Large model cache active updating method and system for data retrieval and computer readable storage medium

Abstract

The invention provides a large model cache active updating method and system for data retrieval, and belongs to the technical field of data management. The method comprises the steps of determining an update target based on update data of a retrieval processing target, and determining a matching deviation target and matching data of the matching deviation target and a retrieval information item by analyzing a matching deviation condition of the update target and a cache information item. The core improvement of the method is as follows: a dynamic time sequence knowledge graph is constructed and maintained, and the graph is associated with a data source, a cache information item and a retrieval context; after the system recognizes a matching deviation target, actively predicting an influence curve of the matching deviation target on future related query and existing cache information item values based on the map and a pre-trained meta-gradient prediction model; and finally, according to the predicted influence degree and update cost, formulating differentiated preventive update, degradation or maintenance strategies for different cache information items. According to the method, a traditional passive response type cache updating mechanism is upgraded into active optimization control based on influence prediction, the problem of insufficient cache timeliness caused by updating delay is solved, and the accuracy and response efficiency of large model data retrieval are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of data management and artificial intelligence technology, specifically relating to a method, system, and computer-readable storage medium for actively updating a large model cache for data retrieval. Background Technology

[0002] With the widespread application of large-scale pre-trained models (hereinafter referred to as "large models") in fields such as medical information consultation, academic research, and intelligent customer service, data retrieval based on large models has become a core interaction method. Caching high-frequency queries and their accurate results is a common technique to improve response speed and computational efficiency. However, data sources in the real world are continuously and dynamically updated, which can cause cached content to lag or even become invalid, resulting in the return of inaccurate or outdated information and damaging the credibility and timeliness of the retrieval system.

[0003] To address the issue of updating cached data, existing technologies have proposed several solutions. For example, patent document CN120256544B discloses a method for applying a large model to data retrieval. The main idea of ​​this method is to monitor the matching deviation between newly added retrieval processing targets (called update targets) and existing cached information items, thereby identifying targets with significant matching deviations from the cache system. Subsequently, by analyzing the frequency and matching status of these target occurrences in historical retrieval records, the effectiveness of the overall cached data is determined, and a differentiated update strategy is formulated. This method, by introducing matching deviation analysis, achieves a certain degree of proactive discovery of caching problems.

[0004] However, the aforementioned existing technical solutions have fundamental limitations: they are essentially still a passive response and post-event statistical mechanism. The system must wait for the matching deviation target to accumulate a sufficient number of user queries (i.e., retrieval processing times) before it can make an update decision based on the statistical results. From the actual update of the data source to the user query behavior reflecting the problem, and then to the system triggering action based on the statistical threshold, there is a significant time delay. During this period, a large amount of related, gradually expired information may already exist in the cache, affecting the user experience. In addition, this solution lacks forward-looking prediction of the impact of data updates and cannot preventively handle related cached information items that have not yet been frequently queried but have been affected by updates, resulting in insufficient intelligence and timeliness of the system. Summary of the Invention

[0005] The purpose of this invention is to overcome the aforementioned deficiencies of the prior art and provide a method and system for proactively updating large model caches for data retrieval. Building upon the existing technology's method of identifying update targets and matching deviation targets to pinpoint data source changes, this invention introduces a prediction and proactive intervention mechanism, upgrading cache management from lagging correction to forward-looking optimization.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for proactively updating a large model cache for data retrieval, comprising the following steps: Step S110: determining the update target and the matching deviation target; Step S120: acquiring historical matching context data; Step S130: constructing and maintaining a temporal knowledge graph; Step S140: performing proactive influence prediction based on the temporal knowledge graph and the meta-gradient prediction model; Step S150: generating and executing a differentiated proactive caching operation strategy.

[0007] Step S110 includes: constructing a retrieval processing target set; when a new target appears in the set, or the description text of an existing target is modified beyond a preset threshold, the target is marked as an updated target. Calculate the similarity between each update target and the text vector of all cached information items; identify update targets whose maximum similarity with the text vector of all cached information items is lower than a preset first similarity threshold as matching deviation targets.

[0008] The temporal knowledge graph includes a node set and an edge set; the node set includes at least a data source node, a cached information item node, and a retrieval context feature node; the edge set represents the relationships between nodes, including source edges, related query edges, and influence transmission edges; the construction and maintenance of the temporal knowledge graph in step S130 specifically includes: S131: Based on the metadata of the data source node, the content of the cached information item, and its attribution relationship, establish source edges connecting the data source node and the cached information item node directly generated by it; S132: Based on the historical matching context data obtained in step S120, obtain historical retrieval logs, extract frequently co-occurring retrieval keywords or context features, form retrieval context feature nodes, and establish related query edges connecting the cached information item node and the retrieval context feature node; S133: Establish influence transmission edges connecting the data source node to the cached information item node that is not directly generated by it but is indirectly affected by its updates.

[0009] The meta-gradient prediction model is a time series prediction model based on an attention mechanism. It is trained by collecting a set of matching deviation targets in historical time periods, a snapshot of the time series knowledge graph at the corresponding time, and the actual change data of the value indicators of relevant cached information items in a subsequent period to form a training sample set. The model is trained under supervision by taking the feature vector of the matching deviation target and the structural features of the relevant sub-graph as inputs and the gradient sequence of the value changes of relevant cached information items at multiple future time points as the prediction target.

[0010] The method according to claim 4, wherein step S140 comprises: S141: Locate and match nodes related to the deviation target in the time-series knowledge graph; S142: Starting from the node related to the matching deviation target, based on the associated query edge and the influence propagation edge, traverse and extract all cached information item nodes within the preset number of hops to form a local subgraph. S143: The feature representation of the matching deviation target and the structured information of the sub-map are used as inputs and fed into the pre-trained meta-gradient prediction model for active influence prediction. The output is a value decay prediction curve for each cached information item. The expression of the prediction curve is V_i(t)=V_i0*exp(-λ_i*t), where V_i0 is the current value and λ_i is the decay rate predicted by the model.

[0011] Step S150 includes: extracting the instantaneous decay rate k_i of the first time point T1 within a preset short-term time period from the prediction curve; estimating the computational resource cost C_c required to verify and update the cached information item C_i; setting a first cost sensitivity coefficient α and a second cost sensitivity coefficient β, where α>β>0; if k_i>α*C_c, then a preventive proactive update strategy is generated and executed for the cached information item, immediately triggering the update process for the content of the cached information item; if β*C_c≤k_i≤α*C_c, then a dynamic degradation marking strategy is generated and executed for the cached information item, lowering the priority of the cached information item in the retrieval results and attaching an information update identifier; if k_i<β*C_c, then a status quo maintenance strategy is generated and executed for the cached information item, keeping its current state unchanged.

[0012] The execution of the preventive proactive update strategy includes: locating the authoritative data source that needs to be updated based on the matching deviation target and the time-series knowledge graph; obtaining the latest information from the data source and generating candidate new cached content; performing quality and relevance checks on the candidate new cached content; and after passing the checks, replacing the original cached information item content and updating its weights on the relevant edges in the time-series knowledge graph.

[0013] The method further includes step S160: after executing the active caching operation strategy, continuously monitor the strategy execution effect; collect user feedback data on subsequent interactions with the processed cached information items; and use the feedback data to perform online incremental learning and optimization of the parameters of the meta-gradient prediction model and the edge weights in the temporal knowledge graph.

[0014] Secondly, this invention provides a large-scale model cache proactive update system for data retrieval, comprising: a data update perception module for monitoring changes in the retrieval processing target and determining the update target and the matching deviation target; a retrieval behavior analysis module for acquiring historical matching data between the matching deviation target and the retrieval information item; a knowledge graph management module for constructing, storing, and maintaining the time-series knowledge graph; a proactive prediction module, which incorporates the meta-gradient prediction model for outputting a prediction value decay impact curve based on the matching deviation target and the knowledge graph; and a strategy decision-making and execution module for generating and executing differentiated proactive caching operation strategies based on the prediction curve and cost calculation.

[0015] Thirdly, the present invention provides a medium storing a computer program that, when executed, can implement the above-described method.

[0016] The beneficial effects of this invention are as follows: 1. Achieved true proactive updates: By modeling data associations through time-series knowledge graphs and using predictive models to proactively assess the impact of updates, the system can identify risks and take action in advance before user queries reveal problems, eliminating the statistical delay window in traditional methods.

[0017] 2. Significantly improved level of decision-making intelligence and refinement: The decision-making basis has shifted from simple historical frequency thresholds to dynamic optimization calculations based on predicted value curves and resource costs, making cache update strategies more economical, reasonable and adaptable.

[0018] 3. The system has the ability to continuously evolve: Through a closed-loop feedback mechanism, the prediction model and knowledge graph are continuously optimized using real effect data after the strategy is executed, forming a self-improving intelligent system.

[0019] 4. Enhanced compatibility and substantial leap forward with existing technologies: This invention fully retains the effective logic of locating the source of data change (matching deviation target) in existing technologies, and builds a brand-new prediction and active control layer on this basis, which is a substantial expansion of its capability dimensions and a significant improvement in performance. Attached Figure Description

[0020] Figure 1 The following is a flowchart of a cache proactive update method provided in one embodiment of the present invention.

[0021] Figure 2 This is a framework diagram of a cache proactive update system provided in one embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention.

[0023] Example 1: Detailed Description of the Method Flow refer to Figure 1 This embodiment details the steps of a large model cache proactive update method for data retrieval according to the present invention.

[0024] Step S110: Determine the update target and the matching deviation target.

[0025] This step aims to detect changes at the data source level and pinpoint their initial impact on the existing caching system. The specific implementation is as follows: The system maintains a set of search and processing targets, which can be derived from a frequently asked questions database, a list of trending topics, or clusters of historically high-frequency queries. When a new target is added to the set, or when the description text of an existing target is modified beyond a preset threshold, the target is marked as an updated target. For example, in the healthcare field, when "the transmission characteristics of the Omega subtype variant BA.2.86" is added to the target database, it becomes an updated target.

[0026] For each update target, it is transformed into a high-dimensional feature vector using the same text encoder as the large model (such as BERT or SentenceTransformer). At the same time, each "cache information item" in the cache database (usually a vectorized representation of a "question-answer" pair or a knowledge fragment) also has its own text vector.

[0027] Next, the cosine similarity between the update target vector and the vectors of all cached information items in the cache library is calculated. Update targets whose maximum similarity with the text vectors of all cached information items is lower than a preset first similarity threshold are identified as mismatch targets. The first similarity threshold is set to θ (for example, θ = 0.35). If the maximum similarity between an update target and all cached information items is lower than θ, it is determined that the knowledge point represented by the target lacks a highly approximate representation in the current caching system, and therefore it is identified as a mismatch target. This step accurately identifies new knowledge gaps or major changes caused by data development.

[0028] Step S120: Obtain historical matching context data.

[0029] To provide context for subsequent predictions, the system extracts recent historical data (such as the past 24 hours or the last 1000 queries) from the retrieval logs. The main statistical analysis focuses on retrieval data related to the matching deviation target identified in step S110. For example, it analyzes the frequency and trends of user queries containing the target keyword or its synonyms and related terms, as well as the contextual characteristics of these queries (such as co-occurring search terms). This data reflects the initial user attention to new knowledge points.

[0030] Step S130: Construct and maintain the time-series knowledge graph.

[0031] Temporal knowledge graphs are a knowledge representation method that models entities, concepts, and their relationships using a graph structure, allowing the attributes of nodes and edges to evolve over time. In this invention, it is used to formally represent the relationships between data sources, cached information items, and retrieval contexts. The temporal knowledge graph includes a set of nodes and a set of edges. The set of nodes includes at least data source nodes, cached information item nodes, and retrieval context feature nodes. The set of edges represents the relationships between nodes, including source edges, related query edges, and influence propagation edges. Each edge is assigned a dynamic weight.

[0032] (1) Definition of nodes in a temporal knowledge graph: Data source nodes: Represent the original source of information, such as a specific API interface, authoritative database, or news website URL. Each node contains attributes such as source ID, name, credibility level, and last update time.

[0033] Cache information item node: Represents a single cached knowledge unit. Node attributes include valuable metrics such as cache ID, content vector, creation time, historical click-through rate, and average user satisfaction rating.

[0034] Retrieval context feature nodes: These represent topics, intents, or keywords abstracted from historical queries. Examples include "hypertension complications" and "vaccination interval." These nodes serve as a bridge connecting user queries with cached knowledge.

[0035] (2) Definition and maintenance of graph edges and dynamic weights: Source edge: Connects the data source node to the cached information item node directly generated by it. The weight W_1 of the source edge is initially set to 1.0. Its existence indicates an explicit source relationship.

[0036] Association query edge: Connects the cached information item node to the retrieval context feature node. The association query edge weight W_2 represents the semantic relevance between the cached information item and the retrieval context. The initial value of the association query edge weight W_2 can be obtained by calculating the semantic similarity between the cached information item content and the feature word set, and then dynamically adjusted according to the user's click behavior on the cached information item in the context (e.g., the higher the click-through rate, the more W_2 is increased).

[0037] Influence transmission edge: This connects the data source node to cached information item nodes that are not directly generated by the data source node but are indirectly affected by its updates, and is key to achieving the prediction function. The influence transmission edge weight W_3 represents the strength of the influence transmission. For example, when "Medical Journal A" publishes a new side effect study on "Drug D" (forming an update target), this may affect not only cached information items that directly describe the side effects of "Drug D", but also cached information items that describe "Drug B used in combination with Drug D". The initial value of the influence transmission edge weight W_3 can be set based on domain rules (such as drug co-occurrence relationships), and then continuously optimized through machine learning: when an update of the data source node S is detected, the system tracks the changes in the value indicators (such as satisfaction) of all cached information items C connected to the data source node S through influence transmission edges over a subsequent period of time, and adjusts W_3 in reverse through regression analysis and other methods to make it more consistent with the actual observed influence strength. It can also dynamically adjust the update influence transmission weight W_3 from the data source node to the cached information item node based on the changes in the value indicators of the updated event sequence and subsequent associated cached information items through causal inference or correlation analysis.

[0038] Maintaining the graph is an ongoing process. As new cached information items are added, new query patterns emerge, and data source update events occur, nodes and edges are dynamically added or deleted, and weights are dynamically updated.

[0039] The construction and maintenance of the time-series knowledge graph in step S130 specifically includes: S131: Based on the metadata of the data source node, the content of the cached information items and their ownership relationships, establish a source edge connecting the data source node and the cached information item node directly generated by it; S132: Based on step S120, obtain historical matching context data, obtain historical search logs, extract frequently co-occurring search keywords or context features, form search context feature nodes, and establish association query edges connecting cached information item nodes and search context feature nodes. S133: Establish an influence propagation edge connecting the data source node to a cached information item node that is not directly generated by the data source node but is indirectly affected by its updates.

[0040] Step S140: Perform proactive impact prediction based on temporal knowledge graph and meta-gradient prediction model.

[0041] Once a matching deviation target T_new is determined in step S110, the system immediately starts the prediction process. Step S140 includes: S141: Location and Association: First, locate and match nodes related to the deviation target T_new in the time-series knowledge graph. Typically, T_new will map to one or more retrieval context feature nodes. Simultaneously, by analyzing the content of T_new, it is possible to associate it with one or more data source nodes that may have published this information.

[0042] S142: Extract the influence subgraph: Starting from these related nodes, based on the association query edges and influence propagation edges, traverse all cached information item nodes within a preset number of hops (2 hops in this embodiment) to form a local subgraph. This subgraph contains all cached information items that may be directly or indirectly affected by T_new.

[0043] S143: Meta-gradient prediction model prediction: The feature representation of T_new and the structured information of the subgraph (such as adjacency matrix, edge weight matrix, node attributes) are fed into the pre-trained meta-gradient prediction model as input.

[0044] The meta-gradient prediction model is a machine learning model specifically designed to predict the impact of an intervention (in this invention, the occurrence of a matching deviation target T_new) on the future state (in this invention, the value of cached information items) in a complex system. It can be built upon architectures such as Graph Neural Networks (GNNs) and spatiotemporal attention mechanisms. In this invention, the meta-gradient prediction model is a time-series prediction model based on an attention mechanism, trained as follows: A training sample set is constructed by collecting a set of matching deviation targets within a historical time period, snapshots of the temporal knowledge graph at corresponding times, and actual changes in the value indicators of relevant cached information items over a subsequent period. The model is trained under supervision, using the feature vector of the matching deviation target and the structural features of the relevant subgraph as input, and the gradient sequence of value changes of relevant cached information items at multiple future time points as the prediction target. The goal of the model learning is to output the gradient of the value change of each cached information item C_i in the subgraph at multiple future time points (t1, t2, ... tk), given a new "target-graph" pair.

[0045] The model outputs a value decay prediction curve for each cached information item C_i, expressed as V_i(t) = V_i0 * exp(-λ_i * t), where V_i0 is the current value and λ_i is the predicted decay rate. The value decay prediction curve quantifies how the utility of each cached information item C_i will decline without any intervention.

[0046] Step S150: Generate and execute a differentiated proactive caching operation strategy. This step translates predictions into concrete actions. For each cached information item C_i in the sub-graph and its prediction curve V_i(t): 1. Calculate key decision indicators: Extract the instantaneous decay rate k_i at the first time point T1 within a preset short-term period (within 24 hours in this embodiment) from the prediction curve. The larger k_i is, the faster the value is lost and the more urgent the situation.

[0047] 2. Assess update costs: Estimate the computational resource costs C_c required to validate and update cached information item C_i, including the overhead of querying the latest data source, calling the large model to regenerate or revise the answer, and performing compliance checks.

[0048] 3. Strategy Decision Engine: Set a first cost sensitivity coefficient α and a second cost sensitivity coefficient β, where α>β>0, to balance revenue and cost.

[0049] If k_i > α*C_c: the predicted rate of value loss is much higher than the update cost, indicating that preventative proactive updating is cost-effective. Therefore, a preventative proactive update strategy is generated and executed for this cached information item, immediately triggering the update process for the cached information item's content. The execution of the preventative proactive update strategy includes: locating the data source that needs updating based on the matching deviation target and the time-series knowledge graph; obtaining the latest information from the data source and generating candidate new cached content; performing quality and relevance checks on the candidate new cached content; and after passing the checks, replacing the original cached information item content and updating its weights on related edges in the time-series knowledge graph.

[0050] If β*C_c≤k_i≤α*C_c: the value loss and update cost are within a comparable range, then a dynamic degradation marking strategy is generated and executed for the cached information item. This lowers the priority of the cached information item in the search results and adds a temporary information update flag. For example, in a search ranking algorithm, the weight of the cached information item C_i is temporarily reduced, causing it to rank lower. Simultaneously, when the cached information item C_i is returned as a result, a temporary information update flag is added, such as: "Information in this field may have undergone significant updates recently; it is recommended to consider the latest data for judgment." If k_i < β*C_c: the prediction impact is negligible, and the benefits of proactive updates do not outweigh the costs. In this case, a status quo strategy is generated and executed for the cached information item to keep its current state unchanged.

[0051] Step S160: Feedback Learning and System Evolution. After executing the active caching operation strategy, continuously monitor the strategy's execution effect; collect user feedback data on subsequent interactions with the processed cached information items; and use the feedback data to perform online incremental learning and optimization of the parameters of the meta-gradient prediction model and the edge weights in the temporal knowledge graph.

[0052] After the strategy is executed, the system enters the monitoring and learning phase. It collects the user's subsequent interaction data on the cached information item C_i after it has been updated or downgraded (such as the click rate under the new ranking, feedback on prompts, and new satisfaction ratings). These actual effect data are compared with the model's prediction of C_i in step S140 to calculate the prediction error. This error signal is used for two aspects: (1) Model fine-tuning: As an incremental learning sample, the meta-gradient prediction model is fine-tuned online to improve its future prediction accuracy. (2) Graph weight correction: Backpropagation error is used to adjust the weights of the edges related to this prediction in the time-series knowledge graph (especially the weights W_3 of the edges that affect the transmission), so that the graph can more realistically reflect the transmission law of the impact of data updates.

[0053] Example 2: Detailed System Architecture refer to Figure 2 A large model cache proactive update system 200 for data retrieval includes the following modules that are interconnected via a bus or network: Data update perception module 210: responsible for monitoring changes in the target set of retrieval and processing, performing text vectorization and similarity calculation, and completing the identification of update targets and matching deviation targets (corresponding to step S110).

[0054] Search behavior analysis module 220: used to obtain historical matching data between the matching deviation target and the search information item, connect to the search log system, and statistically analyze the historical query context data of the matching deviation target (corresponding to step S120).

[0055] Knowledge graph management module 230: Used to construct, store, and maintain the time-series knowledge graph. Provides efficient subgraph traversal and extraction interfaces (supporting steps S130 and S140).

[0056] Active prediction module 240: It has a built-in trained meta-gradient prediction model, which is used to output the prediction value decay impact curve based on the matching deviation target and the knowledge graph. It receives the matching deviation target from module 510 and the relevant sub-graph data from module 530, performs forward inference, and outputs the value decay prediction curve (corresponding to step S140).

[0057] Strategy Decision and Execution Module 250: This module generates and executes differentiated proactive caching operation strategies based on the prediction curve and cost calculations. It includes a strategy decision engine and multiple strategy executors. The decision engine calculates decision indicators and selects strategies based on the prediction curve and cost model; the executors are specifically responsible for calling the cache update interface, modifying cache information item metadata (such as degradation flags), or interacting with the retrieval and ranking module (corresponding to step S150).

[0058] Feedback learning module 260: Collects user interaction data after strategy execution, calculates prediction error, and coordinates the optimization and updating of the model in the active prediction module 540 and the edge weights in the knowledge graph management module 530 (corresponding to step S160).

[0059] Example 3: Computer-readable storage medium This invention also relates to a computer-readable storage medium, such as a solid-state drive, optical disk, or storage node in a distributed storage system, including the aforementioned large model cache proactive update system for data retrieval, for executing the aforementioned large model cache proactive update method for data retrieval. The medium stores a computer program (instructions). When the program is loaded and executed by one or more processors (e.g., server CPU, GPU), it enables the processor to execute any one or more method flows described in Embodiment 1 above, thereby realizing the cache proactive update function provided by this invention.

[0060] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for proactively updating a large model cache for data retrieval, characterized in that, Includes the following steps: Step S110: Determine the update target and the matching deviation target; Step S120: Obtain historical matching context data; Step S130: Construct and maintain the time-series knowledge graph; Step S140: Perform proactive impact prediction based on temporal knowledge graph and meta-gradient prediction model; Step S150: Generate and execute a differentiated proactive caching operation strategy.

2. The method according to claim 1, characterized in that, Step S110 includes: Construct a target set for retrieval and processing. When a new target appears in the set, or the description text of an existing target is modified beyond a preset threshold, the target is marked as an updated target. Calculate the similarity between the text vector of each update target and all cached information items; Update targets whose maximum similarity to the text vectors of all cached information items is lower than a preset first similarity threshold are identified as matching deviation targets.

3. The method according to claim 1, characterized in that, The temporal knowledge graph includes a set of nodes and a set of edges; the set of nodes includes at least data source nodes, cached information item nodes, and retrieval context feature nodes. The edge set represents the relationships between nodes, including source edges, related query edges, and influence propagation edges; The construction and maintenance of the time-series knowledge graph in step S130 specifically includes: S131: Based on the metadata of the data source node, the content of the cached information items and their ownership relationships, establish a source edge connecting the data source node and the cached information item node directly generated by it; S132: Based on step S120, obtain historical matching context data, obtain historical search logs, extract frequently co-occurring search keywords or context features, form search context feature nodes, and establish association query edges connecting cached information item nodes and search context feature nodes. S133: Establish an influence propagation edge connecting the data source node to a cached information item node that is not directly generated by the data source node but is indirectly affected by its updates.

4. The method according to claim 1, characterized in that, The meta-gradient prediction model is a time series prediction model based on an attention mechanism, which is trained in the following way: The training sample set is constructed by collecting the set of matching deviation targets within a historical time period, the snapshot of the temporal knowledge graph at the corresponding time, and the actual change data of the value indicators of relevant cached information items in a subsequent period. The model is trained under supervision by taking the feature vector of the matching deviation target and the structural features of the related sub-map as inputs, and the gradient sequence of the value changes of related cached information items at multiple future time points as the prediction target.

5. The method according to claim 4, characterized in that, Step S140 includes: S141: Locate and match nodes related to the deviation target in the time-series knowledge graph; S142: Starting from the node related to the matching deviation target, based on the associated query edge and the influence propagation edge, traverse and extract all cached information item nodes within the preset number of hops to form a local subgraph. S143: The feature representation of the matching deviation target and the structured information of the sub-map are used as inputs and fed into the pre-trained meta-gradient prediction model for active influence prediction. The output is a value decay prediction curve for each cached information item. The expression of the prediction curve is V_i(t)=V_i0*exp(-λ_i*t), where V_i0 is the current value and λ_i is the decay rate predicted by the model.

6. The method according to claim 5, characterized in that, Step S150 includes: Extract the instantaneous decay rate k_i of the first time point T1 within the future preset short-term time period from the prediction curve; Estimate the computational resource cost C_c required to verify and update cached information item C_i; Set a first cost sensitivity coefficient α and a second cost sensitivity coefficient β, where α>β>0; If k_i>α*C_c, then a preventative proactive update strategy is generated and executed for the cached information item, immediately triggering the update process for the content of the cached information item; If β*C_c≤k_i≤α*C_c, then a dynamic degradation marking strategy is generated and executed for the cached information item, the priority of the cached information item in the retrieval results is reduced, and an information update identifier is attached; If k_i < β*C_c, then a status quo strategy is generated and executed for the cached information item to keep its current state unchanged.

7. The method according to claim 6, characterized in that, The implementation of the preventative proactive update strategy includes: Based on the matching deviation target and the time series knowledge graph, locate the authoritative data source that needs to be updated; Obtain the latest information from the data source and generate candidate new cache content; Perform quality and relevance checks on candidate new cached content; After successful verification, the original cached information item content is replaced, and its weight in the relevant edge of the time-series knowledge graph is updated.

8. The method according to claim 1, characterized in that, The method further includes step S160: After implementing the proactive caching strategy, continuously monitor the effectiveness of the strategy. Collect user feedback data on subsequent interactions with processed cached information items; Using the feedback data, online incremental learning and optimization are performed on the parameters of the meta-gradient prediction model and the edge weights in the temporal knowledge graph.

9. A large model cache proactive update system for data retrieval, used to execute the method described in any one of claims 1-8, characterized in that, include: The data update awareness module is used to monitor changes in the retrieval and processing targets and determine the update targets and the matching deviation targets. The retrieval behavior analysis module is used to obtain historical matching data between the matching deviation target and the retrieval information item; The knowledge graph management module is used to construct, store, and maintain the time-series knowledge graph. The active prediction module, which incorporates the meta-gradient prediction model, is used to output a prediction value decay impact curve based on the matching deviation target and the knowledge graph. The strategy decision-making and execution module is used to generate and execute differentiated proactive caching operation strategies based on the prediction curve and cost calculation.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images