An intelligent matching algorithm based on multi-dimensional knowledge fusion and large model deep thinking

Abstract

This invention discloses an intelligent matching algorithm based on multi-dimensional knowledge fusion and deep thinking of a large model, belonging to the fields of natural language processing and machine learning. The algorithm includes: S1 collecting and preprocessing data around five core indicators; S2 constructing a high-dimensional vector knowledge base for the meat industry; S3 filtering 50-100 candidate orders through a hybrid vector and keyword search; S4 dynamically adjusting weights based on the scenario, and linearly normalizing the scoring and ranking; S5 re-ranking through deep reasoning using an LLM large model, outputting 3-5 accurate recommendations; S6 analyzing the reasons based on feedback from both buyers and sellers, and optimizing the model, knowledge base, and indicator system using LLM to form a closed loop. This invention provides an intelligent matching algorithm based on multi-dimensional knowledge fusion and deep thinking of a large model, achieving full-dimensional accurate matching of buy and sell orders in the meat industry, intelligent adaptation to dynamic scenarios, and controllable transaction risks, significantly improving the matching success rate and industry transaction efficiency.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and machine learning, and in particular to an intelligent matching algorithm based on multi-dimensional knowledge fusion and deep thinking of large models. Background Technology

[0002] In the meat distribution sector, traditional matching technology has long relied on a matching logic of "keywords + basic structured fields," which can only perform simple filtering based on explicit indicators such as "variety, quantity, and price." This is no longer suitable for the diversified and scenario-based transaction needs of the industry. The specific shortcomings are as follows: 1. Incomplete coverage of demand and lack of implicit needs: Traditional technologies only focus on structured fields and do not include core dimensions of the meat industry such as "quarantine standards, transportation timeliness, storage conditions, and supply stability". At the same time, they cannot analyze implicit needs in specific scenarios. For example, in the buyer's demand for "200 catties of free-range black pig belly for high-end catering", the implicit needs of "freshness over price and stable delivery frequency" cannot be identified, resulting in a disconnect between recommended orders and actual needs.

[0003] 2. Rigid matching logic and poor scenario adaptability: The existing solution uses fixed weight scoring, which cannot dynamically adjust the priority of indicators according to the scenario. For example, in the scenario of "emergency replenishment", the importance of "transportation timeliness" should be higher than "price", but the fixed weight will cause "high timeliness but slightly higher price" orders to be ranked behind "low price but delayed" orders, resulting in extremely low response efficiency.

[0004] 3. The contradiction between search efficiency and accuracy: Single keyword search takes more than 10 seconds in a large number of orders, while simple vector search lacks industry knowledge association, resulting in the dilemma of "fast but inaccurate matching" or "accurate but inefficient matching", which cannot meet the dual requirements of efficiency and accuracy of the matching business.

[0005] 4. Insufficient compliance and risk control: Traditional technologies do not incorporate compliance and risk dimensions such as "quarantine qualifications and user creditworthiness" into core matching indicators, which can easily lead to invalid matching where "order fields match but quarantine is not compliant", increasing transaction risks.

[0006] In summary, the meat industry urgently needs an intelligent matching solution that can cover all dimensions of indicators, uncover hidden needs, and adapt to dynamic scenarios to solve the problems of inefficient matching and low accuracy of traditional technologies. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent matching algorithm based on multi-dimensional knowledge fusion and deep thinking of large models, which solves the problems of the existing technology, such as one-sided demand coverage, inability to identify implicit demands, poor scenario adaptability, contradiction between search efficiency and accuracy, and insufficient compliance risk management.

[0008] To achieve the above objectives, this invention provides an intelligent matching algorithm based on multi-dimensional knowledge fusion and deep thinking of large models. Targeting the meat industry's buy-sell order matching scenario, it constructs a six-module architecture including data acquisition, vector knowledge base construction, hybrid retrieval, dynamic scoring, LLM inference, and feedback iteration. Specifically, it includes the following steps: S1. Multi-dimensional data collection and preprocessing: Collect data around 5 major categories of core matching indicators and preprocess the collected data; S2. Construction of a high-dimensional vector knowledge base for the meat industry; S3, Multi-dimensional Hybrid Search: Convert the demand text into a vector for vector search, simultaneously perform keyword search and matching of structured indicators, filter to form a candidate order pool, and control the search response time within 1 second; S4. Dynamic scoring: By identifying the scenario through keywords in the demand text, the indicator weights are automatically adjusted, and the linear normalization algorithm is used to score and initially rank the orders in the candidate order pool. S5. Based on the LLM large model, perform deep reasoning and reordering on the orders in the candidate pool after multi-dimensional hybrid retrieval and dynamic scoring, and output 3-5 accurate matching recommendation orders; S6. Feedback Iteration: Based on the feedback from both buyers and sellers regarding the matching results, the root causes (unidentified implicit needs, uncoordinated conflicting indicators) are analyzed through the LLM large model. The model's semantic analysis weights are optimized, industry knowledge and user behavior features are added to the knowledge base, and the multi-dimensional indicator system and scoring algorithm are dynamically adjusted (such as upgrading "delivery frequency" to a structured indicator). This forms a closed loop of feedback-analysis-optimization, continuously improving the inference accuracy and matching success rate of the large model.

[0009] Preferably, the five core matching indicators in step S1 include basic transaction dimensions, compliance and quarantine dimensions, logistics and storage dimensions, business cooperation dimensions, and risk warning dimensions. Among them, the basic transaction dimension includes meat type, quantity, unit price, delivery and receipt locations, and delivery cycle; the compliance and quarantine dimension includes quarantine certificates, factory inspection reports, non-epidemic area certificates, and vehicle disinfection certificates; the logistics and storage dimension includes transportation methods, transportation timeliness, and packaging standards; the business cooperation dimension includes payment methods, payment terms, and invoice types; and the risk warning dimension is the user's credit score.

[0010] Preferably, the preprocessing method in step S1 specifically includes the following steps: S11. Clean, normalize, and standardize the structured data; S12. Transform the unstructured text data into a 1024-dimensional high-dimensional dense vector using the embeddings general text vector model; S13. Synchronize dynamic data in real time through API interface, connect to user credit rating model, obtain user credit rating, and support multi-dimensional knowledge fusion in dynamic scenarios.

[0011] Preferably, step S2 specifically includes the following steps: S21. Vector storage optimization: The preprocessed 1024-dimensional vector is stored in ES, and the HNSW algorithm is used to build the index, which improves the retrieval speed by 3 times compared with the traditional algorithm. S22. Hybrid Index Design: Simultaneously construct a hybrid index that combines structured field indexes (variety, quarantine qualifications, etc.) and vector indexes (semantic similarity) to achieve dual retrieval channels and resolve the contradiction between "precise matching + efficient retrieval". S23. Knowledge base update mechanism: Establish a knowledge base update mechanism that updates industry knowledge monthly (such as adding quarantine standards and transportation specifications) and optimizes vector models quarterly.

[0012] Preferably, the scoring formula in step S4 is: The indicator score is a quantitative score of 0-10 points, obtained based on the degree of matching between each matching indicator of the order and the buyer's needs.

[0013] Preferably, step S5 specifically includes the following steps: S51. Deep Semantic and Contextualized Needs Mining: For orders in the candidate pool after multi-dimensional hybrid retrieval and dynamic scoring, the natural language understanding capabilities of the LLM large model are used to deeply analyze the implicit needs and contextualized information in the orders. S52, Cross-dimensional knowledge association and conflict coordination: Through the knowledge reasoning capabilities of the LLM large model, cross-dimensional association verification and conflict coordination are performed on the multi-dimensional indicators of candidate orders; S53. Final output of accurate matching results: Based on the multi-dimensional dynamic scoring results and the deep reasoning conclusions of the LLM large model, the candidate orders are finally adjusted and ranked.

[0014] Preferably, the feedback from both the buyer and seller on the matching result in step S6 includes: whether they accept the recommendation, the reason for rejection, and their satisfaction with the performance.

[0015] Therefore, the present invention employs the above-mentioned intelligent matching algorithm based on multi-dimensional knowledge fusion and deep thinking of large models, which has the following beneficial effects: (1) Improve the success rate of matching: Comprehensive indicator coverage and hidden demand mining (such as "freshness first" in high-end catering) reduce matching failures caused by incomplete understanding of demand. In practical applications, it can significantly improve the matching success rate of meat orders. (2) Improve matching efficiency: Fast retrieval within 1 second + initial screening of 50-100 candidate pools + precise recommendation of Top 3-5, shorten the matching cycle of buy and sell orders, reduce communication costs for both parties, and improve the efficiency of industry transactions; (3) Ensure transaction compliance and risk controllability: Incorporate compliance and quarantine dimensions (quarantine tickets, non-epidemic area certificates) and risk warning dimensions (user credit) into core indicators, and combine them with LLM cross-dimensional verification to reduce non-compliant order matching and reduce transaction risks; (4) Adaptable to diverse scenarios: The dynamic weight adjustment and conflict coordination capabilities can adapt to different scenarios such as "instant supply", "bulk purchase" and "high-end catering supply", meet the diverse transaction needs of the meat industry and enhance the versatility of the solution.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is an overall flowchart of an embodiment of the present invention; Figure 2 This is a flowchart illustrating the multi-dimensional data acquisition process according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating the multi-dimensional data preprocessing process according to an embodiment of the present invention. Figure 4 This is a flowchart illustrating the multi-dimensional hybrid retrieval process according to an embodiment of the present invention; Figure 5 This is a flowchart illustrating the multi-dimensional dynamic scoring and ranking process according to an embodiment of the present invention. Figure 6 This is a flowchart of the LLM-based deep inference reordering process according to an embodiment of the present invention; Figure 7 This is a flowchart illustrating the feedback iteration process in an embodiment of the present invention. Detailed Implementation

[0018] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0019] Example 1: High-end catering procurement scenario Demand Input: The buyer's demand is "200 catties of free-range black pig belly for daily supply to high-end catering stores, cold chain transportation, payment term of 7 days, and a quarantine certificate and a certificate of non-epidemic area required".

[0020] The order requirements are processed as follows, such as... Figure 1 As shown: I. Multi-dimensional data collection and preprocessing: Collect basic transaction dimension data in this requirement, such as Figure 2 The specifications are as follows: Variety: Free-range black pork belly; Quantity: 200 jin (converted to 100 kg); Shipping and receiving locations: (buyer's store address and seller's possible shipping address); Delivery period: (determined within a reasonable range based on daily supply needs); Compliance and quarantine data: Quarantine certificate and non-epidemic area certificate required; Logistics and storage data: Cold chain transportation, packaging standards (meeting high-end catering ingredient requirements); Business cooperation data: 7-day payment period, payment method (default or buyer-specified), invoice type (default or buyer-specified); Risk warning data: Seller's user credit rating obtained through API interface.

[0021] The above data is preprocessed, such as... Figure 3 As shown: Structured data such as quantity and payment period are normalized and standardized; unstructured text in the demand description (such as "free-range black pig pork belly", "daily supply of high-end catering stores", "cold chain transportation", etc.) are transformed into 1024-dimensional high-dimensional dense vectors through the embeddings general text vector model.

[0022] II. Construction of High-Dimensional Vector Knowledge Base: A large amount of meat industry buy and sell order data has been preprocessed according to the above preprocessing method, stored in ES and indexed using the HNSW algorithm. At the same time, structured field indexes and vector indexes have been constructed. The knowledge base is maintained by a mechanism of updating industry knowledge monthly and optimizing vector models quarterly.

[0023] III. Multi-dimensional hybrid search, such as Figure 4 As shown Vector retrieval: The vector of the buyer's demand text is retrieved in Elasticsearch to match sales orders and industry knowledge related to "free-range black pig belly, cold chain transportation, and high-end catering supply".

[0024] Keyword search: Matching structured indicators such as "variety = black pork belly, quantity = 100kg, quarantine qualifications = complete (with quarantine certificate and non-epidemic area certificate), transportation method = cold chain transportation, payment period = 7 days".

[0025] Results screening: A pool of 80 candidate orders was formed, with a retrieval response time of 0.8 seconds, which is 15 times faster than the traditional technology (12 seconds).

[0026] IV. Dynamic scoring, such as Figure 5 As shown Scene recognition: By identifying the scene through the demand of "high-end catering, daily supply", the weight of "freshness (cold chain timeliness), supply stability (delivery frequency) and compliance (quarantine qualification)" is automatically increased from the usual 20% to 40%.

[0027] Scoring calculation: A linear normalization algorithm is used to score 80 candidate orders based on 21 indicators in 5 categories, and the top 10 orders are initially ranked.

[0028] V. LLM large-scale model deep inference reordering, such as Figure 6 As shown Deep semantic and scenario-based demand mining: The LLM model identifies the implicit demands of "high meat freshness and priority of supply stability" from the buyer's demand, and extracts the potential advantages of each seller from the seller descriptions of candidate orders (such as some sellers can provide daily cold chain delivery and support returns and exchanges).

[0029] Cross-dimensional knowledge association and conflict coordination: The big model combines "high-end catering demand" and "quarantine qualification level" to prioritize matching orders with "green certified high-quality fresh meat"; for some orders, conflict coordination is carried out to select orders with higher matching degree for potential conflicts in indicators such as payment period and delivery cycle.

[0030] The final output is a precise matching result: combining the dynamic scoring results and LLM deep reasoning conclusions, the top 10 orders are finally adjusted and sorted by weight, and the 3-5 orders with the highest overall suitability are selected as priority matching recommendations to the buyer.

[0031] VI. Feedback and iteration, such as Figure 7 As shown: After receiving a recommended order, the buyer provides feedback such as acceptance or reasons for rejection. If the buyer rejects part of the order, the LLM model analyzes the root cause of the rejection (such as implicit needs not being fully met, or conflicting metrics not being properly coordinated), optimizes the model's semantic parsing weights, supplements the vector knowledge base with relevant industry knowledge and user behavior characteristics, and dynamically adjusts the multi-dimensional indicator system and scoring algorithm to provide more accurate support for subsequent matching business.

[0032] Therefore, this invention adopts an intelligent matching algorithm based on multi-dimensional knowledge fusion and deep thinking of large models. By constructing a full-link architecture of "multi-dimensional indicator system + high-dimensional vector knowledge base + hybrid retrieval + dynamic scoring + LLM deep reasoning + feedback iteration", it systematically solves the core problems of traditional technologies in meat industry matching, such as one-sided demand coverage, difficulty in identifying implicit demands, rigid scenario adaptation, imbalance between efficiency and accuracy, and insufficient compliance risk control. Ultimately, it achieves full-dimensional accurate matching of buy and sell orders, dynamic intelligent adaptation to scenarios, and effective prevention and control of transaction risks, significantly improving the matching success rate and industry transaction efficiency, and meeting the diversified and scenario-based transaction needs of the meat industry.

[0033] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent matching algorithm based on multi-dimensional knowledge fusion and deep thinking of large models, characterized in that, For the meat industry's buy-sell order matching scenario, a six-module architecture is constructed, including data collection, vector knowledge base construction, hybrid retrieval, dynamic scoring, LLM inference, and feedback iteration. Specifically, the following steps are included: S1. Multi-dimensional data collection and preprocessing: Collect data around 5 major categories of core matching indicators and preprocess the collected data; S2. Construction of a high-dimensional vector knowledge base for the meat industry; S3, Multi-dimensional Hybrid Search: Convert the demand text into a vector for vector search, and simultaneously perform keyword search and structured index matching to filter and form a pool of 50-100 candidate orders; S4. Dynamic scoring: By identifying the scenario through keywords in the demand text, the indicator weights are automatically adjusted, and the linear normalization algorithm is used to score and initially rank the orders in the candidate order pool. S5. Based on the LLM large model, perform deep reasoning and reordering on the orders in the candidate pool after multi-dimensional hybrid retrieval and dynamic scoring, and output 3-5 accurate matching recommendation orders; S6. Feedback Iteration: Based on the feedback from both buyers and sellers regarding the matching results, the root causes are analyzed through the LLM large model, the semantic analysis weights of the model are optimized, industry knowledge and user behavior characteristics are added to the knowledge base, and the multi-dimensional indicator system and scoring algorithm are dynamically adjusted to form a closed loop of feedback-analysis-optimization.

2. The intelligent matching algorithm based on multi-dimensional knowledge fusion and deep thinking of large models according to claim 1, characterized in that, The five core matching indicators in step S1 include basic transaction dimensions, compliance and quarantine dimensions, logistics and storage dimensions, business cooperation dimensions, and risk warning dimensions.

3. The intelligent matching algorithm based on multi-dimensional knowledge fusion and large-scale model deep thinking as described in claim 2, characterized in that, The basic transaction dimensions include meat type, quantity, unit price, delivery and receipt locations, and delivery cycle; the compliance and quarantine dimensions include quarantine certificates, factory inspection reports, non-epidemic area certificates, and vehicle disinfection certificates; the logistics and storage dimensions include transportation methods, transportation timeliness, and packaging standards; the business cooperation dimensions include payment methods, payment terms, and invoice types; and the risk warning dimension is the user's credit score.

4. The intelligent matching algorithm based on multi-dimensional knowledge fusion and deep thinking of large models according to claim 1, characterized in that, The preprocessing method in step S1 specifically includes the following steps: S11. Clean, normalize, and standardize the structured data; S12. Transform the unstructured text data into a 1024-dimensional high-dimensional dense vector using the embeddings general text vector model; S13. Synchronize dynamic data in real time via API interface.

5. The intelligent matching algorithm based on multi-dimensional knowledge fusion and large-scale model deep thinking as described in claim 4, characterized in that, Step S2 specifically includes the following steps: S21. Vector storage optimization: Store the preprocessed 1024-dimensional vector into Elasticsearch and use the HNSW algorithm to build an index. S22. Hybrid Index Design: Simultaneously construct a hybrid index that combines structured field indexes and vector indexes; S23. Knowledge base update mechanism: Establish a knowledge base update mechanism that updates industry knowledge monthly and optimizes vector models quarterly.

6. The intelligent matching algorithm based on multi-dimensional knowledge fusion and large-scale model deep thinking as described in claim 1, characterized in that, The scoring formula in step S4 is: The indicator score is a quantitative score of 0-10 points, obtained based on the degree of matching between each matching indicator of the order and the buyer's needs.

7. The intelligent matching algorithm based on multi-dimensional knowledge fusion and large-scale model deep thinking as described in claim 1, characterized in that, Step S5 specifically includes the following steps: S51. Deep Semantic and Contextualized Needs Mining: For orders in the candidate pool after multi-dimensional hybrid retrieval and dynamic scoring, the natural language understanding capabilities of the LLM large model are used to deeply analyze the implicit needs and contextualized information in the orders. S52, Cross-dimensional knowledge association and conflict coordination: Through the knowledge reasoning capabilities of the LLM large model, cross-dimensional association verification and conflict coordination are performed on the multi-dimensional indicators of candidate orders; S53. Final output of accurate matching results: Based on the multi-dimensional dynamic scoring results and the deep reasoning conclusions of the LLM large model, the candidate orders are finally adjusted and ranked.

8. The intelligent matching algorithm based on multi-dimensional knowledge fusion and deep thinking of large models according to claim 1, characterized in that, The feedback from both the buyer and seller regarding the matchmaking results in step S6 includes: whether they accept the recommendation, the reason for rejection, and their satisfaction with the performance.

Images