A rag-based cross-organizational project bidding team intelligent matching method

Abstract

This invention relates to a method for intelligent matching of cross-organizational project bidding teams based on RAG (Rich Artificial Intelligence). The method includes the following steps: collecting hard and soft indicators of personnel; using a large language model to deconstruct the mandatory constraints and preference conditions in the bid documents; employing progressive hard screening, dynamic weighted soft matching, cross-validation, and combinatorial optimization; and automatically outputting the optimal team configuration and estimated score. This RAG-based intelligent matching method for cross-organizational project bidding teams ensures team qualification compliance through progressive hard screening; uses vector semantic matching to replace keyword retrieval for accurate understanding of project experience; dynamically weights adapt to bid document preferences; comprehensively ranks teams for scientific selection; introduces cross-validation to detect job conflicts and personnel duplication; combinatorial optimization selects the highest-scoring team configuration globally; and simulates review to output estimated scores. This reduces team formation time from hours to minutes, significantly increasing the probability of winning bids.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, natural language processing and enterprise resource planning, in particular to a cross-organization project bidding team intelligent matching method based on RAG. BACKGROUND

[0002] In the daily business activities of large and medium-sized enterprises such as provincial group companies and their affiliated multiple branch companies, participating in external project bidding is a core business. In order to stand out in fierce market competition, enterprises often need to form a team that can fully meet the stringent requirements in the bidding documents. These requirements usually specifically cover specific enterprise qualifications, personnel title levels, special certification certificates, and specific field historical project experience, etc., and the pros and cons of team configuration directly determine the success rate.

[0003] At present, in the prior art, the screening of personnel participating in bidding usually relies on the human resource tag system independently maintained by each branch company or manual review of resumes. The specific method is to pre-index the structured fields such as enterprise qualifications, personnel titles, and certificate names with keywords, and when bidding, business personnel will accurately match and search in the internal system according to the hard requirements in the bidding documents, and then manually form a bidding team after summarizing the personnel list that meets the conditions. In actual use, employee qualifications, certificates, and project experience are scattered in different branch company heterogeneous systems or unstructured documents, and cross-company search for compound talents relies on manual inquiry, which is easy to miss the best candidates, thus forming a data island and causing cross-organization scheduling difficulties. Moreover, the deconstruction of bidding document requirements completely relies on manual reading, and the complex scoring standards often implied in the bidding documents of hundreds of pages are easy to be missed by manual review, thus causing disqualification or loss of points, and the traditional database retrieval ability is limited, and the rule-based keyword accurate matching cannot understand the semantic connotation of unstructured project experience, resulting in serious lack of matching accuracy. Therefore, a cross-organization project bidding team intelligent matching method based on RAG is proposed to solve the above problems. SUMMARY

[0004] In view of the defects of the prior art, the present application provides a RAG-based cross-organization project bidding team intelligent matching method, which has the advantages of cross-organization intelligent matching, solves the problem that the employee qualifications, certificates and project experience are scattered in different heterogeneous systems or unstructured documents of different sub-companies, and the search for compound talents across companies relies on manual inquiry, which is easy to miss the best candidates, thereby forming a data island and causing difficulties in cross-organization scheduling, and the requirement for the bid document is completely dependent on manual reading, and the long bidding document often contains complex scoring standards, and manual review is easy to miss the key constraints, thereby causing bid rejection or score loss, and the traditional database retrieval capability is limited, the rule-based keyword accurate matching cannot understand the semantic connotation of unstructured project experience, resulting in serious lack of matching accuracy.

[0005] To achieve the above object, the present application provides the following technical scheme: a RAG-based cross-organization project bidding team intelligent matching method, comprising the following steps: S1. Collect personnel data and personnel data of each individual, and the hard indicators of each individual at least include the first sub-indicator , the second sub-indicator , and the third sub-indicator ; S2. Extract the mandatory constraints in the bidding document based on the large language model , and the mandatory constraints at least include the first sub-constraint , the second sub-constraint , and the third sub-constraint ; S3. Screen out individuals whose hard indicators in the personnel data meet the mandatory constraints , and construct a candidate personnel set ; S4. Select the first N individuals in the candidate personnel set in the preset order, perform simulated review based on the large language model and the original scoring standard of the bidding document, and output the proposed bidding team personnel configuration matrix M and the estimated score .

[0006] Further, the S1 further collects the soft indicators of each individual in the personnel data ; wherein the soft indicators at least include the first sub-soft indicator , the second sub-soft indicator , and the third sub-soft indicator ; S2 further introduces preference conditions obtained from the tender documents based on a large language model. ; Among them, preference conditions At least includes the first child preference condition. Second-sub preference condition Third-sub preference condition ; The S3 constructs a candidate pool. Then, preference conditions are further introduced. Screening mechanism; Selected candidate list Medium soft indicators Meeting the preference conditions Individuals, and build a set of people with project experience matching. ; Before selecting the first N individuals in S4, a preference-based condition is further introduced. Matching personnel with project experience Soft indicators for all individuals The similarity is sorted, and then the top N individuals are selected.

[0007] Furthermore, the mandatory constraints in S4 and preference conditions It is a serial filtering mechanism.

[0008] Furthermore, the mandatory constraints The filtering is progressive; Judgment Formula 1: in, For mandatory constraints The first sub-constraint in, As a hard indicator The first sub-index in; Judgment Formula Two: in, For mandatory constraints The second sub-constraint in As a hard indicator The second sub-index; Judgment Form 3: in, For mandatory constraints The third sub-constraint in As a hard indicator The third sub-index; Finally, a candidate pool is constructed based on all individuals who passed Formula Three and retained their eligibility. .

[0009] Furthermore, the preference conditions Filtering is now dynamic. Set preference conditions First sub-preference condition Second-sub preference condition Third-sub preference condition The initial importance factors are respectively , , The corresponding dynamic weight is , , ; ; in, , , To dynamically adjust factors, they can be based on project matching degree. Calculation, i.e., the current project and the candidate set. The first sub-soft index for each individual Second soft indicator Third sub-soft indicators Similarity; ; in, The embedding vector for the current project is extracted from the tender documents by the large language model. , , The first sub-soft indicator for candidates Second soft indicator Third sub-soft indicators The corresponding set of embedded vectors for historical records. The vector magnitude; Similarity is a comprehensive score of each candidate's preferences, i.e. ; in, , , These are the first sub-soft indicators. Second soft indicator Third sub-soft indicators The quantitative score is calculated using the same formula for the maximum cosine similarity.

[0010] Further, the similarity in step S4 is calculated using the preference comprehensive score, and the specific formula is: ; Wherein, , , The matching scores of the first candidate personnel in the first sub-soft index , the second sub-soft index , and the third sub-soft index , respectively, are calculated using the maximum cosine similarity, and all individuals in the project experience matching personnel set are sorted in descending order according to the value.

[0011] Further, cross-validation and combination optimization are further introduced when the simulation review in S4 is performed; The first N individuals selected are constructed into an initial team personnel configuration matrix according to the post and corresponding job level required by the bidding document. The original scoring standard in the bidding document and the personnel configuration matrix are jointly constructed into a prompt template, which is input into a large language model, and the large language model performs cross-validation to check whether each post personnel in the personnel configuration matrix satisfies the corresponding mandatory constraints and preference conditions , and detects the qualification conflict and personnel repeated count between posts; On the basis of passing the cross-validation, the large language model further performs combination optimization to try different personnel and post mapping schemes in the candidate personnel set, and selects the combination with the highest estimated total score as the final team personnel configuration matrix M; The large language model outputs the scores of each scoring item and the total score as the estimated score; Further, the same prompt template is sampled multiple times to take the average, or a few sample calibration is combined with historical evaluation cases; The calculation formula of the estimated score is: ; Wherein, is the original scoring standard text in the bidding document, is the large language model inference function.

[0012] Compared with the prior art, the present application provides a cross-organization project bidding team intelligent matching method based on RAG, which has the following beneficial effects: ​1. This RAG-based intelligent matching method for cross-organizational project bidding teams achieves automatic global retrieval and rapid screening of compliant personnel by uniformly collecting hard indicators of personnel from the group and its subsidiaries and binding them to a unified personnel data pool. This solves the problem of missing the best candidates due to manual inquiries caused by data silos. 2. This RAG-based intelligent matching method for cross-organizational project bidding teams automatically extracts mandatory constraints from bids through a large language model and uses a progressive judgment-based screening method to achieve automated hierarchical verification of hard indicators and precise location of elimination items, solving the problem of missing key constraints during manual reading that leads to bid rejection or loss of points. 3. This RAG-based intelligent matching method for cross-organizational project bidding teams collects soft indicators and calculates a comprehensive preference score based on vector embedding and dynamic weights. It achieves semantic matching and adaptive ranking of unstructured project experience, solving the problems of traditional keyword retrieval being unable to understand semantic connotations and having low matching accuracy. 4. This RAG-based intelligent matching method for cross-organizational project bidding teams cross-validates and optimizes the initial team configuration through a large language model, thereby generating the globally optimal team configuration and stably outputting the estimated score. This solves the problems that local optimization cannot guarantee the highest overall team score and lacks simulated review. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the structure of a cross-organizational project bidding team intelligent matching method based on RAG proposed in this invention. Detailed Implementation

[0014] 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.

[0015] Example 1: Please refer to Figure 1 This embodiment of a cross-organizational project bidding team intelligent matching method based on RAG includes the following steps: S1. Collect personnel data and personnel data Hard indicators for each individual And hard indicators It must include at least the first sub-indicator. Second sub-indicator Third sub-indicator ; S2. Extracting mandatory constraints from bidding documents based on a large language model. And mandatory constraints At least includes a first child constraint Second sub-constraint Third child constraint ; S3. Filter out personnel data Medium hard indicators To achieve mandatory constraints Individuals, and construct a candidate pool. ; S4. Select candidate sets according to preset order. The first N individuals are used to perform simulated reviews based on the large language model and the original scoring criteria in the tender documents, outputting the proposed bidding team personnel configuration matrix M and the estimated score. .

[0016] It should be noted that the hard indicators At a minimum, it includes the enterprise qualifications or business type of the individual's employer, the individual's position, and the individual's job level, each corresponding to the first sub-indicator. Second sub-indicator Third sub-indicator ; Mandatory constraints It should at least include the project type, job requirements, and job level requirements specified in the tender documents, each corresponding to the first sub-constraint. Second sub-constraint Third child constraint ; When an individual's company qualifications or business type match the project type required in the tender document, the individual's position in the company matches the position requirements in the tender document, and the individual's job level in the company matches the job level requirements in the tender document, it indicates that the individual's hard indicators are met. Complies with the mandatory constraints in the tender documents ; By uniformly collecting basic qualifications, job positions, and job levels of personnel from the group and all its subsidiaries, and binding all individuals to a unified personnel database, we can achieve this goal. This ensures that the screening meets the mandatory constraints. At the same time, it can automatically perform a global search across organizational boundaries, eliminating the need for manual inquiries to each company, effectively avoiding the omission of the best candidates; Automatically extract mandatory constraints such as project type, job requirements, and job level requirements from bidding documents using large language models. The cumbersome process of relying on manual page-by-page reading is converted into machine-readable structured constraints, significantly reducing the risk of waste and loss of points due to manual omission or misreading of key clauses; Utilizing hard constraint screening, ensuring that all selected candidates meet the mandatory requirements of the tender on hard indicators Lay the foundation for subsequent formation of a compliant bidding team; The system automatically completes the rapid screening of compliant candidates from a large number of personnel, sorts them in the pre-set order, and combines the large language model simulation evaluation to output the proposed bidding team configuration matrix M and the estimated score, compressing the traditional manual screening work of several hours or even several days to minutes.

[0017] Among them, the S1 further collects personnel data Soft indicators of each individual ; Among them, the soft indicators At least include the first sub-soft indicators , the second sub-soft indicators , and the third sub-soft indicators ; The S2 further introduces the preference conditions extracted from the tender document based on the large language model ; Among them, the preference conditions At least include the first sub-preference conditions , the second sub-preference conditions , and the third sub-preference conditions ; After the S3 constructs the candidate personnel set , further introduce the preference screening mechanism ; Screen out individuals in the candidate personnel set whose soft indicators meet the preference conditions , and construct the project experience matching personnel set ; Before selecting the top N individuals in the S4, further introduce the similarity sorting of the soft indicators of all individuals in the project experience matching personnel set based on the preference conditions , and then select the top N individuals.

[0018] It should be noted that the soft indicators At least include the individual's historical performance, historical projects, and qualifications, which correspond to the first sub-preference conditions , the second sub-preference conditions , and the third sub-preference conditions ; Preference conditions This includes at least the individual's past performance, project experience corresponding to the tender documents, and the individual's years of service, each corresponding to the first sub-preference condition. Second-sub preference condition Third-sub preference condition ; By collecting soft indicators such as an individual's historical performance, past projects, and qualifications. And combine large language models to extract preference conditions from the bidding documents. The system no longer relies on simple keyword matching, but is based on vector embedding and cosine similarity calculation. It can understand the semantic equivalence between "led the construction of a public resource trading platform in a certain city" and "experience in government information platform", which significantly improves the matching accuracy.

[0019] In addition, the mandatory constraints in S4 and preference conditions It is a serial filtering mechanism.

[0020] It should be noted that mandatory constraints are used first. Filter out those that meet the hard criteria A collection of candidates Furthermore, based on this, and on the basis of preference conditions Filter out soft indicators Qualified personnel, project experience matching personnel set This sequential mechanism ensures that the team is absolutely compliant with the hard requirements such as qualifications, positions, and job levels, and further selects candidates who best fit the tender document's preferences in terms of historical experience, thus achieving a reasonable screening logic of "first ensuring a minimum, then selecting the best".

[0021] And, the mandatory constraints The filtering is progressive; Judgment Formula 1: in, For mandatory constraints The first sub-constraint in, As a hard indicator The first sub-index in; Judgment Formula Two: in, For mandatory constraints The second sub-constraint in As a hard indicator The second sub-index; Judgment Form 3: in, For mandatory constraints The third sub-constraint in As a hard indicator The third sub-index; Finally, a candidate pool is constructed based on all individuals who passed Formula Three and retained their eligibility. .

[0022] It should be noted that by using judgment formula one, judgment formula two, and judgment formula three to verify the project type, qualifications, position, and job level layer by layer, each layer only processes individuals who have passed the previous layer, which greatly reduces the amount of calculation in subsequent steps and is especially suitable for the rapid filtering of massive personnel data at the group level. The three judgments are independent of each other and have a clear order. When an individual is eliminated, the specific constraint that is not met can be accurately located, which makes it easier for business personnel to understand the screening results and conduct manual review. Progressive screening ensures that candidates are ultimately included in the candidate pool. Each individual fully meets the mandatory requirements of the bidding documents in terms of project type, job title, and job level, thus eliminating the possibility of bid rejection due to unqualified personnel. The judgment structure can be flexibly expanded according to the actual requirements of the tender documents, such as adding a fourth judgment "certificate requirements", without changing the overall framework, and has good versatility and maintainability.

[0023] Wherein, the preference condition Filtering is now dynamic. Set preference conditions First sub-preference condition Second-sub preference condition Third-sub preference condition The initial importance factors are respectively , , The corresponding dynamic weight is , , ; ; in, , , To dynamically adjust factors, they can be based on project matching degree. Calculation, i.e., the current project and the candidate set. The first sub-soft index for each individual Second soft indicator Third sub-soft indicatorsSimilarity of ; ; , is the embedding vector of the current project, extracted by the large language model from the bidding document, , , are the first, second and third sub-soft indicators of the candidate personnel respectively , , is the embedding vector set corresponding to the historical record, is the vector length; Similarity is the comprehensive score of the preference of each candidate, that is, ; , , , are the quantitative scores of the first, second and third sub-soft indicators respectively , , , which are calculated by using the same cosine similarity maximum formula.

[0024] It should be noted that different bidding documents have different emphasis on historical performance, historical projects and experience. Through the joint calculation of the initial importance factor , , and the dynamic adjustment factor , , , the system can automatically adjust the weight of each dimension according to the actual matching of the current project and the historical data of the candidate personnel, so that the comprehensive score S is more consistent with the real preference of the bidding document, and the scoring distortion caused by fixed weight is avoided; The cosine similarity is used to calculate the maximum similarity between the embedding vector of the current project and the embedding vector of the historical record of the candidate personnel, which can capture the semantic correlation between "smart city platform construction" and "city brain project", rather than relying on literal matching only, solving the problem that the traditional label system cannot understand the deep meaning of project experience, and significantly improving the accuracy of matching; By weightedly fusing the matching scores of the three dimensions of historical performance, historical project and experience, the one-sidedness of single dimension evaluation is avoided. Since the weight and similarity are both based on quantifiable vector operation, the system can clearly show the score of each candidate in each dimension and the basis of weight allocation, providing transparent recommendation reasons for business personnel and enhancing the credibility and controllability of team building.

[0025] Embodiment two: please refer toFigure 1 On the basis of embodiment one, a RAG-based cross-organizational project bidding team intelligent matching method, the similarity in step S4 is calculated by using the preference comprehensive score, and the specific formula is: ; Among them, , , are the matching scores of the first candidate , the second sub-soft index , the third sub-soft index , and the third sub-soft index , respectively. The maximum cosine similarity is used to sort all individuals in the project experience matching personnel set in descending order of value.

[0026] It should be noted that by descending order sorting all individuals in the project experience matching personnel set by using the preference comprehensive score, the matching results of multi-dimensional soft indexes are aggregated into a single comparable numerical value, so that the subjective judgment of "who is more suitable for the bid preference" is converted into an objective and quantifiable sorting result, which facilitates the system to quickly select the top N optimal individuals in the preset order; When calculating the matching score of each candidate on each soft index, the maximum cosine similarity is used to ensure that the system can capture the most relevant and most persuasive past experience of the candidate, and avoid underestimating excellent talents due to poor matching of a single record; All candidates use the same dynamic weight , , and the same similarity calculation method, which ensures the fairness and comparability of cross-individual scores, and the top N individuals selected on this basis can represent the overall optimal talent pool in soft indexes, providing high-quality input for subsequent simulation review and team combination optimization; Solves the problem that traditional manual screening often relies on subjective impression or fragmented resume comparison, which is prone to bias and time-consuming.

[0027] Finally, the simulation review in S4 further introduces cross-validation and combination optimization; The top N individuals are selected according to the positions and corresponding job levels required by the bidding document, and a personnel configuration matrix of the initial team is constructed; The original scoring standard in the bidding document and the personnel configuration matrix are combined to construct a prompt template, which is input into a large language model, and the large language model performs cross-validation to check whether each position in the personnel configuration matrix satisfies the corresponding mandatory constraints and preference conditions and detect conflicts in qualifications and repeated personnel counts between positions; On the basis of cross-validation, the large language model further performs combination optimization, trying different personnel and position mapping schemes in the candidate personnel set, and selects the combination with the highest estimated total score as the final team personnel configuration matrix M; The large language model outputs the scores of each scoring item and the total score as the estimated score; Further, multiple samplings of the same prompt template are taken to obtain an average, or a few-sample calibration is combined with historical evaluation cases; The calculation formula of the estimated score is: ; Among them, is the original scoring standard text in the tender document, is the inference function of the large language model.

[0028] It should be noted that the initial team configuration matrix is cross-validated by the large language model, which verifies whether each position personnel meets the corresponding mandatory constraints and preference conditions and automatically detects whether the same person is assigned to multiple positions or there is a conflict in qualifications between different positions. This mechanism can identify and exclude non-compliant configurations in advance during the simulation evaluation stage, effectively avoiding the risk of bid rejection due to team configuration problems; On the basis of cross-validation, the large language model further performs combination optimization, trying different personnel and position mapping schemes in the candidate personnel set, and selects the combination with the highest estimated total score as the final team personnel configuration matrix M, which is not limited to the local optimal strategy of "selecting the best for a single position". Instead, it selects the combination with the highest estimated total score as the final team personnel configuration matrix M from a global perspective. This global optimization capability can uncover talent pairing schemes that synergize across positions, such as assigning a senior expert to a key position with higher scoring weights, thereby significantly improving the overall bid score; Using the semantic understanding and logical reasoning capabilities of the large language model, the natural language descriptions in the original scoring standards can be automatically parsed, and the estimated total scores of different configurations can be dynamically calculated during the combination optimization process. This achieves highly intelligent team assembly. By taking multiple samplings of the same prompt template and obtaining an average, the scoring fluctuations caused by the randomness of the large language model output can be reduced. By combining historical evaluation cases for few-sample calibration, the model's scoring scale can be aligned with that of real evaluation experts, further improving the accuracy and credibility of the estimated score and providing more valuable references for bidding decisions; From the input of the tender to the final output of the tender team personnel configuration matrix M and the estimated score, no manual intervention is required throughout, greatly shortening the formation period of the tender team. At the same time, since the output result contains the scores of each scoring item and the total score, business personnel can directly use it as the basis for the "team configuration explanation" and "technical score response" parts in the tender document, realizing the automatic generation of the tender materials.

[0029] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element preceded by "comprises... a" does not, without more limitations, foreclose the existence of additional identical elements in the process, method, article, or apparatus that comprises the recited element.

[0030] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, alternatives, and variations can be made in the embodiments without departing from the spirit and scope of the present application as defined by the appended claims and their equivalents.

Claims

1. A RAG-based cross-organizational project bidding team intelligent matching method, characterized in that, The method comprises the following steps: S1. Collecting personnel data and personnel data Hard indicators of each individual , and the hard indicators at least include a first sub-indicator , a second sub-indicator , a third sub-indicator ; S2. Extracting mandatory constraints in the bidding document based on a large language model , and the mandatory constraints at least include a first sub-constraint , a second sub-constraint , and a third sub-constraint ; S3. Screening out personnel data Medium-hardness indicators Reaching mandatory constraints of individuals, and building a candidate personnel collection ; S4. Select candidate sets according to preset order. The first N individuals are used to perform simulated reviews based on the large language model and the original scoring criteria in the tender documents, outputting the proposed bidding team personnel configuration matrix M and the estimated score. .

2. The RAG-based cross-organizational project bid team intelligent matching method according to claim 1, characterized in that: The S1 further collects personnel data Soft indicators for each individual ; Wherein, the soft index At least comprising a first sub-soft index , a second sub-soft index , a third sub-soft index ; The S2 further introduces a preference condition based on a large language model extracting the bidding document ; wherein the preference condition at least comprises a first sub-preference condition , a second sub-preference condition , a third sub-preference condition ; The S3 constructs candidate pool Further, preference conditions are introduced Screening mechanism; Screening out candidate pools Soft metrics Reaching preference conditions Individuals and building project experience matching pools ; Further introduce preference-based condition before selecting top N individuals in S4 and project experience matching personnel set Soft index of all individuals Sort the similarity, and then select top N individuals.

3. The RAG-based cross-organizational project bid team intelligent matching method according to claim 2, characterized in that: The mandatory constraints in S4 and the preference conditions are serial screening mechanisms.

4. The RAG-based cross-organizational project bid team intelligent matching method according to claim 3, characterized in that: The mandatory constraints The screening is a progressive screening; Formula 1: wherein, is a mandatory constraint is a first sub-constraint in the mandatory constraint, is a hard indicator is a first sub-indicator in the hard indicator; Formula 2: wherein, is a mandatory constraint is a second sub-constraint in the mandatory constraint, is a hard indicator is a second sub-indicator in the hard indicator; Formula 3: wherein, is a mandatory constraint is a third sub-constraint in the constraint, is a hard indicator is a third sub-indicator in the indicator; Finally, a candidate pool is constructed based on all individuals that passed criterion three and retained eligibility .

5. The RAG-based cross-organizational project bid team intelligent matching method according to claim 4, characterized in that: Said preference condition The screening is dynamic screening; Set preference condition First sub-preference condition in the set preference condition Second sub-preference condition Third sub-preference condition The initial importance factors of the first, second and third sub-preference conditions are respectively , , The corresponding dynamic weights are , , ; ； wherein, , , is a dynamic adjustment factor, which can be calculated based on the project matching degree , i.e. the similarity of the first sub-soft index , the second sub-soft index , and the third sub-soft index of each individual in the current project and the candidate pool ; ； wherein, is the embedding vector of the current project, extracted from the tender document by the large language model, , , are the first, second and third sub-soft indicators of the candidate personnel, respectively , , is the embedding vector set corresponding to the historical record, is the vector length. The similarity is the comprehensive score of the preference of each candidate, that is, ； Wherein, , , are the quantitative scores of the first sub-soft index , the second sub-soft index , and the third sub-soft index , respectively, and the same cosine similarity maximum formula is used for calculation.

6. The RAG-based cross-organizational project bid team intelligent matching method according to claim 5, characterized in that: The similarity in the step S4 is calculated by using the comprehensive score of the preference, and a specific formula is as follows: ； in, , , The first The candidates in the first sub-soft indicator Second soft indicator Third sub-soft indicators The matching score is calculated using the maximum cosine similarity, and is applied to all individuals in the set of people with project experience. Sort by value in descending order.

7. The RAG-based cross-organizational project bid team intelligent matching method according to claim 2, characterized in that: In the step S4, cross validation and combination optimization are further introduced in the simulation review; The first N selected individuals are arranged according to the post and corresponding job level required by the bidding document to construct an initial team personnel configuration matrix ; The original scoring criteria in the bidding documents are combined with the staffing matrix The joint configuration prompt template is input into the large language model, cross-validation is performed by the large language model, and the staffing matrix is checked one by one Whether the personnel at each post meet the corresponding mandatory constraints And the preference conditions And detect the qualification conflict between posts and the repeated counting of personnel; On the basis of passing the cross validation, the large language model further performs combination optimization, tries different personnel and post mapping schemes in the candidate personnel set, and selects a combination with the highest estimated total score as a final team personnel configuration matrix M; The large language model outputs scores of each scoring item and a total score as an estimated score; Further, the same prompt template is sampled multiple times to take an average, or a few sample calibration is combined with historical evaluation cases; A calculation formula of the estimated score is as follows: ； wherein, is the original scoring criteria text in the tender document, is the large language model inference function.

Images