A server access method for large models and actual business scenarios

Abstract

The application relates to the technical field of server access of large models and actual business scenes, and discloses a server access method for large models and actual business scenes, which comprises the following steps: configuring access parameters, including a protocol type identifier, an interface calling frequency and a data transmission format; obtaining a business interaction data set under parameter constraints; extracting a parameter identifier vector from the access parameters; performing integration processing on the business interaction data set based on feature correlation to obtain a business data correlation feature vector, the process comprising format conversion, analysis of a correlation reference point, calculation of a correlation direction and integration; performing multidimensional feature collaborative processing on the two vectors to obtain a parameter-business collaborative feature vector, the collaborative processing involving external business knowledge, an attention mechanism and a Transformer structure; and completing access based on the collaborative feature vector, wherein access verification comprises dimension alignment, initial verification and deep verification. The method improves the compatibility, dynamic adaptability and reliability of access.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a server access method for large models and actual business scenarios. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, large models have shown strong capabilities in natural language processing, data analysis, intelligent decision-making and other fields. The integration of large models with actual business scenarios has become a key direction to promote the intelligent upgrading of industries. However, there are significant technical differences and interaction barriers between large model systems and the servers of actual business scenarios, which pose many challenges to the efficient access of the two.

[0003] Actual business scenarios often have diverse protocol types, interface specifications and data formats. For example, transaction data transmission on e-commerce platforms may use HTTP protocol, while core business interactions in financial systems may rely on private protocols. At the same time, the interface call frequency of different business scenarios varies greatly, with high-concurrency scenarios (such as flash sales) and low-frequency interaction scenarios (such as background data statistics) requiring completely different response capabilities from servers. These differences make it easy for large models to encounter protocol incompatibility, data transmission delays, interface call conflicts and other problems during access, seriously affecting the stability and efficiency of business processes.

[0004] In existing technologies, the server access of large models and business scenarios often uses a fixed configuration scheme, i.e., achieving the connection by manually presetting protocol parameters, interface rules and data conversion methods. This approach lacks the ability to perceive and adapt to dynamic data characteristics during business interactions. When business scenarios change (such as interface call frequency adjustment, data format update), parameter configuration and system debugging need to be re-done, resulting in high access costs and poor flexibility. In addition, traditional access methods often ignore the internal relationships between business data and the synergistic relationship between access parameters and business data, and only complete the interaction through simple data format conversion, making it difficult to ensure the accuracy and integrity of data transmission, and thus affecting the understanding and response quality of large models to business needs.

[0005] With the continuous iteration of large model technology and the complexity of business scenarios, existing access methods have been unable to meet the access requirements of high compatibility, high dynamics and high reliability. Therefore, a server access method that can consider access parameters and business data characteristics and achieve multi-dimensional synergistic adaptation is needed to break down the technical barriers between large models and actual business scenarios and improve the interaction efficiency and stability of the two. SUMMARY

[0006] The purpose of the present application is to provide a server access method for large models and actual business scenarios to solve the problems raised in the background.

[0007] To achieve the above object, the application provides a server access method for large models and actual business scenarios, which comprises the following steps:

[0008] configuring access parameters, wherein the access parameters comprise protocol type identification, interface calling frequency and data transmission format;

[0009] obtaining business interaction data of a target business scenario and a large model system under the constraint of the access parameters to obtain a business interaction data set;

[0010] performing feature extraction processing on the access parameters to obtain a parameter identification vector;

[0011] performing feature correlation-based integration processing on the business interaction data set to obtain a business data correlation feature vector;

[0012] performing multi-dimensional feature collaborative processing on the business data correlation feature vector and the parameter identification vector to obtain a parameter-business collaborative feature vector;

[0013] completing server access of a large model and an actual business scenario based on the parameter-business collaborative feature vector.

[0014] Preferably, the feature correlation-based integration processing on the business interaction data set to obtain a business data correlation feature vector comprises the following steps:

[0015] respective business interaction data in the business interaction data set are converted by a format conversion unit to obtain a standard business data unit set;

[0016] analyzing a business data correlation reference point representation vector based on the feature distribution situation of the standard business data unit set;

[0017] calculating the correlation motion direction of each standard business data unit in the standard business data unit set relative to the business data correlation reference point representation vector to obtain a business data correlation direction set;

[0018] based on the business data correlation direction set, the standard business data unit set is integrated towards the business data correlation reference point representation vector to obtain the business data correlation feature vector.

[0019] Preferably, the analysis of the business data correlation reference point representation vector based on the feature distribution situation of the standard business data unit set comprises the following steps:

[0020] each standard business data unit in the standard business data unit set is input into a feature distribution measurement module to obtain a business data distribution measurement coefficient set;

[0021] inputting the set of service data distribution metric coefficients into a feature weight screening control unit to obtain a set of service data distribution weight factors;

[0022] based on the set of service data distribution weight factors, calculating a weighted sum of the set of standard service data units to obtain the service data correlation reference point representation vector.

[0023] Preferably, each standard service data unit in the set of standard service data units is inputted into a feature distribution metric module to obtain a set of service data distribution metric coefficients, including:

[0024] numerically standardizing the standard service data unit to obtain a normalized service data unit;

[0025] calculating the sum of the squares of each feature value in the normalized service data unit divided by the feature range value of the normalized service data unit to obtain the service data distribution metric coefficient.

[0026] Preferably, the service data correlation motion direction is the inverse cosine similarity value between the standard service data unit and the service data correlation reference point representation vector.

[0027] Preferably, the service data correlation feature vector and the parameter identification vector are subjected to multi-dimensional feature collaborative processing to obtain a parameter-service collaborative feature vector, including:

[0028] based on external business knowledge, the service data correlation feature vector and the parameter identification vector are subjected to feature interaction adjustment based on an attention mechanism to obtain a knowledge-optimized parameter-service collaborative feature matrix;

[0029] based on the knowledge-optimized parameter-service collaborative feature matrix, the service data correlation feature vector and the parameter identification vector are respectively subjected to feature adjustment optimization to obtain an optimized service data correlation feature vector and an optimized parameter identification vector;

[0030] calculating the point-by-point division between the optimized service data correlation feature vector and the optimized parameter identification vector to obtain the parameter-service collaborative feature vector.

[0031] Preferably, based on external business knowledge, the service data correlation feature vector and the parameter identification vector are subjected to feature interaction adjustment based on an attention mechanism to obtain a knowledge-optimized parameter-service collaborative feature matrix, including:

[0032] inputting the service data correlation feature vector and the parameter identification vector into a feature collaborative interaction network to obtain a parameter-service collaborative feature matrix;

[0033] input the parameter-business synergy feature matrix into an attention processing unit based on external business knowledge to obtain a knowledge-optimized parameter-business synergy feature matrix.

[0034] Preferably, based on the knowledge-optimized parameter-business synergy feature matrix, the business data association feature vector and the parameter identifier vector are respectively subjected to feature adjustment optimization to obtain an optimized business data association feature vector and an optimized parameter identifier vector, comprising:

[0035] The business data association feature vector is subjected to linear transformation processing to obtain a first query feature vector and a first value feature vector, and the knowledge-optimized parameter-business synergy feature matrix is used as a key matrix, the first query feature vector, the first value feature vector and the key matrix are input into a feature adjustment module based on a Transformer structure to obtain the optimized business data association feature vector.

[0036] The parameter identifier vector is subjected to linear transformation processing to obtain a second query feature vector and a second value feature vector, and the knowledge-optimized parameter-business synergy feature matrix is used as a key matrix, the second query feature vector, the second value feature vector and the key matrix are input into the feature adjustment module based on the Transformer structure to obtain the optimized parameter identifier vector.

[0037] Preferably, based on the parameter-business synergy feature vector, server access of a large model to an actual business scenario is completed, comprising:

[0038] The parameter-business synergy feature vector is input into an access verification module based on a determinator to obtain an access verification result, which is used to indicate whether the large model and the server of the actual business scenario have completed access.

[0039] Preferably, the parameter-business synergy feature vector is input into an access verification module based on a determinator to obtain an access verification result, comprising:

[0040] The parameter-business synergy feature vector is subjected to dimension alignment processing to obtain an aligned feature vector.

[0041] The aligned feature vector is input into a basic verification layer of the determinator for initial checking to obtain an initial checking mark.

[0042] The initial checking mark is input into a deep verification layer of the determinator for secondary checking to obtain the access verification result.

[0043] The parameter-business synergy feature vector is subjected to dimension alignment processing to obtain an aligned feature vector, comprising:

[0044] extracting a preset standard feature dimension parameter;

[0045] calculating a deviation amount of a current dimension parameter of the parameter-service coordination feature vector from the standard feature dimension parameter;

[0046] performing a dimension adjustment operation on the parameter-service coordination feature vector based on the deviation amount to obtain the aligned feature vector.

[0047] Compared with the prior art, the beneficial effects of the present application are:

[0048] The method realizes the standardization starting point of the access process by configuring access parameters such as protocol type identifier, interface call frequency and data transmission format, and obtaining service interaction data based on these parameter constraints. This parameterized configuration mode can accurately adapt to the individual needs of different business scenarios, avoiding protocol conflicts and interface incompatibility problems under the traditional fixed configuration mode, significantly improving the compatibility of large models with various actual business scenario servers, so that the same large model system can be flexibly connected to business platforms in different fields such as e-commerce, finance and medical treatment.

[0049] In the business interaction data processing link, standard business data units are obtained through format conversion, and based on feature distribution trend analysis, correlation reference points are obtained, and data integration is realized through correlation direction calculation, and finally business data correlation feature vectors are obtained. This process fully excavates the internal correlation relationship between business data, and converts scattered and heterogeneous business interaction data into correlation vectors with unified feature dimensions, not only eliminating the interference caused by data format differences, but also strengthening the overall feature expression of business data, so that the large model can more comprehensively and accurately understand the core needs of the business scenario, providing a high-quality data basis for subsequent interaction response.

[0050] The parameter identifier vector is obtained by feature extraction of the access parameter, and multi-dimensional collaborative processing is performed with the business data correlation feature vector, especially the introduction of external business knowledge and attention mechanism for feature interaction adjustment, realizing the deep fusion of access parameters and business data. This collaborative processing method breaks through the limitations of separate processing of parameters and data in traditional methods, so that the access parameters can dynamically adapt to the feature changes of the business data, and the business data can also be optimized according to the parameter constraints, and the collaborative action of the two effectively improves the dynamic adaptation ability of the access process. When the interface call frequency, data transmission format and other parameters of the business scenario change, the system can automatically adjust the feature interaction strategy to ensure the stability of the access process.

[0051] The feature adjustment module introducing the Transformer structure in the feature coordination processing further optimizes the expression accuracy of the business data associated feature vector and the parameter identification vector through the interaction calculation of the query vector, the value vector and the key matrix, so that the collaborative feature vector of the two can more accurately reflect the matching relationship between the business demand and the access parameter. Based on the access verification process of the collaborative feature vector, through the multi-level verification mechanism of dimension alignment, initial verification and deep verification, the reliability of the access result is ensured, the probability of problems such as data transmission error and interface call exception is effectively reduced, and the accuracy and security of the interaction between the large model and the business server are improved.

[0052] The whole method flow forms a complete closed loop from parameter configuration, data acquisition, feature extraction, collaborative optimization to access verification, and each link closely connects and supports each other, which not only guarantees the standardization and operability of the access process, but also has flexibility and robustness to cope with changes in complex business scenarios. Through the method, the large model can quickly and stably access servers in various actual business scenarios, reduce access cost, improve business interaction efficiency, and provide important technical support for large-scale application of large models in industrial fields. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 The working principle diagram of the server access method for large models and actual business scenarios described in the application;

[0054] Figure 2 The flowchart of the business data associated feature vector;

[0055] Figure 3 The flowchart of the business data associated reference point representation vector;

[0056] Figure 4 The flowchart of the parameter-business collaborative feature vector;

[0057] Figure 5 The flowchart of the optimized business data associated feature vector and parameter identification vector. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0059] Please refer to Figures 1-5 The application provides a server access method for large models and actual business scenarios, and the specific implementation steps are as follows:

[0060] configure access parameters, the access parameters including protocol type identification, interface call frequency and data transmission format.

[0061] Obtain the business interaction data set of the target business scenario and the large model system under the constraint of the access parameters.

[0062] Feature extraction processing is performed on the access parameters to obtain a parameter identification vector.

[0063] The business interaction data set is integrated based on feature correlation to obtain a business data correlation feature vector.

[0064] The business data correlation feature vector and the parameter identification vector are processed in multiple dimensions to obtain a parameter-business collaborative feature vector.

[0065] Based on the parameter-business collaborative feature vector, the server access of the large model and the actual business scenario is completed.

[0066] Embodiment 1:

[0067] When the business interaction data set is integrated based on feature correlation to obtain a business data correlation feature vector, a specific process needs to be performed. Each business interaction data in the business interaction data set is processed and passed through a format conversion unit. The format conversion unit has pre-set format conversion rules. After different format business interaction data enters the unit, it will be converted into a unified standard format according to these rules, and then a standard business data unit set is formed.

[0068] Based on the feature distribution situation of the standard business data unit set, analyze the business data correlation reference point representation vector. In this process, first input each standard business data unit in the standard business data unit set into the feature distribution measurement module. In the feature distribution measurement module, each standard business data unit is processed by numerical standardization. The specific operation is to determine the value range of each feature in the unit, subtract the minimum value of each feature value, and divide by the difference between the maximum value and the minimum value. In this way, the feature value is mapped to a specific interval to obtain a standard business data unit.

[0069] After the processing of the standard business data unit is completed, the sum of the squares of each feature value in the standard business data unit is calculated, and then the sum is divided by the feature range value of the standard business data unit. The feature range value here is the difference between the maximum value and the minimum value of the standardized feature. Through this calculation, the business data distribution measurement coefficient is obtained. Each standard business data unit is processed in this way, and finally a business data distribution measurement coefficient set is obtained.

[0070] The business data distribution metric coefficient set is input into the feature weight screening control unit. The unit is internally provided with specific weight screening algorithms and rules, which will evaluate each coefficient in the set. According to the size, distribution, and other factors of the coefficient, each coefficient is assigned a corresponding business data distribution weight factor, thereby obtaining a business data distribution weight factor set.

[0071] Based on the business data distribution weight factor set, the weighted sum of the standard business data unit set is calculated. Specifically, each standard business data unit is multiplied by the corresponding business data distribution weight factor, and then all the multiplied results are added to obtain the sum, which is the business data association reference point representation vector.

[0072] After obtaining the business data association reference point representation vector, the association movement direction of each standard business data unit in the standard business data unit set relative to the vector is calculated. The association movement direction is determined by calculating the inverse cosine similarity value between the standard business data unit and the business data association reference point representation vector. Each standard business data unit performs this calculation, and finally forms a business data association direction set.

[0073] Based on the business data association direction set, the standard business data unit set is integrated towards the business data association reference point representation vector. During integration, a specific integration algorithm is used to adjust and aggregate each standard business data unit according to its association movement direction. For example, for standard business data units with close association movement directions, their positions in the feature space are made closer during integration, while for units with large differences in association movement directions, corresponding adjustments are made, and finally these units are integrated into one whole, obtaining the business data association feature vector. This vector integrates the features of the standard business data unit set and their association relationship, and can more comprehensively reflect the features of business interaction data, providing effective data support for subsequent multi-dimensional feature collaborative processing.

[0074] Embodiment 2:

[0075] The embodiment needs to follow specific processing logic and steps when performing multi-dimensional feature collaborative processing on the service data association feature vector and the parameter identification vector to obtain the parameter-service collaborative feature vector. Based on external business knowledge, the feature interaction adjustment based on the attention mechanism is performed on the service data association feature vector and the parameter identification vector to obtain the knowledge-optimized parameter-service collaborative feature matrix. Specifically, the service data association feature vector and the parameter identification vector are input into the feature collaborative interaction network. The network adopts a specific neural network architecture, internally contains multiple layers of neurons and connection weights, and can perform feature interaction processing on the two input vectors. During network operation, each feature in the service data association feature vector and the parameter identification vector will interact with each other, generate a parameter-service collaborative feature matrix through weighted summation, activation function operation and other operations. This matrix fuses the feature information of the two vectors and reflects the preliminary association between the service data features and the access parameter features.

[0076] The parameter-service collaborative feature matrix is input into the attention processing unit based on external business knowledge. The unit stores pre-arranged external business knowledge, which covers industry rules, business processes, data features and other information of the target business scenario. The attention processing unit will analyze each feature in the parameter-service collaborative feature matrix according to these external business knowledge and calculate the attention weight of each feature. For features closely associated with business knowledge and more important for business understanding and processing, a higher weight is given; for features with lower association and relatively weaker importance, a lower weight is given. In this way, important features are highlighted and secondary features are suppressed, thereby obtaining the knowledge-optimized parameter-service collaborative feature matrix. After the optimization of the business knowledge, the matrix can more accurately reflect the key association features between the business scenario and the access parameters.

[0077] After obtaining the knowledge-optimized parameter-service collaborative feature matrix, the feature adjustment and optimization need to be performed on the service data association feature vector and the parameter identification vector based on the matrix to obtain the optimized service data association feature vector and the optimized parameter identification vector. First, the service data association feature vector is processed and linearly transformed. Linear transformation is realized through a pre-set transformation matrix, which maps the service data association feature vector to a new feature space, thereby obtaining a first query feature vector and a first value feature vector. At the same time, the knowledge-optimized parameter-service collaborative feature matrix is used as a key matrix.

[0078] The first query feature vector, the first value feature vector, and the key matrix are input into a feature adjustment module based on a Transformer structure. The Transformer structure includes components such as self-attention mechanisms and feedforward neural networks. In the feature adjustment module, the first query feature vector is subjected to self-attention calculation with the key matrix to determine the attention weight of each feature position, thereby focusing on the part closely associated with the features in the key matrix. Then, the first value feature vector is weighted and summed according to the attention weight to obtain an adjusted feature representation. After further processing by the feedforward neural network, the feature is subjected to nonlinear transformation and dimension adjustment, and finally an optimized business data association feature vector is obtained.

[0079] The parameter identification vector is processed. The parameter identification vector is subjected to linear transformation processing and converted into a second query feature vector and a second value feature vector through another preset transformation matrix. Again taking the knowledge optimization parameter-business collaboration feature matrix as the key matrix, the second query feature vector, the second value feature vector, and the key matrix are input into the feature adjustment module based on the Transformer structure. In this module, the same process as processing the business data association feature vector is followed, including self-attention calculation, weighted summation, feedforward neural network processing, and other steps to obtain an optimized parameter identification vector.

[0080] The point-by-point division operation between the optimized business data association feature vector and the optimized parameter identification vector is calculated to obtain a parameter-business collaboration feature vector. The point-by-point division operation refers to dividing the elements at corresponding positions of the two vectors to obtain new vector elements. Through this operation, the feature differences and association between the two vectors can be highlighted, and the parameter-business collaboration feature vector can more comprehensively reflect the collaboration relationship between the business data features and the access parameter features. In the entire processing process, each step depends on the results of the previous step, and through the collaborative work of components such as the feature collaboration interaction network, the attention processing unit, and the feature adjustment module based on the Transformer structure, multi-dimensional feature collaboration processing of the business data association feature vector and the parameter identification vector is realized, and finally a parameter-business collaboration feature vector that can effectively support the access of large models and actual business scenario servers is obtained.

[0081] Embodiment 3:

[0082] In this embodiment, when the parameter-service collaborative feature vector is used to access the server of the actual business scenario based on the large model, the parameter-service collaborative feature vector needs to be input into the access verification module based on the determinator to obtain the access verification result. The specific implementation is as follows. The parameter-service collaborative feature vector is subjected to dimension alignment processing to ensure that its dimension matches the preset standard feature dimension. When the dimension alignment processing is performed, the preset standard feature dimension parameter is first extracted. The parameter is preset according to the demand of the connection between the large model and the actual business scenario, and specifies the expected feature dimension quantity and the meaning of each dimension.

[0083] The deviation of the current dimension parameter of the parameter-service collaborative feature vector from the standard feature dimension parameter is calculated. The current dimension parameter here includes the dimension quantity of the feature vector and the value range of each dimension and other information. By comparing the current dimension parameter with the standard feature dimension parameter, the difference between the two is determined, for example, whether the dimension quantity is consistent, whether the value range of each dimension is within the allowed error range, and the like.

[0084] Based on the calculated deviation, the dimension adjustment operation is performed on the parameter-service collaborative feature vector. If the dimension quantity of the current dimension parameter is more than the standard feature dimension parameter, some unimportant dimensions may need to be removed through feature selection or dimension reduction algorithm; if the dimension quantity is less than the standard feature dimension parameter, the missing dimensions need to be supplemented through feature generation or interpolation method. For the deviation of the value range of each dimension, normalization or standardization method can be used for adjustment, so that the adjusted feature vector dimension matches the standard feature dimension parameter, thereby obtaining the aligned feature vector.

[0085] After the dimension alignment processing is completed, the aligned feature vector is input into the basic verification layer of the determinator for initial checking. The basic verification layer of the determinator is internally provided with a series of initial checking rules. These rules mainly check the basic features of the feature vector, such as whether there are obvious outliers in the feature vector, whether the distribution of feature values conforms to the basic statistical law, and the like. The basic verification layer checks the input aligned feature vector one by one, and generates an initial checking mark according to the preset rules. The initial checking mark is used to record the checking result of the aligned feature vector in the basic verification layer, such as which features meet the requirements and which features have potential problems.

[0086] The initial check mark is input into the deep verification layer of the decision maker for secondary verification. The deep verification layer analyzes and verifies the initial check mark from a deeper perspective, and its internal contains more complex verification rules and algorithms. The deep verification layer not only considers the basic features of the feature vector, but also combines the specific needs of the large model and the actual business scenario to evaluate the business logic and parameter matching reflected by the feature vector. For example, check whether the coordination relationship between the business data correlation feature and the access parameter feature meets the requirements of the business scenario, whether the parameter-business coordination feature vector can accurately reflect the interaction needs of the large model and the actual business scenario, etc.

[0087] The deep verification layer finally obtains the access verification result through detailed analysis and complex calculation of the initial check mark. The access verification result is used to indicate whether the large model and the server of the actual business scenario complete the access. If the access verification result indicates that all verification rules and requirements are met, it means that the large model and the server of the actual business scenario access successfully; otherwise, the access fails.

[0088] In the entire implementation process, each step is closely connected. The dimension alignment processing is to ensure that the format of the feature vector meets the requirements of the verification module. The initial check of the basic verification layer is a preliminary screening of the basic features of the feature vector. The secondary verification of the deep verification layer is an in-depth analysis from the perspective of business logic and actual demand, and finally an accurate access verification result is obtained.

[0089] Embodiment 4:

[0090] In this embodiment, when each standard business data unit in the standard business data unit set is input into the feature distribution measurement module to obtain the business data distribution measurement coefficient set, specific processing operations need to be performed for each standard business data unit. Taking the user shopping behavior data of a certain e-commerce platform as an example, assuming that the standard business data unit set contains shopping record data of multiple users, each standard business data unit corresponds to a user's shopping record, which contains features such as product category, purchase quantity, consumption amount, shopping time, etc.

[0091] For one of the standard business data units, i.e. the shopping record data of a certain user, numerical standardization is performed. Taking the consumption amount feature as an example, the value range of this feature in the standard business data unit can be 0 to 10000 yuan. In order to map it to a specific interval, a common standardization method such as min-max standardization is adopted. Specifically, the minimum value of this feature is determined to be 0, and the maximum value is determined to be 10000. The consumption amount value of the user is subtracted by the minimum value 0, and then divided by the difference between the maximum value and the minimum value 10000, to obtain the standardized consumption amount value. Assuming that the consumption amount of the user is 5000 yuan, the standardized value is (5000-0) / (10000-0)=0.5. Similarly, other features such as product category, purchase quantity, and shopping time are standardized. For the product category feature which is a non-numeric type, it can be necessary to first perform encoding processing, such as encoding different product categories into different numerical values, and then standardizing. The purchase quantity feature has a value range of 1 to 100, and is also processed by the min-max standardization method. The shopping time feature can need to be converted into a numerical type of data first, such as the number of days from a certain reference time, and then standardized. Through such processing, the values of each feature in the entire standard business data unit are mapped to a specific interval, such as [0, 1], to obtain a normalized business data unit.

[0092] After the processing of the normalized business data unit is completed, the sum of the squares of the values of each feature in the unit is calculated. Continuing with the above example of the shopping record of the user, assuming that the standardized consumption amount is 0.5, the product category encoding is standardized to 0.3, the purchase quantity is standardized to 0.6, and the shopping time is standardized to 0.4. Then the sum of the squares of the values of each feature is 0.5 2 + 0.3 2 + 0.6 2 + 0.4 2 = 0.25 + 0.09 + 0.36 + 0.16 = 0.86.

[0093] The feature range value of the normalized business data unit is calculated. The feature range value is the difference between the maximum value and the minimum value of each standardized feature. Assuming that in this normalized business data unit, the maximum value of the standardized features is 0.6 (purchase quantity feature) and the minimum value is 0.3 (product category feature), then the feature range value is 0.6-0.3=0.3.

[0094] The sum of the squares of the values of each feature is divided by the feature range value to obtain the business data distribution measurement coefficient. That is, 0.86 / 0.3≈2.87. In this way, the business data distribution measurement coefficient corresponding to this standard business data unit is about 2.87.

[0095] In the same way, each standard business data unit in the standard business data unit set is processed. For example, in another user's shopping record, the consumption amount is 8000 yuan, which is standardized to 0.8, the product category code is standardized to 0.7, the purchase quantity is standardized to 0.9, and the shopping time is standardized to 0.5. The sum of the squares of the respective feature values is 0.8 2 + 0.7 2 + 0.9 2 + 0.5 2 = 0.64 + 0.49 + 0.81 + 0.25 = 2.19. Assuming that the maximum value of the feature of the standard business data unit is 0.9, the minimum value is 0.5, the feature range value is 0.9-0.5=0.4, and the business data distribution measure coefficient is 2.19 / 0.4=5.475.

[0096] For another example, in the shopping record of a third user, the consumption amount is 2000 yuan, which is standardized to 0.2, the product category code is standardized to 0.4, the purchase quantity is standardized to 0.3, and the shopping time is standardized to 0.2. The sum of the squares of the respective feature values is 0.2 2 + 0.4 2 + 0.3 2 + 0.2 2 = 0.04 + 0.16 + 0.09 + 0.04 = 0.33. The maximum value of the feature is 0.4, the minimum value is 0.2, the feature range value is 0.4-0.2=0.2, and the business data distribution measure coefficient is 0.33 / 0.2=1.65.

[0097] After all the standard business data units in the standard business data unit set are processed according to the above steps, the business data distribution measure coefficients obtained for each unit are collected to form a business data distribution measure coefficient set. In this example of an e-commerce platform, the business data distribution measure coefficient set may contain multiple values such as 2.87, 5.475, 1.65, etc., each value corresponding to the business data distribution measure coefficient of a user's shopping record data.

[0098] Through this processing process, the operation of inputting each unit in the standard business data unit set into the feature distribution measure module to obtain the business data distribution measure coefficient set is realized. This process accurately obtains the business data distribution measure coefficient of each standard business data unit through standardization processing, calculation of the sum of the squares, and feature range value, etc. steps, providing basic data for subsequent feature weight selection and calculation of business data association reference point representation vectors. The entire processing process combines specific business scenario examples to detail the operation method and calculation process of each step, ensuring the operability and understandability of the implementation.

[0099] Embodiment 5:

[0100] This embodiment needs to combine specific examples to explain the operation process when analyzing the feature distribution trend of the standard business data unit set to obtain the business data association reference point representation vector. Taking the loan application business of a certain financial institution as an example, it is assumed that the standard business data unit set contains information data of multiple loan applicants, and each standard business data unit corresponds to the information of an applicant, including income status, credit score, loan amount, loan period, etc.

[0101] After obtaining the business data distribution measurement coefficient set, it is input into the feature weight screening control unit. For example, the business data distribution measurement coefficients of three standard business data units in this set are 3.2 for applicant A, 5.8 for applicant B, and 2.1 for applicant C. The feature weight screening control unit internally presets a weight screening rule based on business logic, such as determining the weight factor according to the size of the business data distribution measurement coefficient. The larger the coefficient, the greater the influence of the feature distribution of this data unit on the whole, and the higher the corresponding weight factor should be.

[0102] The feature weight screening control unit evaluates and processes these three coefficients. For the coefficient 3.2 of applicant A, according to the rule, it can be converted into the corresponding business data distribution weight factor 0.3; the coefficient 5.8 of applicant B is larger, which is converted into the weight factor 0.5; the coefficient 2.1 of applicant C is smaller, which is converted into the weight factor 0.2. In this way, the business data distribution weight factor set is obtained, that is, [0.3, 0.5, 0.2].

[0103] Next, based on the business data distribution weight factor set, the weighted sum of the standard business data unit set is calculated to obtain the business data association reference point representation vector. Each standard business data unit contains multiple features. Taking the income status, credit score, loan amount, and loan period as examples, it is assumed that:

[0104] The standardized data of applicant A is: income status 0.6, credit score 0.7, loan amount 0.4, and loan period 0.5;

[0105] The standardized data of applicant B is: income status 0.8, credit score 0.9, loan amount 0.6, and loan period 0.7;

[0106] The standardized data of applicant C is: income status 0.3, credit score 0.4, loan amount 0.2, and loan period 0.3.

[0107] When calculating the weighted sum, each feature is calculated separately:

[0108] The weighted sum of the income status is: 0.6x0.3+0.8x0.5+0.3x0.2=0.18+0.4+0.06=0.64;

[0109] The weighted sum of the credit score is: 0.7x0.3+0.9x0.5+0.4x0.2=0.21+0.45+0.08=0.74;

[0110] The weighted sum of the loan amount is: 0.4x0.3+0.6x0.5+0.2x0.2=0.12+0.3+0.04=0.46;

[0111] The weighted sum of the loan term is: 0.5x0.3+0.7x0.5+0.3x0.2=0.15+0.35+0.06=0.56.

[0112] Combining these weighted sums, the business data correlation reference point representation vector is [0.64, 0.74, 0.46, 0.56]. This vector reflects the overall feature distribution trend of the standard business data unit set in the loan application business, for example, the comprehensive feature value of the income status is 0.64, and the comprehensive feature value of the credit score is 0.74, etc.

[0113] Taking another medical diagnosis business scenario as an example, the standard business data unit set contains the diagnosis data of patients, and each unit has features such as body temperature, blood pressure, heart rate, and blood oxygen saturation. Assuming that the business data distribution metric coefficient set is [4.5, 3.7, 6.2], which corresponds to the business data distribution metric coefficients of three patients. The feature weight screening control unit converts it into the weight factor set [0.25, 0.2, 0.55] according to the rules.

[0114] The standardized feature values of each standard business data unit are as follows:

[0115] Patient A: body temperature 0.7, blood pressure 0.6, heart rate 0.8, blood oxygen saturation 0.9;

[0116] Patient B: body temperature 0.5, blood pressure 0.7, heart rate 0.6, blood oxygen saturation 0.8;

[0117] Patient C: body temperature 0.9, blood pressure 0.8, heart rate 0.9, blood oxygen saturation 0.95.

[0118] Calculate the weighted sum of each feature:

[0119] Body temperature: 0.7x0.25+0.5x0.2+0.9x0.55=0.175+0.1+0.495=0.77;

[0120] Blood pressure: 0.6x0.25+0.7x0.2+0.8x0.55=0.15+0.14+0.44=0.73;

[0121] Heart rate: 0.8x0.25+0.6x0.2+0.9x0.55=0.2+0.12+0.495=0.815;

[0122] Oxygen saturation: 0.9x0.25+0.8x0.2+0.95x0.55=0.225+0.16+0.5225=0.9075.

[0123] The obtained service data association reference point representation vector is [0.77, 0.73, 0.815, 0.9075], which integrates the diagnosis data feature distribution of the three patients, and reflects the overall feature trend. For example, the comprehensive value of body temperature is 0.77, which is close to the standardized value of the normal body temperature range, and the comprehensive value of oxygen saturation is high, reflecting the good state of the overall patient group in this feature.

[0124] In actual operation, the rules of the feature weight screening control unit can be more complex, and the actual meaning and importance of the service data need to be considered. For example, in the financial loan business, the credit score feature can be given a higher weight factor to generate the weight, and in medical diagnosis, oxygen saturation can be a more critical feature and will be given priority consideration in weight screening. Through the processing of the feature weight screening control unit on the service data distribution metric coefficient, combined with the rules of the specific business scenario, a reasonable set of service data distribution weight factors is generated, and then the service data association reference point representation vector is calculated by weighted sum, which can effectively aggregate the feature information of the standard service data unit set, providing key basic data for subsequent calculation of the association motion direction of the standard service data unit relative to the vector and integration processing, ensuring that in the server access process of the large model and the actual business scenario, the overall feature distribution of the business data can be accurately grasped, providing support for the smooth progress of the subsequent steps.

[0125] It should be noted that, in the present document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0126] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A server access method for large models and actual business scenarios, characterized in that, The method comprises the following steps: configuring access parameters, wherein the access parameters comprise a protocol type identifier, an interface calling frequency, and a data transmission format; obtaining business interaction data of a target business scenario and a large model system under the constraint of the access parameters to obtain a business interaction data set; performing feature extraction processing on the access parameters to obtain a parameter identifier vector; performing feature correlation-based integration processing on the business interaction data set to obtain a business data correlation feature vector; performing multi-dimensional feature collaborative processing on the business data correlation feature vector and the parameter identifier vector to obtain a parameter-business collaborative feature vector; and based on the parameter-business collaborative feature vector, completing server access of a large model to an actual business scenario.

2. The server access method for large models and actual business scenarios according to claim 1, characterized in that, The method of performing feature correlation-based integration processing on the business interaction data set to obtain a business data correlation feature vector comprises the following steps: respective business interaction data in the business interaction data set is converted by a format conversion unit to obtain a standard business data unit set; based on a feature distribution situation of the standard business data unit set, a business data correlation reference point representation vector is analyzed; a correlation motion direction of each standard business data unit in the standard business data unit set relative to the business data correlation reference point representation vector is calculated to obtain a business data correlation direction set; based on the business data correlation direction set, the standard business data unit set is integrated towards the business data correlation reference point representation vector to obtain the business data correlation feature vector.

3. The server access method for large models and actual business scenarios according to claim 2, characterized in that, The method of analyzing a business data correlation reference point representation vector based on a feature distribution situation of the standard business data unit set comprises the following steps: each standard business data unit in the standard business data unit set is input into a feature distribution measurement module to obtain a business data distribution measurement coefficient set; the business data distribution measurement coefficient set is input into a feature weight screening control unit to obtain a business data distribution weight factor set; based on the business data distribution weight factor set, a weighted sum of the standard business data unit set is calculated to obtain the business data correlation reference point representation vector.

4. The server access method for large models and actual business scenarios according to claim 3, characterized in that, The method of inputting each standard business data unit in the standard business data unit set into a feature distribution measurement module to obtain a business data distribution measurement coefficient set comprises the following steps: performing numerical standardization processing on the standard business data unit to obtain a normalized business data unit; calculating a sum of squares of each feature value in the normalized business data unit divided by a feature range value of the normalized business data unit to obtain the business data distribution measurement coefficient.

5. The server access method for large models and actual business scenarios according to claim 4, characterized in that, The business data correlation motion direction is an inverse cosine similarity value between the standard business data unit and the business data correlation reference point representation vector.

6. The server access method for large models and actual business scenarios according to claim 5, characterized in that, The method of performing multi-dimensional feature collaborative processing on the business data correlation feature vector and the parameter identifier vector to obtain a parameter-business collaborative feature vector comprises the following steps: based on external business knowledge, performing feature interaction adjustment based on an attention mechanism on the business data correlation feature vector and the parameter identifier vector to obtain a knowledge-optimized parameter-business collaborative feature matrix; and optimizing the business data correlation feature vector and the parameter identifier vector based on the knowledge-optimized parameter-business collaboration feature matrix, to obtain an optimized business data correlation feature vector and an optimized parameter identifier vector; calculating a point-by-position division between the optimized business data correlation feature vector and the optimized parameter identifier vector to obtain the parameter-business collaboration feature vector.

7. The server access method for large models and actual business scenarios according to claim 6, characterized in that, based on external business knowledge, performing feature interaction adjustment based on an attention mechanism on the business data correlation feature vector and the parameter identifier vector to obtain a knowledge-optimized parameter-business collaboration feature matrix, including: inputting the business data correlation feature vector and the parameter identifier vector into a feature collaboration interaction network to obtain a parameter-business collaboration feature matrix; inputting the parameter-business collaboration feature matrix into an attention processing unit based on external business knowledge to obtain the knowledge-optimized parameter-business collaboration feature matrix.

8. The server access method for large models and actual business scenarios according to claim 7, characterized in that, optimizing the business data correlation feature vector and the parameter identifier vector based on the knowledge-optimized parameter-business collaboration feature matrix, to obtain an optimized business data correlation feature vector and an optimized parameter identifier vector, including: performing linear transformation processing on the business data correlation feature vector to obtain a first query feature vector and a first value feature vector, and taking the knowledge-optimized parameter-business collaboration feature matrix as a key matrix, inputting the first query feature vector, the first value feature vector, and the key matrix into a feature adjustment module based on a Transformer structure to obtain the optimized business data correlation feature vector; performing linear transformation processing on the parameter identifier vector to obtain a second query feature vector and a second value feature vector, and taking the knowledge-optimized parameter-business collaboration feature matrix as a key matrix, inputting the second query feature vector, the second value feature vector, and the key matrix into the feature adjustment module based on the Transformer structure to obtain the optimized parameter identifier vector.

9. The server access method for large models and actual business scenarios according to claim 8, characterized in that, based on the parameter-business collaboration feature vector, completing server access of a large model to an actual business scenario, including: inputting the parameter-business collaboration feature vector into an access verification module based on a determiner to obtain an access verification result, the access verification result being used to indicate whether the large model and the server of the actual business scenario have completed access.

10. The server access method for large models and real business scenarios according to claim 9, characterized in that, inputting the parameter-business collaboration feature vector into an access verification module based on a determiner to obtain an access verification result, including: performing dimension alignment processing on the parameter-business collaboration feature vector to obtain an aligned feature vector; inputting the aligned feature vector into a basic verification layer of the determiner to perform initial checking to obtain an initial checking mark; inputting the initial checking mark into a deep verification layer of the determiner to perform secondary checking to obtain the access verification result; the dimension alignment processing on the parameter-business collaboration feature vector to obtain the aligned feature vector, including: extracting a pre-set standard feature dimension parameter; calculating a deviation amount of a current dimension parameter of the parameter-service coordination feature vector from the standard feature dimension parameter; performing a dimension adjustment operation on the parameter-service coordination feature vector based on the deviation amount to obtain the alignment feature vector.

Images