B2b supply and demand intelligent matching method and device based on ai large language model

Abstract

The application relates to the technical field of artificial intelligence, and discloses a B2B supply-demand intelligent matching method and device based on an AI large language model, which comprises the following steps: generating a candidate object similarity calculation result corresponding to a target object according to business demand information of the target object with a matching demand and an enterprise knowledge graph, and adjusting the candidate object similarity calculation result according to real-time market data associated with the business demand information, so as to determine a target candidate object list corresponding to the target object according to the adjusted candidate object similarity calculation result; and inputting object information and business demand information of each target candidate object in the target candidate object list into an AI large language model to generate natural language recommendation instructions of all target candidate objects. Compared with a traditional matching mode, the B2B supply-demand matching accuracy and timeliness are improved, the matching deviation problem caused by market environment changes is overcome, and the matching success rate and transaction conversion efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a B2B intelligent supply and demand matching method and apparatus based on an AI large language model. Background Technology

[0002] B2B (business-to-business) supply and demand matching is the core link in inter-enterprise transactions. Current technologies primarily rely on three methods: First, the traditional B2B platform's "yellow pages" model, where suppliers post product information, and buyers search for targets through keyword searches or category browsing. The platform only provides information display and basic search, and the matching process requires human interaction. Second, rule-based matching systems filter suppliers and buyers through preset hard rules such as product categories, price ranges, and geographical locations. However, both methods have significant limitations: traditional keyword searches and rule engines can only handle simple comparisons of structured fields and cannot understand the complex semantics of natural language. Existing collaborative filtering is also limited in effectiveness due to the long decision-making chain, multiple roles, and sparse behavioral data in B2B transactions, making it difficult to meet the multi-dimensional and refined matching needs of enterprises. Therefore, providing a technical solution that can improve the accuracy of B2B supply and demand matching is particularly important. Summary of the Invention

[0003] This invention provides a B2B supply and demand intelligent matching method and device based on AI large language model. Compared with traditional matching methods, it improves the accuracy and timeliness of B2B supply and demand matching, and overcomes the matching deviation problem caused by changes in the market environment, thereby improving the matching success rate and transaction conversion efficiency.

[0004] To address the aforementioned technical problems, the first aspect of this invention discloses a B2B intelligent supply and demand matching method based on an AI large language model, the method comprising: Obtain business requirement information of target objects with matching needs, and generate candidate object similarity calculation results corresponding to the target objects based on the business requirement information and the pre-built enterprise knowledge graph; Obtain real-time market data associated with the business requirement information, and adjust the similarity calculation result of the candidate objects based on the real-time market data to obtain the adjusted similarity calculation result of the candidate objects; Based on the adjusted similarity calculation results of the candidate objects, a list of target candidate objects corresponding to the target object is determined; The object information of each target candidate object in the target candidate object list and the business requirement information are input into the AI ​​large language model to generate natural language recommendation descriptions for all target candidate objects; the natural language recommendation descriptions for all target candidate objects include the business matching reasons between the target candidate object and the target object.

[0005] As an optional implementation, in the first aspect of the present invention, the target object includes a purchaser or supplier, and the business demand information of the target object includes the target object's product procurement demand information, price constraint demand information, geographical location demand information, qualification threshold demand information, performance preference information, and credit rating demand information. The step of generating candidate object similarity calculation results corresponding to the target object based on the business requirement information and the pre-constructed enterprise knowledge graph includes: The AI ​​large language model performs deep semantic parsing and vector generation operations on the business requirement information to obtain the requirement feature vector corresponding to the business requirement information; the requirement feature vector includes key entity vector, attribute parameter vector, constraint condition vector and implicit preference vector. Based on the demand feature vector, a multi-dimensional similarity calculation operation is performed on the pre-constructed enterprise knowledge graph to obtain the similarity calculation results of the candidate objects corresponding to the target object. The similarity calculation results of the candidate objects include the matching degree of product specifications, the overlap of price range, the adaptability of geographical location, the matching degree of certification qualifications, the matching degree of fulfillment rate, and the compliance degree of credit rating between the target object and each candidate object.

[0006] As an optional implementation, in the first aspect of the present invention, the step of performing deep semantic parsing and vector generation operations on the business requirement information based on the AI ​​large language model to obtain the requirement feature vector corresponding to the business requirement information includes: A structured prompt template is constructed, and the business requirement information is input into the AI ​​large language model, so that the AI ​​large language model performs a key entity recognition operation based on the structured prompt template to obtain multiple target entities in the business requirement information, and performs semantic encoding on all the target entities to obtain the key entity vector corresponding to the business requirement information; all the target entities include at least one of product entities, industry entities, and specification entities; The quantitative parameter information in the business requirement information is input into the AI ​​large language model, so that the AI ​​large language model performs unit normalization and numerical encoding operations on the quantitative parameter information to obtain the attribute parameter vector corresponding to the business requirement information. The restrictive statements in the business requirement information are input into the AI ​​large language model, so that the AI ​​large language model can determine multiple target constraint information in the restrictive statements and map all the target constraint information into a constraint condition vector corresponding to the business requirement information; all the target constraint information includes price range constraint information, geographical range constraint information, time window constraint information, and qualification threshold constraint information. The non-explicit requirement description information in the business requirement information is input into the AI ​​large language model, so that the AI ​​large language model performs contextual semantic reasoning operation on the non-explicit requirement description information to obtain the implicit preference vector corresponding to the business requirement information; the implicit preference vector includes a performance stability preference vector, a cooperation flexibility preference vector, and a credit rating preference vector.

[0007] As an optional implementation, in the first aspect of the present invention, the step of performing a multi-dimensional similarity calculation operation on a pre-constructed enterprise knowledge graph based on the demand feature vector to obtain the similarity calculation result of the candidate object corresponding to the target object includes: Using the key entity vector as the semantic query vector and the attribute parameter vector as the precision constraint vector, an approximate nearest neighbor search is performed in the product-enterprise association subgraph of the pre-constructed enterprise knowledge graph to obtain a candidate object set. The cosine similarity between the key entity vector and the product node entity vector of each candidate object in the candidate object set is calculated as the semantic matching score, and the Euclidean distance between the attribute parameter vector and the product node parameter vector of each candidate object is calculated as the parameter deviation score. The semantic matching score and parameter deviation score between the target object and each candidate object are weighted and fused to obtain the product specification matching degree between the target object and each candidate object. The price range constraint vector in the constraint vector is compared with the product price range vector of each candidate object in the enterprise knowledge graph to calculate the overlap of the price range between the target object and each candidate object. Based on the geographical range constraint vector and time window constraint vector in the constraint condition vector, as well as the transaction address coordinate vector and transaction time window vector of each candidate object in the enterprise knowledge graph, the spherical transaction distance parameter between the target object and each candidate object is calculated, and the geographical location fit between the target object and each candidate object is determined based on the spherical transaction distance parameter between the target object and each candidate object.

[0008] As an optional implementation, in the first aspect of the present invention, the step of performing a multi-dimensional similarity calculation operation on a pre-constructed enterprise knowledge graph based on the demand feature vector to obtain the similarity calculation result of the candidate object corresponding to the target object further includes: Based on the qualification threshold constraint vector in the constraint condition vector and the qualification node vector of each candidate object in the enterprise knowledge graph, the qualification inclusion ratio between the target object and each candidate object is calculated as the certification qualification matching degree between the target object and each candidate object; Based on the performance stability preference vector, the cooperation flexibility preference vector, and the historical performance node vector of each candidate object in the enterprise knowledge graph, a performance weighted matching operation is performed between the target object and each candidate object to obtain the performance rate matching degree between the target object and each candidate object; Based on the credit rating preference vector and the credit rating node vector of each candidate object in the enterprise knowledge graph, a credit rating mapping and matching operation is performed between the target object and each candidate object to obtain the credit rating conformity between the target object and each candidate object.

[0009] As an optional implementation, in the first aspect of the present invention, adjusting the similarity calculation result of the candidate objects based on the real-time market data to obtain the adjusted similarity calculation result of the candidate objects includes: Based on the real-time market data and the corresponding historical benchmark data in the enterprise knowledge graph, the target dimension market deviation degree between the target object and each candidate object is calculated. Then, based on the target dimension market deviation degree between the target object and each candidate object, dynamic adjustment parameter objects and static adjustment parameter objects are determined from the candidate object similarity calculation results. The target dimension market deviation degree includes price dimension market deviation degree, geographic location dimension market deviation degree, and fulfillment rate dimension market deviation degree. Based on the target dimension market deviation between the target object and each candidate object, the dynamic adjustment parameter object between the target object and each candidate object is corrected to obtain the corrected dynamic adjustment parameter object between the target object and each candidate object. A normalization fusion operation is performed on the static adjustment parameter object and the modified dynamic adjustment parameter object between the target object and each candidate object to obtain the adjusted similarity calculation result of the candidate object.

[0010] As an optional implementation, in the first aspect of the present invention, the enterprise knowledge graph is constructed in the following manner: Acquire multi-source enterprise data; the multi-source enterprise data includes enterprise business registration data, product catalog data, historical transaction data, qualification certification data, and credit rating data; The AI ​​large language model is used to perform entity recognition, relation extraction, and attribute alignment operations on the multi-source enterprise data to obtain structured graph triples. The structured graph triples include entity nodes, attribute nodes, and relation edges. The entity nodes include enterprise entities, product entities, and industry entities, and the relation edges include at least one of supply relationships, procurement relationships, competitive relationships, and cooperative relationships. Based on the structured graph triples, an initial enterprise knowledge graph is constructed, and a vectorized encoding operation is performed on each entity node in the initial enterprise knowledge graph to obtain the graph embedding representation corresponding to the initial enterprise knowledge graph. When the latest enterprise data is obtained, the attribute nodes and relationship edges of the initial enterprise knowledge graph are incrementally expanded using the latest enterprise data, and the graph embedding identifier is updated synchronously to obtain the enterprise knowledge graph.

[0011] A second aspect of this invention discloses a B2B intelligent supply and demand matching device based on an AI large language model, the device comprising: The acquisition module is used to acquire business requirement information of target objects that have matching needs; The generation module is used to generate similarity calculation results of candidate objects corresponding to the target object based on the business requirement information and the pre-built enterprise knowledge graph. The acquisition module is also used to acquire real-time market data associated with the business demand information; The adjustment module is used to adjust the similarity calculation result of the candidate objects based on the real-time market data to obtain the adjusted similarity calculation result of the candidate objects; The determination module is used to determine the list of target candidate objects corresponding to the target object based on the adjusted similarity calculation results of the candidate objects; The generation module is further configured to input the object information of each target candidate object in the target candidate object list and the business requirement information into the AI ​​large language model to generate natural language recommendation descriptions for all target candidate objects; the natural language recommendation descriptions for all target candidate objects include the business matching reasons between the target candidate object and the target object.

[0012] As an optional implementation, in the second aspect of the present invention, the target object includes a purchaser or supplier, and the business demand information of the target object includes the target object's product procurement demand information, price constraint demand information, geographical location demand information, qualification threshold demand information, performance preference information, and credit rating demand information. Specifically, the generation module generates the candidate object similarity calculation result corresponding to the target object based on the business requirement information and the pre-built enterprise knowledge graph in the following ways: The AI ​​large language model performs deep semantic parsing and vector generation operations on the business requirement information to obtain the requirement feature vector corresponding to the business requirement information; the requirement feature vector includes key entity vector, attribute parameter vector, constraint condition vector and implicit preference vector. Based on the demand feature vector, a multi-dimensional similarity calculation operation is performed on the pre-constructed enterprise knowledge graph to obtain the similarity calculation results of the candidate objects corresponding to the target object. The similarity calculation results of the candidate objects include the matching degree of product specifications, the overlap of price range, the adaptability of geographical location, the matching degree of certification qualifications, the matching degree of fulfillment rate, and the compliance degree of credit rating between the target object and each candidate object.

[0013] As an optional implementation, in the second aspect of the present invention, the generation module performs deep semantic parsing and vector generation operations on the business requirement information based on the AI ​​large language model to obtain the requirement feature vector corresponding to the business requirement information in the following specific ways: A structured prompt template is constructed, and the business requirement information is input into the AI ​​large language model, so that the AI ​​large language model performs a key entity recognition operation based on the structured prompt template to obtain multiple target entities in the business requirement information, and performs semantic encoding on all the target entities to obtain the key entity vector corresponding to the business requirement information; all the target entities include at least one of product entities, industry entities, and specification entities; The quantitative parameter information in the business requirement information is input into the AI ​​large language model, so that the AI ​​large language model performs unit normalization and numerical encoding operations on the quantitative parameter information to obtain the attribute parameter vector corresponding to the business requirement information. The restrictive statements in the business requirement information are input into the AI ​​large language model, so that the AI ​​large language model can determine multiple target constraint information in the restrictive statements and map all the target constraint information into a constraint condition vector corresponding to the business requirement information; all the target constraint information includes price range constraint information, geographical range constraint information, time window constraint information, and qualification threshold constraint information. The non-explicit requirement description information in the business requirement information is input into the AI ​​large language model, so that the AI ​​large language model performs contextual semantic reasoning operation on the non-explicit requirement description information to obtain the implicit preference vector corresponding to the business requirement information; the implicit preference vector includes a performance stability preference vector, a cooperation flexibility preference vector, and a credit rating preference vector.

[0014] As an optional implementation, in the second aspect of the present invention, the generation module performs a multi-dimensional similarity calculation operation on the pre-constructed enterprise knowledge graph based on the required feature vector to obtain the similarity calculation result of the candidate object corresponding to the target object. Specifically, this includes: Using the key entity vector as the semantic query vector and the attribute parameter vector as the precision constraint vector, an approximate nearest neighbor search is performed in the product-enterprise association subgraph of the pre-constructed enterprise knowledge graph to obtain a candidate object set. The cosine similarity between the key entity vector and the product node entity vector of each candidate object in the candidate object set is calculated as the semantic matching score, and the Euclidean distance between the attribute parameter vector and the product node parameter vector of each candidate object is calculated as the parameter deviation score. The semantic matching score and parameter deviation score between the target object and each candidate object are weighted and fused to obtain the product specification matching degree between the target object and each candidate object. The price range constraint vector in the constraint vector is compared with the product price range vector of each candidate object in the enterprise knowledge graph to calculate the overlap of the price range between the target object and each candidate object. Based on the geographical range constraint vector and time window constraint vector in the constraint condition vector, as well as the transaction address coordinate vector and transaction time window vector of each candidate object in the enterprise knowledge graph, the spherical transaction distance parameter between the target object and each candidate object is calculated, and the geographical location fit between the target object and each candidate object is determined based on the spherical transaction distance parameter between the target object and each candidate object.

[0015] As an optional implementation, in the second aspect of the present invention, the method by which the generation module performs a multi-dimensional similarity calculation operation on the pre-constructed enterprise knowledge graph based on the required feature vector to obtain the similarity calculation result of the candidate object corresponding to the target object further includes: Based on the qualification threshold constraint vector in the constraint condition vector and the qualification node vector of each candidate object in the enterprise knowledge graph, the qualification inclusion ratio between the target object and each candidate object is calculated as the certification qualification matching degree between the target object and each candidate object; Based on the performance stability preference vector, the cooperation flexibility preference vector, and the historical performance node vector of each candidate object in the enterprise knowledge graph, a performance weighted matching operation is performed between the target object and each candidate object to obtain the performance rate matching degree between the target object and each candidate object; Based on the credit rating preference vector and the credit rating node vector of each candidate object in the enterprise knowledge graph, a credit rating mapping and matching operation is performed between the target object and each candidate object to obtain the credit rating conformity between the target object and each candidate object.

[0016] As an optional implementation, in a second aspect of the present invention, the adjustment module adjusts the similarity calculation result of the candidate objects based on the real-time market data, and the method for obtaining the adjusted similarity calculation result of the candidate objects specifically includes: Based on the real-time market data and the corresponding historical benchmark data in the enterprise knowledge graph, the target dimension market deviation degree between the target object and each candidate object is calculated. Then, based on the target dimension market deviation degree between the target object and each candidate object, dynamic adjustment parameter objects and static adjustment parameter objects are determined from the candidate object similarity calculation results. The target dimension market deviation degree includes price dimension market deviation degree, geographic location dimension market deviation degree, and fulfillment rate dimension market deviation degree. Based on the target dimension market deviation between the target object and each candidate object, the dynamic adjustment parameter object between the target object and each candidate object is corrected to obtain the corrected dynamic adjustment parameter object between the target object and each candidate object. A normalization fusion operation is performed on the static adjustment parameter object and the modified dynamic adjustment parameter object between the target object and each candidate object to obtain the adjusted similarity calculation result of the candidate object.

[0017] As an optional implementation, in a second aspect of the invention, the enterprise knowledge graph is constructed in the following manner: Acquire multi-source enterprise data; the multi-source enterprise data includes enterprise business registration data, product catalog data, historical transaction data, qualification certification data, and credit rating data; The AI ​​large language model is used to perform entity recognition, relation extraction, and attribute alignment operations on the multi-source enterprise data to obtain structured graph triples. The structured graph triples include entity nodes, attribute nodes, and relation edges. The entity nodes include enterprise entities, product entities, and industry entities, and the relation edges include at least one of supply relationships, procurement relationships, competitive relationships, and cooperative relationships. Based on the structured graph triples, an initial enterprise knowledge graph is constructed, and a vectorized encoding operation is performed on each entity node in the initial enterprise knowledge graph to obtain the graph embedding representation corresponding to the initial enterprise knowledge graph. When the latest enterprise data is obtained, the attribute nodes and relationship edges of the initial enterprise knowledge graph are incrementally expanded using the latest enterprise data, and the graph embedding identifier is updated synchronously to obtain the enterprise knowledge graph.

[0018] A third aspect of this invention discloses another B2B intelligent supply and demand matching device based on an AI large language model, the device comprising: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the B2B intelligent supply and demand matching method based on AI large language model disclosed in the first aspect of the present invention.

[0019] The fourth aspect of the present invention discloses a computer storage medium storing computer instructions, which, when invoked, are used to execute the B2B intelligent supply and demand matching method based on an AI large language model disclosed in the first aspect of the present invention.

[0020] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: In this embodiment of the invention, based on the business demand information of the target object with matching needs and the enterprise knowledge graph, a similarity calculation result of the candidate object corresponding to the target object is generated. Then, based on real-time market data associated with the business demand information, the similarity calculation result of the candidate object is adjusted to determine the target candidate object list corresponding to the target object. The object information and business demand information of each target candidate object in the target candidate object list are input into an AI large language model to generate natural language recommendation descriptions for all target candidate objects. This improves the accuracy and timeliness of B2B supply and demand matching compared to traditional matching methods, and overcomes the matching deviation problem caused by changes in the market environment, thereby increasing the success rate of matching and transaction conversion efficiency. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating a B2B intelligent supply and demand matching method based on an AI large language model disclosed in an embodiment of the present invention. Figure 2 This is a flowchart illustrating another B2B intelligent supply and demand matching method based on an AI large language model disclosed in an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of a B2B intelligent supply and demand matching device based on an AI large language model disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram of another B2B intelligent supply and demand matching device based on an AI large language model disclosed in an embodiment of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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.

[0024] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0025] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0026] This invention discloses a B2B supply and demand intelligent matching method and device based on an AI large language model. Compared with traditional matching methods, it improves the accuracy and timeliness of B2B supply and demand matching, and overcomes the matching deviation problem caused by changes in the market environment, thereby improving the matching success rate and transaction conversion efficiency.

[0027] Example 1 Please see Figure 1 , Figure 1 This is a flowchart illustrating a B2B intelligent supply and demand matching method based on an AI large language model, as disclosed in an embodiment of the present invention. Optionally, this method can be implemented by an intelligent supply and demand matching device, which can be integrated into an intelligent supply and demand service platform, or it can be a local server or cloud server used to process the B2B intelligent supply and demand matching process based on an AI large language model, etc. The embodiments of the present invention do not impose limitations. Figure 1 As shown, this AI-based large language model-based B2B intelligent supply and demand matching method can include the following operations: 101. Obtain the business requirement information of the target object with matching needs, and generate the similarity calculation results of the candidate object corresponding to the target object based on the business requirement information and the pre-built enterprise knowledge graph.

[0028] In this embodiment of the invention, the target object includes a purchaser or supplier. The business needs information of the target object includes product procurement needs information (such as the name of the product to be purchased, product specifications, product quantity, product quality standards, etc.), price constraint needs information (such as the expected purchase unit price range, the highest acceptable quotation, payment method requirements, etc.), geographical location needs information (such as the expected supply area, the maximum acceptable logistics transportation distance, whether same-city or intra-province delivery is required, etc.), qualification threshold needs information (such as requiring suppliers to have ISO9001 quality certification, specific industry operating licenses, minimum registered capital, etc.), performance preference information (such as the expected delivery cycle, whether partial delivery is supported, whether samples are required, after-sales response time requirements, etc.), and credit rating needs information (such as requiring the supplier's corporate credit rating to be no lower than AA, no major default records in the past year, whether third-party credit guarantees are accepted, etc.).

[0029] Furthermore, the enterprise knowledge graph is constructed in the following ways: Acquire multi-source enterprise data; By using an AI large language model, entity recognition, relation extraction, and attribute alignment operations are performed on multi-source enterprise data to obtain structured graph triples; Based on the structured graph triples, an initial enterprise knowledge graph is constructed, and vectorized encoding is performed on each entity node in the initial enterprise knowledge graph to obtain the graph embedding representation corresponding to the initial enterprise knowledge graph. When the latest enterprise data is obtained, the attribute nodes and relationship edges of the initial enterprise knowledge graph are incrementally expanded using the latest enterprise data, and the graph embedding identifier is updated synchronously to obtain the enterprise knowledge graph.

[0030] In this optional embodiment, the multi-source enterprise data further includes enterprise business registration data, product catalog data, historical transaction data, qualification certification data, and credit rating data, as well as structured graph triples including entity nodes, attribute nodes, and relationship edges; wherein, entity nodes include enterprise entities, product entities, and industry entities, and relationship edges include at least one of supply relationships, procurement relationships, competitive relationships, and cooperative relationships.

[0031] 102. Obtain real-time market data related to business needs, and adjust the similarity calculation results of candidate objects based on the real-time market data to obtain the adjusted similarity calculation results of candidate objects.

[0032] In this embodiment of the invention, optional real-time market data may include industry supply and demand, raw material price fluctuation data, logistics capacity data, regional market activity data, and seasonal demand change data, etc.

[0033] 103. Based on the adjusted candidate object similarity calculation results, determine the target candidate object list corresponding to the target object.

[0034] In this embodiment of the invention, the target candidate object list can be further sorted from high to low according to the adjusted candidate object similarity calculation results, and then a list consisting of candidate objects that meet or exceed a preset number or a similarity threshold. The preset number and / or similarity threshold can be customized by the target object.

[0035] 104. Input the object information and business requirement information of each target candidate object in the target candidate object list into the AI ​​large language model to generate natural language recommendation descriptions for all target candidate objects.

[0036] In this embodiment of the invention, optionally, the natural language recommendation description of all target candidate objects may include the business matching reason between the target candidate object and the target object, and may also include at least one of the following: a summary of the core advantages of the target candidate object, a summary of the historical cooperation record with the target object, and a warning of potential cooperation risks.

[0037] As can be seen, implementing this embodiment of the invention can generate similarity calculation results for candidate objects corresponding to the target object based on the target object's business needs information and enterprise knowledge graph. These results are then adjusted based on real-time market data to determine a list of target candidate objects corresponding to the target object, and natural language recommendation descriptions are generated for each target candidate object in the list, allowing the target object to query in real time. This improves the accuracy and timeliness of B2B supply and demand matching compared to traditional matching methods, and overcomes the matching deviation problem caused by changes in the market environment, thereby increasing the success rate of matching and transaction conversion efficiency.

[0038] Example 2 Please see Figure 2 , Figure 2 This is a flowchart illustrating another B2B intelligent supply and demand matching method based on an AI large language model, as disclosed in this embodiment of the invention. Optionally, this method can be implemented by an intelligent supply and demand matching device, which can be integrated into an intelligent supply and demand service platform, or it can be a local server or cloud server used to process the B2B intelligent supply and demand matching process based on an AI large language model, etc. This embodiment of the invention does not impose any limitations. Figure 2 As shown, this AI-based large language model-based B2B intelligent supply and demand matching method can include the following operations: 201. Obtain the business requirement information of the target object with matching needs, and perform deep semantic parsing and vector generation operations on the business requirement information based on the AI ​​large language model to obtain the requirement feature vector corresponding to the business requirement information.

[0039] In this embodiment of the invention, the demand feature vector includes a key entity vector, an attribute parameter vector, a constraint vector, and an implicit preference vector.

[0040] 202. Based on the requirement feature vector, perform multi-dimensional similarity calculation on the pre-built enterprise knowledge graph to obtain the similarity calculation results of the candidate objects corresponding to the target object.

[0041] In this embodiment of the invention, the similarity calculation results of the candidate objects include the matching degree of product specifications, the overlap of price range, the adaptability of geographical location, the matching degree of certification qualifications, the matching degree of fulfillment rate, and the compliance degree of credit rating between the target object and each candidate object.

[0042] 203. Obtain real-time market data related to business needs, and adjust the similarity calculation results of candidate objects based on the real-time market data to obtain the adjusted similarity calculation results of candidate objects.

[0043] 204. Based on the adjusted similarity calculation results of the candidate objects, determine the list of target candidate objects corresponding to the target object.

[0044] 205. Input the object information and business requirement information of each target candidate object in the target candidate object list into the AI ​​large language model to generate natural language recommendation descriptions for all target candidate objects.

[0045] In this embodiment of the invention, for other descriptions of steps 201-202, please refer to the detailed description of steps 102 and 103 in Embodiment 1. This embodiment of the invention will not repeat them.

[0046] As can be seen, implementing the embodiments of the present invention can perform deep semantic analysis on business demand information through an AI large language model, deconstructing relevant demand information into four types of feature vectors: key entities, attribute parameters, constraints, and implicit preferences. This effectively captures implicit demands that are difficult to identify through traditional keyword matching, improving the accuracy of supply and demand matching. At the same time, it integrates multi-dimensional similarity calculation of enterprise knowledge graphs with dynamic adjustments based on real-time market data, ensuring that the matching results conform to static enterprise profiles and dynamic market environments, overcoming the lag of static matching. Finally, it uses the AI ​​large language model to generate natural language recommendation explanations containing reasons for business matching, enhancing the interpretability of recommendation results, reducing the decision-making cost of target objects, and improving the intelligence level and transaction conversion efficiency of B2B supply and demand matching.

[0047] In an optional embodiment, step 201 above, which involves performing deep semantic parsing and vector generation operations on the business requirement information based on the AI ​​large language model to obtain the requirement feature vector corresponding to the business requirement information, includes: A structured prompt template is constructed, and business requirement information is input into the AI ​​large language model. The AI ​​large language model then performs key entity recognition based on the structured prompt template to obtain multiple target entities in the business requirement information. All target entities are semantically encoded to obtain the key entity vector corresponding to the business requirement information. The quantitative parameter information in the business requirement information is input into the AI ​​big language model, so that the AI ​​big language model performs unit normalization and numerical encoding operations on the quantitative parameter information to obtain the attribute parameter vector corresponding to the business requirement information. The restrictive statement information in the business requirement information is input into the AI ​​big language model so that the AI ​​big language model can determine multiple target constraint information in the restrictive statement information and map all target constraint information into the constraint condition vector corresponding to the business requirement information. The non-explicit requirement statements in the business requirement information are input into the AI ​​large language model, so that the AI ​​large language model can perform contextual semantic reasoning on the non-explicit requirement statements to obtain the implicit preference vector corresponding to the business requirement information.

[0048] In this optional embodiment, optionally, all target entities include at least one of product entities, industry entities, and specification entities. Further, all target constraint information includes price range constraint information, geographical scope constraint information, time window constraint information, and qualification threshold constraint information, as well as implicit preference vectors including performance stability preference vector, cooperation flexibility preference vector, and credit rating preference vector.

[0049] For example, the above operations can be understood as follows: (1) Key Entity Identification Operation For example, when the target object inputs the business requirement information as "purchase 10,000 M10×30 stainless steel hex bolts for building steel structure engineering, requiring compliance with GB / T 5782 standard", the AI ​​large language model can identify: the product entity as "stainless steel hex bolts", the industry entity as "building steel structure engineering", the specification entities as "M10×30" and "GB / T5782", and use semantic encoding capabilities to map the above entities into multi-dimensional key entity vectors respectively.

[0050] (2) Attribute parameter vector For example, when the business requirement information includes "unit price not exceeding 5 yuan / piece" and "purchase quantity 10,000 pieces", the AI ​​large language model can convert "5 yuan / piece" and the candidate object's price "50 yuan / kg" in the knowledge base into "unit price per thousand pieces" for normalization comparison (e.g., 5 yuan / piece → 5,000 yuan / thousand pieces, 50 yuan / kg → assuming 200 pieces per kilogram, then it is 250 yuan / thousand pieces), and encode the normalized value [5,000, 250] into an attribute parameter vector.

[0051] (3) Constraint vector mapping For example, when the business requirement is "urgently need to purchase a batch of high-strength concrete admixtures, with a water reduction rate of no less than 25%, and a dosage controlled between 1% and 2% of the cementitious materials; the supplier must have a building material product quality inspection report; delivery can only be accepted within G province; delivery must be completed within 7 days; and payment is only accepted via bank acceptance bills," the AI ​​large language model identifies and extracts multiple target constraint information from the restrictive expressions in the requirement (such as statements guided by qualifiers like "no less than," "only," "must," and "only accept"). Specifically, it extracts geographical constraint information from "delivery can only be accepted within G province," time window constraint information from "delivery must be completed within 7 days," qualification threshold constraint information from "the supplier must have a building material product quality inspection report," payment method constraint information from "payment is only accepted via bank acceptance bills," and product performance range constraint information from "water reduction rate of no less than 25%, dosage controlled between 1% and 2%."

[0052] Subsequently, the AI ​​large language model uniformly maps all the above target constraint information into a constraint condition vector. Taking the geographical range constraint information as an example, since the demand is limited to within Province G, the constraint weight corresponding to enterprises within Province G is mapped to 1 in the geographical dimension of the constraint condition vector, and the constraint weight corresponding to enterprises outside Province G is mapped to 0. Taking the time window constraint information as an example, the demand requires delivery within 7 days, so in the time dimension of the constraint condition vector, candidates with a historical average delivery cycle ≤ 7 days are mapped to 1, and candidates with a delivery cycle exceeding 7 days are mapped to 0. Taking the qualification threshold constraint information as an example, candidates holding a building material product quality inspection report are mapped to 1 in the qualification dimension, and those without such a report are mapped to 0. Taking the payment method constraint information as an example, candidates supporting bank acceptance bills are mapped to 1 in the payment dimension, and those that do not support them are mapped to 0. Taking the product performance range constraint information as an example, candidates with a water reduction rate ≥ 25% and a dosage between 1% and 2% are mapped to 1 in the performance dimension, and those that do not meet either condition are mapped to 0. Finally, the AI ​​large language model encodes the above five types of constraint information into a multi-dimensional constraint vector, with each dimension taking values ​​in the range of [0,1], for subsequent matching calculations with the actual attributes of candidate objects.

[0053] (4) Contextual semantic reasoning operation For example, when business requirements include non-explicit statements such as "hope for long-term and stable cooperation", "accept 60-day monthly payment", and "prefer to choose old suppliers with a history of cooperation", the AI ​​big language model can generate the following through contextual reasoning: performance stability preference vector (corresponding to "long-term and stable cooperation" inferring a high weight requirement for the supplier's historical performance rate), cooperation flexibility preference vector (corresponding to "accept 60-day monthly payment" inferring a high tolerance for payment method flexibility), and credit rating preference vector (corresponding to "prefer old suppliers" inferring an implicit requirement for the supplier's credit rating to be no lower than A).

[0054] As can be seen, this optional embodiment can decompose business requirement information into four dimensions: key entities, quantifiable parameters, restrictive statements, and non-explicit requirements, and generate corresponding entity vectors, parameter vectors, constraint vectors, and implicit preference vectors, respectively, thus achieving multi-granular, structured semantic representation of business requirements. This allows for the accurate capture of explicit entity information such as products, industries, and specifications, as well as the effective quantification of hard constraints such as price ranges and time windows. Furthermore, implicit preferences such as performance stability and cooperation flexibility can be mined through semantic reasoning, significantly improving the semantic completeness and matching accuracy of the requirement vector, and providing a more comprehensive and accurate data foundation for subsequent intelligent recommendations and decisions.

[0055] In another optional embodiment, step 202 above, which involves performing a multi-dimensional similarity calculation on the pre-built enterprise knowledge graph based on the demand feature vector to obtain the similarity calculation result of the candidate object corresponding to the target object, includes: Using key entity vectors as semantic query vectors and attribute parameter vectors as precision constraint vectors, an approximate nearest neighbor search is performed on the product-enterprise association subgraph of the pre-constructed enterprise knowledge graph to obtain a set of candidate objects. The cosine similarity between the key entity vector and the product node entity vector of each candidate object in the candidate object set is calculated as the semantic matching score, and the Euclidean distance between the attribute parameter vector and the product node parameter vector of each candidate object is calculated as the parameter deviation score. The semantic matching score and parameter deviation score between the target object and each candidate object are weighted and fused to obtain the product specification matching degree between the target object and each candidate object. The overlap between the price range constraint vector in the constraint vector and the product price range vector of each candidate object in the enterprise knowledge graph is calculated to obtain the price range overlap between the target object and each candidate object. Based on the geographical range constraint vector and time window constraint vector in the constraint condition vector, as well as the transaction address coordinate vector and transaction time window vector of each candidate object in the enterprise knowledge graph, the spherical transaction distance parameter between the target object and each candidate object is calculated, and the geographical location fit between the target object and each candidate object is determined based on the spherical transaction distance parameter between the target object and each candidate object.

[0056] In this optional embodiment, for example, the above-mentioned three-dimensional similarity calculation operations can be understood as follows: (1) The product specification matching degree between the target object and each candidate object. For example, first, using the key entity vector as the query vector, perform an approximate nearest neighbor search in the enterprise knowledge graph to filter out a set of candidate objects. Then, calculate the two scores separately and fuse them with weights: Product specification matching degree = a*cos(V) 目标实体 V 候选实体 )+b*(1-∥V 目标属性 -V 候选属性 ∥2 / ∥V 目标属性 ∥2); where V 目标实体 V is the key entity vector. 候选实体 V is the product node entity vector of the candidate object. 目标属性 V is a vector of attribute parameters. 候选属性 Let be the product node parameter vector of the candidate object, where a and b are the preset weights, and ∥.∥2 is the L2 norm.

[0057] It's important to note that in calculating product specification matching, the semantic matching score uses cosine similarity, while the parameter deviation score uses Euclidean distance. The choice between these two methods stems from the different nature of the data being processed: Key entity vectors are semantic vectors generated by an AI large-scale language model. Their vector magnitudes do not carry actual business meaning; only the vector direction reflects the semantic content. Therefore, cosine similarity is used to determine semantic closeness by measuring the cosine of the angle between two vectors. This method is insensitive to vector magnitudes and accurately reflects semantic consistency without being affected by differences in vector length. Attribute parameter vectors, on the other hand, are quantified parameters obtained after unit normalization and numerical encoding. The magnitude of each dimension directly corresponds to specific business indicators, such as unit price and quantity. Their magnitudes and the absolute differences between each component have clear physical meaning. Therefore, Euclidean distance is used to measure the numerical deviation between the requirement parameters and the candidate object parameters. This method is sensitive to the absolute differences in each dimension and accurately reflects whether the parameters match. Weighting and fusing the two methods ensures accurate alignment at the semantic level and precise matching at the quantified parameter level, resulting in a more reliable product specification matching.

[0058] (2) The overlap of price ranges between the target object and each candidate object For example, the demand price range [P] l P u [S] and the price range of candidate items l S u Imagine two line segments on a number line, and calculate the proportion of the overlapping portion to the length of the demand interval (i.e., the price interval overlap R): R = max(0, min(P) u S u ) max(P l S l )) / P u -P l .

[0059] (3) Geographical location fit between the target object and each candidate object For example, under the constraints of the geographical range constraint vector and the time window constraint vector in the constraint condition vector, the spherical transaction distance d is first calculated using the Haversine formula, and then mapped to the fitness degree through Gaussian decay. At the same time, the time window is introduced for determination. ; Where d is the geographical distance between the target object and the candidate object, R is the average radius of the Earth, φ1 and φ2 are the latitude vectors of the transaction addresses of the target object and the candidate object respectively (which need to be converted to radians, and △φ is the difference between the two), and λ1 and λ2 are the longitude vectors of the transaction addresses of the target object and the candidate object respectively (which need to be converted to radians, and △λ is the difference between the two).

[0060] Then, the geographic location fit A=e -d^2 / 2h^2 *1 （T交货≤T窗口） , among which, T 交货 T is the promised delivery time for the candidate. 窗口 For the longest delivery time window required by the target object, 1 (T交货≤T窗口) It is an indicator function (also called a judgment function).

[0061] Furthermore, based on the demand feature vector, a multi-dimensional similarity calculation operation is performed on the pre-constructed enterprise knowledge graph to obtain the similarity calculation results of the candidate objects corresponding to the target object, which also includes: Based on the qualification threshold constraint vector in the constraint condition vector and the qualification node vector of each candidate object in the enterprise knowledge graph, calculate the qualification inclusion ratio between the target object and each candidate object, which is used as the certification qualification matching degree between the target object and each candidate object. Based on the performance stability preference vector, the cooperation flexibility preference vector, and the historical performance node vector of each candidate object in the enterprise knowledge graph, a performance weighted matching operation is performed between the target object and each candidate object to obtain the performance rate matching degree between the target object and each candidate object. Based on the credit rating preference vector and the credit rating node vector of each candidate object in the enterprise knowledge graph, a credit rating mapping and matching operation is performed between the target object and each candidate object to obtain the credit rating conformity between the target object and each candidate object.

[0062] In this optional embodiment, for example, the above-mentioned three-dimensional similarity calculation operations can be understood as follows: (1) Authentication qualification matching degree between the target object and each candidate object The qualification threshold constraint vector records the certification qualifications required by the target object for the candidate object, such as ISO 9001, ISO 14001, CE certification, etc., and can be expressed in the form of a multi-dimensional binary vector or a weighted vector. Each candidate object in the enterprise knowledge graph is associated with a corresponding qualification node vector, recording the various certification qualifications actually held by the candidate object. During calculation, the qualification threshold constraint vector of the target object and the qualification node vector of the candidate object are compared dimension-by-dimensionally. The number of qualifications actually held by the candidate object that fall into the target object's required qualification set is counted, and the ratio of the number of qualifications held to the total number of required qualifications of the target object is used as the qualification coverage ratio. M qualification = |D demand |D candidate | / |D demand |; Wherein, |D-Requirement|D-Candidate| represents the number of intersections between the qualifications required by the target and those actually held by the candidate, and |D-Requirement| represents the total number of qualification requirements proposed by the target. A higher ratio indicates a more complete coverage of candidate qualifications and a higher degree of matching between certifications. For example, if the target requires both ISO 9001 and CE certifications, candidate A holds both, resulting in a coverage ratio of 1.0; candidate B holds only ISO 9001, resulting in a coverage ratio of 0.5; and candidate C holds neither, resulting in a coverage ratio of 0, and is therefore eliminated.

[0063] (2) Matching degree of fulfillment rate between the target object and each candidate object The performance stability preference vector reflects the target entity's preference weights for stability indicators such as supplier on-time delivery rate and order completion rate, while the cooperation flexibility preference vector reflects the target entity's preference weights for the candidate entity's flexibility in areas such as batch adjustments, delivery date changes, and return / exchange responses. Each candidate entity in the enterprise knowledge graph is associated with a historical performance node vector, recording historical behavioral data such as on-time delivery rate, order fulfillment rate, number of defaults, and average response time in past transactions. During calculation, firstly, the candidate entity's historical performance node vector and performance stability preference vector are weighted and multiplied to obtain a stability matching score; then, the historical performance node vector and cooperation flexibility preference vector are weighted and multiplied to obtain a flexibility matching score; finally, the two scores are linearly fused according to the weight ratio preset by the demand side. M 履约 =γ*P 稳定 *H 候选 / ∥P 稳定 ∥*∥H 候选 ∥+δ*P 灵活 *H 候选 / ∥P 灵活 ∥*∥H 候选 ∥; Among them, P 稳定 Let P be the performance stability preference vector. 灵活 H is the cooperation flexibility preference vector. 候选 Let γ be the historical fulfillment node vector of the candidate object, and δ be the preset weight coefficients of the target object, satisfying γ+δ=1. For example, when the target object values ​​on-time delivery more, γ=0.7, δ=0.3, etc.

[0064] (3) The credit rating compatibility between the target object and each candidate object. The credit rating preference vector records the target entity's minimum requirement for the candidate entity's credit rating, such as requiring a credit rating of at least A, and is expressed as an ordered rating vector or threshold vector. Each candidate entity in the enterprise knowledge graph is associated with a credit rating node vector, recording the enterprise's credit rating in the internal assessment system of a third-party credit reporting agency or platform, such as AAA, AA, A, BBB, etc. During calculation, a rating mapping matching strategy is used: the candidate entity's credit rating node vector is mapped to a numerical rating at the same scale as the target entity's credit rating preference vector. If the candidate entity's credit rating is not lower than the target entity's minimum threshold, positive points are awarded based on the rating difference exceeding the threshold; the higher the rating, the more points are awarded. If it is lower than the threshold, it is directly judged as ineligible. The specific calculation method can be: ; Among them, L 候选 L is the numerical value that maps the credit rating of the candidate. 需求L is the minimum credit rating value required by the target object. 最高 With L 最低 These are the highest and lowest credit rating values ​​in the credit rating system.

[0065] As can be seen, this optional embodiment can perform multi-dimensional similarity calculations on the enterprise knowledge graph, covering six matching dimensions: product specifications, price range, geographical location, certification qualifications, performance, and credit rating. This achieves comprehensive object screening, from semantic understanding to numerical verification, and from soft preferences to hard constraints. Specifically, the weighted fusion of semantic matching and parameter bias balances intent understanding and accuracy control; the calculation of interval overlap and spherical distance ensures the objective accuracy of price and geographical dimensions; and the mapping of qualification inclusion ratio and credit rating reduces the entry of non-compliant objects into the candidate set through hard constraints. This significantly improves the overall matching quality and recommendation reliability between candidate and target objects.

[0066] In another optional embodiment, step 203 above, adjusting the candidate object similarity calculation result based on real-time market data to obtain an adjusted candidate object similarity calculation result, includes: Based on real-time market data and corresponding historical benchmark data in the enterprise knowledge graph, the target dimension market deviation between the target object and each candidate object is calculated. Based on the target dimension market deviation between the target object and each candidate object, the dynamic adjustment parameter object and the static adjustment parameter object are determined from the candidate object similarity calculation results. Based on the target dimension market deviation between the target object and each candidate object, the dynamic adjustment parameter object between the target object and each candidate object is corrected to obtain the corrected dynamic adjustment parameter object between the target object and each candidate object. A normalization fusion operation is performed on the static adjustment parameter objects between the target object and each candidate object, as well as the corrected dynamic adjustment parameter objects, to obtain the adjusted candidate object similarity calculation results.

[0067] In this optional embodiment, the target dimension market deviation includes price dimension market deviation, geographic location dimension market deviation, and fulfillment rate dimension market deviation.

[0068] Furthermore, it should be noted that the market deviation of the target dimension can be: δ=|V 实时 V 基准 | / V 基准 ; Among them, V 实时 V provides real-time market data. 基准The historical benchmark data for this dimension in the knowledge graph (such as the average price over the past 90 days) indicates that the candidate object deviates more severely from the current market situation in this dimension, and requires a greater reduction in weighting.

[0069] The process of determining the dynamic and static adjustment parameter objects can be understood as follows: candidate objects with large market deviations in the price and geographical location dimensions are classified as dynamic adjustment parameter objects, because these two dimensions are greatly affected by market fluctuations and need to be corrected in real time; relatively stable indicators such as fulfillment rate, product specification matching degree, certification qualification matching degree, and credit rating compliance degree are classified as static adjustment parameter objects, because these dimensions will not change drastically in the short term, and the initial calculation results can be directly used.

[0070] Additionally, it should be noted that the architecture and construction process of the AI ​​large language model in this solution can be understood as follows: This large language model can be based on the Transformer architecture, consisting of multiple stacked Transformer Blocks. Each layer includes a multi-head self-attention mechanism, a feedforward neural network, layer normalization, and residual connections. Training is divided into two stages: pre-training and fine-tuning. The pre-training stage is unsupervised learning. The model is input into a massive text corpus, which is segmented into token sequences. These sequences are mapped to high-dimensional vectors through an embedding layer, and then passed through Transformer Blocks layer by layer for contextual semantic encoding. Finally, the model predicts the probability distribution of the next token, using the cross-entropy loss function as the optimization objective. Through backpropagation, the model's total number of parameters can reach billions to hundreds of billions, the hidden layer dimension can reach tens of thousands, and the number of attention heads can reach nearly a hundred. This stage enables the model to learn the grammar, semantics, and logical relationships of the language.

[0071] In this approach, the semantic understanding capabilities endowed by pre-training directly support the generation of demand feature vectors. Specifically, the Embedding layer maps key entities of the demand side, such as ISO9001 and a delivery period of 30 days, into high-dimensional semantic vectors, which serve as semantic query vectors for approximate nearest neighbor retrieval in the product-enterprise association subgraph. The multi-head self-attention mechanism captures the implicit relationships between various dimensions of demand features, such as the trade-off between low price and fast delivery, and is used to generate a weighted fusion product specification matching degree. The context window encodes the complete demand description, constraint vector, and preference vector of the demand side in one go, ensuring that contextual information is not lost during multi-dimensional calculations.

[0072] The fine-tuning phase falls under supervised learning, typically combined with reinforcement learning with human feedback (RLHF). Based on the pre-trained model, supervised fine-tuning is performed using manually labeled instruction response data. The generated results are then scored and ranked using a reward model, and reinforcement learning algorithms are used to further optimize the output quality. In this approach, fine-tuning enables the model to automatically perform six-dimensional matching calculations based on demand feature vectors, including product specification matching, price range overlap, geographical location suitability, certification matching, fulfillment rate matching, and credit rating compliance. Simultaneously, RLHF teaches the model how to prioritize when price range overlap is high but fulfillment rate matching is low, ultimately outputting a recommendation ranking that better aligns with business preferences. The real-time market bias adjustment phase is similar to online learning, dynamically correcting existing matching results using real-time market data. This is equivalent to adding a real-time error correction module to the model, making the final recommendation more closely aligned with current market realities.

[0073] As can be seen, this optional embodiment can effectively overcome the problem of static knowledge graph matching results being out of sync with current market conditions by dynamically correcting the similarity calculation results of candidate objects through the introduction of real-time market data. Simultaneously, by identifying parameters significantly affected by market fluctuations as dynamically adjustable objects and correcting them in real time, while retaining relatively stable parameters as statically adjustable objects, it ensures that the matching results are sensitive to market changes and reduces ranking distortion caused by over-adjustment. Finally, the adjusted results obtained after normalization and fusion make the final recommendation more aligned with the current market reality, significantly improving the timeliness and accuracy of candidate object selection.

[0074] Example 3 Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of a B2B intelligent supply and demand matching device based on an AI large language model, as disclosed in an embodiment of the present invention. Figure 3 As shown, this AI-based large language model-based B2B intelligent supply and demand matching device may include: The acquisition module 301 is used to acquire business requirement information of target objects with matching needs; The generation module 302 is used to generate similarity calculation results of candidate objects corresponding to the target object based on business requirement information and a pre-built enterprise knowledge graph. The acquisition module 301 is also used to acquire real-time market data associated with business needs information; The adjustment module 303 is used to adjust the similarity calculation results of candidate objects based on real-time market data to obtain the adjusted similarity calculation results of candidate objects. The determination module 304 is used to determine the list of target candidate objects corresponding to the target object based on the adjusted candidate object similarity calculation results; The generation module 302 is also used to input the object information and business requirement information of each target candidate object in the target candidate object list into the AI ​​large language model to generate natural language recommendation descriptions for all target candidate objects.

[0075] In this embodiment of the invention, the natural language recommendation description for all target candidate objects includes the business matching reason between the target candidate object and the target object.

[0076] Furthermore, the enterprise knowledge graph is constructed in the following ways: Acquire multi-source enterprise data; multi-source enterprise data includes enterprise registration data, product catalog data, historical transaction data, qualification certification data, and credit rating data; Using an AI large language model, entity recognition, relation extraction, and attribute alignment operations are performed on multi-source enterprise data to obtain structured graph triples. The structured graph triples include entity nodes, attribute nodes, and relation edges. Among them, entity nodes include enterprise entities, product entities, and industry entities, and relation edges include at least one of supply relations, procurement relations, competitive relations, and cooperative relations. Based on the structured graph triples, an initial enterprise knowledge graph is constructed, and vectorized encoding is performed on each entity node in the initial enterprise knowledge graph to obtain the graph embedding representation corresponding to the initial enterprise knowledge graph. When the latest enterprise data is obtained, the attribute nodes and relationship edges of the initial enterprise knowledge graph are incrementally expanded using the latest enterprise data, and the graph embedding identifier is updated synchronously to obtain the enterprise knowledge graph.

[0077] It is evident that implementation Figure 3 The described AI-based large language model-driven B2B intelligent supply and demand matching device can generate similarity calculation results for candidate objects based on the target object's business needs information and enterprise knowledge graph. It then adjusts these results based on real-time market data to determine a list of target candidate objects and generates natural language recommendations for each candidate object in the list, allowing the target object to query in real time. Compared to traditional matching methods, this improves the accuracy and timeliness of B2B supply and demand matching and overcomes matching deviations caused by changes in the market environment, thereby increasing the success rate of matching and transaction conversion efficiency.

[0078] In an optional embodiment, the target object includes a purchaser or supplier, and the target object's business needs information includes the target object's product procurement needs, price constraint needs, geographical location needs, qualification threshold needs, performance preference information, and credit rating needs. Specifically, the generation module 302 generates the similarity calculation results of the candidate objects corresponding to the target object based on business requirement information and a pre-built enterprise knowledge graph in the following ways: Based on the AI ​​large language model, deep semantic parsing and vector generation operations are performed on the business requirement information to obtain the requirement feature vector corresponding to the business requirement information. Based on the demand feature vector, perform multi-dimensional similarity calculation on the pre-built enterprise knowledge graph to obtain the similarity calculation results of the candidate objects corresponding to the target object.

[0079] In this optional embodiment, the demand feature vector includes a key entity vector, an attribute parameter vector, a constraint vector, and an implicit preference vector; the candidate object similarity calculation results include the product specification matching degree, price range overlap degree, geographical location adaptation degree, certification qualification matching degree, fulfillment rate matching degree, and credit rating compliance degree between the target object and each candidate object.

[0080] It is evident that implementation Figure 3 The described AI-based intelligent B2B supply and demand matching device can perform deep semantic analysis on business demand information through the AI ​​large language model. It deconstructs relevant demand information into four types of feature vectors: key entities, attribute parameters, constraints, and implicit preferences. This effectively captures implicit demands that are difficult to identify by traditional keyword matching, improving the accuracy of supply and demand matching. At the same time, it integrates multi-dimensional similarity calculation of enterprise knowledge graphs with real-time market data dynamic adjustment, so that the matching results conform to static enterprise profiles and dynamic market environments, overcoming the lag of static matching. Finally, it uses the AI ​​large language model to generate natural language recommendation explanations containing business matching reasons, enhancing the interpretability of recommendation results, reducing the decision-making cost of target objects, and improving the intelligence level and transaction conversion efficiency of B2B supply and demand matching.

[0081] In another optional embodiment, the generation module 302 performs deep semantic parsing and vector generation operations on the business requirement information based on the AI ​​large language model to obtain the requirement feature vector corresponding to the business requirement information. Specifically, this includes: A structured prompt template is constructed, and business requirement information is input into the AI ​​large language model. The AI ​​large language model then performs key entity recognition based on the structured prompt template to obtain multiple target entities in the business requirement information. All target entities are semantically encoded to obtain the key entity vector corresponding to the business requirement information. The quantitative parameter information in the business requirement information is input into the AI ​​big language model, so that the AI ​​big language model performs unit normalization and numerical encoding operations on the quantitative parameter information to obtain the attribute parameter vector corresponding to the business requirement information. The restrictive statement information in the business requirement information is input into the AI ​​big language model so that the AI ​​big language model can determine multiple target constraint information in the restrictive statement information and map all target constraint information into the constraint condition vector corresponding to the business requirement information. The non-explicit requirement statements in the business requirement information are input into the AI ​​large language model, so that the AI ​​large language model can perform contextual semantic reasoning on the non-explicit requirement statements to obtain the implicit preference vector corresponding to the business requirement information.

[0082] In this optional embodiment, all target entities include at least one of product entities, industry entities, and specification entities; all target constraint information includes price range constraint information, geographical range constraint information, time window constraint information, and qualification threshold constraint information; the implicit preference vector includes performance stability preference vector, cooperation flexibility preference vector, and credit rating preference vector.

[0083] It is evident that implementation Figure 3 The described AI-based large language model-driven B2B intelligent supply and demand matching device decomposes business demand information into four dimensions: key entities, quantifiable parameters, restrictive statements, and implicit requirements. It then generates corresponding entity vectors, parameter vectors, constraint vectors, and implicit preference vectors, achieving multi-granular, structured semantic representation of business demands. This allows for the precise capture of explicit entity information such as products, industries, and specifications, as well as the effective quantification of hard constraints like price ranges and time windows. Furthermore, semantic reasoning can uncover implicit preferences such as performance stability and cooperation flexibility, significantly improving the semantic completeness and matching accuracy of demand vectors. This provides a more comprehensive and accurate data foundation for subsequent intelligent recommendations and decision-making.

[0084] In another optional embodiment, the generation module 302 performs multi-dimensional similarity calculation operations on the pre-built enterprise knowledge graph based on the required feature vector, and obtains the similarity calculation results of the candidate objects corresponding to the target object in the following specific ways: Using key entity vectors as semantic query vectors and attribute parameter vectors as precision constraint vectors, an approximate nearest neighbor search is performed on the product-enterprise association subgraph of the pre-constructed enterprise knowledge graph to obtain a set of candidate objects. The cosine similarity between the key entity vector and the product node entity vector of each candidate object in the candidate object set is calculated as the semantic matching score, and the Euclidean distance between the attribute parameter vector and the product node parameter vector of each candidate object is calculated as the parameter deviation score. The semantic matching score and parameter deviation score between the target object and each candidate object are weighted and fused to obtain the product specification matching degree between the target object and each candidate object. The overlap between the price range constraint vector in the constraint vector and the product price range vector of each candidate object in the enterprise knowledge graph is calculated to obtain the price range overlap between the target object and each candidate object. Based on the geographical range constraint vector and time window constraint vector in the constraint condition vector, as well as the transaction address coordinate vector and transaction time window vector of each candidate object in the enterprise knowledge graph, the spherical transaction distance parameter between the target object and each candidate object is calculated, and the geographical location fit between the target object and each candidate object is determined based on the spherical transaction distance parameter between the target object and each candidate object.

[0085] In this optional embodiment, the generation module 302 further performs multi-dimensional similarity calculation operations on the pre-built enterprise knowledge graph based on the required feature vector to obtain the similarity calculation results of the candidate objects corresponding to the target object. Specifically, this method also includes: Based on the qualification threshold constraint vector in the constraint condition vector and the qualification node vector of each candidate object in the enterprise knowledge graph, calculate the qualification inclusion ratio between the target object and each candidate object, which is used as the certification qualification matching degree between the target object and each candidate object. Based on the performance stability preference vector, the cooperation flexibility preference vector, and the historical performance node vector of each candidate object in the enterprise knowledge graph, a performance weighted matching operation is performed between the target object and each candidate object to obtain the performance rate matching degree between the target object and each candidate object. Based on the credit rating preference vector and the credit rating node vector of each candidate object in the enterprise knowledge graph, a credit rating mapping and matching operation is performed between the target object and each candidate object to obtain the credit rating conformity between the target object and each candidate object.

[0086] It is evident that implementation Figure 3 The described AI-based large language model-driven B2B intelligent supply and demand matching device can perform multi-dimensional similarity calculations on enterprise knowledge graphs, covering six matching dimensions: product specifications, price range, geographical location, certification qualifications, performance, and credit rating. This achieves comprehensive object screening, from semantic understanding to numerical verification, and from soft preferences to hard constraints. Specifically, the weighted fusion of semantic matching and parameter bias balances intent understanding and accuracy control; interval overlap and spherical distance calculations ensure the objective accuracy of price and geographical dimensions; and the mapping of qualification inclusion ratio and credit rating reduces the entry of non-compliant objects into the candidate set through hard constraints. This significantly improves the overall matching quality and recommendation reliability between candidate and target objects.

[0087] In another optional embodiment, the adjustment module 303 adjusts the candidate object similarity calculation results based on real-time market data, and the specific methods for obtaining the adjusted candidate object similarity calculation results include: Based on real-time market data and corresponding historical benchmark data in the enterprise knowledge graph, the target dimension market deviation between the target object and each candidate object is calculated. Based on the target dimension market deviation between the target object and each candidate object, the dynamic adjustment parameter object and the static adjustment parameter object are determined from the candidate object similarity calculation results. Based on the target dimension market deviation between the target object and each candidate object, the dynamic adjustment parameter object between the target object and each candidate object is corrected to obtain the corrected dynamic adjustment parameter object between the target object and each candidate object. A normalization fusion operation is performed on the static adjustment parameter objects between the target object and each candidate object, as well as the corrected dynamic adjustment parameter objects, to obtain the adjusted candidate object similarity calculation results.

[0088] In this optional embodiment, the target dimension market deviation includes price dimension market deviation, geographic location dimension market deviation, and fulfillment rate dimension market deviation.

[0089] It is evident that implementation Figure 3 The described AI-based large language model-driven B2B intelligent supply and demand matching device can dynamically correct the similarity calculation results of candidate objects by introducing real-time market data, effectively overcoming the problem of static knowledge graph matching results being out of touch with current market conditions. Simultaneously, by identifying parameters significantly affected by market fluctuations as dynamically adjustable objects and correcting them in real time, while retaining relatively stable parameters as static adjustment objects, it ensures that the matching results are sensitive to market changes and reduces ranking distortion caused by over-adjustment. Finally, the adjusted results obtained after normalization and fusion make the final recommendation more aligned with the current market reality, significantly improving the timeliness and accuracy of candidate object selection.

[0090] Example 4 Please see Figure 4 , Figure 4 This is a schematic diagram of another B2B intelligent supply and demand matching device based on an AI large language model disclosed in an embodiment of the present invention. Figure 4 As shown, this AI-based large language model-based B2B intelligent supply and demand matching device may include: Memory 401 storing executable program code; Processor 402 coupled to memory 401; The processor 402 calls the executable program code stored in the memory 401 to execute the steps in the B2B intelligent supply and demand matching method based on AI large language model described in Embodiment 1 or Embodiment 2 of the present invention.

[0091] Example 5 This invention discloses a computer storage medium storing computer instructions. When these computer instructions are invoked, they are used to execute the steps in the AI-based large language model-based B2B intelligent supply and demand matching method described in Embodiment 1 or Embodiment 2 of this invention.

[0092] Example 6 This invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to perform the steps in the AI-based large language model-based B2B intelligent supply and demand matching method described in Embodiment 1 or Embodiment 2.

[0093] The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0094] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0095] Finally, it should be noted that the B2B intelligent supply and demand matching method and apparatus based on AI large language model disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention, and are only used to illustrate the technical solutions of the present invention, not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A B2B intelligent supply and demand matching method based on an AI large language model, characterized in that, The method includes: Obtain business requirement information of target objects with matching needs, and generate candidate object similarity calculation results corresponding to the target objects based on the business requirement information and the pre-built enterprise knowledge graph; Obtain real-time market data associated with the business requirement information, and adjust the similarity calculation result of the candidate objects based on the real-time market data to obtain the adjusted similarity calculation result of the candidate objects; Based on the adjusted similarity calculation results of the candidate objects, a list of target candidate objects corresponding to the target object is determined; The object information of each target candidate object in the target candidate object list and the business requirement information are input into the AI ​​large language model to generate natural language recommendation descriptions for all target candidate objects; the natural language recommendation descriptions for all target candidate objects include the business matching reasons between the target candidate object and the target object.

2. The B2B intelligent supply and demand matching method based on an AI large language model according to claim 1, characterized in that, The target objects include purchasers or suppliers, and the business needs information of the target objects includes product procurement needs, price constraint needs, geographical location needs, qualification threshold needs, performance preference information, and credit rating needs. The step of generating candidate object similarity calculation results corresponding to the target object based on the business requirement information and the pre-constructed enterprise knowledge graph includes: The AI ​​large language model performs deep semantic parsing and vector generation operations on the business requirement information to obtain the requirement feature vector corresponding to the business requirement information; the requirement feature vector includes key entity vector, attribute parameter vector, constraint condition vector and implicit preference vector. Based on the demand feature vector, a multi-dimensional similarity calculation operation is performed on the pre-constructed enterprise knowledge graph to obtain the similarity calculation results of the candidate objects corresponding to the target object. The similarity calculation results of the candidate objects include the matching degree of product specifications, the overlap of price range, the adaptability of geographical location, the matching degree of certification qualifications, the matching degree of fulfillment rate, and the compliance degree of credit rating between the target object and each candidate object.

3. The B2B intelligent supply and demand matching method based on an AI large language model according to claim 2, characterized in that, The step of performing deep semantic parsing and vector generation operations on the business requirement information based on the AI ​​large language model to obtain the requirement feature vector corresponding to the business requirement information includes: A structured prompt template is constructed, and the business requirement information is input into the AI ​​large language model, so that the AI ​​large language model performs a key entity recognition operation based on the structured prompt template to obtain multiple target entities in the business requirement information, and performs semantic encoding on all the target entities to obtain the key entity vector corresponding to the business requirement information; all the target entities include at least one of product entities, industry entities, and specification entities; The quantitative parameter information in the business requirement information is input into the AI ​​large language model, so that the AI ​​large language model performs unit normalization and numerical encoding operations on the quantitative parameter information to obtain the attribute parameter vector corresponding to the business requirement information. The restrictive statements in the business requirement information are input into the AI ​​large language model, so that the AI ​​large language model can determine multiple target constraint information in the restrictive statements and map all the target constraint information into a constraint condition vector corresponding to the business requirement information; all the target constraint information includes price range constraint information, geographical range constraint information, time window constraint information, and qualification threshold constraint information. The non-explicit requirement description information in the business requirement information is input into the AI ​​large language model, so that the AI ​​large language model performs contextual semantic reasoning operation on the non-explicit requirement description information to obtain the implicit preference vector corresponding to the business requirement information; the implicit preference vector includes a performance stability preference vector, a cooperation flexibility preference vector, and a credit rating preference vector.

4. The B2B intelligent supply and demand matching method based on an AI large language model according to claim 3, characterized in that, The step of performing multi-dimensional similarity calculation on the pre-constructed enterprise knowledge graph based on the demand feature vector to obtain the similarity calculation result of the candidate object corresponding to the target object includes: Using the key entity vector as the semantic query vector and the attribute parameter vector as the precision constraint vector, an approximate nearest neighbor search is performed in the product-enterprise association subgraph of the pre-constructed enterprise knowledge graph to obtain a candidate object set. The cosine similarity between the key entity vector and the product node entity vector of each candidate object in the candidate object set is calculated as the semantic matching score, and the Euclidean distance between the attribute parameter vector and the product node parameter vector of each candidate object is calculated as the parameter deviation score. The semantic matching score and parameter deviation score between the target object and each candidate object are weighted and fused to obtain the product specification matching degree between the target object and each candidate object. The price range constraint vector in the constraint vector is compared with the product price range vector of each candidate object in the enterprise knowledge graph to calculate the overlap of the price range between the target object and each candidate object. Based on the geographical range constraint vector and time window constraint vector in the constraint condition vector, as well as the transaction address coordinate vector and transaction time window vector of each candidate object in the enterprise knowledge graph, the spherical transaction distance parameter between the target object and each candidate object is calculated, and the geographical location fit between the target object and each candidate object is determined based on the spherical transaction distance parameter between the target object and each candidate object.

5. The B2B intelligent supply and demand matching method based on an AI large language model according to claim 4, characterized in that, The step of performing a multi-dimensional similarity calculation operation on the pre-constructed enterprise knowledge graph based on the demand feature vector to obtain the similarity calculation result of the candidate object corresponding to the target object also includes: Based on the qualification threshold constraint vector in the constraint condition vector and the qualification node vector of each candidate object in the enterprise knowledge graph, the qualification inclusion ratio between the target object and each candidate object is calculated as the certification qualification matching degree between the target object and each candidate object; Based on the performance stability preference vector, the cooperation flexibility preference vector, and the historical performance node vector of each candidate object in the enterprise knowledge graph, a performance weighted matching operation is performed between the target object and each candidate object to obtain the performance rate matching degree between the target object and each candidate object; Based on the credit rating preference vector and the credit rating node vector of each candidate object in the enterprise knowledge graph, a credit rating mapping and matching operation is performed between the target object and each candidate object to obtain the credit rating conformity between the target object and each candidate object.

6. The B2B intelligent supply and demand matching method based on an AI large language model according to any one of claims 2-5, characterized in that, The step of adjusting the similarity calculation result of the candidate objects based on the real-time market data to obtain the adjusted similarity calculation result of the candidate objects includes: Based on the real-time market data and the corresponding historical benchmark data in the enterprise knowledge graph, the target dimension market deviation degree between the target object and each candidate object is calculated. Then, based on the target dimension market deviation degree between the target object and each candidate object, dynamic adjustment parameter objects and static adjustment parameter objects are determined from the candidate object similarity calculation results. The target dimension market deviation degree includes price dimension market deviation degree, geographic location dimension market deviation degree, and fulfillment rate dimension market deviation degree. Based on the target dimension market deviation between the target object and each candidate object, the dynamic adjustment parameter object between the target object and each candidate object is corrected to obtain the corrected dynamic adjustment parameter object between the target object and each candidate object. A normalization fusion operation is performed on the static adjustment parameter object and the modified dynamic adjustment parameter object between the target object and each candidate object to obtain the adjusted similarity calculation result of the candidate object.

7. The B2B intelligent supply and demand matching method based on an AI large language model according to any one of claims 1-5, characterized in that, The enterprise knowledge graph is constructed in the following way: Acquire multi-source enterprise data; the multi-source enterprise data includes enterprise business registration data, product catalog data, historical transaction data, qualification certification data, and credit rating data; The AI ​​large language model is used to perform entity recognition, relation extraction and attribute alignment operations on the multi-source enterprise data to obtain structured graph triples; The structured graph triple includes entity nodes, attribute nodes, and relationship edges; wherein, the entity nodes include enterprise entities, product entities, and industry entities, and the relationship edges include at least one of supply relationships, procurement relationships, competitive relationships, and cooperative relationships; Based on the structured graph triples, an initial enterprise knowledge graph is constructed, and a vectorized encoding operation is performed on each entity node in the initial enterprise knowledge graph to obtain the graph embedding representation corresponding to the initial enterprise knowledge graph. When the latest enterprise data is obtained, the attribute nodes and relationship edges of the initial enterprise knowledge graph are incrementally expanded using the latest enterprise data, and the graph embedding identifier is updated synchronously to obtain the enterprise knowledge graph.

8. A B2B intelligent supply and demand matching device based on an AI large language model, characterized in that, The device includes: The acquisition module is used to acquire business requirement information of target objects that have matching needs; The generation module is used to generate similarity calculation results of candidate objects corresponding to the target object based on the business requirement information and the pre-built enterprise knowledge graph. The acquisition module is also used to acquire real-time market data associated with the business demand information; The adjustment module is used to adjust the similarity calculation result of the candidate objects based on the real-time market data to obtain the adjusted similarity calculation result of the candidate objects; The determination module is used to determine the list of target candidate objects corresponding to the target object based on the adjusted similarity calculation results of the candidate objects; The generation module is further configured to input the object information of each target candidate object in the target candidate object list and the business requirement information into the AI ​​large language model to generate natural language recommendation descriptions for all target candidate objects; the natural language recommendation descriptions for all target candidate objects include the business matching reasons between the target candidate object and the target object.

9. A B2B intelligent supply and demand matching device based on an AI large language model, characterized in that, The device includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the B2B supply and demand intelligent matching method based on an AI large language model as described in any one of claims 1-7.

10. A computer storage medium, characterized in that, The computer storage medium stores computer instructions, which, when invoked, are used to execute the B2B intelligent supply and demand matching method based on an AI large language model as described in any one of claims 1-7.

Images