An intelligent information technology consulting method

Abstract

This invention discloses an intelligent information technology consulting method, relating to the field of intelligent information consulting technology. This method uses a large language model to parse the client's natural language requirements into an undetermined IT architecture attribute graph. Based on an IT component compatibility knowledge graph, it uses a graph neural network to predict links and detect conflicts on candidate connection edges, obtaining a technically feasible solution space and a quantitative value for the degree of conflict. Within the technically feasible solution space, a multi-agent reinforcement learning module, including cost, performance, and security compliance agents, is used in conjunction with real-time market data to perform multi-objective game optimization, outputting a Pareto optimal implementation path report. The quantitative value of the degree of conflict and the Pareto reward deviation value are fed back to the large language model to drive strategy reconstruction, forming an automatic closed loop. This method is used to improve the engineering feasibility and multi-objective balance of information consulting solutions.

Description

Technical Field

[0001] This invention relates to the field of information-based intelligent consulting technology, specifically to an intelligent information technology consulting method. Background Technology

[0002] In the field of intelligent IT consulting services, the digital transformation of large enterprises and institutions places high demands on the planning, reconstruction, and upgrading of underlying IT architecture. Traditional IT consulting relies on the experience of senior architects, which suffers from problems such as long consulting cycles, high labor costs, and difficulty in real-time tracking of open-source components and cloud-native technology stack version changes. With the development of large language model technology, a method has emerged in the industry that uses large language models to directly generate IT consulting solution text based on the client's natural language requirements. This method has become a common technical path in this field. Specifically, it typically involves receiving the client's IT construction requirements text in natural language form, inputting the requirements text along with preset prompt word templates into a trained large language model, and then having the large language model generate the solution based on parameterized language generalizations. The current approach involves generating consulting report text, including technical component recommendations, architecture descriptions, and implementation suggestions, word by word based on probability distribution, and then formatting and outputting the generated text to the client. Structurally, this approach belongs to a single-model direct generation mode. Because this type of method lacks an independent engineering logic verification step between natural language requirements and the consulting report text, and also lacks a step for mathematical optimization calculations for multiple objectives such as cost, performance, and security, it is essentially a text continuation process based on statistical probability. Therefore, it is difficult to guarantee the true compatibility of the underlying protocols and versions between the recommended components, and it is also difficult to achieve multi-objective balance under dynamic market data conditions. This results in insufficient engineering feasibility and a lack of multi-objective optimization. To address these issues, we propose an intelligent information technology consulting method. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent information technology consulting method.

[0004] The technical problem this invention aims to solve is that existing methods for directly generating IT consulting solutions based on large language models lack an independent engineering logic verification step between natural language requirements and consulting report text. This easily leads to the recommendation of IT component combinations with incompatible underlying protocols, versions, or interfaces, resulting in an illusion of architectural dependency—seemingly professional but actually impractical. Furthermore, existing methods lack mathematical optimization steps targeting multiple objectives such as cost, performance, security, and compliance. They cannot accurately perform multi-objective dynamic optimization under dynamic market data changes such as cloud vendor computing power prices and open-source component maintenance costs. There is also no feedback mechanism for verification results to the generation process, making it difficult to improve solution quality through system iteration and thus hindering the output of high-quality IT implementation path reports that are both technically feasible and economically reasonable.

[0005] To solve the above-mentioned technical problems, the present invention provides an intelligent information technology consulting method, comprising:

[0006] The system receives IT infrastructure requirements text input by customers in natural language form, and parses the requirements text into a pending IT architecture attribute graph using a large language model. The pending IT architecture attribute graph includes a set of candidate IT component nodes, a set of candidate connection edges, and a node feature matrix.

[0007] Based on the IT component compatibility knowledge graph, a graph neural network is used to predict the links of candidate connection edges in the attribute graph of the undetermined IT architecture, and the link prediction score of each candidate connection edge is obtained. Candidate connection edges with link prediction scores not lower than a preset compatibility threshold are retained to obtain a technically feasible solution space. Candidate connection edges with link prediction scores lower than the preset compatibility threshold are marked as conflict edges, and a conflict degree quantification value corresponding to the conflict edge is generated.

[0008] Within the technically feasible solution space, a multi-agent reinforcement learning module, including a cost control agent, a performance optimization agent, and a safety and compliance agent, is used to perform multi-objective game optimization in conjunction with real-time market data to obtain a Pareto front solution set. Based on the objective priority preferences reflected in the requirement text, representative solutions are selected from the Pareto front solution set and transformed into a Pareto optimal implementation path report. The reward function of each agent includes a penalty term composed of the quantified value of the conflict degree.

[0009] The conflict intensity quantification value and the reward deviation value of the Pareto optimal implementation path report relative to the Pareto front solution set are fed back to the large language model, driving the large language model to adjust the node selection strategy and edge generation strategy of the undetermined IT architecture attribute graph, regenerate a new round of undetermined IT architecture attribute graph, and repeat the link prediction and multi-objective game optimization process until the conflict intensity quantification value drops below the preset convergence threshold and the reward deviation value reaches the preset convergence standard, and outputs an information consulting report.

[0010] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0011] 1. This invention uses graph neural networks based on IT component compatibility knowledge graphs to perform link prediction and conflict detection on the attribute graph of a given architecture. It can block incompatible component combinations of underlying protocols, versions or interfaces at the engineering logic level during the solution generation stage, eliminate the illusion of architecture dependency, and improve the engineering feasibility of consulting solutions.

[0012] 2. This invention uses multi-agent reinforcement learning to conduct multi-objective game theory on cost, performance, security and compliance within the technically feasible solution space constructed by graph neural networks, and combines real-time market data for dynamic optimization. It can output a dynamic information implementation path report that achieves Pareto optimality, thus solving the problem of missing multi-objective non-stationary optimization.

[0013] 3. This invention feeds back the conflict detection results of the graph neural network and the Pareto reward bias value of multi-agent reinforcement learning to the large language model for policy reconstruction, forming an automatic closed loop from natural language needs to high-feasibility consultation reports. This allows the accuracy of the consultation plan to continuously improve with iteration, while maintaining high generation efficiency. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the entire process of the method in the embodiment of the present invention;

[0015] Figure 2 This is a schematic diagram of the illusion elimination module in an embodiment of the present invention;

[0016] Figure 3 This is a schematic diagram of multi-agent optimization and closed-loop feedback in an embodiment of the present invention. Detailed Implementation

[0017] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0018] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0019] System overall architecture:

[0020] The intelligent information technology consulting method provided in this invention consists of a requirement analysis and attribute graph construction module, an architecture illusion elimination module, a multi-objective game optimization module, and a closed-loop feedback reconstruction module. These four modules work collaboratively around a continuously updated IT component compatibility knowledge graph database. The requirement analysis and attribute graph construction module, based on a large language model, transforms the natural language requirements input by the client into a pending IT architecture attribute graph. The architecture illusion elimination module, based on a graph neural network and connected to the knowledge graph database, performs link prediction and conflict detection on the pending IT architecture attribute graph. The multi-objective game optimization module, based on multi-agent reinforcement learning, performs multi-objective game optimization within the technically feasible solution space output by the architecture illusion elimination module. The closed-loop feedback reconstruction module is connected to the aforementioned three modules and feeds back the verification and optimization results to the requirement analysis and attribute graph construction module, driving the large language model to reconstruct the generation strategy, forming an automatic closed loop. The data sources of the knowledge graph database include cloud vendor interface documents, open-source community version dependency tables, and dynamic market data collected by the real-time computing power pricing collection unit.

[0021] Requirements analysis and construction of attribute diagrams for undetermined IT architecture based on large language models:

[0022] The requirement parsing and attribute graph construction module first segments the natural language requirement text input by the customer into sentences and extracts key information, identifying business scenario keywords, performance indicator keywords, and constraint condition keywords involved in the requirement text. Then, it performs semantic matching between the extracted key information and a pre-set candidate IT component library. Combining the few-shot prompting learning capability of the large language model, it generates a list of candidate IT components related to the semantics of the requirement, which serves as a node set. Furthermore, based on its linguistic understanding of common IT architecture patterns, the large language model determines whether there is a potential integration necessity for each pair of nodes in the candidate component list. If so, it generates a candidate edge between the two nodes, forming an edge set. Then, for each candidate node, the large language model is invoked to extract the component's attribute description in the context of customer requirements, such as the required version range and deployment environment requirements, and this is encoded into a feature vector and filled into the feature matrix. Finally, the node set, edge set, and feature matrix are encapsulated in graph data format to output a complete attribute graph of the undetermined IT architecture. As input to the architecture illusion elimination module, where Represents the set of candidate IT component nodes. This represents the set of candidate connection edges between nodes. This represents the node feature matrix, where each row corresponds to the attribute vector of a node.

[0023] Graph neural network-based illusion cancellation mechanism:

[0024] The core function of the architecture illusion elimination module is to create a property diagram of a given IT architecture. Link prediction and conflict detection are performed on candidate edges to determine whether the two components connected by each candidate edge are compatible under real engineering conditions, thereby blocking incompatible component combinations from a logical level.

[0025] Specifically, the first step is to read the attribute diagram of the pending IT architecture. The initial feature vector of each node is used as the node embedding of the first layer of the graph neural network. Subsequently, for each layer of the graph neural network... For nodes and its neighbor node set Each node in Calculate the attention coefficient:

[0026] ;

[0027] in, Indicates the first Layer nodes Embedded vector, This represents the learnable weight matrix. This represents a learnable attention weight vector. This represents a vector concatenation operation. This represents a linear activation function with leakage correction. Represents a node For nodes The normalized attention coefficient;

[0028] The node embedding is updated by weighted aggregation of neighbor node features based on the attention coefficient:

[0029] ;

[0030] in, Represents a non-linear activation function. Indicates the updated number Layer nodes The above calculation process is repeated layer by layer until the preset number of network layers is reached, and the output layer embedding vectors of the nodes at both ends of the candidate edge are obtained.

[0031] For any candidate edge in the attribute graph Link prediction scores are calculated based on the output layer embedding of the two-end nodes:

[0032] ;

[0033] in, , Represents a node ,node The output layer embedding vector, This represents a learnable scoring matrix specifically for link prediction. Represents a node With nodes The predicted probability of a true compatible connection exists;

[0034] Score Compatibility threshold with preset If a comparison is made, Not less than If the edge is a compatible edge, it is retained. Below If an edge is identified as a conflict edge, it is removed from the attribute graph, and the reason for the conflict is recorded, such as version mismatch or protocol incompatibility. The subgraph composed of all retained edges is used as the technically feasible solution space and output to the multi-objective game optimization module. The number of conflict edges and the difference between the link prediction score of each conflict edge and the preset compatibility threshold are used together to determine the quantification value of the conflict degree. The output is sent to the closed-loop feedback reconstruction module;

[0035] The training of graph neural networks is supervised offline using the following loss function:

[0036] ;

[0037] in, This represents the set of positive edges in the training samples that have been verified as compatible. This represents the set of negative edge examples that have been verified as conflicting. The loss function is optimized for graph neural network parameters based on historical project data and knowledge graph labeled data.

[0038] Multi-objective game optimization mechanism based on multi-agent reinforcement learning:

[0039] Within the technically feasible solution space output by the architecture illusion elimination module, the multi-objective game optimization module introduces three agents with independent decision-making objectives, denoted as follows: (Cost control intelligent agent) (Performance optimization agent) (Safety and compliance intelligent agent) achieves Pareto optimality search for multiple objectives of cost, performance and safety and compliance through reinforcement learning game theory. This stage adopts an improved multi-agent deep deterministic policy gradient algorithm architecture.

[0040] First, independent policy network parameters and value network parameters are initialized for each agent. Then, the candidate architecture subgraph in the technically feasible solution space is encoded into a joint state observation vector of the environment by combining dynamic data such as cloud vendor computing power prices and open-source component maintenance costs obtained from the real-time market data acquisition unit. Each agent, based on its own local observations Each outputs actions through its own policy network. Among them, intelligent agents Actions indicate the choice of cloud resource specifications and deployment methods; intelligent agents Actions represent the choice between concurrency and caching strategies; the agent The action indicates the selection of data encryption and compliance component configurations;

[0041] Environment based on joint action Update the state and calculate the local reward for each agent:

[0042] ;

[0043] in, , , These represent the normalized values ​​of cost savings, performance improvements, and safety compliance achievements calculated based on the current joint operations, respectively. , , This represents the pre-set weighting coefficient for the corresponding target. This represents a quantitative value indicating the degree of architectural conflict transmitted by the architectural illusion elimination module. This represents the pre-set conflict penalty coefficient. Represents intelligent agents The total reward obtained at the current time step;

[0044] Each agent updates the centralized value network based on joint observation and joint actions:

[0045] ;

[0046] in, Represents intelligent agents The centralized action value function, This indicates the joint observation at the next time step. This indicates the actions of each agent in the next instant. Indicates the learning rate. This represents the discount factor, and each agent adjusts its own policy network parameters according to the updated value function.

[0047] The above-mentioned processes of state encoding, action output, reward calculation and value network update are repeatedly executed until the preset number of iterations or joint reward convergence is reached. The Pareto dominance relationship of the multiple sets of joint action schemes that have finally converged is judged. That is, if there is no other scheme that can increase the reward of at least one agent without reducing the reward of any agent, the scheme is included in the Pareto front solution set. From the Pareto front solution set, representative schemes are selected according to the target priority preferences reflected by the customer in the requirement text, and transformed into a specific dynamic information implementation path report and output.

[0048] Requirement-Architecture-Policy Automatic Closed-Loop Feedback Network:

[0049] To achieve fully automated closed-loop processing, the closed-loop feedback refactoring module will incorporate the conflict detection results output by the architecture illusion elimination module. The Pareto reward deviation value output by the multi-objective game optimization module is fed back to the demand analysis and attribute graph construction module, driving it to reconstruct the consulting strategy;

[0050] Specifically, the degree of conflict is first quantified. Together with the reward bias of the current joint scheme relative to the Pareto front, they constitute a closed-loop feedback correction signal. This correction signal is then added as a penalty term to the policy reconstruction objective function of the large language model.

[0051] ;

[0052] in, This represents the original text generation loss of the large language model. This represents the relative reward score of the current joint scheme within the Pareto front solution set, normalized to an interval. , , This represents the pre-defined weighting coefficients for the two types of penalty terms. This represents the optimization objective of the reconstructed large language model strategy.

[0053] Large language models are based on modified objective functions The node selection and edge generation strategies for the undetermined IT architecture attribute graph are adjusted to generate a new round of undetermined IT architecture attribute graphs. These new attribute graphs are then re-input into the architecture illusion elimination module and the multi-objective game optimization module for verification and optimization to determine the outcome of this round. Whether to reduce to below the preset threshold and If the preset convergence criteria are not met, the iteration along the closed loop continues. If the convergence criteria are met, the closed loop iteration is terminated. The current technically feasible architecture is integrated with the Pareto optimal implementation path to generate and output the final high-feasibility information consulting report.

[0054] Alternative implementation methods:

[0055] In another implementation of the illusion elimination module, instead of using a graph attention network, a hybrid approach combining a graph convolutional network and a rule engine can be adopted to achieve link prediction and conflict detection. Specifically, the graph convolutional network first performs convolutional aggregation and updates on the embedding vectors of the nodes at both ends of the candidate connection edge to obtain the convolutional link prediction score of the candidate connection edge. Then, the judgment result corresponding to the score is logically operated on with the verification result of an explicit compatibility rule base compiled by experts in advance, and the logical intersection of the two is taken as the final judgment criterion for inclusion in the technically feasible solution space. This approach can improve the accuracy of conflict detection when knowledge graph data is insufficient.

[0056] In another implementation of the multi-objective game optimization module, instead of using the multi-agent deep deterministic policy gradient algorithm, a multi-objective evolutionary algorithm based on non-dominated sorting genetic algorithm can be used to solve the Pareto front. That is, cost savings, performance improvement and safety compliance achievement are used as the individual fitness evaluation dimensions, respectively. The Pareto optimal solution set is searched iteratively through genetic operations such as population initialization, non-dominated sorting, crowding calculation, and crossover mutation. This method can also achieve multi-objective equilibrium optimization in scenarios where the dynamic market data update frequency is low, and can be used as a supplement or verification method to the aforementioned reinforcement learning scheme.

[0057] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. An intelligent information technology consulting method, characterized in that, include: The system receives information technology construction requirements text input by customers in natural language form, and parses the requirements text into a pending IT architecture attribute graph through a large language model. The pending IT architecture attribute graph includes a set of candidate IT component nodes, a set of candidate connection edges, and a node feature matrix. Based on the IT component compatibility knowledge graph, a graph neural network is used to predict the links of candidate connection edges in the attribute graph of the undetermined IT architecture, and the link prediction score of each candidate connection edge is obtained. Candidate connection edges with link prediction scores not lower than a preset compatibility threshold are retained to obtain a technically feasible solution space. Candidate connection edges with link prediction scores lower than the preset compatibility threshold are marked as conflict edges, and a conflict degree quantification value corresponding to the conflict edge is generated. Within the technically feasible solution space, a multi-agent reinforcement learning module, including a cost control agent, a performance optimization agent, and a safety and compliance agent, is used to perform multi-objective game optimization in conjunction with real-time market data to obtain a Pareto front solution set. Based on the objective priority preferences reflected in the requirement text, representative solutions are selected from the Pareto front solution set and transformed into a Pareto optimal implementation path report. The reward function of each agent includes a penalty term composed of the conflict degree quantification value. The conflict intensity quantification value and the reward deviation value of the Pareto optimal implementation path report relative to the Pareto front solution set are fed back to the large language model, driving the large language model to adjust the node selection strategy and edge generation strategy of the undetermined IT architecture attribute graph, regenerate a new round of undetermined IT architecture attribute graph, and repeat the link prediction step and the multi-objective game optimization step until the conflict intensity quantification value drops below the preset convergence threshold and the reward deviation value reaches the preset convergence standard, and outputs an information consulting report.

2. The intelligent information technology consulting method according to claim 1, characterized in that: Link prediction is performed on candidate connection edges in the attribute graph of the undetermined IT architecture using a graph neural network, including: Obtain the initial feature vectors of the nodes at both ends of the candidate connection edge, and use them as the node embedding vectors of the first layer of the graph neural network; For each layer of the graph neural network, for any node and each neighbor node in the set of neighbor nodes of that node, the intermediate vector obtained based on the embedding vector of that node and the neighbor node and the learnable weight matrix is ​​processed by a linear activation function with leakage correction and normalization to obtain the attention coefficient of that neighbor node to that node. The embedding vectors of neighboring nodes are weighted and aggregated according to the attention coefficients, and combined with a nonlinear activation function to update the embedding vector of the node in the next layer. The above calculation process is repeated until the preset number of network layers is reached to obtain the output layer embedding vectors of the nodes at both ends of the candidate connection edge. The intermediate values ​​calculated based on the output layer embedding vectors of the two endpoints of the candidate connection edge and the link prediction score matrix are processed by a nonlinear activation function to obtain the link prediction score of the candidate connection edge.

3. The intelligent information technology consulting method according to claim 1, characterized in that: Candidate connection edges marked as conflict edges are recorded with conflict reasons, which include at least one of version mismatch and protocol incompatibility. The conflict severity quantification value is determined based on the number of candidate connection edges marked as conflict edges and the difference between the link prediction score of each conflict edge and the preset compatibility threshold.

4. The intelligent information technology consulting method according to claim 1, characterized in that: The actions of the cost control agent represent the selection of cloud resource specifications and deployment methods; the actions of the performance optimization agent represent the selection of concurrency processing strategies and caching strategies; the actions of the security and compliance agent represent the selection of data encryption methods and compliance component configurations; and the real-time market data includes cloud vendor computing power prices and open-source component maintenance costs.

5. The intelligent information technology consulting method according to claim 1, characterized in that: The reward function for each agent is determined by multiplying the agent’s cost savings, performance improvement and safety compliance achievement by the corresponding preset weight coefficients, summing the results, and then subtracting the conflict degree quantification value multiplied by the preset conflict penalty coefficient.

6. The intelligent information technology consulting method according to claim 1, characterized in that: The Pareto front solution set is determined by: judging the Pareto dominance relationship of multiple joint action schemes, and including the schemes that have no other schemes that can increase the reward of at least one agent without reducing the reward of any agent into the Pareto front solution set.

7. The intelligent information technology consulting method according to claim 1, characterized in that: Driving the large language model to adjust the node selection strategy and edge generation strategy of the undetermined IT architecture attribute graph includes: using the conflict degree quantification value and the reward deviation value as penalty terms, which together with the text generation loss of the large language model constitute the strategy reconstruction objective function, and the large language model adjusts the parameters of the node selection strategy and the edge generation strategy based on the strategy reconstruction objective function.

8. The intelligent information technology consulting method according to claim 1, characterized in that: Link prediction of candidate connection edges in the attribute graph of the undetermined IT architecture using a graph neural network includes: using a graph convolutional network to perform convolutional aggregation and update on the embedding vectors of the nodes at both ends of the candidate connection edge to obtain the convolutional link prediction score of the candidate connection edge; obtaining the rule verification result of the candidate connection edge from a pre-built explicit compatibility rule base; performing logical operation on the judgment result corresponding to the convolutional link prediction score and the rule verification result; and representing the logical operation result as a compatible candidate connection edge and reserving it in the technically feasible solution space.

9. The intelligent information technology consulting method according to claim 1, characterized in that: The optimization of the technically feasible solution space involves multi-objective game theory, including: taking cost savings, performance improvement and safety compliance as individual fitness evaluation dimensions, and performing iterative search through population initialization, non-dominated sorting, crowding calculation and crossover mutation operations to obtain the Pareto front solution set.

10. An intelligent information technology consulting system, characterized in that, include: The requirement parsing and attribute graph construction module is used to receive the information construction requirement text input by the customer in natural language form, and parse the requirement text into a pending IT architecture attribute graph through a large language model. The pending IT architecture attribute graph includes a set of candidate IT component nodes, a set of candidate connection edges, and a node feature matrix. The architecture illusion elimination module is connected to the requirement analysis and attribute graph construction module. It is used to perform link prediction on candidate connection edges in the attribute graph of the undetermined IT architecture based on the IT component compatibility knowledge graph and through graph neural network to obtain the technically feasible solution space and the quantification value of the degree of conflict. A multi-objective game optimization module, connected to the architecture illusion elimination module, is used to perform multi-objective game optimization within the technically feasible solution space by using a multi-agent reinforcement learning network including a cost control agent, a performance optimization agent, and a security compliance agent, combined with real-time market data, and output a Pareto optimal implementation path report. The closed-loop feedback reconstruction module is connected to the architecture illusion elimination module, the multi-objective game optimization module, and the requirement analysis and attribute graph construction module, respectively. It is used to feed back the conflict degree quantification value and the reward deviation value of the Pareto optimal implementation path report relative to the Pareto front solution set to the requirement analysis and attribute graph construction module, thereby driving the large language model to adjust the node selection strategy and edge generation strategy of the undetermined IT architecture attribute graph.

Images