An intelligent source seeking decision method based on a large language model

Abstract

The application discloses an intelligent source seeking decision method based on a large language model, and belongs to the technical field of data processing. The time and space dislocation problem of dynamic data is eliminated through an augmented graph data alignment method based on information entropy attenuation, ensuring the real-time performance and high fidelity of the data. Meanwhile, the large language model is used to realize deep semantic analysis of non-standard demand, and the adaptive navigation capability of reinforcement learning is combined to realize dynamic evolution and automatic optimization of the evaluation strategy according to the characteristics of the business scene. Finally, the visual technology of supply and demand form matching is used to provide intelligent decision support with precision, adaptability and high transparency, so as to significantly improve the source seeking efficiency and decision quality of complex service procurement and effectively reduce the performance risk.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to an intelligent source-finding decision-making method based on a large language model. Background Technology

[0002] In today's wave of digital economy and modernization of enterprise supply chain management, especially for large state-owned engineering design institutes and high-end equipment manufacturing enterprises, the focus of procurement management is gradually shifting from traditional compliance control to supply chain value creation. Unlike standardized material procurement, engineering service procurement is characterized by non-standardized demand descriptions, dynamic service processes, and multi-dimensional evaluation indicators. In traditional procurement models, supply and demand matching mainly relies on manual experience or simple information forms. Procurement personnel need to manually screen suitable suppliers from a massive supplier database, which is not only time-consuming and labor-intensive but also highly susceptible to subjective factors, making it difficult to achieve efficient and accurate sourcing decisions in large-scale concurrent business.

[0003] Therefore, existing technologies rely on traditional natural language processing methods, extracting keywords from procurement announcements and performing inverted index matching in supplier databases. However, this approach is limited by its inability to understand the deeper business intent and logical constraints behind natural language. For example, when procurement requirements are expressed as ambiguous text implying specific technical parameters or service scenarios, traditional systems often lack the ability to reason about contextual semantics, making it difficult to identify qualified but not fully matched high-quality potential suppliers, resulting in a high false negative rate. Even methods that partially incorporate pre-trained language models are mostly limited to static calculations of text similarity, lacking the ability to analyze complex business logic. Summary of the Invention

[0004] The purpose of this application is to provide an intelligent source-finding decision-making method based on a large language model, which solves the problems of high false negative rate and lack of ability to parse complex business logic in existing technologies.

[0005] This application is achieved through the following technical solution:

[0006] An intelligent source-finding decision-making method based on a large language model includes:

[0007] Obtain multi-dimensional supplier data and preprocess the multi-dimensional supplier data to obtain preprocessed multi-dimensional supplier data;

[0008] In response to the procurement requirements input by the procurement personnel, a pre-trained large language model is invoked to parse the procurement requirements, obtain a structured requirement feature tree, and perform an initial screening of all suppliers based on the requirement feature tree to obtain the first candidate supplier.

[0009] Based on the demand feature tree and the preprocessed multi-dimensional supplier data corresponding to the first candidate supplier, the business intent fit is obtained;

[0010] A reinforcement learning algorithm is used to analyze the feature vectors corresponding to the demand feature tree, determine the weighting parameters, and use the weighting parameters to weight the preprocessed multi-dimensional data of the first candidate supplier and the business intent fit, thereby generating a comprehensive score.

[0011] The first candidate supplier is screened a second time based on the comprehensive score to obtain the second candidate supplier. The second candidate supplier is then reviewed and ranked using a large language model to obtain the ranked second candidate supplier.

[0012] The comprehensive score data of the second candidate supplier after sorting is converted into radar data charts, and the radar data charts are displayed in sorting order to realize intelligent source sourcing decision-making based on a large language model.

[0013] In one possible implementation, multi-dimensional supplier data is acquired and preprocessed to obtain preprocessed multi-dimensional supplier data, including:

[0014] Obtain the static data corresponding to the supplier; wherein, the static data includes basic information, qualification certification, historical performance and user reviews;

[0015] The dynamic data corresponding to the supplier is obtained by using an augmented graph data alignment method based on information entropy decay; wherein, the dynamic data includes an activity index calculated from the login frequency and browsing depth of the most recent days, a responsiveness index calculated from the average time difference between receiving a notification and responding, and a cooperation index based on the average response time and word count richness in instant messaging tools; the dynamic data and static data together constitute the supplier's multi-dimensional data;

[0016] The supplier multi-dimensional data is preprocessed to obtain preprocessed supplier multi-dimensional data.

[0017] In one possible implementation, the supplier multi-dimensional data is preprocessed to obtain preprocessed supplier multi-dimensional data, including:

[0018] The static data in the supplier's multi-dimensional data is normalized to obtain the normalized static data.

[0019] The normalized static data and the original dynamic data are used together as preprocessed supplier multi-dimensional data.

[0020] In one possible implementation, in response to the procurement requirements input by the procurement personnel, a pre-trained large language model is invoked to parse the procurement requirements and obtain a structured requirement feature tree, including:

[0021] In response to the procurement requirements input by the procurement personnel, a pre-trained large language model is invoked as the core of parsing. A preset template is used to guide the large language model to perform entity extraction and prompt analysis, and the procurement requirements are converted into a structured requirement feature tree.

[0022] In one possible implementation, all suppliers are initially screened based on the demand feature tree to obtain a first candidate supplier, including:

[0023] Based on the preset hard indicator types, the target hard indicators in the demand feature tree are determined;

[0024] Based on multi-dimensional data from suppliers, suppliers that do not meet the target hard indicators are eliminated, resulting in the first candidate supplier.

[0025] In one possible implementation, the business intent fit is obtained based on the demand feature tree and the preprocessed multi-dimensional supplier data corresponding to the first candidate supplier, including:

[0026] Determine the target data in the multi-dimensional data of the first candidate supplier that corresponds to the features in the demand feature tree;

[0027] The cosine similarity between the feature vector formed by the features in the demand feature tree and the feature vector corresponding to the target data is obtained to obtain the business intent fit.

[0028] In one possible implementation, a reinforcement learning algorithm is used to analyze the feature vectors corresponding to the demand feature tree to determine the weighting parameters, including:

[0029] The feature vector formed by the features in the demand feature tree is used as the state space of the reinforcement learning algorithm. The reinforcement learning algorithm is used to select action vectors to obtain weighted weight parameters. The action vectors are selected in the action space, which is a multi-dimensional solution space composed of the value ranges of multiple weighted weight parameters. The process of selecting action vectors is constrained to the point that the sum of each weighted weight parameter is 1.

[0030] In one possible implementation, the preprocessed multi-dimensional data of the first candidate supplier and its business intent fit are weighted using the weighted weighting parameters to generate a comprehensive score:

[0031] S_total=W_sem×S_sem+W_cred×S_cred+W_perf×S_perf+W_eval×S_eval+W_dyn×S_dyn;

[0032] Wherein, S_total is the overall score, S_sem is the business intent fit, W_sem is the first weighted weight parameter, S_cred is the qualification certification data in the normalized static data, W_cred is the second weighted weight parameter, S_perf is the historical performance data in the normalized static data, W_perf is the third weighted weight parameter, S_eval is the user evaluation score data in the normalized static data, W_eval is the fourth weighted weight parameter, S_dyn is the dynamic data in the normalized static data, and W_dyn is the fifth weighted weight parameter.

[0033] In one possible implementation, the first candidate supplier is screened a second time based on the comprehensive score to obtain a second candidate supplier, and a large language model is used to review and rank the second candidate supplier to obtain the ranked second candidate supplier, including:

[0034] The first candidate supplier is selected through Top-K screening based on the comprehensive score to obtain the second candidate supplier; where K is a preset value.

[0035] A large language model is used to compare specific constraints in the demand feature tree with the unstructured descriptions corresponding to suppliers, generating a logical verification score; wherein, the specific constraints refer to unstructured features.

[0036] The second candidate suppliers are arranged in descending order of their logical verification scores to obtain the sorted second candidate suppliers.

[0037] In one possible implementation, the indicator data corresponding to the comprehensive score of the ranked second candidate supplier are converted into radar data charts, and the radar data charts are displayed in sorted order, including:

[0038] The data of various indicators corresponding to the comprehensive score of the second candidate supplier after the ranking are obtained as target business intent fit, target qualification certification data, target historical performance data, target user evaluation score data, and target dynamic data.

[0039] The target business intent alignment, target qualification certification data, target historical performance data, target user evaluation score data, and target dynamic data are constructed into a radar data chart containing five axes, and the radar data chart is displayed in sorted order.

[0040] Compared with the prior art, this application has the following advantages and beneficial effects:

[0041] This application discloses an intelligent sourcing decision-making method based on a large language model. It eliminates the spatiotemporal misalignment problem of dynamic data through an augmented graph data alignment method based on information entropy decay, ensuring data real-time performance and high fidelity. Simultaneously, it utilizes a large language model to achieve deep semantic analysis of non-standard requirements and combines this with the adaptive navigation capabilities of reinforcement learning to realize the dynamic evolution and automatic optimization of evaluation strategies according to business scenario characteristics. Finally, through visualization technology for matching supply and demand patterns, it provides intelligent decision support that combines accuracy, adaptability, and high transparency, thereby significantly improving the sourcing efficiency and decision quality of complex service procurement and effectively reducing performance risks. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the exemplary embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0043] Figure 1 A flowchart illustrating an intelligent source-finding decision-making method based on a large language model, provided for embodiments of this application;

[0044] Figure 2 This is a schematic diagram of a data alignment topology based on heterogeneous augmented graphs provided in an embodiment of this application;

[0045] Figure 3 A 3D dimension-reduced schematic diagram illustrating the distribution of requirements and suppliers in semantic space for embodiments of this application;

[0046] Figure 4 A comparison of scene adaptive weight dynamic allocation based on RL provided in the embodiments of this application;

[0047] Figure 5 A schematic diagram of the initial cold start search (before RL optimization) provided for an embodiment of this application;

[0048] Figure 6 This is a schematic diagram of RL adaptive navigation search provided in an embodiment of this application;

[0049] Figure 7 A radar chart comparing supplier capability profiles based on RL weights, provided for embodiments of this application. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the embodiments and accompanying drawings. The illustrative embodiments and descriptions of this application are only for explaining this application and are not intended to limit this application.

[0051] like Figure 1 As shown in the embodiments of this application, an intelligent source-finding decision-making method based on a large language model is provided, including:

[0052] S101. Obtain multi-dimensional supplier data and preprocess the multi-dimensional supplier data to obtain preprocessed multi-dimensional supplier data.

[0053] S102. In response to the procurement requirements input by the procurement personnel, the pre-trained large language model is invoked to parse the procurement requirements, obtain a structured requirement feature tree, and perform an initial screening of all suppliers based on the requirement feature tree to obtain the first candidate supplier.

[0054] S103. Based on the demand feature tree and the preprocessed multi-dimensional supplier data corresponding to the first candidate supplier, obtain the business intent fit.

[0055] S104. The reinforcement learning algorithm is used to analyze the feature vector corresponding to the demand feature tree, determine the weighting parameters, and use the weighting parameters to weight the preprocessed multi-dimensional data of the first candidate supplier and the business intent fit, and generate a comprehensive score.

[0056] S105. The first candidate supplier is screened a second time based on the comprehensive score to obtain the second candidate supplier. The second candidate supplier is then reviewed and sorted using a large language model to obtain the sorted second candidate supplier.

[0057] S106. Convert the data of each indicator corresponding to the comprehensive score of the second candidate supplier after sorting into a radar data chart, and display the radar data chart in sorting order to realize intelligent source sourcing decision based on a large language model.

[0058] This application discloses an intelligent sourcing decision-making method based on a large language model. It eliminates the spatiotemporal misalignment problem of dynamic data by using an augmented graph data alignment method based on information entropy decay, ensuring data real-time performance and high fidelity. Simultaneously, it utilizes a large language model to achieve deep semantic analysis of non-standard requirements and combines this with the adaptive navigation capabilities of reinforcement learning to realize the dynamic evolution and automatic optimization of evaluation strategies according to business scenario characteristics. Finally, through visualization technology for matching supply and demand patterns, it provides intelligent decision support that combines accuracy, adaptability, and high transparency, thereby significantly improving the sourcing efficiency and decision quality of complex service procurement and effectively reducing performance risks.

[0059] In one possible implementation, multi-dimensional supplier data is acquired and preprocessed to obtain preprocessed multi-dimensional supplier data, including:

[0060] The system acquires static data corresponding to the supplier, including basic information, qualifications, historical performance, and user reviews. It then uses an augmented graph data alignment method based on information entropy decay to acquire dynamic data corresponding to the supplier. This dynamic data includes activity metrics calculated from login frequency and browsing depth over recent days, responsiveness metrics calculated from the average time difference between receiving a notification and responding, and cooperation metrics based on average response time and word count in instant messaging tools. The dynamic and static data together constitute the supplier's multi-dimensional data.

[0061] The supplier evaluation center can be defined as the core computing node, and the interactive log server and LBS positioning system that generate dynamic data can be defined as edge sensing nodes. Several virtual buffer nodes are inserted into the data transmission path between the sensing nodes and the computing nodes to simulate the time-varying and time-lag processes in network transmission and data cleaning. The transmission process of dynamic data is the process of data flowing through these virtual nodes. During this process, the system uses the information entropy decay function built into the virtual nodes to process the data in real time. That is, according to the lag time step of the data in the virtual node chain, the confidence of the data is exponentially reduced and corrected, so as to realize the timeliness alignment of the data at the same time as the transmission is completed, eliminate the misleading of historical expired behavior data on the current performance intention assessment, and finally output the effective dynamic data after spatiotemporal alignment and confidence correction to the evaluation model.

[0062] For example, the system first retrieves raw data for each dimension from the system backend logs. The activity metric (S_act) is derived by directly counting and normalizing the supplier's system login frequency and page view depth within a recent specific period. The responsiveness metric (S_res) quantifies the supplier's responsiveness to the task by calculating the time difference between receiving the purchase notification and responding. The cooperation metric (S_coop) is derived by comprehensively counting the supplier's average response time and the richness of their response content in instant messaging tools. After obtaining these three basic sub-metrics, the system uses the information entropy decay function of the aforementioned augmented graph to correct the timeliness of each metric, resulting in corrected sub-metric values ​​(denoted as ). _act, _res and (_coop), and finally, a linear weighted fusion formula is used to generate dynamic multi-dimensional supplier data D5. The fusion calculation formula is as follows:

[0063] D5=w_act _act+w_res _res+w_coop _coop;

[0064] Among them, w_act, w_res and w_coop are preset weight coefficients corresponding to activity, responsiveness and cooperation, respectively, and the sum of the three is 1. This is used to transform discrete multi-source dynamic behavior data into a unified standard scalar value D5, where D5 represents dynamic data.

[0065] The supplier multi-dimensional data is preprocessed to obtain preprocessed supplier multi-dimensional data.

[0066] In one possible implementation, the supplier multi-dimensional data is preprocessed to obtain preprocessed supplier multi-dimensional data, including:

[0067] The static data in the supplier's multi-dimensional data is normalized to obtain the normalized static data.

[0068] The normalized static data and the original dynamic data are used together as preprocessed supplier multi-dimensional data.

[0069] For example, a panoramic data perception topology based on "fast and slow dual channels" is constructed. The system defines the supplier evaluation center as the core computing node and constructs a heterogeneous data perception network: by integrating the supplier management system (SRM), contract system, and credit database through ETL tools, a direct connection edge with zero latency is established to obtain low-frequency static data such as basic files and qualification certifications, and by using a large language model to correct and complete unstructured qualification scans, forming a static rigid channel of static performance base (D1–D4); by connecting the LBS (Location-Based Services) positioning service and the system log server through an API (Application Programming Interface) gateway, high-frequency interaction traces such as login timestamps, page dwell time, inquiry response latency, and IM (Instant Messaging) reply frequency are captured. To address network jitter and cleaning delays in such data transmission, directed edges containing time-varying communication latency are established to form a dynamic elastic channel for dynamic intention data flow.

[0070] The system deploys augmented graph data alignment units based on information entropy decay. To address the temporal misalignment problem of dynamic data, the system reconstructs the dynamic elastic channel using augmented graph techniques from graph theory: along the transmission path of dynamic data, several "virtual agents" are inserted according to the maximum time delay limit. These virtual nodes do not generate new data; they map the delayed dynamic data to different components in the augmented state matrix, thus forming a virtual node chain. Unlike the lossless transmission of traditional augmented graphs, this system embeds a decay function into the virtual nodes. Each time data passes through a virtual node (i.e., each unit of time delay), the confidence level of the information it carries automatically undergoes an exponential decay. This ensures that the dynamic willingness index (D5) ultimately entering the evaluation model is an "effective willingness value" corrected for timeliness weighting and physical proximity (LBS), thus constructing an information entropy decay mechanism.

[0071] A "4+1" five-dimensional evaluation index system was constructed. The static data index is based on basic information (D1), qualification certification (D2), historical performance (D3), and user evaluation (D4) as the benchmark for performance capability. The dynamic willingness dimension (D5) combines online activity and offline physical proximity, including activity index calculated based on login frequency and browsing depth in the past 30 days, responsiveness index calculated based on the average time difference from "receiving notification" to "responding", and cooperation index based on average response time and word count in instant messaging tools.

[0072] Dynamic scoring and synthesis based on time decay. Min-Max standardization is applied to dimensions D1-D4, and static normalization is applied to the [0,1] interval. For the dynamic index (D5), implicit time decay processing based on information entropy has been completed in the augmented graph transport layer. The system directly performs linear fusion of the aligned five-dimensional vectors to generate the final supplier activity index.

[0073] For low-frequency static data such as basic files and qualification certifications, zero-latency direct connections are established using ETL tools to form a static performance foundation. For high-frequency dynamic data such as login logs and LBS positioning, directed edges containing time-varying communication delays are established to form a dynamic willingness data stream. This architecture effectively solves the time base misalignment problem caused by traditional single topologies when processing multi-source heterogeneous data by separating data transmission paths of different frequencies at the logical level, providing a structured data foundation for subsequent accurate profiling. An improved augmented graph data processing unit is deployed in the dynamic elastic channel. Several virtual buffer proxy nodes are inserted according to the maximum time delay limit of the transmission link, and an information entropy decay function is innovatively implanted in the virtual nodes, so that the confidence of the data flowing through the nodes automatically decays exponentially with the lag time step. This technology not only achieves strict alignment of asynchronous data on the logical clock, but also automatically downweights outdated behavioral data from the underlying algorithm level, ensuring that the final generated dynamic willingness index can truly reflect the supplier's effective response potential at the current moment.

[0074] In one possible implementation, in response to the procurement requirements input by the procurement personnel, a pre-trained large language model is invoked to parse the procurement requirements and obtain a structured requirement feature tree, including:

[0075] In response to the procurement requirements input by the procurement personnel, a pre-trained large language model is invoked as the core of parsing. A preset template is used to guide the large language model to perform entity extraction and prompt analysis, and the procurement requirements are converted into a structured requirement feature tree.

[0076] Specifically, the procurement requirements are first filled into a preset prompt template, which includes role definitions, extraction rules, and output format constraints. Then, the large language model performs semantic analysis on the text based on an attention mechanism to extract key entities such as service objects, technical parameters, business terms, and qualification requirements. Finally, the extracted entities are mapped according to a predefined hierarchical logic to generate a structured requirement feature tree (such as JSON format data) containing technical feature branches, business feature branches, and qualification feature branches.

[0077] In one possible implementation, all suppliers are initially screened based on the demand feature tree to obtain a first candidate supplier, including:

[0078] Based on the preset hard indicator types, the target hard indicators in the demand feature tree are determined;

[0079] Based on multi-dimensional data from suppliers, suppliers that do not meet the target hard indicators are eliminated, resulting in the first candidate supplier.

[0080] In one possible implementation, the business intent fit is obtained based on the demand feature tree and the preprocessed multi-dimensional supplier data corresponding to the first candidate supplier, including:

[0081] Determine the target data in the multi-dimensional data of the first candidate supplier that corresponds to the features in the demand feature tree;

[0082] The cosine similarity between the feature vector formed by the features in the demand feature tree and the feature vector corresponding to the target data is obtained to obtain the business intent fit.

[0083] An example is intent deconstruction driven by prompt engineering. It receives non-standard, colloquial, or long texts containing complex logic from procurement personnel and processes them as natural language requirement input. It then invokes a pre-trained Large Language Model (LLM) as the parsing core to call the LLM cognitive engine. Specific prompt templates (such as "role setting + task instructions + output constraints") are designed to guide the model in performing entity extraction and logical analysis through prompt engineering. The system precisely identifies extraction dimensions such as "service object entities" (e.g., municipal pipelines), "core technical parameters" (e.g., pressure resistance ratings), "hard qualification thresholds" (e.g., Class A qualifications), and "business preferences" (e.g., schedule priority), transforming the unstructured text into a structured output of a JSON-formatted Requirement Feature Tree.

[0084] Initial screening based on hard logic. Hard indicators are extracted by parsing the necessary constraints from the JSON tree; Boolean logic filtering is then applied to the supplier pool based on hard thresholds (such as "registered capital > 10 million" or "ISO certification"). This directly filters out non-compliant entities, ensuring that subsequent calculations only target the valid candidate pool, significantly reducing unnecessary computation.

[0085] Intent fit calculation based on soft semantics. A vectorization tool (Embedding Model) is used to map the core requirement descriptions and technical parameters after LLM parsing, generating high-dimensional semantic vectors of core fields. The cosine similarity of the spatial distance between the requirement vector and the candidate supplier capability vector is calculated. This calculation result is defined as "Business Intent Fit" as a soft indicator to aid decision-making.

[0086] For example, firstly, feature serialization is performed, serializing the structured requirement feature tree obtained from the previous steps and concatenating it into a standard text description string that conforms to natural language logic. Next, fixed-dimensional mapping is performed, inputting this standard text description string into a pre-trained vectorization tool (Embedding Model, such as BERT or Text-Embedding-Ada). Internally, this model handles variable-length inputs through padding and truncation mechanisms, and uses a pooling layer to uniformly map texts of different lengths to a pre-defined fixed high-dimensional vector space, thus outputting a requirement feature vector V_req with fixed dimensions. Similarly, the capability description text of candidate suppliers is mapped to a capability vector V_sup of the same dimension using the same model, and cosine similarity is calculated based on the two fixed and identical-dimensional vectors.

[0087] By using a pre-trained Large Language Model (LLM) as the cognitive core, and through specific role settings and task instruction Prompt templates, entity extraction and logical analysis are performed on the colloquial and vague requirement text input by procurement personnel. This accurately identifies key dimensions such as service objects, core parameters, hard thresholds, and business preferences, transforming vague business requirements into JSON-formatted state-space feature vectors that can be read by the algorithm model. This solves the industry pain point of difficulty in quantifying requirements in non-standard service procurement.

[0088] In one possible implementation, a reinforcement learning algorithm is used to analyze the feature vectors corresponding to the demand feature tree to determine the weighting parameters, including:

[0089] The feature vector formed by the features in the demand feature tree is used as the state space of the reinforcement learning algorithm. The reinforcement learning algorithm is used to select action vectors to obtain weighted weight parameters. The action vectors are selected in the action space, which is a multi-dimensional solution space composed of the value ranges of multiple weighted weight parameters. The process of selecting action vectors is constrained to the point that the sum of each weighted weight parameter is 1.

[0090] In one possible implementation, the preprocessed multi-dimensional data of the first candidate supplier and its business intent fit are weighted using the weighted weighting parameters to generate a comprehensive score:

[0091] S_total=W_sem×S_sem+W_cred×S_cred+W_perf×S_perf+W_eval×S_eval+W_dyn×S_dyn;

[0092] Wherein, S_total is the overall score, S_sem is the business intent fit, W_sem is the first weighted weight parameter, S_cred is the qualification certification data in the normalized static data, W_cred is the second weighted weight parameter, S_perf is the historical performance data in the normalized static data, W_perf is the third weighted weight parameter, S_eval is the user evaluation score data in the normalized static data, W_eval is the fourth weighted weight parameter, S_dyn is the dynamic data in the normalized static data, and W_dyn is the fifth weighted weight parameter.

[0093] For example, a dual-track fusion calculation model based on "hard power + soft intent" is constructed. The system establishes a linear weighting strategy to fuse the "business intent fit" calculated from the second aspect with the "five-dimensional profile indicators" constructed from the first aspect, generating a comprehensive score S_total for ranking. The calculation formula for S_total is: S_total = W_sem×S_sem+ W_cred×S_cred + W_perf×S_perf + W_eval×S_eval + W_dyn×S_dyn. In the formula, S_sem is the semantic score calculated from the second aspect based on embedding; S_cred, S_perf, and S_eval are the qualification certification, historical performance, and user evaluation scores from the first aspect based on ETL cleaning, respectively; S_dyn is the dynamic intent score from the first aspect based on augmented graph topology alignment; and W_sem to W_dyn represent the weight coefficient vectors of each dimension, which determine the search focus of the evaluation system in the multi-dimensional space.

[0094] Configure a cold start mechanism for search probes based on LLM scene awareness. To address the lack of historical interaction data during system initialization, the system is configured with parameter initialization logic based on a large language model. This logic first reads the JSON requirement feature tree generated in the second part, parses the project type labels, budget range, and urgency features, and then inputs these features into a pre-trained large language model (LLM). Using preset expert prompt word templates, it outputs an initial weight vector W_0, thereby setting the initial search probe direction for different types of procurement tasks based on general expert knowledge.

[0095] Before using reinforcement learning algorithms, a weight navigation self-evolution mechanism based on reinforcement learning (RL) can be deployed. The system incorporates a lightweight contextual bandit algorithm as an adaptive navigator to dynamically optimize weight parameters based on business feedback. This mechanism defines the feature vector of the procurement project as the state space, the direction and magnitude of fine-tuning the current weight vector as the action space, and the initial hit rate of the recommendation list and subsequent fulfillment feedback (i.e., assigning one value after fulfillment and another value for non-fulfillment) as the reward function. When performing recommendation tasks, the system adjusts the weights based on the current state output action and updates the policy network after receiving business feedback. By continuously correcting the metric scale in the multi-dimensional scoring space, the system achieves continuous evolution and automatic optimization of the optimal supplier search path.

[0096] In the initial stage where the system lacks historical interaction data, the general knowledge base of LLM is used to parse project features and output initial weight vectors, achieving "expert-level" policy initialization. Subsequently, a lightweight contextual multi-armed slot machine algorithm is introduced as an adaptive navigator, defining the project feature tree as the state space and the weight fine-tuning as the action space. By capturing business feedback in real time to update the policy network, the evaluation model can continuously evolve online from "general rules" to "optimal solutions for specific scenarios". By executing the weight adjustment actions output by the reinforcement learning agent, the distance metric in the multi-dimensional evaluation space is corrected in real time, transforming the original single search logic based on Euclidean geometric distance into a complex optimization logic based on weighted feature distance. This technology can dynamically distort the evaluation space according to the characteristics of the scenario, shifting the center of gravity of the search path from a single semantic geometric center to a feature-rich region with stronger comprehensive practical capabilities. This automatically filters out risky suppliers with "high semantic matching degree but lack of qualification and performance support" from the massive candidate pool, achieving safe selection.

[0097] In one possible implementation, the first candidate supplier is screened a second time based on the comprehensive score to obtain a second candidate supplier, and a large language model is used to review and rank the second candidate supplier to obtain the ranked second candidate supplier, including:

[0098] The first candidate supplier is selected through Top-K screening based on the comprehensive score to obtain the second candidate supplier; where K is a preset value.

[0099] It is worth noting that if the number of first-candidate suppliers is less than K, all first-candidate suppliers can be selected as second-candidate suppliers.

[0100] A large language model is used to compare specific constraints in the demand feature tree with the unstructured descriptions corresponding to suppliers, generating a logical verification score; wherein, the specific constraints refer to unstructured features.

[0101] The second candidate suppliers are arranged in descending order of their logical verification scores to obtain the sorted second candidate suppliers.

[0102] In one possible implementation, the indicator data corresponding to the comprehensive score of the ranked second candidate supplier are converted into radar data charts, and the radar data charts are displayed in sorted order, including:

[0103] The data of various indicators corresponding to the comprehensive score of the second candidate supplier after the ranking are obtained as target business intent fit, target qualification certification data, target historical performance data, target user evaluation score data, and target dynamic data.

[0104] The target business intent alignment, target qualification certification data, target historical performance data, target user evaluation score data, and target dynamic data are constructed into a radar data chart containing five axes, and the radar data chart is displayed in sorted order.

[0105] For example, configuring a deep logical review and secondary sorting mechanism based on LLM can be done based on the comprehensive score S_total calculated from the third aspect. Suppliers are initially sorted in descending order, and the top K (Top-K) are selected as the initial set. Then, the semantic reasoning capabilities of the Large Language Model (LLM) are invoked to perform a refined review of this set. This review focuses on comparing specific constraints in the JSON requirement feature tree generated from the second aspect (such as location matching requirements and deep logical relevance of past cases) with the unstructured descriptions of the suppliers. This generates a logical verification score independent of vector similarity, and the final sorting of the candidate list is fine-tuned accordingly to correct semantic biases that may arise from relying solely on vector distance.

[0106] Specifically, configuring an LLM-based deep logical review and secondary sorting mechanism can include the following four implementation steps:

[0107] 1. Verification Data Assembly and Context Construction: The system first extracts nodes with strong logical constraints from the JSON requirement feature tree generated in the second aspect. These nodes include "hard threshold constraints" (e.g., the location must be the project location, registered capital > 50 million) and "deep logical constraints" (e.g., past cases must include construction experience in high-altitude and cold environments). Simultaneously, the system retrieves the corresponding "qualification certificate text" and "detailed historical project case text" from the unstructured database of each supplier in the initial selection set (Top-K). The "constraints" and "texts to be verified" are then assembled into a verification context.

[0108] 2. Verification Prompt Engineering: The system utilizes verification prompt engineering to construct dedicated logic verification instructions. These instructions include a rule input layer that transforms constraints in JSON into logical judgment rules, a fact input layer that inputs the supplier's unstructured descriptive text, and a chain of thought (CoT) that requires the LLM to perform step-by-step reasoning before outputting results. For example, an instruction could be set as: "Please first analyze the supplier's case study regarding the construction environment, determine whether it meets the definition of indoor testing, and then provide a judgment conclusion based on the logical rules."

[0109] 3. Perform Deep Semantic Entailment Inference: Invoke a pre-trained large language model to perform inference for each candidate supplier. This step aims to correct logical pitfalls that simple vector retrieval (Embedding) cannot handle, such as location logic verification. LLM does not simply match keywords to place names, but analyzes whether the "service coverage radius" or "site distribution" in the supplier's profile logically covers the "project implementation location" in the requirements; Case relevance verification. LLM deeply analyzes the "technical difficulties" and "solutions" descriptions in the supplier's past cases to determine whether they have truly solved pain points similar to the current procurement needs (e.g., distinguishing between "having done similar projects" and "only mentioning similar keywords"), eliminating "pseudo-match" suppliers who only pile up keywords but lack actual relevant experience.

[0110] 4. Generating Logical Verification Score and Final Reordering: The LLM outputs a quantified logical verification score (S_logic) based on the reasoning results, with a confidence range of 0 to 1. If the LLM determines that there is a "fatal flaw" (such as falsified qualifications or logical conflicts in key parameters), it assigns S_logic = 0; if the judgment is highly consistent with the case details, it assigns a high score close to 1. The system then weights and merges this logical verification score with the comprehensive score (S_total) calculated from a third aspect to calculate the final score S_final = (in , The list is pre-weighted and reordered according to S_final to output the final recommendation results.

[0111] A multi-dimensional capability profile radar chart generation module is constructed. To intuitively present the overall quality distribution of suppliers, the system maps the final evaluation results to a five-dimensional coordinate system to generate a dynamic radar chart. The five axes of the radar chart strictly correspond to the key variables in the third aspect weighted model, covering semantic intent matching degree, qualification certification level, historical performance, user evaluation feedback, and dynamic service willingness. By rendering the area and shape of the closed region covered by the lines connecting the score points of each axis, the system intuitively shows whether the supplier belongs to the "balanced overall strength type" or the "specific technology fit type," thereby providing transparent and interpretable decision support for procurement personnel in combination with the logical review results of LLM.

[0112] The ideal weight strategy generated by RL (Reinforcement Learning) is transformed into a visualized ideal demand profile, which is then overlaid and rendered with the supplier's actual capability profile. This technology abandons simple score comparison and instead helps decision-makers quickly identify suitable suppliers that perfectly cover the requirements in the critical axis and have acceptable shortcomings in the secondary axis by intuitively displaying the fit and coverage relationship between the two polygonal profiles in terms of shape. This greatly improves the transparency and interpretability of complex decision-making processes.

[0113] In one specific embodiment, this system was deployed and applied to the digital procurement cloud platform of a municipal engineering design and research institute in China. This embodiment selected a full set of business data spanning five years, from January 2020 to December 2024, as the basis for cold start and training. This dataset covers 12 core service categories, including geological exploration, environmental monitoring, and engineering consulting, and has integrated over 3,850 active suppliers, cleaning and structuring 9,240 non-standardized procurement requirements. Based on this, the system used ETL (Extract Transform Load) tools to extract and construct 36,500 high-quality "requirement-performance" closed-loop feedback sample pairs for pre-training and policy initialization of the reinforcement learning model.

[0114] Before processing the 36,500 sets of "demand-fulfillment" data in this embodiment, the system first constructed, as follows: Figure 2 The diagram shows a "data alignment topology based on a heterogeneous augmented graph." The SRM (Supplier Relationship Management) system and the contract management database are connected to the computing center via a "zero-latency direct connection," ensuring that the supplier's basic qualifications (D2) and historical performance (D3) are read in real time, building a robust static performance foundation. For the dynamic behavior log (D5), the system identifies a maximum cleaning lag of approximately 15 minutes in the data transmission link. Therefore, the system inserts three "virtual buffer nodes" (such as...) into the topology. Figure 2 (As shown by the dashed node on the right). When LBS location data flows through the virtual node chain, the built-in information entropy decay function... Outdated data is automatically downweighted. For example, the confidence level of a supplier's "active 30 minutes ago" signal is reduced after attenuation, thus ensuring that the D5 metric input to the RL model reflects the valid and true intentions at the current moment, rather than historical noise.

[0115] This example focuses on a typical procurement task for an "indoor evaporation test (stability testing and debugging) for a scientific research project." In this example, the procurement personnel input the original requirement text into the system: "To conduct an indoor evaporation test for a scientific research project, stability testing and debugging of the evaporation temperature and humidity environmental control system are required, with a budget of 10,000." First, the LLM parsing engine is invoked to convert the unstructured text into a JSON feature tree, accurately extracting the category tag "technical testing service," as well as key constraints such as "environmental control system" and "accuracy ±0.5℃." Then, the system uses Embedding technology to transform this requirement into a high-dimensional vector. For example... Figure 3Embedding technology transforms text into a 1536-dimensional high-dimensional vector. To facilitate human visual understanding, the system uses PCA (Principal Component Analysis) to project it into a three-dimensional space. The dense red, green, and blue scatter points in the graph represent service clusters of massive historical data in the database. The system projects the procurement demand vector v1 into space and, at the micro-semantic level, finds that supplier A (v2, whose description includes "high-precision environmental control"), located in the green cluster, has the closest geometric distance to the demand vector (cosine similarity 0.92), while supplier B (v3, whose description includes "design consulting") is farther away. This demonstrates that the system can accurately identify the degree of relevance to business intent, completing the initial semantic screening.

[0116] In further evaluation, the system defines the JSON feature tree parsed in step one (containing the "stability test" type, the "scientific research" tag, and the "high precision" parameter) as the state input for the reinforcement learning (RL) model. The system's built-in lightweight contextual bandit algorithm, acting as an adaptive navigator, reveals that the current state has extremely high reliability requirements for the test results, while being less sensitive to the supplier's geographical location due to its support for remote monitoring or single on-site visits. Figure 3 W_dyn in this context corresponds to the spatiotemporal dynamic intention index (D5) mentioned in the first aspect. This index is a comprehensive spatiotemporal feature generated by the system based on heterogeneous augmented graph topology, after information entropy decay processing of the supplier's online interaction behavior and offline physical proximity. For example... Figure 4 As shown in the figure, the blue bars illustrate the agent's action space decision-making based on the current state: the agent autonomously judges that "talking the talk but not walking the walk" (merely having good semantics) is risky, and that "territorial restrictions" may wrongly exclude high-quality service providers. Therefore, it executes a specific strategy adjustment, actively reducing the weights of semantic fit (W_sem) and spatiotemporal dynamic intention (W_dyn), and significantly increasing the weights of historical performance and user evaluation. This dynamic adjustment based on reinforcement learning reflects the system's deep perception and response to business scenarios through the output of Actions.

[0117] Before establishing a weighting strategy that emphasizes "hard power," the system first performed a macro-level search within the multi-dimensional evaluation space. For example... Figure 5The diagram (Phase 1: Initial Cold Start) shows the simulation results without reinforcement learning optimization (i.e., without weight adjustment). At this stage, the metric in the multidimensional scoring space is not yet corrected, and the system filters solely based on geometric "semantic distance." The central yellow star in the diagram represents the procurement demand anchor point, while the surrounding dots correspond to suppliers. The size of each dot represents the supplier's "hard power score" (covering qualifications, performance, and user reviews). Smaller dots indicate poorer qualifications (potentially shell companies); larger dots indicate stronger qualifications (industry leaders). As the diagram shows, due to the singular metric, the recommended set (red nodes), while closely centered, contains many tiny nodes. These nodes represent "newcomers" or "shell suppliers" with extremely high semantic matching but lacking qualifications and performance support, posing a significant risk of non-compliance. This vividly illustrates the shortcomings of traditional technologies, which, lacking a search path evolution mechanism, easily recommend "high-risk" suppliers with impressive technical documentation but lacking actual qualification guarantees.

[0118] And by using reinforcement learning, execution Figure 3 The aforementioned weight adjustment corrected the metric in the multidimensional scoring space (changing from standard Euclidean distance to a weighted distance emphasizing hard metrics), thereby triggering the continuous evolution of the optimal supplier search path. For example... Figure 6 As shown in the diagram (Phase 2: Reinforcement Learning Adaptive Navigation Search), the results of the evolved search path are illustrated. The selected recommendation set (blue nodes) has shifted significantly: the system abandons the small, seemingly insignificant nodes in the central region and expands outwards, accurately locking onto large nodes (representing mature suppliers with high qualifications and reputation). This result demonstrates that the system successfully filters out low-quality, risky points and achieves safer selection.

[0119] The optimal supplier (e.g., supplier A) selected through RL optimization is further refined and optimized by the system, and the final decision support chart is output.

[0120] like Figure 7As shown in the supplier capability profile radar chart, this chart visually demonstrates why supplier A is the best choice. The gray dashed polygon represents the ideal requirement profile generated by the RL agent, the red solid line represents the ultimately winning supplier A, and the green dashed line represents supplier C, which was eliminated in the second round of selection (note that supplier C was also selected in the first step due to semantic proximity). Given the extremely high requirements for data authenticity and process rigor in the "scientific research stability test," the gray ideal profile shows a significant outward expansion in the user evaluation and historical performance axes, representing high weight and high requirements; while it is relatively inward-shrinking in the semantic fit and spatiotemporal dynamic intention axes, representing low sensitivity. Observing the green profile, we can see that although supplier C is extremely prominent in the "semantic intent fit" axis, indicating a high degree of overlap between the keywords in its bid document and the requirements, this accurately corresponds to... Figure 4 The small "spheres" close to the center but tiny in size represent suppliers with good semantics but weak capabilities. However, their "historical performance" and "user reviews" axes show a clear collapse, failing to fill the gray ideal demand outline. Therefore, the system classifies this supplier as a risky entity with excellent packaging but insufficient overall strength. In contrast, while Supplier A's red outline is somewhat concave in the "semantic" and "spatial-temporal willingness" axes, limited by geographical location and document description, this value still covers the boundary of the concave ideal outline, meeting the minimum threshold set by RL. More importantly, Supplier A performs extremely well in the "user reviews" and "historical performance" axes, which are most valued by RL. The red outline perfectly covers and even overflows the center area of ​​the gray ideal outline, accurately corresponding to... Figure 4 The "blue sphere" was precisely locked in the image. This visual representation of contour fitting clearly shows that although supplier A is not an all-rounder, the shape of its capabilities is highly complementary to the shape of the project's requirements. By adjusting the RL weights, the system successfully escaped the traditional trap of keyword matching and selected this cost-effective and pragmatic partner who is most suitable for the current scientific research task.

[0121] Through the above interconnected charts, procurement personnel can clearly see the complete logical chain of how the system filters out this "high-performance, high-willingness" supplier step by step from massive amounts of data, based on the specific requirements of the "stability test" scenario.

[0122] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0123] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0124] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0125] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0126] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0127] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An intelligent source-finding decision-making method based on a large language model, characterized in that, include: Obtain multi-dimensional supplier data and preprocess the multi-dimensional supplier data to obtain preprocessed multi-dimensional supplier data; In response to the procurement requirements input by the procurement personnel, a pre-trained large language model is invoked to parse the procurement requirements, obtain a structured requirement feature tree, and perform an initial screening of all suppliers based on the requirement feature tree to obtain the first candidate supplier. Based on the demand feature tree and the preprocessed multi-dimensional supplier data corresponding to the first candidate supplier, the business intent fit is obtained; A reinforcement learning algorithm is used to analyze the feature vectors corresponding to the demand feature tree, determine the weighting parameters, and use the weighting parameters to weight the preprocessed multi-dimensional data of the first candidate supplier and the business intent fit, thereby generating a comprehensive score. The first candidate supplier is screened a second time based on the comprehensive score to obtain the second candidate supplier. The second candidate supplier is then reviewed and ranked using a large language model to obtain the ranked second candidate supplier. The comprehensive score data of the second candidate supplier after sorting is converted into radar data charts, and the radar data charts are displayed in sorting order to realize intelligent source sourcing decision-making based on a large language model.

2. The intelligent source-finding decision method based on a large language model according to claim 1, characterized in that, Obtain multi-dimensional supplier data and preprocess the multi-dimensional supplier data to obtain preprocessed multi-dimensional supplier data, including: Obtain the static data corresponding to the supplier; wherein, the static data includes basic information, qualification certification, historical performance and user reviews; The dynamic data corresponding to the supplier is obtained by using an augmented graph data alignment method based on information entropy decay; wherein, the dynamic data includes an activity index calculated from the login frequency and browsing depth of the most recent days, a responsiveness index calculated from the average time difference between receiving a notification and responding, and a cooperation index based on the average response time and word count richness in instant messaging tools; the dynamic data and static data together constitute the supplier's multi-dimensional data; The supplier multi-dimensional data is preprocessed to obtain preprocessed supplier multi-dimensional data.

3. The intelligent source-finding decision-making method based on a large language model according to claim 2, characterized in that, The supplier multi-dimensional data is preprocessed to obtain preprocessed supplier multi-dimensional data, including: The static data in the supplier's multi-dimensional data is normalized to obtain the normalized static data. The normalized static data and the original dynamic data are used together as preprocessed supplier multi-dimensional data.

4. The intelligent source-finding decision-making method based on a large language model according to claim 1, characterized in that, In response to the procurement requirements input by the procurement personnel, a pre-trained large language model is invoked to parse the procurement requirements and obtain a structured requirement feature tree, including: In response to the procurement requirements input by the procurement personnel, a pre-trained large language model is invoked as the core of parsing. A preset template is used to guide the large language model to perform entity extraction and prompt analysis, and the procurement requirements are converted into a structured requirement feature tree.

5. The intelligent source-finding decision method based on a large language model according to claim 1, characterized in that, Based on the aforementioned demand feature tree, all suppliers are initially screened to obtain the first candidate suppliers, including: Based on the preset hard indicator types, the target hard indicators in the demand feature tree are determined; Based on multi-dimensional data from suppliers, suppliers that do not meet the target hard indicators are eliminated, resulting in the first candidate supplier.

6. The intelligent source-finding decision-making method based on a large language model according to claim 1, characterized in that, Based on the demand feature tree and the preprocessed multi-dimensional supplier data corresponding to the first candidate supplier, the business intent fit is obtained, including: Determine the target data in the multi-dimensional data of the first candidate supplier that corresponds to the features in the demand feature tree; The cosine similarity between the feature vector formed by the features in the demand feature tree and the feature vector corresponding to the target data is obtained to obtain the business intent fit.

7. The intelligent source-finding decision-making method based on a large language model according to claim 1, characterized in that, The feature vectors corresponding to the demand feature tree are analyzed using a reinforcement learning algorithm to determine the weighting parameters, including: The feature vector formed by the features in the demand feature tree is used as the state space of the reinforcement learning algorithm. The reinforcement learning algorithm is used to select action vectors to obtain weighted weight parameters. The action vectors are selected in the action space, which is a multi-dimensional solution space composed of the value ranges of multiple weighted weight parameters. The process of selecting action vectors is constrained to the point that the sum of each weighted weight parameter is 1.

8. The intelligent source-finding decision-making method based on a large language model according to claim 1, characterized in that, The preprocessed multi-dimensional data of the first candidate supplier and its business intent fit are weighted using the aforementioned weighting parameters to generate a comprehensive score: S_total=W_sem×S_sem+W_cred×S_cred+W_perf×S_perf+W_eval×S_eval+W_dyn×S_dyn; Wherein, S_total is the overall score, S_sem is the business intent fit, W_sem is the first weighted weight parameter, S_cred is the qualification certification data in the normalized static data, W_cred is the second weighted weight parameter, S_perf is the historical performance data in the normalized static data, W_perf is the third weighted weight parameter, S_eval is the user evaluation score data in the normalized static data, W_eval is the fourth weighted weight parameter, S_dyn is the dynamic data in the normalized static data, and W_dyn is the fifth weighted weight parameter.

9. The intelligent source-finding decision-making method based on a large language model according to claim 1, characterized in that, The first candidate supplier is screened a second time based on the comprehensive score to obtain the second candidate supplier. A large language model is then used to review and rank the second candidate supplier to obtain the ranked second candidate supplier, including: The first candidate supplier is selected through Top-K screening based on the comprehensive score to obtain the second candidate supplier; where K is a preset value. A large language model is used to compare specific constraints in the demand feature tree with the unstructured descriptions corresponding to suppliers, generating a logical verification score; wherein, the specific constraints refer to unstructured features. The second candidate suppliers are arranged in descending order of their logical verification scores to obtain the sorted second candidate suppliers.

10. The intelligent source-finding decision method based on a large language model according to claim 1, characterized in that, The comprehensive score data corresponding to the second candidate supplier after ranking is converted into radar data charts, and the radar data charts are displayed in sorted order, including: The data of various indicators corresponding to the comprehensive score of the second candidate supplier after the ranking are obtained as target business intent fit, target qualification certification data, target historical performance data, target user evaluation score data, and target dynamic data. The target business intent alignment, target qualification certification data, target historical performance data, target user evaluation score data, and target dynamic data are constructed into a radar data chart containing five axes, and the radar data chart is displayed in sorted order.

Images