed items are grouped according to the problem conclusion of each of the failed items to obtain at least one failed item group, and the multiple failed items in each failed item group have the same problem conclusion; For each group of failed items that includes multiple failed items, the multiple failed items are merged into a target failed item, the target failed item includes the problem conclusion, a set of modification suggestions consisting of multiple modification suggestions, a set of review criteria consisting of multiple review criteria, and a set of result fragment indexes consisting of multiple result fragment indexes; Replace the corresponding multiple failed items in the set of failed items for each of the aforementioned targets with the failed items for each dimension, and update the dimension review result based on the replaced set of failed items for each dimension.

8. The intelligent review method for tender documents according to claim 3, characterized in that, Step S30 includes: The dimension review instruction corresponding to each review dimension and one or more review inputs are combined to obtain the corresponding dimension review task; Obtain a preset concurrency level, and execute multiple dimension review tasks in batches according to the preset concurrency level until each dimension review task is completed or terminated abnormally. In each of the dimensional review tasks, the review inputs are combined with the dimensional review instructions and then input into the target large model to obtain the model review results corresponding to each of the review inputs. The review results of each model are associated with the corresponding review dimensions.

9. The intelligent review method for tender documents according to claim 8, characterized in that, Execute multiple dimension review tasks in batches and in parallel, including: Obtain the task execution status of each of the aforementioned dimension review tasks, and determine the abnormal dimension review tasks that are experiencing execution errors based on the task execution status; Terminate the abnormal dimension review task and do not retry the abnormal dimension review task. Associate the task execution status corresponding to the abnormal dimension review task with the corresponding review dimension to obtain an abnormal task record. Continue executing the other dimension review tasks besides the abnormal dimension review task, until the other dimension review tasks are completed or terminated abnormally; Associate the model review result corresponding to the completed dimension review task with the corresponding review dimension, and output the abnormal task record.

10. A tender document intelligent review system, which executes the tender document intelligent review method as described in any one of claims 1 to 9 during operation, characterized in that, include: The task management module is used to receive tender documents and review configurations, generate review task data, and output the review task data to the review module. The review configurations include multiple review dimensions. The review module is connected to the task management module and is used to receive the review task data, call the target large model to perform the tender document review, generate the model review result, and output the model review result to the result management module. The result management module is connected to the review module and is used to receive the model review results, form the dimension review results corresponding to each of the review dimensions, and summarize the review results of each of the dimensions to generate the tender document review conclusion.

Images