A method and system for preparing a tender and bid document based on a generative large model

Abstract

The application discloses a kind of based on generative large model's tender and tender file compiling method and system, belong to artificial intelligence technical field, it includes parsing target tender document, extracts hierarchical scoring system structure and each end-level scoring rule;Based on the tender evaluation data of historical winning bid project, the marginal gain value that each scoring item score each change a unit produces to the marginal probability of winning bid is quantified using statistical analysis method, the winning bid sensitivity parameter of each scoring item is obtained, the winning bid sensitivity parameter is different from the nominal score marked in tender document, reflects the actual influence of scoring item;According to two-dimensional index, scoring item is divided into three categories and matched with differentiated generation strategy, multi-agent deep argumentation reinforcement is carried out for high sensitivity weak item, high sensitivity advantage item is highlighted by strong binding of qualification and requirement, low sensitivity routine item is executed compliance minimal output, realize that resource is accurately inclined to high value link.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, specifically to a method and system for writing tender and bid documents based on generative large models. Background Technology

[0002] Bidding and tender document preparation systems rely on the natural language understanding and generation capabilities of large language models to automatically parse bidding documents, intelligently write tender documents, and assist in content compliance review—making them decision-making tools for this purpose. However, existing solutions for generating tender documents using large models generally employ an equal generation strategy. Regardless of the scoring rules or the importance of each clause in the bidding document, the entire document is treated as input, generating all responses with the same depth and granularity. This is tantamount to assuming that all scoring points have the same impact on the winning bid and that bidders are equally competitive on each scoring point. Consequently, the computational power and length of the tender document are evenly distributed across all clauses. This generation method has significant shortcomings in practical applications: firstly, in bidding projects using a comprehensive scoring method with highly unbalanced scoring weights, high-leverage clauses cannot be adequately reasoned and deeply argued, leading to significant point deductions; secondly, the overwhelming advantages of bidders on certain scoring points, such as exclusive patents or typical performance records, cannot be explicitly highlighted and strongly linked, and evaluation experts may completely overlook them, causing the advantages to be submerged and ultimately severely impacting the probability of winning the bid. Summary of the Invention

[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a method and system for writing bidding and tender documents based on a generative large model. Addressing the problem that existing solutions' equal generation strategy, which treats all scoring items equally, leads to the loss of points for high-leverage clauses and the suppression of inherent advantages, this solution proposes a sensitivity-adaptive generation mechanism for scoring items. First, the model parses the bidding documents, extracting the hierarchical scoring structure and detailed scoring rules at each final level. Then, based on historical winning bid announcements, a bidding preference prediction model is constructed to quantify the marginal contribution of changes in the score of each scoring item to the probability of winning the bid—that is, the winning leverage coefficient. Combining the semantic matching degree and uncertainty of the bidder's own qualification materials with the scoring rules, the competitiveness matching entropy of each scoring item is calculated. Based on these two quantitative indicators, scoring items are divided into three categories: high leverage and weak inherently, high leverage and strong inherently, and low leverage. These categories trigger three differentiated generation modes: adversarial deep argumentation generation mode, advantage anchoring strong highlighting mode, and compliance minimalist mode. This transforms the generation behavior of the large model from treating all items equally to a dynamic response, ensuring that every bit of computing power is invested in the key aspects that contribute the most to the probability of winning the bid.

[0004] Meanwhile, this invention also constructs a multi-agent attack and defense review chain to conduct internal simulated bidding-style reinforcement learning closed-loop verification of high-leverage clauses. Dynamic chapter topology reconstruction reorganizes differentiated content blocks into logically coherent complete bid documents. The bidding evaluation cognitive graph continuously learns implicit bidding preferences from historical data and dynamically updates the leverage coefficient. Private incremental security injection ensures consistent content style and data authenticity and traceability through efficient fine-tuning and retrieval enhancement generation. Finally, an adaptive decision generation system with the goal of maximizing the probability of winning the bid is constructed, which effectively improves the quality of bid documents and the success rate of winning the bid.

[0005] To achieve the above objectives, the present invention provides the following technical solution, comprising the following steps:

[0006] Step 1: Analyze the target tender document and extract the hierarchical scoring system structure and the detailed scoring rules for each final level;

[0007] Step 2: Based on the evaluation data of historical winning bids, statistical analysis methods are used to quantify the marginal gain value of each unit change in the score of each scoring item on the probability of winning the bid, so as to obtain the bidding sensitivity parameter of each scoring item. The bidding sensitivity parameter is different from the nominal score marked in the bidding document and reflects the actual influence of the scoring item.

[0008] Step 3: Compare the matching degree between the bidder's own qualification materials and each scoring criterion, and calculate the competitiveness fit value of each scoring item based on the matching uncertainty.

[0009] Step 4: Combining the sensitivity parameters of the winning bid with the competitiveness fit value, all scoring items are divided into three categories: high sensitivity with weak competitiveness, high sensitivity with competitive advantage, and low sensitivity with normal.

[0010] Step 5: Dynamically adjust the generation depth, argument strength, and content length of the large model according to the category to which the scoring item belongs, and implement differentiated content generation strategies for different categories of scoring items;

[0011] Step Six: For highly sensitive scoring items related to weak competitiveness, use an in-depth argumentation generation mode to output the response content; for highly sensitive scoring items related to competitive advantages, use an advantage anchoring and highlighting generation mode to output the response content; for low-sensitivity, conventional scoring items, use a compliance-based, simplified generation mode to output the response content.

[0012] Step 7: Integrate and arrange the responses to each scoring item according to the standard format of the tender document, and output the complete tender document;

[0013] As a further preferred option of this technical solution: when parsing the tender documents, according to the hierarchical structure of the scoring method, the scoring categories, scoring items and final scoring details are extracted level by level, and the corresponding scores, review methods and veto clauses for each scoring item are marked simultaneously.

[0014] As a further preferred embodiment of this technical solution: the bid-winning sensitivity parameter is obtained by statistically analyzing the marginal changes in the scores of each scoring item in historical projects and the bid-winning results; the higher the sensitivity, the greater the actual marginal impact of the score fluctuation of that scoring item on the probability of winning the bid, and its value is independent of the original score weight marked in the tender document.

[0015] As a further preferred option of this technical solution, the rules for dynamically adjusting the large model generation strategy are as follows: the higher the sensitivity of the scoring item, the more generation computing power is allocated, the richer the argumentation dimensions of the response content, and the higher the content length of a single scoring item; the lower the sensitivity of the scoring item, the more the generated content focuses on meeting compliance requirements, and the more the generated content and computing resource consumption are controlled.

[0016] As a further preferred embodiment of this technical solution: the deep argumentation generation mode is as follows: for the requirements of this scoring item, a multi-agent self-game generation process is initiated. After the initial draft of the response is generated in the first round, the simulated evaluation expert agent conducts multiple rounds of reverse questioning and requests for supplementary evidence until the closure judgment condition is met, at which point the generation process ends and the response content, which includes multi-dimensional supporting materials, implementation paths, and risk response plans, is output. The closure judgment condition is: the simulated evaluation expert agent does not raise any new valid questions in two consecutive rounds, and all questions raised have been supported by corresponding supporting materials.

[0017] The advantage anchoring and highlighting generation mode is as follows: the vectorized evidence of the bidder's core qualifications, patent achievements, and typical performance is accurately matched and bound to the requirements vector of the scoring rules, and the structured tag-style highlighting is presented in the title, abstract and first paragraph of the response content.

[0018] The compliance-compliant simplified generation mode is as follows: complete the basic response according to the scoring requirements, satisfying the formal compliance requirements while controlling the content length to not exceed a preset threshold.

[0019] As a further preferred option of this technical solution: after generating the response content for high-sensitivity scoring items, a multi-role simulated evaluation and review process is initiated. The evaluation expert AI agent simulates scoring the response content according to the real scoring rules and raises questions, while the compliance review AI agent simultaneously scans for the risk of rejection. If the simulated score is lower than the preset threshold or there is a risk of rejection, the response content is reconstructed from the generation stage until the simulated score reaches or exceeds the preset threshold and there is no risk of rejection, thus completing the verification and forming a closed loop of generation, review, and optimization.

[0020] As a further preferred option of this technical solution: when integrating and generating content, according to the standard format and logical hierarchy of the tender document, the differentiated content modules are reorganized into a complete tender document with a coherent structure and unified format, and the format specification verification is completed simultaneously.

[0021] As a further preferred option of this technical solution: construct a bid evaluation knowledge base, store the bid evaluation rules and implicit review preferences of historical projects, continuously iterate and update the bid evaluation data of newly added winning projects, dynamically calibrate the bid-winning sensitivity parameters of each scoring item, and adapt to the review characteristics of different industries and different bidders.

[0022] As a further preferred option of this technical solution: by combining localized model fine-tuning with search enhancement generation, the bidder's own qualification materials, historical project cases and exclusive writing style are injected into the large model generation process to ensure that the output content has a consistent style, the data is authentic and traceable, and at the same time meets the data security control requirements.

[0023] As a further preferred embodiment of this technical solution: the system for executing the bidding document preparation method includes:

[0024] The tender document parsing module is used to parse the target tender document and extract the hierarchical scoring structure and final-level scoring details.

[0025] The sensitivity measurement module for scoring items is used to calculate the sensitivity parameters of each scoring item based on historical bidding and evaluation data using statistical analysis methods.

[0026] The competitiveness assessment module is used to compare the bidder's qualification materials with the scoring rules to obtain the competitiveness fit value for each scoring item.

[0027] The dynamic matching module for generating strategies combines the sensitivity parameter of winning the bid with the competitiveness fit value to divide the scoring items into different categories, and matches the corresponding generation depth, argumentation strength and length configuration for each category;

[0028] The differentiation generation module is used to generate the response content for each scoring item one by one according to the matched generation configuration;

[0029] The multi-agent review and verification module is used to conduct multi-role simulated evaluation and review of the generated content of high-sensitivity scoring items, and trigger content reconstruction based on the review results to achieve closed-loop verification of generation-review-optimization.

[0030] The content integration and output module is used to integrate the responses to each scoring item into a complete tender document with a unified style and coherent structure, and then output it.

[0031] The bid evaluation knowledge base iterative update module is used to store historical bid evaluation rules and implicit preferences, and is continuously updated with new winning bid data to dynamically calibrate the bid sensitivity parameters.

[0032] Compared with the prior art, the beneficial effects of the present invention are:

[0033] 1. Addressing the common problem in existing generative large-scale model bid document writing schemes that employ an equal generation strategy, resulting in insufficient justification for high-leverage scoring items leading to point deductions and the obscuring of bidders' core competitive advantages, thus limiting the improvement of bid document quality and success rate, this solution establishes a bid-winning sensitivity measurement mechanism. Based on historical evaluation data, it calculates the actual impact of each scoring item on the bid-winning result, moving beyond the superficial constraints of nominal scores. Combining semantic matching degree and matching uncertainty, it calculates the competitiveness fit value, accurately identifying the score improvement potential for each scoring item. Based on dual-dimensional indicators, scoring items are divided into three categories and matched with differentiated generation strategies. For high-sensitivity weak items, in-depth multi-agent analysis and reinforcement are conducted; for high-sensitivity advantageous items, strong binding of qualifications and requirements is implemented to highlight them; and for low-sensitivity routine items, compliant and simplified output is executed. This achieves precise allocation of generation resources to high-value aspects, effectively improving the bid document's scoring competitiveness and bid-winning probability.

[0034] 2. To address the issues of existing bid document generation solutions lacking internal quality closed-loop verification, unable to dynamically adapt to different review preferences, and struggling to balance content style and data security, resulting in unstable generation quality and weak scenario adaptability, this solution constructs a multi-agent, dual-role review and verification mechanism. This mechanism forms a complete closed loop of generation, review, and optimization through simulated expert review and bid rejection risk scanning, proactively identifying potential points deductions and bid rejection risks. An iterative bid evaluation knowledge base is built, dynamically calibrating the bid-winning sensitivity parameters through momentum updates, continuously accumulating implicit review preferences, and adapting to the review characteristics of different industries and bidders. Localized retrieval enhances generation and lightweight style fine-tuning, achieving a unified output style and traceable content authenticity while ensuring data security and controllability, comprehensively improving the reliability and scenario adaptability of bid document generation. Attached Figure Description

[0035] Figure 1 This is an overall flowchart of a method and system for writing tender and bid documents based on a generative large model, according to the present invention.

[0036] Figure 2 This is a diagram illustrating the tender document analysis method and system for writing tender and bid documents based on a generative large model, as described in this invention.

[0037] Figure 3 This is a flowchart illustrating the modules of a bidding and tender document writing system based on a generative large model, as described in this invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Example

[0040] Please see Figures 1-3 This is a schematic diagram of some embodiments of a method and system for writing tender and bid documents based on a generative large model, as described in this application.

[0041] Example 1

[0042] This embodiment illustrates the overall pipeline architecture and end-to-end execution process of this method. The bidding document writing method based on a generative large model relies on seven functional modules that work together sequentially. These seven modules are: bidding document parsing module, scoring item sensitivity measurement module, competitiveness assessment module, generation strategy dynamic matching module, differentiated generation module, multi-agent review and verification module, and content integration and output module. They are also supported by an evaluation knowledge base iterative update module and a localized data security module.

[0043] The overall execution process is as follows: the tender document parsing module receives the target tender document, extracts the hierarchical scoring system structure and the final scoring details, and outputs a list of structured scoring items;

[0044] The scoring item sensitivity quantification module calculates the bidding sensitivity parameters of each scoring item based on the evaluation data of historical winning projects, and quantifies the actual impact of each scoring item on the bidding result.

[0045] The competitiveness assessment module compares the bidder's own qualification materials with each scoring rule and calculates the competitiveness fit value for each scoring item.

[0046] The dynamic matching module for generating strategies combines the sensitivity parameter of winning the bid with the competitiveness adaptation value to divide the scoring items into three categories and match the corresponding generation depth, argumentation strength and length configuration.

[0047] The differentiation generation module relies on a generative large model to generate the response content for each scoring item one by one according to the matched configuration;

[0048] The multi-agent review and verification module conducts multi-role simulated evaluation and review for highly sensitive scoring items, forming a closed-loop verification of generation, review, and optimization;

[0049] The content integration and output module integrates and arranges the response content of all scoring items according to the standard format of the tender document, and outputs the complete tender document after format verification.

[0050] The evaluation knowledge base iterative update module continuously accumulates historical evaluation data and implicit review preferences, and dynamically calibrates the winning sensitivity parameters of each scoring item; the localized data security module injects bidder-specific data into the generation process through local fine-tuning and retrieval enhancement technologies, ensuring data security and style consistency.

[0051] The overall data flow is as follows: tender document parsing, bid-winning sensitivity measurement, competitiveness assessment, dynamic matching of generation strategies, differentiated generation, multi-agent review and verification, and content integration and output. Among them, bid-winning sensitivity measurement and bid evaluation knowledge base iterative updates form a two-way data interaction, and a feedback re-entry loop is formed between differentiated generation and multi-agent review and verification.

[0052] Example 2

[0053] This module targets tender documents in PDF, Word, and other formats, and extracts a complete hierarchical scoring system through document structure parsing and semantic entity extraction techniques. The specific execution steps are as follows:

[0054] Document structure restoration: Analyze the table of contents, headings, and paragraph formats of the tender documents, locate the core chapters such as scoring methods, evaluation criteria, and detailed review, and restore the tree-like hierarchical structure of the scoring system, which is divided into three levels from top to bottom: scoring categories, scoring items, and final scoring details.

[0055] Assume the chapter structure of the tender document is a tree structure. Where N is the set of chapter nodes, This is a set of parent-child relationship edges. For each chapter node... Record its title level and text content The recursive extraction algorithm executes as follows: starting from the root node, iterates through the tree. When it detects that the title level of the current node is shallower than the previous node, it backtracks to the parent node of the corresponding level. After traversal, all leaf nodes are extracted from the tree structure as the final-level scoring criteria.

[0056] The field structure is extracted by simultaneously extracting the corresponding fields for each final-level scoring rule: the original text of the rule, the bid score, the review method, whether it is a veto clause, and the required supporting documents. Among these, veto clauses are identified through keyword matching and semantic classification. Clauses containing mandatory semantics such as invalid bid, rejected bid, or failure to pass are marked as high-risk veto items. ,otherwise .

[0057] The scoring item standardization mapping involves uniformly encoding all extracted scoring items and establishing a mapping table of scoring item ID, category, nominal score, and review rules, which serves as a unified input benchmark for subsequent quantitative evaluation and content generation.

[0058] After parsing, a standardized hierarchical rating structure is output. Let M be the total number of final-level rating items obtained from the parsing, and let the set of rating items be... This ensures that all scoring items are complete and without duplication, and that the hierarchical classification is accurate.

[0059] Example 3: Scoring Sensitivity Measurement Module

[0060] This module uses a dataset of historical successful bids and employs partial correlation analysis to quantify the marginal gain of changes in the score of individual scoring items on the probability of winning the bid, thus obtaining a sensitivity parameter for each scoring item. This parameter is independent of the nominal score stated in the tender documents and reflects the actual influence of the scoring item on the bid outcome.

[0061] Symbol definition

[0062] Let the total number of historical items be N, and there be M scoring items; the score of the j-th scoring item in the i-th item is... The winning bid result for the i-th project is (A score of 1 is assigned to the winning bidder, and 0 to the unsuccessful bidder). Controlling for the influence of the scores of the remaining M-1 scoring items, calculate the partial correlation coefficient between the j-th score and the winning bid result. .

[0063] Formula for calculating partial correlation coefficient: ;

[0064] in: Pearson correlation coefficient between the winning bid result and the score of the j-th scoring item; : The correlation coefficient between the winning bid result and the overall score of the remaining evaluation items; : The correlation coefficient between the score of the j-th rating item and the combined score of the remaining rating items.

[0065] To avoid the instability of estimation based on a single dataset, this module employs a cross-validation ensemble estimation method. The historical dataset is randomly divided into K folds (K = 5 or 10), with each fold serving as the validation set and the remainder as the training set. K sets of partial correlation coefficient estimates are then calculated. The final partial correlation coefficient is the mean of the K estimated values.

[0066] The sensitivity parameter conversion formula calculates the bid sensitivity parameter based on the partial correlation coefficient. The partial correlation coefficient is obtained by using a nonlinear mapping. Sensitivity values ​​converted to the [0,1] interval: ;

[0067] When the partial correlation coefficient is negative, it indicates that the score of this scoring item is negatively correlated with the probability of winning the bid. Cut off to 0.

[0068] The value ranges from [0,1], with larger values ​​indicating a stronger actual influence of the scoring item on the bidding result. The sensitivity obtained in this way can separate the nominal relationship of the scores in the bidding documents and truly reflect the implicit weighting preferences in the bidding process.

[0069] Example: The technical solution in a certain tender document is nominally worth 40 points, but historical data analysis reveals that its sensitivity parameter... =0.72, indicating that this scoring item has a strong practical influence on the bidding result; while the project management plan has a nominal score of 5 points, and the sensitivity parameter... =0.12, indicating that the change in its score has little impact on the winning result.

[0070] The historical bidding dataset input into this module must cover bidding projects of the same industry and type, with a sample size of no less than 200 records, to ensure that the statistical results are statistically significant.

[0071] Example 4: Competitiveness Assessment Module

[0072] This module compares the bidder's own qualification material library with the scoring rules for each scoring item, and calculates the competitiveness fit value from two dimensions: semantic matching degree and matching uncertainty, thereby quantifying the bidder's competitive advantage level in that scoring item.

[0073] The semantic matching degree calculation involves performing a semantic search between the text of the i-th scoring rule and the bidder's qualification material database, recalling the top-K most relevant qualification certificate texts, calculating the semantic similarity of each, and taking the maximum value to obtain the semantic matching degree. Cosine similarity was used for calculation.

[0074] ;in: : The semantic vector of the scoring rules text; : The semantic vector of the qualification certificate text; : This is a collection of qualification certificates for the recall.

[0075] The value ranges from [0,1], and the higher the value, the better the match between the qualifications and the scoring requirements.

[0076] Matching uncertainty calculation: Calculate matching uncertainty based on the similarity dispersion of recall qualifications and the completeness of supporting materials. : ;

[0077] in, The standard deviation of the similarity of the recall qualifications reflects the stability of the matching results. The larger the standard deviation, the greater the difference in the degree of matching between each candidate qualification and the scoring requirements, and the higher the matching uncertainty. : The number of existing supporting documents; : This refers to the quantity of material types required in the scoring criteria. When At that time, take . The value ranges from [0,1]. The higher the value, the greater the uncertainty of the matching result and the lower the reliability of the competitiveness.

[0078] The competitiveness fit value is synthesized by combining the matching degree and uncertainty to obtain the competitiveness fit value of the i-th scoring item. : ;

[0079] The use of multiplicative synthesis instead of weighted summation is based on the technical consideration that matching degree and uncertainty have a limiting effect on competitiveness evaluation. No matter how high the matching degree, if the matching degree is highly uncertain... If it is close to 1, then A value close to 0 indicates a significantly lower overall fit, consistent with the evaluation logic that untraceable evidence of advantage does not constitute effective competitiveness; conversely, regardless of how certain the match, if the fit itself is low,... The overall fit value also approaches 0, reflecting the principle that dominant evidence unrelated to the scoring requirements does not contribute to the score. The value ranges from [0,1], and the higher the value, the stronger the bidder's competitive advantage in that scoring item.

[0080] Example 5: Dynamic Matching Module for Generation Strategy

[0081] This module uses the sensitivity parameter of winning bids. Fit with competitiveness Based on two core dimensions, all scoring items are divided into three categories, and the corresponding generation strategy parameters are dynamically matched to realize that the generation behavior of the large model responds differently according to the value of the scoring items, replacing the traditional equal generation mode.

[0082] Classification threshold is dynamically determined and sensitivity threshold is set. and fit threshold Both thresholds are dynamically determined based on the distribution of scoring items in the current bidding project, rather than fixed values, to ensure the adaptability and objectivity of the classification results.

[0083] Sensitivity threshold The sensitivity parameters of the winning bid for all scoring items Sort the values ​​from highest to lowest, and take the values ​​corresponding to the top 30% of the sorted positions as the high sensitivity threshold: ;

[0084] Where M is the total number of scoring items, It is a rounding function. This represents the sensitivity value of the k-th item after sorting. This method ensures that high-sensitivity scoring items account for approximately 30% of all scoring items, which aligns with the applicability of the Pareto principle in bidding evaluation scenarios.

[0085] Fit threshold Take the median of the competitiveness fit values ​​for all rating items: This method divides the scoring items into two groups with relatively strong and weak competitiveness, avoiding threshold shifts caused by skewed data distribution.

[0086] Due to differences in scoring criteria and bidder qualification requirements between different bidding projects, and Adaptive to change, rather than fixed values. For example, calculations in municipal engineering projects... , ; calculated in another information technology project , The threshold is dynamically adapted to the distribution of project data.

[0087] The three-category classification rule for the scoring items is based on the dynamic threshold mentioned above, and the classification rules are as follows:

[0088] Highly sensitive and weakly competitive categories ;satisfy and This type of scoring item has a significant impact on the winning bid, but the bidder's own qualifications are relatively weak. It is the core link with the greatest potential for score improvement and requires the most resources to be invested in in-depth analysis.

[0089] Highly sensitive competitive advantage category ;satisfy and These scoring items have a significant impact on the winning bid and are where bidders have a clear advantage; they are a crucial factor in differentiating scores and therefore require bidders to highlight their competitive advantages.

[0090] Low sensitivity general class ;satisfy These scoring items have a limited impact on the bidding results, with compliance requirements as the core objective, and adopt a minimum resource generation strategy.

[0091] Three categories satisfy And they are mutually exclusive.

[0092] The generation strategy dynamically adjusts the rules, dynamically adjusting the generation depth, argument strength, and content length ratio of the large model for different categories of scoring items. (Generation depth coefficient) The adjustment formula is as follows:

[0093] ;

[0094] in: : is the generation depth coefficient of the i-th rating item; : This is the baseline generation depth, with a value ranging from 0.8 to 1.2. It controls the uniform depth in the equal generation mode and can be set to 1.0. : This is a sensitivity adjustment coefficient, ranging from 0.5 to 0.8, which controls the intensity of the influence of the winning bid sensitivity on the generation depth; : This is the fitness adjustment coefficient, ranging from 0.5 to 0.8. It controls the influence of the competitiveness fitness value on the generation depth. The lower the fitness value, the lower the generation depth. The larger the value, the higher the generation depth, reflecting the strategic logic of focusing on strengthening weak links.

[0095] Generate depth coefficients The coefficient directly corresponds to the number of reasoning steps, the number of argumentation dimensions, and the proportion of content in the large model: the higher the coefficient, the more reasoning steps, the richer the argumentation dimensions, and the higher the proportion of content in a single scoring item, thus achieving a differentiated tilt of computing power and content towards high-value scoring items.

[0096] Example 6: Differentiation Generation Module

[0097] Based on the above embodiments, this embodiment describes the specific implementation of the three types of generation modes of the differentiated generation module.

[0098] Deep argument generation mode, targeting For highly sensitive and weakly competitive scoring items, a multi-agent self-game generation process is adopted, with multiple rounds of questioning and iterative reinforcement to strengthen the sufficiency of the argument. The specific process is as follows:

[0099] The first round of generation involves creating a draft response based on the scoring criteria and existing qualification materials, clearly defining the response framework and core arguments. The generated prompt template includes the original scoring criteria, nominal scores, review methods, and a summary of the bidder's qualification materials for reference.

[0100] The reverse challenge involves a simulated expert agent raising questions about the initial response draft from three dimensions: compliance, feasibility, and advancement. The agent is required to provide supplementary supporting materials, implementation plans, and risk mitigation measures. The simulated expert agent's prompt template includes scoring details and the current response content, requiring it to conduct the review from the perspective of a real expert.

[0101] Through iterative updates, the large model addresses each question by supplementing and refining its responses, adding supporting evidence and practical details.

[0102] The closure determination is made when no new valid questions are raised in two consecutive rounds, and all questions raised have been supported by corresponding supporting materials. The argument is then deemed closed, and the final response is output.

[0103] The criteria for determining whether supporting evidence has been obtained are: the revised draft contains specific responses to the challenge, and the responses cite specific qualification materials, data, or cases as supporting evidence. After each iteration, the system records the iteration number and the number of challenges resolved.

[0104] The response output of this model includes multi-dimensional supporting materials, implementation paths, risk response plans, and expected results, making up for qualification deficiencies to the greatest extent and increasing the probability of scoring.

[0105] Advantage anchoring highlights the generation mode, targeting The highly sensitive and competitive scoring items are generated using a structured approach that emphasizes qualifications and strong binding requirements. The specific process is as follows:

[0106] Element matching aligns and binds the bidder's core qualifications, patent achievements, and typical performance records with each requirement of the scoring criteria, establishing a mapping relationship between requirements, evidence, and value. Specifically, the scoring criteria are divided into several requirement points, and the advantageous evidence is divided into several evidence sentences. Semantic similarity calculation is used to establish a mapping between requirement points and evidence sentences.

[0107] The structured presentation uses structured tags in the title, abstract, and first paragraph of the response to highlight the core advantages. The main body of the response uses a three-part structure of performance name, project scale, and corresponding capabilities to strengthen the argument.

[0108] Value enhancement involves explaining the actual value that this advantage can bring to the tendering party in conjunction with the project requirements, thereby strengthening the evaluation experts' perception of the advantage.

[0109] Compliance-compliant minimalist generation mode, targeting The low-sensitivity, standard scoring items are generated using a point-based response format, with strict control over length. Specific rules are as follows:

[0110] This response addresses each of the scoring requirements individually, without providing any further elaboration.

[0111] The length of a single rating item response shall not exceed the preset length threshold. Length threshold and nominal score of scoring items Positive correlation: ;

[0112] in This is an empirical coefficient. The overall maximum length is controlled within the range of 300-600 words. In the standard mode, the word count of a single sub-item response adapts to the nominal score, usually controlled within the range of 300-600 words. This ensures that the response content meets formal compliance requirements, with no missing response points and no expressive errors. This mode, while guaranteeing compliance scores, minimizes the computational cost and length of low-value scoring items.

[0113] Example 7: Multi-Agent Review and Verification Module

[0114] This module conducts multi-role simulated review of the generated content for all high-sensitivity scoring items, using internal closed-loop verification to proactively identify and mitigate the risk of losing points and potential rejection. The specific process is as follows:

[0115] A dual-role parallel review process is implemented, simultaneously launching a simulated expert agent and a compliance review agent. The evaluation expert agent scores page by page, outputting a simulated score. The formula for calculating the simulated score is:

[0116] ;

[0117] in: The weight of the k-th rating dimension; The score for the i-th rating item on the k-th dimension is given by the sum of the weights of each dimension being 1. , , The nominal score is assigned to each scoring item.

[0118] The compliance review agent scans the matching degree of rejection clauses and outputs the risk value of bid rejection. The formula for calculating the risk value of rejected bids is: ;

[0119] in: It is a set of veto rules; This is a set of semantic fragments of the current response content; This is a semantic matching function with a value range of [0,1].

[0120] Reconstruction trigger determination simulation scoring threshold The threshold is dynamically determined based on the importance of the scoring items. For highly sensitive scoring items, the threshold is set at 75% to 80% of the nominal score. For low-sensitivity scoring items, the threshold distribution is dynamically determined, taking the lower quartile of the risk value of historical rejected bids. ;

[0121] Where F represents the set of historical rejected bid cases. This is the lower quartile function.

[0122] when Or satisfy This triggers content reconstruction, returning to the differentiation generation module to regenerate the response content; while simultaneously satisfying... or At that time, it is considered to have passed the review.

[0123] Closed-loop iterative control: The maximum number of reconstructions is set to 3. If the threshold requirement is still not met after 3 attempts, a risk warning and a suggestion for manual review are output to ensure that the system can promptly notify the user to intervene when automatic iteration fails to converge.

[0124] Example 8: Content Integration and Output Module

[0125] This module reorganizes and arranges the differentiated scoring items according to the standard format of the tender documents, forming a logically coherent and uniformly formatted complete tender document. The specific steps are as follows:

[0126] Chapter topology mapping

[0127] Establish a mapping relationship between the scoring items and the chapters of the tender document, and classify the scoring items of business, technology and credit into the corresponding chapters, and arrange the content modules according to the logical order of the chapters;

[0128] Transition generation

[0129] The large model is designed to generate transition sentences between adjacent content modules, introductory paragraphs, unified structural content, and the beginning content of the next chapter. The goal is to generate natural transition sentences that connect the preceding and following sections without introducing new content.

[0130] Format specification verification

[0131] The entire document was formatted in accordance with the requirements of the tender documents, including font size, paragraph format, page numbering, header and footer numbering, to ensure that the tender documents fully comply with the format requirements.

[0132] Final output generation

[0133] Generate ready-to-use Word / PDF format bid documents, and output a scoring item comparison table, marking the response position, argument strength and estimated score of each scoring item, to facilitate manual review.

[0134] Example 9: Knowledge Base Iterative Update Module

[0135] This module builds an annotation knowledge base, continuously accumulates historical bidding data and implicit review preferences, and dynamically iterates the bidding sensitivity parameters of each scoring item to achieve continuous evolution of the generation strategy.

[0136] Knowledge base composition

[0137] The knowledge base stores three types of core data:

[0138] Historical winning bid data: detailed scoring rules, score distribution, and structured evaluation data of winning bids for each historical project;

[0139] Local project-specific preferences: extracted from past winning bids, performance records, and evaluation characteristics of expert reviewers;

[0140] Evaluation rules for different industries and different bidders (difference labels).

[0141] Sensitivity parameters are updated dynamically.

[0142] Each time a new batch of winning bid data is added, the bid-winning sensitivity parameters of the scoring items are calibrated using a momentum update method. The update formula is: ;

[0143] in: Sensitivity parameters before the update; The new sensitivity was recalculated based on a new batch of data; The momentum coefficient ranges from 0.85 to 0.95. By updating the momentum, the stability of historical data and the timeliness of new data are balanced. The parameter gradually converges with continuous iteration of the sample size, and the stability of the sensitivity parameter and its accuracy in reflecting review preferences are continuously improved.

[0144] Scenario-based adaptation allows for the partitioning of the knowledge base by tags such as industry, tenderer, and project type. It outputs corresponding sensitivity parameters for different scenarios, adapting to the review characteristics of different fields. For example, the distribution of sensitivity parameters differs significantly between municipal engineering projects and information technology projects; scenario-based adaptation allows for the separate invocation of corresponding sub-knowledge bases for calculation.

[0145] Example 10

[0146] This module combines localized model fine-tuning with enhanced search generation, incorporating bidders' own qualification materials, historical project cases, and unique writing style into the large model generation process. This ensures consistent content style and traceable data authenticity while guaranteeing data security. The specific implementation method is as follows:

[0147] Local knowledge base construction: All qualification materials, historical bid documents, patent certificates, performance certificates and other documents provided by the bidder are parsed and vectorized to build a dedicated knowledge base. All data is stored on the bidder's local server and is not transmitted to the external network.

[0148] Retrieval-enhanced generation injection: When generating response content for each scoring item, matching supporting materials and reference texts are first retrieved from the local knowledge base. The search results are then injected as context to generate prompts. The large model generates content based on the retrieved real data, ensuring that all statements are supported by authentic materials and are traceable and verifiable. Recall quantity of retrieval-enhanced generation. Select 5 to 10, sort the search results in descending order of semantic similarity, and then truncate the results. strip.

[0149] Lightweight Style Fine-Tuning: Based on samples of historical bid documents from bidders, lightweight LoRA (Low-Rank Adaptation) fine-tuning is performed on the large model. LoRA fine-tuning introduces a low-rank decomposition matrix next to the weight matrix of the pre-trained model. During fine-tuning, only the parameters of the low-rank matrix are updated, keeping the original pre-trained weights unchanged, thereby achieving efficient model adaptation. The goal of fine-tuning is to minimize the differences in language style, argumentation habits, and expression methods between the generated content and historical winning bid documents, ensuring that the output writing style and expression habits are consistent with the original document style of the bidders.

[0150] Data security management: All data remains within the local environment throughout the entire process, supporting private deployment and offline operation, meeting the security management requirements for confidential data and trade secrets.

[0151] Example 11

[0152] This embodiment illustrates the system architecture implementation of the present invention. The bidding document writing system based on a generative large model executes the bidding document writing method as described in the preceding embodiments. The system adopts a modular design, comprising seven core modules and two supporting modules, and supports various deployment methods including private deployment, cloud SaaS deployment, and edge node deployment.

[0153] The seven core modules are: Tender Document Parsing Module: responsible for parsing the target tender document, extracting the hierarchical scoring structure and final-level scoring details, and outputting a standardized list of scoring items; Scoring Item Sensitivity Measurement Module: responsible for calculating the bid-winning sensitivity parameters of each scoring item based on historical bid-winning evaluation data using partial correlation statistical methods; Competitiveness Assessment Module: responsible for comparing bidder qualifications with scoring details and calculating the competitiveness fit value for each scoring item; Dynamic Matching of Generation Strategies Module: responsible for matching the corresponding generation depth, argumentation strength, and length configuration for different scoring items by combining the bid-winning sensitivity parameters and competitiveness fit values; Differentiated Generation Module: responsible for generating the response content for each scoring item one by one according to the matched generation configuration, relying on the generative large model; Multi-Agent Review and Verification Module: responsible for conducting multi-role simulated bid-winning review of the generated content of high-sensitivity scoring items, triggering content reconstruction, and realizing closed-loop verification of generation-review-optimization; Content Integration and Output Module: responsible for integrating the content of each scoring item according to the standard format, and outputting a complete tender document after format verification.

[0154] The two main support modules are:

[0155] The bid evaluation knowledge base iterative update module is responsible for storing historical bid evaluation rules and implicit review preferences, continuously updating them as new winning bid data is added, and dynamically calibrating the bid sensitivity parameters.

[0156] Localized security injection module: responsible for local knowledge base management, retrieval enhancement generation, and style fine-tuning, ensuring data security and content style consistency.

[0157] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

[0158] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for writing tender and bid documents based on a generative large model, characterized in that: Includes the following steps: Step 1: Analyze the target tender document and extract the hierarchical scoring system structure and the detailed scoring rules for each final level; Step 2: Based on the evaluation data of historical winning bids, statistical analysis methods are used to quantify the marginal gain value of each unit change in the score of each scoring item on the probability of winning the bid, so as to obtain the bidding sensitivity parameter of each scoring item. The bidding sensitivity parameter is different from the nominal score marked in the bidding document and reflects the actual influence of the scoring item. Step 3: Compare the matching degree between the bidder's own qualification materials and each scoring criterion, and calculate the competitiveness fit value of each scoring item based on the matching uncertainty. Step 4: Combining the sensitivity parameters of the winning bid with the competitiveness fit value, all scoring items are divided into three categories: high sensitivity with weak competitiveness, high sensitivity with competitive advantage, and low sensitivity with normal. Step 5: Dynamically adjust the generation depth, argument strength, and content length of the large model according to the category to which the scoring item belongs, and implement differentiated content generation strategies for different categories of scoring items; Step Six: For highly sensitive scoring items related to weak competitiveness, use an in-depth argumentation generation mode to output the response content; for highly sensitive scoring items related to competitive advantages, use an advantage anchoring and highlighting generation mode to output the response content; for low-sensitivity, conventional scoring items, use a compliance-based, simplified generation mode to output the response content. Step 7: Integrate and arrange the responses to each scoring item according to the standard format of the tender document, and output the complete tender document.

2. The method for writing tender and bid documents based on a generative large model according to claim 1, characterized in that: When analyzing the tender documents, the scoring categories, sub-items, and final scoring details are extracted level by level according to the hierarchical structure of the scoring method. The corresponding score, review method, and veto clause are marked simultaneously for each scoring item.

3. The method for writing tender and bid documents based on a generative large model according to claim 2, characterized in that: The bid-winning sensitivity parameter is obtained by statistically analyzing the marginal changes in the scores of each scoring item in historical projects and the bid-winning results. The higher the sensitivity, the greater the actual marginal impact of the score fluctuation of that scoring item on the probability of winning the bid. Its value is independent of the original score weight marked in the tender document.

4. The method for writing tender and bid documents based on a generative large model according to claim 3, characterized in that: The rules for dynamically adjusting the large model generation strategy are as follows: the higher the sensitivity of a scoring item to winning the bid, the more generation computing power is allocated, the richer the argumentation dimensions of the response content, and the higher the content length of a single scoring item; the lower the sensitivity of a scoring item to winning the bid, the more the generated content is centered on meeting compliance requirements, and the control of the generated length and computing resource consumption.

5. The method for writing tender and bid documents based on a generative large model according to claim 4, characterized in that: The in-depth argumentation generation mode is as follows: For the requirements of this scoring item, a multi-agent self-game generation process is initiated. After the initial draft of the response is generated in the first round, the simulated expert agent conducts multiple rounds of reverse questioning and requests for supplementary evidence until the closure judgment condition is met, at which point the generation process ends, and the response content, including multi-dimensional supporting materials, implementation paths, and risk response plans, is output. The closure judgment condition is: the simulated expert agent does not raise any new valid questions in two consecutive rounds, and all questions raised have been supported by corresponding supporting materials. The advantage anchoring and highlighting generation mode is as follows: the vectorized evidence of the bidder's core qualifications, patent achievements, and typical performance is accurately matched and bound to the requirements vector of the scoring rules, and the structured tag-style highlighting is presented in the title, abstract and first paragraph of the response content. The compliance-compliant minimalist generation mode is as follows: complete the basic response according to the scoring requirements, meet the formal compliance requirements, and control the content length to not exceed the preset threshold.

6. The method for writing tender and bid documents based on a generative large model according to claim 5, characterized in that: After generating the response content for high-sensitivity scoring items, a multi-role simulated bidding review process is initiated. The bidding expert AI agent simulates scoring the response content according to the real scoring rules and raises questions, while the compliance review AI agent simultaneously scans for the risk of rejection. If the simulated score is lower than the preset threshold or there is a risk of rejection, the response content is reconstructed from the generation stage until the simulated score reaches or exceeds the preset threshold and there is no risk of rejection, thus completing the verification and forming a closed loop of generation, review, and optimization.

7. The method for writing tender and bid documents based on a generative large model according to claim 6, characterized in that: When integrating the generated content, the differentiated content modules are reorganized into a complete tender document with a coherent structure and unified style, in accordance with the standard format and logical hierarchy of the tender document, and the format specification verification is completed simultaneously.

8. The method for writing tender and bid documents based on a generative large model according to claim 7, characterized in that: Build a bid evaluation knowledge base to store the bid evaluation rules and implicit review preferences of historical projects. It continuously iterates and updates the bid evaluation data of newly won projects, dynamically calibrating the bid sensitivity parameters of each scoring item to adapt to the review characteristics of different industries and different bidders.

9. The method for writing tender and bid documents based on a generative large model according to claim 8, characterized in that: By combining localized model fine-tuning with search-enhanced generation, the bidder's own qualification materials, historical project cases, and unique writing style are injected into the large model generation process, ensuring that the output content has a consistent style, the data is authentic and traceable, and at the same time meets data security control requirements.

10. A bidding and tender document writing system based on a generative large model, characterized in that: The system for executing the bidding document preparation method as described in any one of claims 1 to 9 includes: The tender document parsing module is used to parse the target tender document and extract the hierarchical scoring structure and final-level scoring details. The sensitivity measurement module for scoring items is used to calculate the sensitivity parameters of each scoring item based on historical bidding and evaluation data using statistical analysis methods. The competitiveness assessment module is used to compare the bidder's qualification materials with the scoring rules to obtain the competitiveness fit value for each scoring item. The dynamic matching module for generating strategies combines the sensitivity parameter of winning the bid with the competitiveness fit value to divide the scoring items into different categories, and matches the corresponding generation depth, argumentation strength and length configuration for each category; The differentiation generation module is used to generate the response content for each scoring item one by one according to the matched generation configuration; The multi-agent review and verification module is used to conduct multi-role simulated evaluation and review of the generated content of high-sensitivity scoring items, and trigger content reconstruction based on the review results to achieve closed-loop verification of generation-review-optimization. The content integration and output module is used to integrate the responses to each scoring item into a complete tender document with a unified style and coherent structure, and then output it. The bid evaluation knowledge base iterative update module is used to store historical bid evaluation rules and implicit preferences, and is continuously updated with new winning bid data to dynamically calibrate the bid sensitivity parameters.

Images