Bid review methods, apparatus, computer equipment and readable storage media

Abstract

This invention provides a method, apparatus, computer equipment, and readable storage medium for reviewing tender documents, relating to the field of data processing technology. The method includes: acquiring the tender document and corresponding bidding document; acquiring a review feature library containing review items and an applicability matrix, where each review item includes a review identifier and review requirements, and matrix elements are cross-located by review identifiers and construction scheme identifiers and associated with applicable states, verification points, and constraint parameters; identifying original text fragments of bidding evidence that meet the review requirements in the bidding document and generating a list of review points containing location information; identifying scheme evidence fragments in the tender document based on a construction scheme knowledge base and constructing a set of schemes to be reviewed; traversing the review points and schemes to be reviewed, locating target elements through the matrix, completing the review, and outputting review data containing review points, schemes, and review results. This invention enables automatic review and verifiable results of tender responses in scenarios with long documents and multiple schemes.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, apparatus, computer equipment, and readable storage medium for reviewing tender documents. Background Technology

[0002] In engineering bidding activities, tender documents typically outline requirements for construction organization design, construction plans, key technical measures, and resource allocation in the form of technical specifications, review methods, and scoring criteria. Bidders are required to respond to each of these requirements in their tender documents. During the bid evaluation stage, reviewers need to assess the compliance and merits of multiple bidders' tender documents within a short period and establish verifiable review criteria. Because engineering tender documents are generally characterized by multiple sources and heterogeneity, length, complex terminology, and scattered clauses, the review process often relies on retrieving, locating, and aligning evidence from original text fragments to support subsequent judgments and reviews.

[0003] With the development of information technology, computer-aided review systems have emerged in the industry. Common approaches include keyword-matching retrieval models, general text retrieval architectures, and direct use of large language models for single-stage reasoning and output of review conclusions.

[0004] Although the aforementioned intelligent review technologies have improved retrieval and auxiliary judgment capabilities to some extent, the following problems still exist in the practical application of engineering tender document review: First, review evidence often relies mainly on weak anchor points such as page numbers and chapter titles for positioning, making it difficult to form a stable correspondence between evidence and conclusion. This results in the need to repeatedly scan a large amount of original text during the review process, which is inefficient and prone to introducing bias. Second, in engineering review, different review requirements usually have different applicability under different construction schemes and may need to meet different verification points and constraints. Existing systems are unable to generate reproducible verification paths based on the combination relationship between review items and construction schemes, which in turn makes it difficult to verify the consistency between the review process and the output results.

[0005] Therefore, existing technologies make it difficult to form a verifiable chain of evidence during the tender document review process, affecting review efficiency and consistency of results, which has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] The purpose of this invention is to provide a method, apparatus, computer device, and readable storage medium for reviewing tender documents, in order to solve the aforementioned technical problems in the prior art.

[0007] On the one hand, in order to achieve the above objectives, the present invention provides a method for reviewing tender documents.

[0008] The evaluation method for this tender document includes: obtaining the tender document and the corresponding tender document; obtaining a pre-set evaluation feature library, wherein the evaluation feature library includes an applicability matrix and multiple evaluation items, each evaluation item including an evaluation identifier and evaluation requirements; the applicability matrix includes multiple matrix elements cross-located by evaluation identifiers and construction scheme identifiers, and the matrix elements are used to define the applicability status of an evaluation item under a construction scheme and the verification points and constraint parameters associated with the applicability status; in the tender document, identifying the original text fragments of tender evidence that meet the evaluation requirements, and constructing a list of evaluation points in the tender document, wherein the list of evaluation points includes multiple evaluation points, each evaluation point including the original text fragments of tender evidence, the location information of the original text fragments of tender evidence in the tender document, and the evaluation identifier corresponding to the evaluation requirements met by the original text fragments of tender evidence; obtaining a pre-set construction scheme knowledge base, wherein the construction scheme knowledge base includes multiple scheme items, and the method The case involves identifying construction scheme identifiers and construction scheme descriptions. In the tender documents, original fragments of tender evidence that satisfy the construction scheme description are identified, and a set of schemes to be reviewed is constructed. This set includes multiple schemes to be reviewed, each containing original fragments of tender evidence, their location information within the tender document, and the construction scheme identifier corresponding to the construction scheme description satisfied by the original fragments of tender evidence. The review point list and the set of schemes to be reviewed are traversed, and matrix elements are located in the applicability matrix using the review identifiers and construction scheme identifiers to obtain target elements. Schemes to be reviewed are then reviewed based on the target elements to obtain review results for each review point in the review point list. Review data for the tender documents is generated based on the review results for each review point in the review point list. This review data includes multiple review data units, each containing a review point, a scheme to be reviewed, and a review result.

[0009] Furthermore, in the tender document, the steps for identifying the original text fragments of tender evidence that meet the review requirements include: segmenting the tender document to obtain multiple tender text units; determining the location information of each tender text unit in the tender document; calculating the tender vector features of the tender text units; extracting the tender keywords of the tender text units; writing the tender text units and their location information into the tender retrieval space, and constructing indexes using the tender vector features and tender keywords respectively; constructing a tender retrieval vector set and a tender retrieval term set using the review requirements, and performing retrieval in the tender retrieval space based on the tender retrieval vector set and the tender retrieval term set respectively; using a reciprocal sorting fusion algorithm to weight and sort the retrieval results of the tender retrieval vector set and the tender retrieval term set, and extracting a preset number of tender text units to form a tender text candidate set; and selecting and extracting tender text units whose location information meets the preset position limitation conditions from the tender text candidate set to obtain the original text fragments of tender evidence.

[0010] Furthermore, the review items also include review logic types. After obtaining the original text fragment of the bidding evidence, the method further includes: obtaining a structured review question template corresponding to the review logic type; extracting data content from the review requirements and adding it to the data area of ​​the structured review question template to generate a structured review question used to guide the large language model in making a judgment; obtaining a first-level prompt word template for the large language model's reasoning; injecting the original text fragment of the bidding evidence and the structured review question into the data area of ​​the first-level prompt word template respectively to generate the first-level prompt words; inputting the first-level prompt words into the large language model, enabling the large language model to perform deep semantic verification under evidence constraints, and outputting the response result and response reason for whether the original text fragment of the bidding evidence meets the review requirements; wherein, the review points also include the response result and response reason.

[0011] Furthermore, the review items also include review logic types. After constructing the review point list of the tender document, the method further includes: mapping the review points to preset score dimensions according to the review logic types corresponding to the review points; classifying all review points in the review point list according to the score dimensions to obtain the review point set under each score dimension; obtaining a preset score pool configuration, wherein the score pool configuration includes the score pool ratio corresponding to each score dimension; determining the distributable total score of the score dimension according to the score pool ratio; within each score dimension, allocating scores to each review point in the review point set according to the distributable total score to obtain the individual score of each review point; wherein, the review data also includes the total review score corresponding to the tender document, which is calculated based on the review results and individual scores of each review point.

[0012] Furthermore, the tender document includes a single bidder's tender document, and the construction plan description includes a first description and a second description. The steps for identifying original text fragments of tender evidence that satisfy the construction plan description within the tender document include: segmenting the tender document to obtain multiple tender text units; determining the location information of each tender text unit within the tender document; extracting tender keywords from the tender text units; writing the tender text units and their location information into the tender retrieval space, and constructing an index using the tender keywords; constructing a tender retrieval term set using the first description, and performing a retrieval in the tender retrieval space based on the tender retrieval term set to obtain a candidate set of tender texts; and filtering original text fragments of tender evidence from the candidate set of tender texts according to the second description.

[0013] Furthermore, the second description content includes personnel description, machinery and equipment description, material description, method and process description, and environmental description. The step of screening original text fragments of bidding evidence in the candidate bidding text set based on the second description content includes: obtaining a second-level prompt word template for inference in the large language model; injecting the candidate bidding text set and the second description content into the data area of ​​the second-level prompt word template to form second-level prompt words; and inputting the second-level prompt words into the large language model, so that the large language model performs deep semantic verification under evidence constraints and outputs original text fragments of bidding evidence.

[0014] Furthermore, the applicable status includes inapplicable, fully applicable, and conditionally applicable. The steps for reviewing the proposed solution based on the target element and obtaining the review results corresponding to each review point in the review point list include: when the applicable status of the target element is fully applicable, determining that the proposed solution corresponding to the target element passes the review item corresponding to the target element; when the applicable status of the target element is inapplicable, determining that the proposed solution corresponding to the target element does not pass the review item corresponding to the target element; when the applicable status of the target element is conditionally applicable, determining whether there are verification points in the proposed solution corresponding to the target element and whether the constraint parameters are satisfied if they exist, and determining whether the proposed solution corresponding to the target element passes the review item corresponding to the target element based on the judgment result.

[0015] On the other hand, in order to achieve the above objectives, the present invention provides a tender document review device.

[0016] The tender document review device includes: a first acquisition module for acquiring the tender document and the corresponding tender document; a second acquisition module for acquiring a pre-set review feature library, wherein the review feature library includes an applicability matrix and multiple review items, each review item including a review identifier and review requirements; the applicability matrix includes multiple matrix elements cross-located by review identifiers and construction scheme identifiers, and the matrix elements are used to define the applicability status of a review item under a construction scheme and the verification points and constraint parameters associated with the applicability status; a first identification module for identifying original text fragments of tender evidence that meet the review requirements in the tender document and constructing a review point list for the tender document, wherein the review point list includes multiple review points, each review point including original text fragments of tender evidence, the location information of the original text fragments of tender evidence in the tender document, and the review identifier corresponding to the review requirements met by the original text fragments of tender evidence; and a third acquisition module for acquiring a pre-set construction scheme knowledge base, wherein the construction scheme knowledge base includes multiple schemes. The system comprises four modules: a first module for identifying construction schemes and a second module for identifying original bid evidence fragments that satisfy the construction scheme description in the bid document; a third module for constructing a set of schemes to be reviewed, which includes multiple schemes to be reviewed, each containing original bid evidence fragments, their location information within the bid document, and the construction scheme identifier corresponding to the construction scheme description satisfied by the original bid evidence fragments; a fourth module for traversing the review point list and the set of schemes to be reviewed, using the review identifier and construction scheme identifier to locate matrix elements in the applicability matrix to obtain the target element; a fifth module for reviewing the schemes to be reviewed based on the target element, obtaining the review results corresponding to each review point in the review point list; and a sixth module for generating review data for the bid document based on the review results corresponding to each review point in the review point list, wherein the review data includes multiple review data units, each including a review point, a scheme to be reviewed, and a review result.

[0017] On the other hand, to achieve the above objectives, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above method.

[0018] On the other hand, to achieve the above objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above method.

[0019] The present invention provides a method, apparatus, computer equipment, and readable storage medium for reviewing tender documents. It acquires tender documents and tender documents, identifies a list of review points with location information in the tender documents based on a pre-set review feature library, and identifies a set of schemes to be reviewed in the tender documents based on a construction scheme knowledge base. Then, it uses review identifiers and construction scheme identifiers to locate target elements in an applicability matrix, and completes the review of the schemes to be reviewed, outputting review data. Through this invention, review requirements, original tender document evidence, tender response evidence, matrix applicability relationships, and review results can be structurally aligned and output, thereby forming a verifiable chain of evidence in scenarios with long documents and scattered clauses, improving review consistency and review efficiency. Attached Figure Description

[0020] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0021] Figure 1 A flowchart of the bid document review method provided in Embodiment 1 of the present invention;

[0022] Figure 2 This is a block diagram of the bid document review device provided in Embodiment 2 of the present invention;

[0023] Figure 3 This is a hardware structure diagram of a computer device provided in Embodiment 3 of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0025] Example 1

[0026] Embodiment 1 of this invention provides a method for reviewing tender documents. This method enables the creation of a traceable review point list within the tender documents and, during the bidding stage, combines construction scheme identification and applicability matrix index routing to achieve a verifiable review output of the bid response. Specifically, Figure 1 The flowchart of the bid document review method provided in Embodiment 1 of the present invention is as follows: Figure 1 As shown, the method includes the following steps S101 to S108.

[0027] Step S101: Obtain the tender documents and the corresponding bidding documents.

[0028] The bidding documents include technical specifications, scoring criteria, review methods, contract conditions, Q&A or addenda, etc.; the tender documents include construction organization design, special construction plan, resource allocation plan, schedule plan, quality and safety measures, similar performance, etc.

[0029] Optionally, bid documents and tender documents can be obtained through electronic bidding platforms, file servers, or project databases, and the obtained documents can be formatted uniformly, such as converting PDF / Word to text, and retaining page numbers, heading levels, paragraph numbers, and other layout information for subsequent location information generation.

[0030] Step S102: Obtain the preset review feature library.

[0031] Optionally, the review feature library is divided from top to bottom into domain categories (such as basic engineering, surveying, design, etc.), feature dimensions (such as geological conditions, process parameters, etc.), and specific review items. The review feature library includes an applicability matrix and multiple review items; each review item includes a review identifier and review requirements; the applicability matrix includes multiple matrix elements cross-located by review identifiers and construction scheme identifiers. These matrix elements define the applicability of a review item under a given construction scheme and associate the verification points and constraint parameters of that applicability.

[0032] The review identifier is used to uniquely number review items, facilitating retrieval routing, data alignment, and result summarization. Review requirements are fields in natural language or structured expression, describing the technical points that need to be met. Construction scheme identifiers identify a specific type of construction scheme, such as foundation pit support, dewatering, formwork support, hoisting and transportation, and welding processes. Applicability status includes inapplicable, fully applicable, and conditionally applicable, indicating whether the review item needs to be reviewed under this construction scheme and the review method. Verification points characterize the key measures, parameters, resources, or procedural constraints that need to be verified in the bidding evidence during the review process. Constraint parameters include retrieval constraints such as chapter scope, evidence quantity threshold, and similarity threshold; location restrictions such as searching only in chapters like construction organization design, special schemes, and quality and safety; and satisfaction judgment thresholds such as requiring the simultaneous appearance of personnel, equipment, and process elements.

[0033] Step S103: Identify the original text fragments of bidding evidence that meet the review requirements in the bidding documents, and construct a list of review points in the bidding documents.

[0034] The review point list includes multiple review points. Each review point includes a fragment of the original bidding evidence, the location information of the original fragment in the bidding document, and the review identifier corresponding to the review requirements met by the original fragment.

[0035] In this embodiment, the original text of the bidding evidence is used as the source evidence for the review point. The location information may include page number, chapter title path, paragraph number, line number or character range, so that the review point can be verified and located.

[0036] Optionally, the original text fragment of the bidding evidence corresponding to a review requirement can be a single text unit in the bidding document, or multiple text units in different locations within the bidding document. The original text fragment of the bidding evidence is retained for the same review identifier, and the source location and confidence level of the evidence are recorded for subsequent interpretation and verification.

[0037] Step S104: Obtain the pre-set construction plan knowledge base.

[0038] The construction plan knowledge base includes multiple plan items, each containing a construction plan identifier and a construction plan description. The construction plan description can be a set of semantic descriptions of the plan, such as key processes, typical equipment, key materials, constraints, and scope of application, which are used as evidence fragments in the tender documents to identify the construction plan.

[0039] Step S105: Identify original text fragments of bidding evidence that meet the description of the construction plan in the bidding documents, and construct a set of plans to be reviewed.

[0040] The set of proposals to be reviewed includes multiple proposals to be reviewed. Each proposal to be reviewed includes: a fragment of the original bidding evidence, the location information of the original fragment in the bidding document, and the construction scheme identifier corresponding to the construction scheme description satisfied by the original fragment.

[0041] In this embodiment, the original text fragments of bidding evidence can be paragraphs, table texts, clause descriptions, etc., related to a certain construction plan in the bidding document. Optionally, for the same construction plan identifier, the collected original text fragments of bidding evidence can be a single text unit in the bidding document, or multiple text units in different locations in the bidding document, to support the determination of evidence constraints for subsequent review items.

[0042] Step S106: Traverse the review point list and the set of schemes to be reviewed, and locate the target element.

[0043] Specifically, for each combination of review point (review identifier) ​​and the scheme to be reviewed (construction scheme identifier), cross-location is performed in the applicability matrix to obtain the corresponding matrix element as the target element. The target element includes the applicability status under the combination and is associated with the verification points and constraint parameters that can be used for review.

[0044] Step S107: Review the proposed solution based on the target element to obtain the review results for each review point in the review point list.

[0045] In this embodiment, the review includes, but is not limited to: determining the compliance of the original text of the bid evidence to see if it covers the review requirements; determining or scoring the quality of the bid response based on merit; and outputting the reasons for the determination under the constraints of evidence, along with the referenced metadata.

[0046] Optionally, when the target element is associated with verification points and constraint parameters, a retrieval instruction is generated based on the verification points / review requirements and retrieval constraint parameters to obtain evidence paragraphs in the retrieval space of the tender document, and semantic verification and conclusion generation are completed under the evidence constraints, thereby avoiding the review difficulties caused by relying solely on weak anchors or full-text scanning.

[0047] Step S108: Generate review data for the tender documents.

[0048] The review data comprises multiple review data units, each corresponding to a target element, including review points, the scheme to be reviewed, and review results. Review points include fragments of the original bidding evidence and their location information, as well as corresponding review identifiers, which can be linked to review items in the feature library. Schemes to be reviewed include fragments of the original bidding evidence and their location information, as well as corresponding construction scheme identifiers, which can be linked to review schemes in the knowledge base, thus achieving the formation of a verifiable evidence chain.

[0049] In the tender document review method provided in this embodiment, the tender document and the tender document are obtained. Based on a pre-set review feature library, a list of review points with location information is generated from the tender document. Based on a construction scheme knowledge base, a set of schemes to be reviewed is generated from the tender document. Then, the target elements are located in the applicability matrix using review identifiers and construction scheme identifiers. Based on this, the review of the schemes to be reviewed is completed, and review data is output. Using the tender document review method provided in this embodiment, review requirements, original tender document evidence, tender response evidence, matrix applicability relationships, and review results can be structurally aligned and output. This forms a verifiable chain of evidence in scenarios with long documents and scattered clauses, improving review consistency and review efficiency.

[0050] Optionally, in one embodiment, the step of identifying the original text fragments of bidding evidence that meet the review requirements in the bidding document includes: segmenting the bidding document to obtain multiple bidding text units; determining the location information of each bidding text unit in the bidding document; calculating the bidding vector features of the bidding text units; extracting the bidding keywords of the bidding text units; writing the bidding text units and their location information into the bidding retrieval space, and constructing indexes using the bidding vector features and bidding keywords respectively; constructing a bidding retrieval vector set and a bidding retrieval term set using the review requirements, and performing retrieval in the bidding retrieval space based on the bidding retrieval vector set and the bidding retrieval term set respectively; weighting and sorting the retrieval results of the bidding retrieval vector set and the bidding retrieval term set using a reciprocal sorting fusion algorithm, and extracting a preset number of bidding text units to form a bidding text candidate set; and filtering and extracting bidding text units whose location information meets the preset position limitation conditions from the bidding text candidate set to obtain the original text fragments of bidding evidence.

[0051] Specifically, when identifying original text fragments of bidding evidence that meet the review requirements in the bidding documents, the layout of the unstructured documents in the bidding documents is analyzed, and hierarchical recognition technology is used to extract text such as titles, paragraphs and table information. Based on the principle of semantic integrity, the extracted text stream is intelligently segmented and divided. During the extraction process, each text unit is assigned metadata including chapter path, page number, page coordinates, paragraph number, title level path, and start and end character offset as positioning information, so that the subsequently output original text fragments of bidding evidence have the ability to trace back to the original text.

[0052] Furthermore, bidding vector features can be calculated for each bidding text unit. For example, a pre-trained text embedding model can be used to map each bidding text unit into a high-dimensional dense vector to obtain bidding vector features. These bidding vector features can be used to characterize the semantic representation of the bidding text unit. Bidding keywords can also be extracted from each bidding text unit, which can be used to characterize the explicit terms and high-frequency points of the bidding text unit. Subsequently, the bidding text units and their location information are written into the bidding retrieval space, and indexes are constructed based on the bidding vector features and bidding keywords respectively, enabling the bidding retrieval space to simultaneously support both semantic retrieval and keyword retrieval channels.

[0053] Based on this, a tender retrieval vector set and a tender retrieval term set are constructed using the review requirements. Retrieval is then performed in the tender retrieval space to obtain two types of retrieval results. Next, a reciprocal sorting fusion algorithm is used to weight and sort the two types of retrieval results, and a predetermined number of tender text units are extracted from the sorted results to form a candidate tender text set. Finally, the candidate tender text set is further filtered to extract tender text units whose location information meets preset location constraints. These location constraints can be, for example, limited to target sections such as the scoring criteria section, the technical specifications section, or the pre-appendices to the instructions to bidders, thereby obtaining tender evidence fragments that more closely match the context of the review requirements.

[0054] Optionally, the calculation formula used in the reciprocal sorting fusion algorithm is as follows:

[0055]

[0056] in, This indicates the tender text unit to be scored. Confidence score; This represents the weighting coefficient of the retrieval method, used to adjust the contribution of different retrieval methods to the final ranking; It is a smoothing constant used to reduce the sensitivity of the calculated score difference of the top-ranked positions, for example, a value of 60; The ranking position of the tender text unit d in the retrieval method (counting from 1). This formula calculates a weighted sum of the reciprocals of the rankings of the two retrieval results, so that tender text units that rank higher in both retrieval paths receive a higher fusion ranking score, thereby achieving effective integration of heterogeneous retrieval results.

[0057] The tender document review method provided in this application establishes both vector and keyword indexes for each tender text unit and uses a reciprocal sorting fusion algorithm to merge and sort the results from different retrieval channels. This ensures that the original supporting text fragments corresponding to the review requirements maintain a high recall and relevance even in long documents, with inconsistent terminology, or scattered expressions. At the same time, by combining location information and location constraints to filter irrelevant segments, the output tender evidence original text fragments have stronger location and verifiability, providing a stable evidentiary basis for the subsequent construction of the review point list.

[0058] Optionally, in one embodiment, the review item further includes a review logic type. After obtaining the original text fragment of the bidding evidence, the method further includes: obtaining a structured review question template corresponding to the review logic type; extracting data content from the review requirements and adding it to the data area of ​​the structured review question template to generate a structured review question for guiding the large language model to make a judgment; obtaining a first-level prompt word template for the large language model's reasoning; injecting the original text fragment of the bidding evidence and the structured review question into the data area of ​​the first-level prompt word template respectively to generate a first-level prompt word; inputting the first-level prompt word into the large language model, so that the large language model performs deep semantic verification under the evidence constraints, and outputs a response result and response reason for whether the original text fragment of the bidding evidence meets the review requirements; wherein, the review point further includes a response result and a response reason.

[0059] Specifically, in this embodiment, the review items also include review logic types. After obtaining the original text fragments of the bidding evidence, the determination process for whether the review requirements are met can be further structured based on the review logic type. The review logic type can be used to distinguish different judgment methods, such as compliance judgment, merit-based scoring, or composite review, so as to select a matching structured review question template. The structured review question template may include a data area and an instruction area. The data area is used to carry the key data content extracted from the review requirements, and the instruction area is used to specify the judgment items, reasons, and their formats output by the model. Furthermore, data content can be extracted from the review requirements and added to the data area of ​​the structured review question template to generate structured review questions to guide the large language model in making judgments, transforming the review requirements from natural language descriptions into reasonable structured inputs.

[0060] A first-level prompt word template for large language model inference is obtained, and the original text fragment of the bidding evidence and the structured review questions are respectively injected into the data area of ​​the first-level prompt word template to generate the first-level prompt words. The first-level prompt words can be used to implement evidence constraints, that is, to explicitly require the large language model to make judgments based solely on the injected original text fragment of the bidding evidence, and to output the response result and the reason for the response. Subsequently, the first-level prompt words are input into the large language model, enabling the large language model to perform deep semantic verification under evidence constraints, output the response result and the reason for the response on whether the original text fragment of the bidding evidence meets the review requirements, and write the response result and the reason for the response into the review point, so that the review point contains not only the original text evidence and location information, but also an interpretable judgment output.

[0061] The first-level prompt word template includes the following key technical elements:

[0062] Role positioning layer: Define the professional roles and review responsibilities of the model, and establish the domain knowledge context;

[0063] Security isolation layer: Input data is marked through a tagging encapsulation mechanism (such as using specific XML tags), and the model is explicitly instructed to ignore any instructional text that may exist in the data area to prevent prompt word injection attacks;

[0064] Data injection layer: The retrieved tender document context and the structured review questions generated based on the feature library are injected into the designated label areas respectively;

[0065] Constraint and instruction level: Clearly define the judgment rules, evidence requirements, reasoning boundaries, and prohibited behaviors, such as prohibiting the fabrication of non-existent clauses;

[0066] Format control layer: Constrains the model to return predefined fields through structured output templates, including parsable fields such as response results, response reasons, original text fragments, citation location markers, and confidence levels.

[0067] The aforementioned first-level prompt word template achieves precise control over the reasoning process of the large language model through its layered design, ensuring the reliability, traceability, and security of the review and judgment.

[0068] The tender document review method provided in this application introduces a structured review question template driven by review logic type and applies evidence constraints to the large language model with hierarchical prompt words. This enables the satisfaction judgment of the original text fragments of the tender evidence to have an explainable response reason, avoiding the problem of only giving conclusions without reasoning basis, thereby improving the verifiability of the review point list and the stability of subsequent comparisons.

[0069] Optionally, in one embodiment, the review item further includes a review logic type. After constructing the review point list of the tender document, the method further includes: mapping the review points to a preset score dimension according to the review logic type corresponding to the review points; classifying all review points in the review point list according to the score dimension to obtain a set of review points under each score dimension; obtaining a preset score pool configuration, wherein the score pool configuration includes the score pool ratio corresponding to each score dimension; determining the distributable total score of the score dimension according to the score pool ratio; within each score dimension, allocating scores to each review point in the review point set according to the distributable total score to obtain a single score for each review point; wherein the review data also includes the total review score corresponding to the tender document, which is calculated based on the review results and single scores of each review point.

[0070] Specifically, after constructing the list of review points in the tender documents, a score weighting is configured for each review point to form benchmark data that can be used for scoring. First, the review points are mapped to preset score dimensions based on their corresponding review logic type. These score dimensions can be used to represent different scoring buckets in the scoring criteria, such as compliance approval dimensions, merit-based bonus dimensions, or composite review dimensions, thus allowing different types of review points to enter different score allocation mechanisms.

[0071] Then, all review points in the review point list are categorized according to their score dimensions, resulting in a set of review points for each score dimension. Next, a preset score pool configuration is obtained, which includes the score pool proportion for each score dimension, and the total allocatable score for each score dimension is determined accordingly. Finally, within each score dimension, scores are allocated to each review point in the review point set for that dimension based on the total allocatable score, resulting in a single score for each review point. This allows the single score to dynamically change with the actual number of review points.

[0072] Based on this, the individual scores of the review points can be combined with the review results output during the bidding stage to calculate the total review score of the bid documents. The total review score can then be output as part of the review data, thereby achieving a closed-loop output from the review point list to the scoring benchmark and the review results.

[0073] The bid review method provided in this application maps review points to score dimensions according to review logic type, and dynamically allocates scores based on the score pool ratio and the number of review points in each dimension. This allows the scoring benchmark to be adaptively adjusted according to the review point identification results of different bid documents, avoiding unfairness or incomparability issues introduced by fixed score rules when the number of review points changes, thereby improving the interpretability and consistency of the scoring output.

[0074] Optionally, in one embodiment, the tender document includes a single bidder's tender document, and the construction plan description includes a first description and a second description. The step of identifying original text fragments of tender evidence that satisfy the construction plan description within the tender document includes: segmenting the tender document to obtain multiple tender text units; determining the location information of each tender text unit within the tender document; extracting tender keywords from the tender text units; writing the tender text units and their location information into a tender retrieval space, and constructing an index using the tender keywords; constructing a tender retrieval term set using the first description, and performing a retrieval in the tender retrieval space based on the tender retrieval term set to obtain a candidate set of tender texts; and filtering original text fragments of tender evidence from the candidate set of tender texts according to the second description.

[0075] Specifically, in this embodiment, the tender documents are limited to those of a single bidder, thus isolating the tender search spaces of different bidders. The construction plan description includes a first description and a second description. The first description includes the core description of the plan, and the second description includes a detailed description of the plan. Optionally, the first description includes core technological features, mechanical equipment requirements, and key construction methods, while the second description includes personnel, mechanical equipment, materials, methods and processes, and environmental descriptions. When identifying original text fragments of tender evidence that satisfy the construction plan description in the tender documents, a two-stage identification path of first searching and then filtering is adopted to reduce the computational overhead of full-volume accurate judgment of long documents.

[0076] The process begins by segmenting the bid document into multiple bid text units and determining the location information for each unit to support subsequent retrieval of the original bid evidence fragments. Then, bid keywords are extracted from each bid text unit, and the bid text units and their location information are written into the bid retrieval space. An index is built using these bid keywords, enabling the bid retrieval space to support rapid retrieval based on terminology and explicit expressions.

[0077] Based on this, a bid retrieval term set is constructed using the first descriptive content, and a search is performed in the bid retrieval space based on the bid retrieval term set to obtain a candidate set of bid texts. The first descriptive content is used to carry relatively coarse-grained scheme trigger descriptions, such as scheme names, typical components, or common terms, to quickly recall potentially relevant paragraphs. Subsequently, based on the second descriptive content, the original text fragments of bid evidence are further filtered in the candidate set of bid texts. The second descriptive content is used to carry more fine-grained scheme limiting information, so that the final output original text fragments of bid evidence are closer to the substantive description of the construction scheme, rather than just containing scheme titles or generalized expressions.

[0078] The bid document review method provided in this application splits the construction scheme identification into a two-stage process: candidate recall driven by the first description content and candidate screening driven by the second description content. This allows for the acquisition of highly relevant scheme evidence fragments with a controllable retrieval scale even in scenarios where bid documents are lengthy and scheme descriptions are scattered. This improves the accuracy of subsequent cross-review based on review points and construction schemes and reduces computational redundancy.

[0079] Optionally, in one embodiment, the second description includes personnel description, machinery and equipment description, material description, method and process description, and environmental description. The step of filtering original text fragments of bidding evidence from the candidate bidding text set based on the second description includes: obtaining a second-level prompt word template for inference using a large language model; injecting the candidate bidding text set and the second description into the data area of ​​the second-level prompt word template to form second-level prompt words; and inputting the second-level prompt words into the large language model, enabling the large language model to perform deep semantic verification under evidence constraints and output original text fragments of bidding evidence.

[0080] Specifically, the personnel description characterizes the proposed personnel configuration, job responsibilities, or qualification requirements; the machinery and equipment description characterizes the proposed equipment model, quantity, or capacity parameters; the materials description characterizes the configuration of key materials, components, or consumables; the methods and processes description characterizes the construction process route, key procedures, or control methods; and the environmental description characterizes the construction conditions, site constraints, or environmental protection measures. Through these fine-grained descriptions, the second description content can cover the core resources and technological elements of the construction plan.

[0081] When selecting original text fragments of bidding evidence from the candidate bidding text set based on the second description content, a semantic verification of evidence constraints using a large language model is introduced. Specifically, a second-level prompt word template for inference using the large language model is obtained, and the candidate bidding text set and the second description content are respectively injected into the data area of ​​the second-level prompt word template to form second-level prompt words. The second-level prompt words are used to explicitly require the large language model to judge whether the candidate paragraphs meet the personnel, equipment, materials, processes, and environmental elements limited by the second description content based solely on the candidate text set, thereby outputting more credible original text fragments of bidding evidence.

[0082] Subsequently, the second-level prompt words are input into the large language model, enabling the large language model to perform deep semantic verification under evidence constraints and output the original text fragment of the bidding evidence. This avoids misidentification caused by relying solely on keyword hits, such as hitting the title but having vague content or hitting partial terms but missing scheme elements.

[0083] The bid document review method provided in this application constrains the fine-grained elements of the construction plan with the second description of personnel, machinery and equipment, materials, methods and processes, and environment, and filters the candidate set by combining the semantic verification of the large language model of evidence constraints. This makes the identified original text fragments of bid evidence more reflective of the bidder's real resource allocation and process measures, thereby providing more accurate plan input for subsequent review based on the applicability matrix.

[0084] Optionally, in one embodiment, the applicable state includes inapplicable, fully applicable, and conditionally applicable. The step of reviewing the solution to be reviewed based on the target element and obtaining the review results corresponding to each review point in the review point list includes: when the applicable state of the target element is fully applicable, determining that the solution to be reviewed corresponding to the target element passes the review item corresponding to the target element; when the applicable state of the target element is inapplicable, determining that the solution to be reviewed corresponding to the target element does not pass the review item corresponding to the target element; when the applicable state of the target element is conditionally applicable, determining whether there are verification points in the solution to be reviewed corresponding to the target element and whether the constraint parameters are satisfied when they exist, and determining whether the solution to be reviewed corresponding to the target element passes the review item corresponding to the target element based on the determination result.

[0085] Specifically, when reviewing a proposed solution based on a target element, the applicable status of the target element in the applicability matrix directly drives the generation rules of the review conclusion. The target element is obtained by cross-locating the review identifier and the construction solution identifier, and is associated with verification points and constraint parameters to trigger detailed judgments under applicable conditions. When the applicable status of a target element is fully applicable, the proposed solution corresponding to the target element is determined to have passed the review item corresponding to the target element. Full applicability means that the review item is a general requirement that should be met under this construction solution, and no additional triggering conditions are needed to include it in the review conclusion generation. When the applicable status of a target element is inapplicable, the proposed solution corresponding to the target element is determined to have failed the review item corresponding to the target element. Inapplicability means that the review item and the construction solution have no applicable relationship, or industry rules clearly state that the combination is invalid, thus directly giving a failure conclusion and avoiding retrieval and reasoning for meaningless combinations. When the applicable status of a target element is conditionally applicable, it is further determined whether there are verification points in the proposed solution corresponding to the target element, and if so, whether the constraint parameters are met. Verification points are used to characterize key measures or elements that need to be explicitly reflected in the solution text; constraint parameters are used to limit the judgment boundaries of verification points, such as limiting the search scope, limiting the candidate size, or limiting the constraints that must be met. Based on the judgment results of whether verification points exist and whether constraint parameters are met, it is determined whether the solution to be reviewed corresponding to the target element passes the review item corresponding to the target element, thereby achieving an executable judgment under the applicable conditions.

[0086] The tender document review method provided in this application formalizes the combination relationship between review items and construction schemes into an applicable state through an applicability matrix. When the conditions are applicable, verification points and constraint parameters are further bound, so that the review conclusions can be generated according to consistent rules in different construction scheme contexts. This reduces the deviation caused by the same review item being judged in different schemes with a one-size-fits-all approach, while improving the interpretability and verifiability of the review results.

[0087] Optionally, in one embodiment, the total review score for the tender document is calculated using the following formula:

[0088]

[0089] The following formula is used to calculate the single-point score for a review point:

[0090]

[0091] The judgment flag is a binary variable (0 or 1), determined by the review logic type, and the score weight is dynamically allocated by the score pool.

[0092] A review logic type corresponds to a channel score pool under a score dimension:

[0093]

[0094] Individual score weights in a review logic type:

[0095]

[0096] Optionally, in one embodiment, the method further includes the following structured evaluation and analysis steps: using a large language model to aggregate the review data of the tender documents from multiple dimensions, generating a structured review report that includes an overall evaluation of the solution, a structured score highlight analysis, targeted improvement suggestions, and suggestions for supplementing missing measures, thereby achieving a deep profile and quantitative diagnosis of the technical solution's quality. Furthermore, all review output data and its supporting evidence (original text fragments, location metadata) are structured and encapsulated, ensuring that each score has traceable original text evidence, supporting manual sampling and review.

[0097] In the structured review report generation stage, this embodiment uses multi-dimensional aggregation analysis of cue word structure to guide the large language model for in-depth evaluation. This cue word structure template includes the following technical elements:

[0098] Professional role positioning: Define the role of the evaluation expert and clarify the analysis task objectives to activate the model's engineering review and report writing capabilities;

[0099] Security isolation mechanism: The review data is isolated and marked by labeling to prevent instructional text in the data area from interfering with the analysis logic;

[0100] Construction scheme context injection: The identified construction scheme name, type and process characteristics are injected as the analysis benchmark to ensure that the evaluation content is closely related to the actual scheme adopted;

[0101] Structured injection of review data: The review judgment results, structured score details, judgment reasons and evidence chains are injected in a structured format to provide complete data support for multi-dimensional analysis;

[0102] Analysis dimension constraints: By specifying analysis requirements, the model is guided to generate targeted and logically coherent evaluation content;

[0103] Solution Focus Constraints: It is explicitly required that only the construction plan adopted by the bidder be analyzed, avoiding the generation of irrelevant plans for comparison or enumeration in the model, so as to ensure the relevance and accuracy of the analysis;

[0104] Structured output specifications: The analysis report is structured and parsable through a predefined JSON output format (including fields such as overall evaluation, structured score highlights, improvement suggestions, and supplementary measures suggestions).

[0105] The aforementioned prompt word structure template aggregates and associates scattered review and judgment data with construction scheme characteristics. Combined with clear analysis dimension constraints and scheme focusing mechanisms, it enables the large language model to generate professional, objective, and targeted review and analysis reports, achieving in-depth profiling and quantitative diagnosis of the quality of technical solutions.

[0106] In summary, the bid document review method provided in this embodiment achieves the following technical effects:

[0107] For tender documents, by combining semantic retrieval based on dense vectors with probabilistic retrieval based on keyword frequency statistics, and using the inverse ranking fusion algorithm to rearrange the results, we can effectively capture deep semantic relationships and implicit compliance requirements in engineering documents, improve the coverage of evidence in the vector space, and significantly reduce retrieval omissions caused by inconsistent terminology.

[0108] For tender documents, in a unified index, texts from different bidders are prone to clustering in a high-dimensional vector space, leading to misjudgments in similarity ranking. This invention significantly reduces the probability of non-target texts entering the nearest neighbor set by isolating them at the object level index, thereby improving the stability of similarity ranking.

[0109] The construction scheme identification adopts a computational path control of coarse screening and candidate set fine judgment, which limits the deep semantic verification of long documents to the candidate set, reduces context injection redundancy, and improves the predictability of system response time.

[0110] The design of the feature library and applicability matrix transforms the scattered natural language review points into a structured and quantifiable feature judgment matrix and verification constraint parameters, reducing the complexity of the transformation path from review rules to executable verification and improving the consistency and reproducibility of the judgment path.

[0111] By decoupling the rule-based judgment and the selection evaluation, and supporting the flexible allocation and proportion adjustment of the score pool, a comparable quantitative result can be formed on the quality of the bid response within a unified calculation framework, thereby improving the interpretability and verifiability of the scoring output.

[0112] The evidence chain information (including the original text location, metadata identifier, and logical deduction process) is fully preserved in the retrieval, matching, and judgment stages, making the final review output data traceable, supporting random inspection and error location, and providing reusable structured sample data for subsequent feature library iterations.

[0113] The review results, structured evaluation values, or evaluation data are structured technical analysis results generated by the computer system based on preset technical constraints. They are used to assist manual or external bidding evaluation systems in making decisions, rather than directly replacing the final formation of the bidding evaluation conclusion.

[0114] Example 2

[0115] Corresponding to Embodiment 1 above, Embodiment 2 of the present invention provides a tender document review device. The technical features and corresponding technical effects can be referred to Embodiment 1 above, and will not be repeated in this embodiment. Figure 2 This is a block diagram of the bid document review device provided in Embodiment 2 of the present invention, as shown below. Figure 2 As shown, the device includes a first acquisition module 201, a second acquisition module 202, a first identification module 203, a third acquisition module 204, a second identification module 205, a traversal module 206, a review module 207, and a first generation module 208.

[0116] The system comprises the following modules: a first acquisition module 201, used to acquire the tender document and the corresponding bidding document; a second acquisition module 202, used to acquire a pre-set review feature library, wherein the review feature library includes an applicability matrix and multiple review items, each review item including a review identifier and review requirements; the applicability matrix includes multiple matrix elements cross-located by the review identifier and construction scheme identifier, the matrix elements defining the applicability status of a review item under a construction scheme and the verification points and constraint parameters associated with the applicability status; a first identification module 203, used to identify, within the bidding document, original text fragments of bidding evidence that meet the review requirements, and construct a review point list for the bidding document, wherein the review point list includes multiple review points, each review point including the original text fragment of bidding evidence, the location information of the original text fragment of bidding evidence in the bidding document, and the review identifier corresponding to the review requirements met by the original text fragment of bidding evidence; and a third acquisition module 204, used to acquire a pre-set construction scheme knowledge base, wherein the construction scheme knowledge base includes multiple scheme items, each scheme item including the construction party... The tender document includes a case identifier and a construction plan description; a second identification module 205 is used to identify the original text fragment of the tender evidence that satisfies the construction plan description in the tender document, and construct a set of plans to be reviewed, wherein the set of plans to be reviewed includes multiple plans to be reviewed, and each plan to be reviewed includes the original text fragment of the tender evidence, the location information of the original text fragment of the tender evidence in the tender document, and the construction plan identifier corresponding to the construction plan description satisfied by the original text fragment of the tender evidence; a traversal module 206 is used to traverse the review point list and the set of plans to be reviewed, and use the review identifier and the construction plan identifier to locate matrix elements in the applicability matrix to obtain target elements; a review module 207 is used to review the plans to be reviewed according to the target elements to obtain the review results corresponding to each review point in the review point list; a first generation module 208 is used to generate review data of the tender document according to the review results corresponding to each review point in the review point list, wherein the review data includes multiple review data units, and each review data unit includes the review points, the plans to be reviewed, and the review results.

[0117] Optionally, in one embodiment, the first identification module includes: a first segmentation unit, used to segment the tender document to obtain multiple tender text units; a first determination unit, used to determine the location information of each tender text unit in the tender document; a first calculation unit, used to calculate the tender vector features of the tender text units; a first extraction unit, used to extract the tender keywords of the tender text units; a first construction unit, used to write the tender text units and their location information into a tender retrieval space, and to construct indexes using the tender vector features and the keywords respectively; a first retrieval unit, used to construct a tender retrieval vector set and a tender retrieval term set using the review requirements, and to perform retrieval in the tender retrieval space based on the tender retrieval vector set and the tender retrieval term set respectively; a first processing unit, used to perform weighted sorting of the retrieval results of the tender retrieval vector set and the tender retrieval term set using a reciprocal sorting fusion algorithm, and to extract a preset number of tender text units to form a tender text candidate set; and a first filtering unit, used to filter and extract tender text units whose location information meets preset position limitation conditions from the tender text candidate set to obtain the original text fragment of the tender evidence.

[0118] Optionally, in one embodiment, the review item further includes a review logic type, and the device further includes: a fourth acquisition module, configured to acquire a structured review question template corresponding to the review logic type; extract data content from the review requirements and add it to the data area of ​​the structured review question template to generate a structured review question for guiding the large language model to make a judgment; acquire a first hierarchical prompt word template for the large language model to reason; a second generation module, configured to inject the original text fragment of the bidding evidence and the structured review question into the data area of ​​the first hierarchical prompt word template respectively to generate a first hierarchical prompt word; and a first processing module, configured to input the first hierarchical prompt word into the large language model, so that the large language model performs deep semantic verification under evidence constraints, and outputs a response result and a response reason for whether the original text fragment of the bidding evidence meets the review requirements; wherein, the review point further includes the response result and the response reason.

[0119] Optionally, in one embodiment, the review item further includes a review logic type, and the device further includes: a mapping module, used to map the review point to a preset score dimension according to the review logic type corresponding to the review point; a classification module, used to classify all review points in the review point list according to the score dimension to obtain a set of review points under each score dimension; a fifth acquisition module, used to acquire a preset score pool configuration, wherein the score pool configuration includes a score pool ratio corresponding to each score dimension; a first determination module, used to determine the distributable total score of the score dimension according to the score pool ratio; and an allocation module, used to allocate scores to each review point in the review point set according to the distributable total score within each score dimension to obtain a single score for each review point; wherein the review data also includes the total review score corresponding to the tender document, and the total review score is calculated based on the review results and single scores of each review point.

[0120] Optionally, in one embodiment, the tender document includes a single bidder's tender document, the construction plan description includes a first description and a second description, and the second identification module includes: a second segmentation unit for segmenting the tender document to obtain multiple tender text units; a second determination unit for determining the location information of each tender text unit in the tender document; a second extraction unit for extracting tender keywords from the tender text units; a second construction unit for writing the tender text units and their location information into a tender retrieval space and constructing an index using the tender keywords; a second retrieval unit for constructing a tender retrieval term set using the first description, and performing a retrieval in the tender retrieval space based on the tender retrieval term set to obtain a tender text candidate set; and a second filtering unit for filtering the original text fragments of the tender evidence in the tender text candidate set according to the second description.

[0121] Optionally, in one embodiment, the second description content includes personnel description, machinery and equipment description, material description, method and process description, and environmental description. When the second screening unit filters the original text fragment of the bidding evidence according to the bidding text candidate set, the specific steps performed include: obtaining a second hierarchical prompt word template for the reasoning of the large language model; injecting the bidding text candidate set and the second description content into the data area of ​​the second hierarchical prompt word template to form second hierarchical prompt words; and inputting the second hierarchical prompt words into the large language model, so that the large language model performs deep semantic verification under evidence constraints and outputs the original text fragment of the bidding evidence.

[0122] Optionally, in one embodiment, the applicable state includes inapplicable, fully applicable, and conditionally applicable. The review module includes: a first review unit, configured to determine that the solution to be reviewed corresponding to the target element passes the review item corresponding to the target element when the applicable state of the target element is fully applicable; a second review unit, configured to determine that the solution to be reviewed corresponding to the target element does not pass the review item corresponding to the target element when the applicable state of the target element is inapplicable; and a third review unit, configured to determine whether the verification point exists in the solution to be reviewed corresponding to the target element and whether the constraint parameter is satisfied when the applicable state of the target element is conditionally applicable, and to determine whether the solution to be reviewed corresponding to the target element passes the review item corresponding to the target element based on the determination result.

[0123] Example 3

[0124] This embodiment also provides a computer device, such as a smartphone, tablet computer, laptop computer, desktop computer, rack server, blade server, tower server, or cabinet server (including a standalone server or a server cluster composed of multiple servers), etc., capable of executing programs. Figure 3 As shown, the computer device 01 in this embodiment includes, but is not limited to, a memory 012 and a processor 011 that can be interconnected via a system bus, such as... Figure 3 As shown. It should be noted that, Figure 3 Only a computer device 01 with component memory 012 and processor 011 is shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0125] In this embodiment, the memory 012 (i.e., the readable storage medium) includes flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory 012 may be an internal storage unit of the computer device 01, such as the hard disk or memory of the computer device 01. In other embodiments, the memory 012 may also be an external storage device of the computer device 01, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 01. Of course, the memory 012 may include both the internal storage unit and its external storage device of the computer device 01. In this embodiment, the memory 012 is typically used to store the operating system and various reference software installed on the computer device 01, such as the program code of the tender document review device in Embodiment 2. In addition, memory 012 can also be used to temporarily store various types of data that have been output or will be output.

[0126] In some embodiments, processor 011 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip. Processor 011 is typically used to control the overall operation of computer device 01. In this embodiment, processor 011 is used to run program code stored in memory 012 or process data, such as a method for reviewing tender documents.

[0127] Example 4

[0128] This embodiment also provides a computer-readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, server, App reference store, etc., which stores computer programs. When the program is executed by a processor, it implements the corresponding functions. The computer-readable storage medium of this embodiment is used to store a tender document review device, and when executed by a processor, it implements the tender document review method of Embodiment 1.

[0129] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0130] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0131] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0132] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for reviewing tender documents, characterized in that, include: Obtain the tender documents and the corresponding bidding documents; Obtain a pre-set review feature library, wherein the review feature library includes an applicability matrix and multiple review items, the review items include review identifiers and review requirements, the applicability matrix includes multiple matrix elements that are cross-located by the review identifiers and construction scheme identifiers, the matrix elements are used to define the applicability status of a review item under a construction scheme and the verification points and constraint parameters associated with the applicability status, the applicability status includes not applicable, fully applicable, and conditionally applicable; In the tender document, identify the original text fragments of tender evidence that meet the review requirements, and construct a review point list for the tender document. The review point list includes multiple review points, and each review point includes the original text fragments of tender evidence, the location information of the original text fragments of tender evidence in the tender document, and the review identifier corresponding to the review requirements met by the original text fragments of tender evidence. Obtain a pre-set construction plan knowledge base, wherein the construction plan knowledge base includes multiple plan items, and each plan item includes a construction plan identifier and a construction plan description; In the tender document, identify the original text fragment of the tender evidence that satisfies the description of the construction plan, and construct a set of plans to be reviewed. The set of plans to be reviewed includes multiple plans to be reviewed. Each plan to be reviewed includes the original text fragment of the tender evidence, the location information of the original text fragment of the tender evidence in the tender document, and the construction plan identifier corresponding to the construction plan description satisfied by the original text fragment of the tender evidence. Traverse the list of review points and the set of schemes to be reviewed, and use the review identifier and the construction scheme identifier to locate the matrix element in the applicability matrix to obtain the target element; The proposed solution is reviewed based on the target element to obtain the review results for each review point in the review point list. This includes: when the target element is fully applicable, determining that the proposed solution corresponding to the target element passes the review item corresponding to the target element; when the target element is not applicable, determining that the proposed solution corresponding to the target element fails the review item corresponding to the target element; when the target element is conditionally applicable, determining whether the verification point exists in the proposed solution corresponding to the target element and whether the constraint parameter is satisfied if it exists, and determining whether the proposed solution corresponding to the target element passes the review item corresponding to the target element based on the determination result. The review data of the tender document is generated based on the review results corresponding to each review point in the review point list. The review data includes multiple review data units, and each review data unit includes the review point, the scheme to be reviewed, and the review result.

2. The method for reviewing tender documents according to claim 1, characterized in that, The steps for identifying the original text fragments of bidding evidence that meet the evaluation requirements in the bidding documents include: The tender document is segmented to obtain multiple tender text units; Determine the location information of each of the tender text units in the tender document; Calculate the tender vector features of the tender text unit; Extract the bidding keywords from the bidding text unit; The tender text unit and its location information are written into the tender retrieval space, and indexes are constructed using the tender vector features and the tender keywords respectively; Using the aforementioned review requirements, a tender retrieval vector set and a tender retrieval term set are constructed. Based on the tender retrieval vector set and the tender retrieval term set, a retrieval is performed in the tender retrieval space. The retrieval results of the tender retrieval vector set and the tender retrieval term set are weighted and sorted using a reciprocal sorting fusion algorithm, and a preset number of tender text units are extracted to form a candidate tender text set; and The bidding text units that meet the preset location limitation conditions are selected and extracted from the candidate bidding text set to obtain the original bidding evidence fragment.

3. The method for reviewing tender documents according to claim 2, characterized in that, The review items also include review logic types. After obtaining the original text fragment of the tender evidence, the method further includes: Obtain the structured review question template corresponding to the review logic type; The data content is extracted from the review requirements and added to the data area of ​​the structured review question template to generate structured review questions to guide the large language model in making judgments; Obtain the first hierarchical prompt word template for inference in the large language model; The original text of the bidding evidence and the structured review questions are respectively injected into the data area of ​​the first hierarchical prompt word template to generate the first hierarchical prompt words; The first hierarchical prompt word is input into the large language model, which performs deep semantic verification under evidence constraints and outputs the response result and reason for whether the original text of the bidding evidence meets the review requirements. The review points also include the response results and the reasons for the response.

4. The method for reviewing tender documents according to claim 1, characterized in that, The review items also include review logic types. After constructing the review point list of the tender document, the method further includes: The review points are mapped to preset score dimensions according to the review logic type corresponding to the review points; All review points in the review point list are categorized according to the score dimension to obtain a set of review points under each score dimension. Obtain a preset score pool configuration, wherein the score pool configuration includes the score pool ratio corresponding to each score dimension; The total allocatable score for the score dimension is determined based on the score pool ratio. Within each score dimension, scores are allocated to each of the review points in the review point set based on the total allocatable score, resulting in a single score for each review point. The review data also includes the total review score corresponding to the tender document, which is calculated based on the review results and individual scores of each review point.

5. The method for reviewing tender documents according to claim 1, characterized in that, The tender documents include a single bidder's tender document, and the construction plan description includes a first description and a second description. The steps for identifying original fragments of tender evidence that satisfy the construction plan description within the tender documents include: The bid document is segmented to obtain multiple bid text units; Determine the location information of each of the bid text units in the bid document; Extract the bidding keywords from the bid text unit; The bid text unit and its location information are written into the bid retrieval space, and an index is constructed using the bid keywords; A bid retrieval term set is constructed using the first description content; a search is performed in the bid retrieval space based on the bid retrieval term set to obtain a candidate set of bid texts; and Based on the second description, the original text fragments of the bid evidence are selected from the bid text candidate set.

6. The method for reviewing tender documents according to claim 5, characterized in that, The second description includes personnel description, machinery and equipment description, material description, process description, and environmental description. The step of selecting the original text fragments of the bid evidence from the candidate bid text set based on the second description includes: Obtain the second-level cue word template for inference in a large language model; The candidate set of bid texts and the second description content are respectively injected into the data area of ​​the second hierarchical prompt word template to form the second hierarchical prompt words; and The second layered prompt word is input into the large language model, which performs deep semantic verification under evidence constraints and outputs the original text fragment of the bidding evidence.

7. A tender document review device, characterized in that, include: The first acquisition module is used to acquire the tender documents and the corresponding bidding documents; The second acquisition module is used to acquire a preset review feature library, wherein the review feature library includes an applicability matrix and multiple review items, the review items include review identifiers and review requirements, the applicability matrix includes multiple matrix elements that are cross-located by the review identifiers and construction scheme identifiers, the matrix elements are used to define the applicability status of a review item under a construction scheme and the verification points and constraint parameters associated with the applicability status, the applicability status includes not applicable, fully applicable, and conditionally applicable; The first identification module is used to identify, in the tender document, the original text fragment of the tender evidence that meets the review requirements, and to construct a review point list of the tender document. The review point list includes multiple review points, and each review point includes the original text fragment of the tender evidence, the location information of the original text fragment of the tender evidence in the tender document, and the review identifier corresponding to the review requirements met by the original text fragment of the tender evidence. The third acquisition module is used to acquire a pre-set construction plan knowledge base, wherein the construction plan knowledge base includes multiple plan items, and the plan item includes a construction plan identifier and a construction plan description; The second identification module is used to identify, in the tender document, the original text fragment of the tender evidence that satisfies the construction scheme description, and construct a set of schemes to be reviewed, wherein the set of schemes to be reviewed includes multiple schemes to be reviewed, and the schemes to be reviewed include the original text fragment of the tender evidence, the location information of the original text fragment of the tender evidence in the tender document, and the construction scheme identifier corresponding to the construction scheme description satisfied by the original text fragment of the tender evidence; The traversal module is used to traverse the list of review points and the set of schemes to be reviewed, and use the review identifier and the construction scheme identifier to locate the matrix element in the applicability matrix to obtain the target element. The review module is used to review the proposed solution based on the target element and obtain the review results corresponding to each review point in the review point list. It includes: a first review unit, used to determine that the proposed solution corresponding to the target element passes the review item corresponding to the target element when the target element's applicability status is fully applicable; a second review unit, used to determine that the proposed solution corresponding to the target element fails the review item corresponding to the target element when the target element's applicability status is not applicable; and a third review unit, used to determine whether the verification point exists in the proposed solution corresponding to the target element and whether the constraint parameter is satisfied when the target element's applicability status is conditionally applicable, and to determine whether the proposed solution passes the review item corresponding to the target element based on the determination result. The generation module is used to generate review data for the tender document based on the review results corresponding to each review point in the review point list. The review data includes multiple review data units, and each review data unit includes the review point, the scheme to be reviewed, and the review result.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images