A method and system for cost-constrained retrieval optimization based on MCTS

Abstract

The application provides a cost constraint retrieval optimization method and system based on MCTS, and relates to the technical field of information retrieval. After receiving a user query, the application first selects a prompt template and personalized parameters according to the query to enhance the request and receives a total budget constraint. Then, parallel preprocessing and iterative planning are performed, the maximum number of iterations is calculated, and the computing resources are allocated. Next, the MCTS retrieval of the cost constraint is executed, the search tree is constructed through multiple iterations, and the selection and expansion of the tree structure are completed according to the node value and the cost in each iteration. After reaching the maximum number of iterations, a preliminary result set is obtained, and the preliminary result set is reordered based on the personalized parameters, and the final retrieval list is output. Finally, according to the user interaction feedback, the selection of the prompt template and the parameters is adaptively optimized, the accuracy and the personalized efficiency of the information retrieval can be improved through intelligent retrieval, cost constraint and feedback optimization.

Description

Technical Field

[0001] This application relates to the field of information retrieval technology, and in particular to a cost-constrained retrieval optimization method and system based on MCTS. Background Technology

[0002] In the field of information retrieval, especially in applications such as Retrieval Augmentation (RAG), a key issue is how to achieve efficient retrieval optimization under limited computing resources or economic budgets. These cost-constrained retrieval optimization methods aim to balance the accuracy of retrieval results with execution costs, and have significant application prospects for building efficient and scalable AI applications, such as intelligent question-answering systems and knowledge management systems.

[0003] In existing technologies, research has attempted to utilize Monte Carlo Tree Search (MCTS) to solve retrieval optimization problems under cost constraints. For example, some schemes employ an MCTS-based strategy framework, constructing a search tree through multiple iterations to find the optimal combination of information blocks in a sequential manner, and directly integrating budget constraints into the optimization process. In the more generalized constrained Markov decision process programming, MCTS is also used as a fundamental method to seek valuable strategies under a given cost threshold through online search.

[0004] However, these existing methods still have shortcomings when dealing with the complexity of retrieval scenarios. They often fail to fully consider the unique characteristics of different user queries, lack targeted personalized processing, and are also insufficient in comprehensively considering the relevance between information blocks and dynamically balancing cost constraints with search depth and accuracy during iterative retrieval. Therefore, existing technologies struggle to adaptively and efficiently complete high-quality retrieval under cost constraints. Summary of the Invention

[0005] The purpose of this application is to provide a cost-constrained retrieval optimization method and system based on MCTS, so as to solve the problem in the prior art that it is difficult to adaptively and efficiently complete high-quality retrieval under cost constraints.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a cost-constrained retrieval optimization method based on MCTS, comprising:

[0007] Receive user query requests, select prompt templates and personalized configuration parameters based on the query requests, enhance the query requests based on the prompt templates, and receive total budget constraints;

[0008] Based on the enhanced query request and total budget constraint, parallel preprocessing and iterative planning are performed to calculate the maximum number of iterations and allocate parallel computing resources.

[0009] The MCTS retrieval process with integrated cost constraints is implemented by constructing a search tree structure through multiple iterations, and the selection and expansion of the search tree structure are completed in each iteration based on node value evaluation and cost constraints.

[0010] After reaching the maximum number of iterations, a preliminary result set is determined, and the preliminary result set is reordered based on the personalized configuration parameters to output the final search result list.

[0011] Based on user feedback on the search results list, the selection of the prompt template and personalized configuration parameters is adaptively optimized.

[0012] Optionally, the MCTS retrieval process with integrated cost constraints constructs a search tree structure through multiple iterations, and in each iteration, selects and expands the search tree structure based on node value evaluation and cost constraints, including:

[0013] The MCTS retrieval process with integrated cost constraints is executed, and the search tree is initialized. The root node in the search tree represents the initial state, and each node in the search tree corresponds to a block combination.

[0014] In each iteration, starting from the root node, the optimal child node is recursively selected until the leaf node is reached, based on the value evaluation results of the child nodes.

[0015] When a leaf node is reached, the estimated cost of the current path is calculated. If the estimated cost does not exceed the total budget constraint, the leaf node is expanded to generate one or more feasible child nodes. If the estimated cost exceeds the total budget constraint, the expansion of the current path is terminated.

[0016] For newly expanded nodes, perform stochastic simulations to obtain reward estimates, and determine the final cost of the corresponding simulated path based on a more accurate calculation method;

[0017] The estimated reward and final cost obtained from the simulation are backpropagated along the search path to update the statistical information of each node on the path, including the number of visits, cumulative reward, and cumulative cost.

[0018] Optionally, the step of adaptively optimizing the selection of the prompt template and personalized configuration parameters based on user interaction feedback with the search results list includes:

[0019] Collect user interaction feedback data on the search results list;

[0020] Based on the interactive feedback data, the relevance calculation model between the prompt template and the query request is optimized to improve the accuracy of prompt template selection in subsequent searches;

[0021] Based on the interactive feedback data, the matching degree calculation model between personalized configuration parameters and query requests is optimized.

[0022] Optionally, the step of performing parallel preprocessing and iterative planning based on the enhanced query request and total budget constraint, calculating the maximum number of iterations, and allocating parallel computing resources includes:

[0023] Based on the enhanced query request, multiple processing threads or processes are launched in parallel to allocate computing resources;

[0024] Based on the total budget constraint and the preset single iteration cost, the maximum number of iterations in the retrieval process is calculated. The maximum number of iterations is the quotient of the total budget constraint and the single iteration cost, and is the minimum of the preset upper limit of the number of iterations.

[0025] Based on the maximum number of iterations and the number of processing threads or processes, a corresponding number of iteration tasks are allocated to each processing thread or process to execute the MCTS retrieval process in parallel.

[0026] Optionally, the step of recursively selecting the optimal child node up to the leaf node, starting from the root node and based on the value evaluation results of the child nodes in each iteration, includes:

[0027] In each iteration, starting from the root node, the value of each child node is evaluated based on the utility function, and the utility value corresponding to each child node is generated.

[0028] The child node with the highest utility value is selected as the optimal child node for recursive selection until a leaf node is reached. The utility function integrates the node's value estimate, exploration incentive, and cost penalty term consisting of cost weight and remaining budget. The value estimate is the initial value estimate adjusted by the reorderer parameters.

[0029] Optionally, the steps of receiving a user query request, selecting a prompt template and personalized configuration parameters based on the query request, enhancing the query request based on the prompt template, and receiving a total budget constraint include:

[0030] Receive query requests from users;

[0031] Based on the query request, calculate and select the prompt template that is most relevant to the query request from the predefined prompt template set, and reconstruct and enhance the query request based on the prompt template;

[0032] Synchronously calculate and select the personalized configuration parameters that best match the query request from a predefined set of personalized configurations;

[0033] It receives the total budget constraint from the external interface.

[0034] Optionally, after reaching the maximum number of iterations, a preliminary result set is determined, and the preliminary result set is reordered based on the personalized configuration parameters to output a final search result list, including:

[0035] The search process terminates after the maximum number of iterations is reached.

[0036] Based on the statistical information of each node, the node with the highest cumulative reward is selected from the search tree, and the block combination corresponding to the node is determined as the preliminary result set.

[0037] The reorderer is invoked to reorder the blocks in the preliminary result set according to the reorderer parameters in the personalized configuration parameters, thereby generating the final search result list.

[0038] Secondly, this application provides a cost-constrained retrieval optimization system based on MCTS, comprising:

[0039] The receiving module is used to receive user query requests, select a prompt template and personalized configuration parameters according to the query request, enhance the query request according to the prompt template, and receive the total budget constraint;

[0040] The planning module is used to perform parallel preprocessing and iterative planning based on the enhanced query requests and total budget constraints, calculate the maximum number of iterations, and allocate parallel computing resources.

[0041] The retrieval module is used to execute the MCTS retrieval process with integrated cost constraints. It constructs a search tree structure through multiple iterations and selects and expands the search tree structure in each iteration based on node value evaluation and cost constraints.

[0042] The output module is used to determine a preliminary result set after reaching the maximum number of iterations, and to reorder the preliminary result set based on the personalized configuration parameters, and output the final search result list.

[0043] The optimization module is used to adaptively optimize the selection of the prompt template and personalized configuration parameters based on the user's interactive feedback on the search results list.

[0044] Thirdly, this application provides an electronic device, comprising:

[0045] Memory, used to store computer programs;

[0046] A processor is configured to execute the computer program to implement the steps of the cost-constrained retrieval optimization method based on MCTS as described in the first aspect above.

[0047] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the cost-constrained retrieval optimization method based on MCTS as described in the first aspect above.

[0048] The cost-constrained retrieval optimization method based on MCTS provided in this application enhances the accuracy and personalization of queries by receiving user queries and intelligently selecting prompt templates and personalized configuration parameters. Furthermore, it combines total budget constraints with parallel preprocessing and iterative planning to achieve reasonable allocation of computing resources and global optimization of the iteration process. By executing MCTS retrieval with integrated cost constraints, the method dynamically selects and expands search paths based on node value and cost constraints in each iteration, enabling efficient exploration of high-quality information combinations with limited resources and avoiding getting trapped in local optima. Finally, the method reorders the preliminary results based on personalized parameters to ensure that the output results better meet user needs. Simultaneously, it continuously adaptively optimizes template and parameter selection based on user feedback, forming a closed-loop learning mechanism. This results in a significant improvement in retrieval performance and system adaptability under controllable costs.

[0049] Furthermore, a search tree is initialized with the root node representing the initial state, and each node corresponds to a combination of information blocks. In each iteration, starting from the root node, the optimal path is recursively selected based on the value evaluation of child nodes until a leaf node is reached. Upon reaching a leaf node, the estimated cost of the current path is calculated. If the cost does not exceed the total budget, a new child node is generated; otherwise, the path is terminated to control the cost. Subsequently, random simulations are performed on the new nodes to obtain a reward estimate, and the final cost of the simulated path is accurately calculated. Finally, the reward value and cost are backpropagated along the search path to update the statistical information such as the number of visits, cumulative reward, and cumulative cost of each node, providing a data foundation for intelligent decision-making in subsequent iterations. By deeply embedding cost evaluation into each iteration of MCTS, dynamic self-management of the retrieval process under strict budget constraints is achieved, effectively avoiding resource waste. Its mechanism based on backpropagation of reward and cost information enables the search tree to continuously accumulate global knowledge, guiding the search direction to converge towards high-value, low-cost areas, thereby overcoming the short-sightedness of traditional methods such as greedy algorithms and achieving better retrieval results within a limited cost. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A flowchart illustrating a cost-constrained retrieval optimization method based on MCTS provided in this application embodiment;

[0052] Figure 2 A flowchart illustrating a specific embodiment of a cost-constrained retrieval optimization method based on MCTS provided in this application;

[0053] Figure 3 A schematic diagram illustrating a specific embodiment of a cost-constrained retrieval optimization method based on MCTS provided in this application.

[0054] Figure 4 This is a schematic diagram of the structure of a cost-constrained retrieval optimization system based on MCTS, provided in an embodiment of this application. Detailed Implementation

[0055] To overcome the shortcomings of existing technologies, this application proposes a cost-constrained retrieval optimization method based on MCTS. The core of this scheme lies in constructing the retrieval process as a globally explorable search tree problem under budget constraints: through parallel preprocessing and iterative planning, the total budget constraint is dynamically transformed into the maximum number of iterations and resource allocation strategies; then, an integrated cost-constrained MCTS retrieval is executed, involving selection, expansion, simulation, and backpropagation steps in each iteration; finally, a closed-loop learning system is formed by combining personalized parameter reordering and adaptive optimization based on user feedback. This method replaces the local decision-making of greedy algorithms with global sampling exploration of MCTS and replaces rigid fixed thresholds with dynamic cost evaluation, thereby enabling proactive optimization within a given budget. This fundamentally solves the problems of limited vision, rigid resource allocation, and insufficient exploration capabilities in existing technologies, significantly improving retrieval accuracy and system adaptive efficiency under complex queries.

[0056] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0057] The core of this application is to provide a cost-constrained retrieval optimization method based on MCTS, and a flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0058] S101. Receive a user query request, select a prompt template and personalized configuration parameters according to the query request, enhance the query request according to the prompt template, and receive the total budget constraint.

[0059] Optionally, step S101 may specifically include the following steps:

[0060] S1011, Receive the query request input by the user.

[0061] A query request is a question or instruction text entered by the user that requires the system to retrieve information.

[0062] S1012. Based on the query request, calculate and select the prompt template with the highest relevance to the query request from the predefined prompt template set, and reconstruct and enhance the query request based on the prompt template.

[0063] Among them, the prompt template is a predefined text frame containing specific instructions or contextual structures, used to reshape or enrich the original query so that it can be better understood by the system.

[0064] S1013. Simultaneously calculate and select the personalized configuration parameters that best match the query request from the predefined set of personalized configurations.

[0065] The personalized configuration parameters are a set of user-specific settings that control retrieval and sorting behavior. The total budget constraint is a numerical limit representing the upper limit of the total computing resources that the system can consume to complete this query.

[0066] S1014, and receive the total budget constraint from the external interface.

[0067] In this embodiment of the application, the user's original query request text is first received from the user interface. For example, the user enters "explain the basic principles of quantum computing and its potential applications".

[0068] Next, in step S1012, the newly received original query is processed. Based on the relevance calculation model, the original query is compared with each template in a predefined set of suggestion templates to calculate a similarity score, and the template with the highest score is selected. The formula is as follows:

[0069]

[0070] in, This represents the optimal suggestion template for the original query q, where P is a predefined set of suggestion templates. It queries the relevance of p under q.

[0071] Next, the original query is populated into the selected template to complete the query reconstruction and enhancement.

[0072] For example, if the selected template is "Please summarize the core ideas and important applications of the following concepts in a clear and easy-to-understand way: [QUERY]", the system will fill the user's query into the [QUERY] field, forming an enhanced query: "Please summarize the core ideas and important applications of the following concepts in a clear and easy-to-understand way: Explain the basic principles of quantum computing and its potential applications".

[0073] Simultaneously, step S1013 executes a process parallel to step S1012. Based on the matching degree calculation model, the current query is matched against various configurations in the predefined set of personalized configurations, selecting the most suitable personalized configuration parameters for the current query. The matching degree calculation model formula is as follows:

[0074]

[0075] in, This query represents the optimal configuration under q, where C is the configuration set. It queries the rating of configuration c under q.

[0076] For example, for the query "explain the basic principles and potential applications of quantum computing," the matching degree calculation model will identify its semantic features of "knowledge dissemination" and "concept explanation." Based on this, the personalized configuration selected by the system will give higher information source weight to "encyclopedia websites" and "authoritative science popularization sources," because these sources can usually provide accurate and easy-to-understand definitions and examples. At the same time, the configuration will set a sorting rule of "explaining the core concepts and principles first, and then introducing derivative applications," so as to conform to the general cognitive logic of people understanding an unfamiliar technical topic from basic to application.

[0077] Finally, in step S1014, the total budget constraint value for this query task is received from an external management interface, such as a resource scheduling system. This value is typically dynamically allocated by the system based on the current overall load or query priority. For example, the system allocates a budget of 1000 "computation units" for this query.

[0078] Specifically, the consultant enters the initial query "Hybrid Office and Innovation" into the platform's search box. Upon receiving this query, the system initiates two parallel processes: First, through a dynamic prompting process, based on relevance calculations, it selects a prompt template from the template library suitable for analyzing causal relationships: "System analysis of the driving factors, empirical evidence, and potential challenges of [topic] on the enterprise [dimension]". The system then fills in "Hybrid Office Model" as the [topic] and "Innovation Capability" as the [dimension], generating an enhanced query: "System analysis of the driving factors, empirical evidence, and potential challenges of the hybrid office model on the enterprise's innovation capability". Simultaneously, the configuration agent matches a personalized configuration based on the query content (research nature) and the user's identity (consultant): the information source weight favors authoritative academic journals and reports from top consulting firms, and the sorting rule is preset to "Evidence Strength Priority". Finally, the platform's resource management system allocates a total computing budget of 1500 units for this research task.

[0079] S102. Based on the enhanced query request and total budget constraint, perform parallel preprocessing and iterative planning, calculate the maximum number of iterations, and allocate parallel computing resources.

[0080] Optionally, step S102 may specifically include the following steps:

[0081] S1021. Based on the enhanced query request, multiple processing threads or processes are launched in parallel to allocate computing resources.

[0082] The enhanced query request is a processed query text containing richer semantics and explicit instructions.

[0083] S1022. Based on the total budget constraint and the preset single iteration cost, calculate the maximum number of iterations in the retrieval process.

[0084] The maximum number of iterations is the minimum of the quotient of the total budget constraint and the cost per iteration, and the minimum of the preset upper limit for the number of iterations. The total budget constraint is the total amount of computational resources the system is allowed to consume for this query. The cost per iteration refers to the average amount of computational resources required to execute one basic loop of Monte Carlo Tree Search (MCTS). The maximum number of iterations is the maximum number of MCTS iterations that can be executed under the total budget constraint. The upper limit for the number of iterations is a safety threshold used to prevent excessive iterations, even when the cost per iteration is low, from causing the program to run for too long.

[0085] S1023. Based on the maximum number of iterations and the number of processing threads or processes, allocate a corresponding number of iteration tasks to each processing thread or process to execute the MCTS retrieval process in parallel.

[0086] In this context, a processing thread or process is the basic unit in a computer that can execute tasks simultaneously, similar to multiple workers working at the same time.

[0087] In this embodiment, firstly, through step S1021, after receiving the enhanced query request, the parallel processing function of the computer operating system is invoked, such as creating a thread pool or forked processes, to simultaneously start multiple computing units. The core advantage of this parallelization strategy lies in significantly shortening the overall processing time, and its principle can be expressed by the following formula:

[0088]

[0089] in, This represents the time required for parallel processing. This represents the time required for sequential processing, while P represents the number of processors running in parallel. The purpose of this step is to pre-allocate computing resources to prepare for subsequent efficient parallel computing.

[0090] Secondly, in step S1022, critical iterative planning is performed. The maximum number of iterations is calculated using the following formula:

[0091]

[0092] Where I is the final determined maximum number of iterations, and Budget is the total budget constraint. It is the average cost of performing one MCTS iteration. This is the system's preset upper limit for the number of iterations. This formula ensures that the number of iterations neither exceeds the budget nor increases indefinitely due to low cost per iteration, always remaining within a reasonable safety threshold.

[0093] For example, if the total budget is 1000 units and the cost per iteration is 20 units, then =50. If If the value is 40, then the final value is I = min(50, 40) = 40 times.

[0094] Finally, in step S1023, the outputs of the first two steps are integrated and distributed. Based on the number of threads or processes P started in step S1021 and the final maximum number of iterations I determined in step S1022, the total number of iteration tasks is distributed as evenly as possible to each computing unit. The number of iteration tasks allocated to each computing unit is... In this way, each thread or process knows exactly how many rounds of MCTS search need to be performed, allowing them to work in parallel and greatly reducing the overall computation time.

[0095] Specifically, based on enhanced queries and a budget of 1500 units, the system performs resource planning. First, it launches five worker processes in parallel, laying the foundation for efficient computation. Second, the system performs iterative planning: the preset average cost of a single MCTS iteration is known. With 30 units, the theoretical maximum number of iterations is calculated to be This is due to the system's preset maximum number of iterations. The maximum number of iterations for this search is determined to be 50, based on the formula I=min(50,60)=50. Finally, the system distributes these 50 iterations evenly across 5 worker processes, with each process responsible for executing 10 MCTS iterations. With the planning complete, the five processes are ready and will begin the search in parallel.

[0096] This application achieves a significant improvement in computing resource utilization and processing efficiency, ensuring that subsequent core retrieval processes are both efficient and within cost limits.

[0097] S103. Execute the MCTS retrieval process with integrated cost constraints, construct the search tree structure through multiple iterations, and complete the selection and expansion of the search tree structure in each iteration based on node value evaluation and cost constraints.

[0098] Optionally, step S103 may specifically include the following steps:

[0099] S1031. Execute the MCTS retrieval process with integrated cost constraints, initialize the search tree, where the root node in the search tree represents the initial state, and each node in the search tree corresponds to a block combination.

[0100] A search tree is a tree-like structure used to represent all possible paths. The root node represents the starting state of the search, where no information blocks have been selected yet, and the leaf nodes represent the ends of the search paths. Each node corresponds to a specific combination of information blocks. A combination of blocks refers to a set of information fragments selected from the knowledge base.

[0101] S1032. In each iteration, starting from the root node, the optimal child node is recursively selected until the leaf node is reached, based on the value evaluation results of the child nodes.

[0102] Specifically, step S1032 includes the following process: In each iteration, starting from the root node, the value of each child node is evaluated based on the utility function to generate the utility value corresponding to each child node; the child node with the highest utility value is selected as the optimal child node for recursive selection until the leaf node is reached. The utility function integrates the node's value estimate, exploration incentive, and cost penalty term consisting of cost weight and remaining budget. The value estimate is the initial value estimate adjusted by the reorderer parameters.

[0103] S1033. When a leaf node is reached, the estimated cost of the current path is calculated. If the estimated cost does not exceed the total budget constraint, the leaf node is expanded to generate one or more feasible child nodes. If the estimated cost exceeds the total budget constraint, the expansion of the current path is terminated.

[0104] The estimated cost refers to the accumulated cost from the root node to the current leaf node.

[0105] S1034. For the newly expanded nodes, perform random simulations to obtain reward estimates, and determine the final cost of the corresponding simulated path based on a more accurate calculation method.

[0106] In this context, stochastic simulation refers to randomly selecting subsequent paths from the current node to the destination to quickly assess the potential value of that path. The reward estimate is a score obtained through stochastic simulation, representing the potential benefit of that path. The final cost is a more accurate calculation of the resources consumed by the stochastically simulated path.

[0107] S1035. The estimated reward and final cost obtained from the simulation are backpropagated along the search path to update the statistical information of each node on the path, including the number of visits, cumulative reward and cumulative cost.

[0108] Backpropagation refers to transmitting the reward and cost information from the leaf nodes back to the root node, updating the statistical data of all nodes along the path. The statistical information mainly includes the number of times a node was visited, the total cumulative reward received, and the cumulative cost incurred; this information is used to guide future search directions.

[0109] In this embodiment of the application, firstly, through step S1031, the system initializes an empty search tree for the current retrieval task. The root node of this tree does not contain any specific information block, but only represents the start state of the retrieval task.

[0110] Secondly, through step S1032, such as Figure 2 As shown, in each iteration, the system starts from the root node and uses a comprehensive utility function to guide the path selection down to the leaf nodes. The specific form of this utility function is:

[0111]

[0112] in, Represents a node The utility value, It is the initial value estimate, i.e., the utility value, after the parameters of the personalized reorderer in step S101 have been adjusted. It represents the number of times the root node has been visited. It is a node The number of visits, c is the exploration constant, λ is the user-configured cost weight, and B is the remaining budget. It is the estimated cost of the current path of this node. It is an exploration incentive that encourages nodes with fewer visits; This is a cost penalty term. Each node selects the child node with the highest utility value, recursively, until a leaf node is reached. For example, in a node, child node A has high value but also high cost, while child node B has moderate value but low cost. Due to the cost penalty term, the utility function will tend to choose B when the budget is tight.

[0113] Next, in step S1033, when a leaf node is reached, the system calculates the estimated path cost from the root node to that leaf node. The calculation formula is as follows:

[0114]

[0115] in Represents a node Total cost Let k represent the length of the i-th block, and k be the number of blocks.

[0116] The estimated cost is compared to the total budget constraint. If the estimated cost does not exceed the budget, the system expands this leaf node, generating one or more new feasible child nodes, i.e., attempting to add new information blocks to grow the tree. If the estimated cost exceeds the budget, the expansion of this path is terminated. This ensures that the search does not waste resources on over-budget paths.

[0117] Then, in step S1034, for the newly expanded nodes, the system performs a fast random simulation to obtain a reward estimate. After the simulation, the system determines the final cost of the simulated path based on a more accurate calculation method, as follows:

[0118]

[0119] in, Represents a node The final cost, i.e., the cost of all information blocks on the path. Sum the number of markers. This indicates the number of tags in the i-th block. This step verifies the accuracy of the estimated cost.

[0120] Finally, in step S1035, the simulated reward estimate and the verified final cost are backpropagated along the search path to update the statistics of all ancestor nodes on the path, including the number of visits, cumulative reward, and cumulative cost. This updated information will be used for utility calculation in the next iteration, enabling the search process to continuously optimize itself.

[0121] Specifically, the core retrieval loop begins. The search tree is initialized, with the root node representing the starting point of the research. For example, in a typical iteration, the current iteration state is as follows:

[0122] The remaining budget B = 1200 units, which is the initial 1500 minus the 300 already consumed;

[0123] Root node access count =15 times;

[0124] The system presets the exploration constant c = 1.5;

[0125] The cost weight λ = 0.008 is obtained based on the personalized configuration.

[0126] At the current node, the system needs to select from three child nodes. The node information is as follows:

[0127] Node A: "Employee Satisfaction Survey Report" block; Value estimation predicted by the block selection module. =0.85; Estimated cost =80 units, estimated cost is calculated based on block length, this report is approximately 8000 words, calculated at 100 words / unit; historical visit count =3 times.

[0128] Node B: "Internal Communication Tool Usage Data" block; Value Estimation =0.78; Estimated cost =45 units (4500-word report); historical visit count =8 times.

[0129] Node C: "Cross-team collaboration case study" block; Value estimation =0.82; Estimated cost =60 units (6000-word case study); historical visit count =2 times.

[0130] Calculate the utility values ​​for nodes A, B, and C respectively:

[0131]

[0132]

[0133]

[0134] Although node A has the highest estimated value, node B has the highest efficiency due to its lowest cost, so the system selects node B to continue the search downwards.

[0135] The total cost of the path from the root node to the current leaf node (through node B) is 285 + 45 = 330 units, where 285 is the existing cost of the path, i.e., the result accumulated from previous iterations, and 45 is the cost of the current node B. Since 330 units < the total budget of 1500 units, the expansion condition is met, and the system expands the leaf node, generating a new child node.

[0136] Next, a random simulation was performed on the newly expanded nodes: the simulated path added an "Innovation Indicator Data" block (cost 38 units) and a "Patent Output Statistics" block (cost 52 units); the simulated reward estimate was 92 points (calculated based on the content relevance model); and the final cost verification was performed, with the total path cost being 330 + 38 + 52 = 420 units.

[0137] The reward value of 92 and the final cost of 420 are then propagated backward along the path:

[0138] Node B: Access count +1, cumulative reward +92, cumulative cost +420;

[0139] Parent node: Visit count +1, cumulative reward +92, cumulative cost +420;

[0140] Root node: Access count increased to 16, cumulative rewards updated.

[0141] This concludes the current iteration. The updated statistics will affect the utility calculation in the next iteration.

[0142] This application implements a retrieval mechanism for intelligent global exploration within a limited budget by constructing a search tree and deeply integrating cost constraints into every core component of MCTS. It not only balances retrieval quality and cost consumption through a utility function but also avoids resource waste through cost detection. Finally, through information backpropagation, the search process continuously optimizes itself, thereby efficiently discovering high-value, low-cost information combinations and overcoming the shortcomings of traditional methods that are prone to getting trapped in local optima and rigid resource allocation.

[0143] S104. After reaching the maximum number of iterations, determine the preliminary result set, reorder the preliminary result set based on the personalized configuration parameters, and output the final search result list.

[0144] Optionally, step S104 may specifically include the following steps:

[0145] S1041. After reaching the maximum number of iterations, the search process is terminated.

[0146] S1042. Based on the statistical information of each node, select the node with the highest cumulative reward from the search tree, and determine the block combination corresponding to the node as the preliminary result set.

[0147] The node with the highest cumulative reward is the one with the largest cumulative reward statistic among all nodes in the search tree. It represents the most valuable combination of information blocks with the best average performance throughout the entire exploration process. The preliminary result set refers to the specific combination of information blocks corresponding to this optimal node.

[0148] S1043. Invoke the reorderer to reorder the blocks in the preliminary result set according to the reorderer parameters in the personalized configuration parameters, and generate the final search result list.

[0149] The reorderer is a dedicated algorithm module responsible for rearranging a set of information blocks according to specific rules. Reorderer parameters are settings included in the personalized configuration to guide the reorderer's operation. The final search results list is an ordered list of information presented to the user after processing by the reorderer.

[0150] In this embodiment, step S1041 first continuously monitors the number of iterations in the MCTS retrieval process. When the number of completed iterations reaches the maximum number of iterations pre-calculated in step S102, the system issues a command to terminate the entire search process. This means that no new selection, expansion, simulation, and backtracking loops are performed, and the structure of the search tree and the statistical information within the nodes are thus fixed.

[0151] Next, in step S1042, data analysis begins on the finalized search tree. It traverses all nodes in the entire search tree, reading the key statistical information of the cumulative reward recorded at each node. The system compares all these cumulative reward values ​​and selects the node with the highest value. This node is considered the most valuable and reliable combination of information blocks explored within a limited budget and number of iterations. Subsequently, the system extracts the specific combination of information blocks represented by this node and determines it as the preliminary result set.

[0152] Finally, in step S1043, the reordering module is invoked. This module receives two inputs: first, the preliminary result set obtained in the previous step, which is a set of information blocks that have not been finally sorted; and second, the personalized configuration parameters selected for the user's query in step S101, which include the reordering parameters guiding the sorting. The reordering module reorders the information blocks in the preliminary result set according to the rules defined by these parameters. After sorting, the final search result list is generated and can be displayed to the user.

[0153] Specifically, the search process terminates after the system detects that the five parallel processes have completed a total of 50 iterations. The system then traverses the entire constructed search tree, searching for the node with the highest cumulative reward. Assuming the information block combination corresponding to this node is: {hybrid office model, employee autonomy, cross-team project incubation, innovation performance indicators}, this combination is determined as the preliminary result set. Subsequently, the reorderer is invoked, which reorders the four information blocks in this preliminary set based on the selected personalized configuration parameters, particularly the "evidence strength priority" sorting rule. Finally, the system outputs an ordered list of search results, as shown in Table 1.

[0154] Table 1. List of search results

[0155]

[0156] Table 1 places the most compelling evidence first, which meets the requirements for the strength of evidence when consultants write research reports, ensuring the high relevance and practicality of the final output.

[0157] This application ensures the controllability and efficiency of the process by terminating the search at appropriate times, and intelligently selects the best-performing information combination from a large number of exploration results based on objective cumulative reward data. Finally, it introduces a re-ranking mechanism based on personalized parameters to transform the initial result of "technically optimal" into the final presentation of "most user-satisfactory", thereby further enhancing the relevance and personalized experience of the search results at the output end. This makes the entire system not only focus on how to find high-quality information, but also on how to deliver this information in a way that best meets the needs of users.

[0158] S105. Based on the user's interactive feedback on the search results list, adaptively optimize the selection of the prompt template and personalized configuration parameters.

[0159] Optionally, step S105 may specifically include the following steps:

[0160] S1051. Collect user interaction feedback data on the search results list.

[0161] Interactive feedback data refers to all direct or indirect behavioral records generated by users after receiving the search results list, such as clicking on a result, ignoring a result, dwell time, and explicit likes, dislikes, and other comments.

[0162] S1052. Based on the interactive feedback data, optimize the correlation calculation model between the prompt template and the query request to improve the accuracy of prompt template selection in subsequent searches.

[0163] The relevance calculation model is a mathematical model used to evaluate the degree of matching between a suggestion template and a specific query request, and its output is a relevance score.

[0164] S1053. Based on the interactive feedback data, simultaneously optimize the matching degree calculation model between personalized configuration parameters and query requests.

[0165] Among them, the matching degree calculation model is a mathematical model used to evaluate the degree of fit between a set of personalized configuration parameters and a specific query request, so as to improve the accuracy of personalized configuration parameter selection in subsequent retrieval and form a self-improving closed-loop learning system.

[0166] In this embodiment, firstly, through step S1051, after presenting the final search results list to the user, the system continuously and silently collects user interaction data. For example, the system records which result in the list the user clicked, whether they quickly skipped some results, how long they stayed on a result details page, or whether they explicitly gave a "like" or "dislike" rating. This raw behavioral data is aggregated to form interactive feedback data used to optimize the system.

[0167] Secondly, in step S1052, the collected feedback data is used to optimize the selection capability of the prompt template. Specifically, the system collects each query request, the selected prompt template at that time, and the corresponding user feedback. This data is then used to fine-tune or update the relevance calculation model between the prompt template and the query request. For example, if a certain type of query, such as "compare A and B," receives positive feedback after using the "comparative analysis" template, the model will learn this association. In the future, when encountering similar queries, the model will be more inclined to choose the "comparative analysis" template, thereby improving the accuracy of template selection.

[0168] Simultaneously, step S1053 utilizes the same batch of feedback data in parallel to optimize the selection capability of personalized configuration parameters. It uses the query request, the currently applied configuration parameters, and user feedback to fine-tune or update the matching model between personalized configuration parameters and the query request. For example, if it is found that a user is always more satisfied with the results given by the configuration that prioritizes displaying official documentation when faced with technical queries, the model will learn this user preference. Through the collaborative optimization of steps S1052 and S1053, more suitable prompt templates and configuration parameters can be matched for queries in subsequent processing, thus forming a closed-loop learning system that continuously learns from experience and constantly improves itself.

[0169] Specifically, in the final draft of the research report, the consultant extensively cited evidence from the results list regarding "innovation performance indicators" and "cross-team project incubation," explicitly marking these materials as "highly valuable." The system collected this positive feedback. Subsequently, it used this feedback to optimize the relevance calculation model of the prompt templates, strengthening the association between templates related to "analyzing driving factors and evidence" and keywords such as "hybrid office" and "innovation." Simultaneously, it also optimized the matching degree calculation model of the configuration parameters, making it more inclined to automatically select the "evidence strength priority" sorting rule when processing "corporate innovation" queries for this consultant or similar professional users in the future. Through this interaction, the system completed a closed-loop learning process, becoming more intelligent and accurate when serving similar research tasks in the future.

[0170] This application dynamically optimizes two core computational models—prompt template selection and personalized configuration parameter matching—by collecting real user interaction feedback and using it as valuable training data. This enables the system to no longer statically execute rules but to self-adjust and evolve based on users' actual satisfaction and behavioral preferences. Ultimately, it constructs an adaptive closed-loop learning system that becomes increasingly intelligent and personalized with use, significantly improving user experience and long-term system performance.

[0171] The following is a complete example for steps S101 to S105, such as Figure 3 As shown:

[0172] A user entered the query "Planning a 7-day in-depth cultural tour of Japan, with a moderate budget" into a travel planning application. Upon receiving the query, the system initiated two parallel processes: First, based on a relevance calculation model, it identified the key intent "in-depth cultural tour" and the constraints "7 days" and "moderate budget" in the query. It then selected the most suitable prompt template from the template library, enhancing the brief user input into a detailed instruction: "Plan a 7-day travel itinerary for a traveler with a strong interest in traditional Japanese culture. The focus should include historical sites, traditional craft experiences, and local cultural activities, with an overall budget kept at a moderate level. Please provide city suggestions, daily activity arrangements, and approximate budget allocation." Simultaneously, based on a matching degree calculation model and the user's past travel records, it displayed a preference for independent travel and off-the-beaten-path attractions, loading a set of personalized configuration parameters for the user. When selecting attractions, it prioritized options with "low tourist density" and "highly interactive experiences." Furthermore, based on current computing resources, the system allocated a computational budget of 1500 units for this planning task.

[0173] Based on this enhanced, clear query and total budget, parallel preprocessing and iterative planning began. It first launched three computational threads to share the complex planning task. Next, based on the total budget of 1500 units and the estimated cost of a single planning iteration of 30 units, the system calculated a theoretical number of iterations of 50. Since the system's maximum iteration limit was set at 60, the maximum number of iterations for this search was ultimately determined to be 50. Then, the system roughly evenly distributed these 50 iterations among the three threads, with two threads each performing 17 iterations and one thread performing 16 iterations.

[0174] Subsequently, three threads execute the integrated cost-constrained MCTS search in parallel to explore the optimal travel options. The search tree is initialized, with the root node representing the "start of the trip". In each iteration, the algorithm starts from the root node and uses a utility function at different decision points, such as whether to choose "Kyoto" or "Nara" as the first stop, or whether to arrange a "tea ceremony experience" or a "museum visit" on the second day. The optimal branch is selected, and this selection process comprehensively considers the estimated experience value of the option, the degree to which it has been explored, and the impact of the required cost on the total budget.

[0175] When a path is explored to its limit, such as when a preliminary 7-day itinerary is drafted, the system checks whether its total estimated cost exceeds the medium budget constraint. For example, does the total cost of transportation and entrance fees exceed the medium budget constraint? If not, new possible activities are expanded at that node. For the newly expanded nodes, such as adding a "kimono rental" option to the itinerary, a quick simulation evaluation is performed. This involves randomly combining other activities to obtain an experience score and calculating a more accurate simulated total cost. This score and cost are then backpropagated to update the statistical data for each node along the entire search path.

[0176] After all 50 iterations are completed, the system stops searching. It then analyzes the entire search tree, finding the node with the highest cumulative experience reward. This node corresponds to a preliminary 7-day Japanese cultural tour plan with optimal overall evaluation, such as a "Kyoto-Nara-Osaka" route combination, and identifies this as the preliminary result set. Next, the system invokes a reordering engine to fine-tune the attractions in the plan based on the user's personalized configuration and preference for "low tourist density." For example, it might replace a popular but crowded temple with a quieter but equally culturally rich garden, thus generating a final personalized travel plan list for the user.

[0177] The user carefully read the itinerary, showing particular interest in the "traditional morning markets in Nara" and "nighttime geisha performances in Kyoto," clicking to view details and saving them, but ignoring the system-recommended "Universal Studios Japan" option. After collecting this interaction data, the system optimized its prompt template relevance model, enabling it to more accurately match templates emphasizing "localization" and "traditional experiences" for future "in-depth cultural tour" queries. Simultaneously, it also optimized the configuration matching model, further reinforcing the user's preference for "niche" and "immersive" experiences. When the user plans a similar trip next time, the system will be able to provide more accurate recommendations, forming a closed loop of continuous learning and optimization.

[0178] Figure 4 This is a schematic diagram illustrating a specific implementation of a cost-constrained retrieval optimization system based on MCTS, as provided in this application. Figure 4 The system may include:

[0179] The receiving module 41 is used to receive user query requests, select a prompt template and personalized configuration parameters according to the query request, enhance the query request according to the prompt template, and receive the total budget constraint;

[0180] Planning module 42 is used to perform parallel preprocessing and iterative planning based on the enhanced query request and total budget constraint, calculate the maximum number of iterations and allocate parallel computing resources;

[0181] The retrieval module 43 is used to perform the MCTS retrieval process with integrated cost constraints. It constructs a search tree structure through multiple iterations and selects and expands the search tree structure in each iteration based on node value evaluation and cost constraints.

[0182] The output module 44 is used to determine a preliminary result set after the maximum number of iterations is reached, and to reorder the preliminary result set based on the personalized configuration parameters, and output the final search result list.

[0183] The optimization module 45 is used to adaptively optimize the selection of the prompt template and personalized configuration parameters based on the user's interactive feedback on the search result list.

[0184] The cost-constrained retrieval optimization system based on MCTS in this application is used to implement the aforementioned cost-constrained retrieval optimization method based on MCTS. Therefore, the specific implementation of the cost-constrained retrieval optimization system based on MCTS can be found in the embodiment section of the cost-constrained retrieval optimization method based on MCTS above. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0185] This application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of any of the above-described cost-constrained retrieval optimization methods based on MCTS.

[0186] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described cost-constrained retrieval optimization methods based on MCTS.

[0187] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0188] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the cost constraint retrieval optimization method based on MCTS.

[0189] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0190] The above provides a detailed description of the cost-constrained retrieval optimization method and system based on MCTS provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A cost-constrained retrieval optimization method based on MCTS, characterized in that, include: Receive user query requests, select prompt templates and personalized configuration parameters based on the query requests, enhance the query requests based on the prompt templates, and receive total budget constraints; Based on the enhanced query request and total budget constraint, parallel preprocessing and iterative planning are performed to calculate the maximum number of iterations and allocate parallel computing resources. The MCTS retrieval process with integrated cost constraints is implemented by constructing a search tree structure through multiple iterations, and the selection and expansion of the search tree structure are completed in each iteration based on node value evaluation and cost constraints. After reaching the maximum number of iterations, a preliminary result set is determined, and the preliminary result set is reordered based on the personalized configuration parameters to output the final search result list. Based on user feedback on the search results list, the selection of the prompt template and personalized configuration parameters is adaptively optimized. The MCTS retrieval process with integrated cost constraints constructs a search tree structure through multiple iterations, and in each iteration, selects and expands the search tree structure based on node value evaluation and cost constraints, including: The MCTS retrieval process with integrated cost constraints is executed, and the search tree is initialized. The root node in the search tree represents the initial state, and each node in the search tree corresponds to a block combination. In each iteration, starting from the root node, based on the value evaluation results of the child nodes, the optimal child node is recursively selected up to the leaf node, including: In each iteration, starting from the root node, the value of each child node is evaluated based on the utility function, and the utility value corresponding to each child node is generated. The child node with the highest utility value is selected as the optimal child node and recursively selected until a leaf node is reached. The utility function integrates the node's value estimate, exploration incentive, and cost penalty term consisting of cost weight and remaining budget. The value estimate is the initial value estimate after adjustment by the reorderer parameters. When a leaf node is reached, the estimated cost of the current path is calculated. If the estimated cost does not exceed the total budget constraint, the leaf node is expanded to generate one or more feasible child nodes. If the estimated cost exceeds the total budget constraint, the expansion of the current path is terminated. For newly expanded nodes, perform random simulations to obtain reward estimates, and determine the final cost of the corresponding simulation path based on the number of tags in each information block in the simulation path; The estimated reward and final cost obtained from the simulation are backpropagated along the search path to update the statistical information of each node on the path, including the number of visits, cumulative reward, and cumulative cost.

2. The method according to claim 1, characterized in that, The step of adaptively optimizing the selection of the prompt template and personalized configuration parameters based on user interaction feedback on the search results list includes: Collect user interaction feedback data on the search results list; Based on the interactive feedback data, the relevance calculation model between the prompt template and the query request is optimized to improve the accuracy of prompt template selection in subsequent searches; Based on the interactive feedback data, the matching degree calculation model between personalized configuration parameters and query requests is optimized.

3. The method according to claim 1, characterized in that, The process of performing parallel preprocessing and iterative planning based on the enhanced query request and total budget constraint, calculating the maximum number of iterations, and allocating parallel computing resources includes: Based on the enhanced query request, multiple processing threads or processes are launched in parallel to allocate computing resources; Based on the total budget constraint and the preset single iteration cost, the maximum number of iterations in the retrieval process is calculated. The maximum number of iterations is the quotient of the total budget constraint and the single iteration cost, and is the minimum of the preset upper limit of the number of iterations. Based on the maximum number of iterations and the number of processing threads or processes, a corresponding number of iteration tasks are allocated to each processing thread or process to execute the MCTS retrieval process in parallel.

4. The method according to claim 1, characterized in that, The process of receiving a user query request, selecting a prompt template and personalized configuration parameters based on the query request, enhancing the query request based on the prompt template, and receiving a total budget constraint includes: Receive query requests from users; Based on the query request, calculate and select the prompt template that is most relevant to the query request from the predefined prompt template set, and reconstruct and enhance the query request based on the prompt template; Synchronously calculate and select the personalized configuration parameters that best match the query request from a predefined set of personalized configurations; It receives the total budget constraint from the external interface.

5. The method according to claim 1, characterized in that, After reaching the maximum number of iterations, a preliminary result set is determined, and the preliminary result set is reordered based on the personalized configuration parameters to output a final search result list, including: The search process terminates after the maximum number of iterations is reached. Based on the statistical information of each node, the node with the highest cumulative reward is selected from the search tree, and the block combination corresponding to the node is determined as the preliminary result set. The reorderer is invoked to reorder the blocks in the preliminary result set according to the reorderer parameters in the personalized configuration parameters, thereby generating the final search result list.

6. A cost-constrained retrieval optimization system based on MCTS, characterized in that, include: The receiving module is used to receive user query requests, select a prompt template and personalized configuration parameters according to the query request, enhance the query request according to the prompt template, and receive the total budget constraint; The planning module is used to perform parallel preprocessing and iterative planning based on the enhanced query requests and total budget constraints, calculate the maximum number of iterations, and allocate parallel computing resources. The retrieval module is used to execute the MCTS retrieval process with integrated cost constraints. It constructs a search tree structure through multiple iterations and selects and expands the search tree structure in each iteration based on node value evaluation and cost constraints. The output module is used to determine a preliminary result set after reaching the maximum number of iterations, and to reorder the preliminary result set based on the personalized configuration parameters, and output the final search result list. The optimization module is used to adaptively optimize the selection of the prompt template and personalized configuration parameters based on the user's interactive feedback on the search results list. The MCTS retrieval process with integrated cost constraints constructs a search tree structure through multiple iterations, and in each iteration, selects and expands the search tree structure based on node value evaluation and cost constraints, including: The MCTS retrieval process with integrated cost constraints is executed, and the search tree is initialized. The root node in the search tree represents the initial state, and each node in the search tree corresponds to a block combination. In each iteration, starting from the root node, based on the value evaluation results of the child nodes, the optimal child node is recursively selected up to the leaf node, including: In each iteration, starting from the root node, the value of each child node is evaluated based on the utility function, and the utility value corresponding to each child node is generated. The child node with the highest utility value is selected as the optimal child node and recursively selected until a leaf node is reached. The utility function integrates the node's value estimate, exploration incentive, and cost penalty term consisting of cost weight and remaining budget. The value estimate is the initial value estimate after adjustment by the reorderer parameters. When a leaf node is reached, the estimated cost of the current path is calculated. If the estimated cost does not exceed the total budget constraint, the leaf node is expanded to generate one or more feasible child nodes. If the estimated cost exceeds the total budget constraint, the expansion of the current path is terminated. For newly expanded nodes, perform random simulations to obtain reward estimates, and determine the final cost of the corresponding simulation path based on the number of tags in each information block in the simulation path; The estimated reward and final cost obtained from the simulation are backpropagated along the search path to update the statistical information of each node on the path, including the number of visits, cumulative reward, and cumulative cost.

7. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the cost-constrained retrieval optimization method based on MCTS as described in any one of claims 1 to 5 when executing the computer program.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the cost-constrained retrieval optimization method based on MCTS as described in any one of claims 1 to 5.

Images