An intelligent context caching method, system, device and medium

Abstract

The application discloses an intelligent context caching method, system, device and medium, relates to the technical field of artificial intelligence, and is used for solving the technical problems of unaddressable retrieval, enterprise inadaptation and low token efficiency in the prior art. The method is applied to an enterprise-level Agentic AIOps platform and comprises the following steps: Tool Call is used to enable an Agentic Layer to communicate with a Context Engine; the Agentic Layer comprises an LLM, a plurality of Agents, a plurality of Tools and a plurality of Skills; the Context Engine is used to intelligently process multi-modal data returned by a target Tool; the intelligent processing comprises two-dimensional decision score calculation, Persona ACL permission management and control, intelligent decision making, virtual file system (VFS) storage, multi-modal image injection and pointer handoff zero-copy. Therefore, the application can achieve the purposes of multi-Agent cooperation, addressable retrieval, enterprise adaptation and improved token efficiency through two-dimensional decision score calculation, Persona ACL permission management and control, intelligent decision making, VFS storage, multi-modal image injection and pointer handoff zero-copy.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and provides an intelligent context caching method, system, device, and medium. Background Technology

[0002] Currently, while Artificial Intelligence for IT Operations (AIOps) platforms employing a "data post-processing + Large Language Model (LLM) parsing" model possess strong stability, interpretability, and ease of governance, they also suffer from significant limitations such as passive processing, static context, and a lack of agentic capabilities. Therefore, with the evolution of agentic AI technology, portable skills and execution tools are gradually becoming core development directions. Skills utilize portable formats, supporting community sharing, dynamic installation, and natural language descriptions; agents can automatically discover and combine them based on tasks. Tools include shells, file operations, external APIs, and browsers. While agents can achieve autonomous action through tool calls, they face challenges such as context bloat and memory management. Furthermore, existing open-source local agents (such as OpenClaw) and typical memory systems often employ a "file-first" strategy, automatically writing key facts into Markdown logs and combining hybrid retrieval (vectors + keywords) for persistence. Before each round of interaction, relevant memories are injected into the context through "active loading" to prevent LLM from forgetting to call the memory_search tool. However, this can easily lead to token waste and insufficient support for enterprise-level structured massive data.

[0003] Furthermore, when using enterprise-grade agentic AI to process complex and heterogeneous data, existing context management solutions clearly have the following drawbacks: (1) When using "direct injection and truncation" for context management, SQL results, log messages, etc. can be directly concatenated with prompt words or simply truncated to a fixed window. However, in massive data scenarios, this can easily lead to context overflow, increased illusion rate, and a surge in token consumption. (2) When using "general compression and persistent memory" for context management, some open-source agent frameworks adopt a "file-first" strategy, which relies on global summaries and pre-loaded completion information. This results in a contradiction between excessive loading and wasted tokens and insufficient loading and missing information. The compression process is prone to losing key details of fault assessment. (3) When using "regular retrieval-Augmented Generation (RAG)" for context management, it combines SQLite and vector index long-term memory persistence, or Claude-like memory management. However, all of these lack enterprise-level intelligent decision-making mechanisms and cannot handle structured massive data, shared state among multiple agents, and business governance needs. It can be seen that the existing context management scheme has disadvantages such as non-intelligent decision-making, high compression loss, unaddressable retrieval, enterprise incompatibility and low token efficiency; (4) When using “context compression memory of production-grade Agentic framework” for context management, although it adopts a three-layer memory architecture and three-level progressive context compression, it still cannot meet the enterprise-level limitations such as the context management needs of heterogeneous massive data in multi-tenant enterprise AIOps. Summary of the Invention

[0004] This application provides an intelligent context caching method, system, device, and medium to solve technical problems in the prior art such as unaddressable retrieval, enterprise incompatibility, and low token efficiency.

[0005] On the one hand, a smart context caching method is provided for application in an enterprise-level Agentic AIOps platform, the method comprising: Tool Call is used to enable the Agentic Layer to communicate with the ContextEngine; wherein, the Agentic Layer includes a Large Language Model (LLM), multiple Agents, multiple Tools, and multiple Skills; The Context Engine is used to intelligently process the multimodal data returned by the target tool. The intelligent processing includes tool data input, two-dimensional decision score calculation, Persona ACL permission control, intelligent decision-making, virtual file system (VFS) storage, multimodal profile injection, and pointer handoff zero-copy.

[0006] The steps for intelligently processing the multimodal data returned by the target tool using a context engine include: The system employs a context engine to receive multimodal data returned by the target tool; this multimodal data includes log fragments, SQL datasets, alarm details, and interface receipts. Based on the multimodal data, calculate the scale score and semantic relevance score; Based on the scale score and semantic relevance score, intelligent decision-making is performed on the multimodal data to obtain decision results; wherein, the intelligent decision-making is used to determine whether to perform context caching, compression, segmentation and reuse; Based on the decision result, the multimodal data is cached in context; The multimodal data is processed in layers to generate a JSON profile; The JSON profile is sent to the Agentic Layer using a pointer handoff.

[0007] The step of performing context caching on the multimodal data based on the decision result includes: If the decision result is that the size score is less than the preset size score and the semantic relevance score is greater than the preset relevance score, then the direct injection method is used for context caching. If the decision result is that the scale score is not less than the preset scale score, or the semantic relevance score is not greater than the preset relevance score, then the Virtual File System (VFS) is used for context caching, and a structured profile is generated.

[0008] The step of using a Virtual File System (VFS) for context caching includes: Using an address protocol, virtual file paths obtained from the global data base are cached in the virtual file system VFS; wherein, the global data base includes the distributed object storage system MinIO, the graph database Neo4j, and the vector database Weaviate; Fine-grained access control is implemented for enterprise-level identity permissions in the Persona ACL layer of the Virtual File System (VFS). In response to user access commands, the virtual file path is obtained from the Virtual File System (VFS) through the Model Context Protocol (MCP) Tool interface.

[0009] The step of performing layered processing on the multimodal data to generate a JSON profile includes: In the lightweight layer, the Pandas data analysis library is used to perform statistical analysis on the multimodal data and obtain statistical results. At the structural layer, the schema metadata of the multimodal data is extracted; At the semantic layer, the text content of the multimodal data is converted into semantic vectors, and the business themes of the multimodal data are identified; wherein, the business themes include faults, performance, configuration, and security. At the business layer, the semantic vector is matched with the enterprise business knowledge vector library to obtain the matching result; Based on statistical results, schema metadata, semantic vectors, business themes, and matching results, a structured JSON profile is generated.

[0010] After making intelligent decisions based on the multimodal data according to the scale score and semantic relevance score, and obtaining the decision results, the method further includes: Retrieve the global history vector associated with the current task from the global vector memory; The global history vector is segmented to obtain multiple historical segments; Calculate the semantic relevance score for each historical segment separately; For each historical segment whose semantic relevance score is greater than the preset relevance score, a corresponding micro-summary is generated. The weight of each micro-summary is determined based on its semantic relevance score. Based on the weight of each micro-summary, sort the micro-summaries in descending order; Based on the weighted sorting results, generate a list of compressed links and backup index pointers; Obtain the compression history context based on the current task, compression link, and list of backup index pointers.

[0011] The step of generating a list of compressed links and backup index pointers based on the weight sorting results includes: Based on the weight ranking results, multiple micro-summaries with weights greater than a preset weight threshold are concatenated to generate compressed links. Based on the weight sorting results, the VFS access pointers of at least one micro-digest with a weight not greater than a preset weight threshold are placed into the list of standby index pointers.

[0012] The semantic relevance score is calculated using the following formula: relevance=cosine_similarity(aggregate_vec, Goal_vec) Where relevance is the semantic relevance score; cosine_similarity is the cosine similarity; aggregate_vec is the semantic aggregation vector of historical segments; and Goal_vec is the target vector of the current operation and maintenance task.

[0013] On the one hand, an intelligent context caching system is provided for application in an enterprise-level Agentic AIOps platform, the intelligent context caching system comprising: The Agentic Layer is used to communicate with the ContextEngine via Tool Call; the Agentic Layer includes a Large Language Model (LLM), multiple Agents, multiple Tools, and multiple Skills. The Context Engine is used to intelligently process the multimodal data returned by the target tool; the intelligent processing includes tool data input, two-dimensional decision score calculation, Persona ACL permission control, intelligent decision-making, virtual file system (VFS) storage, multimodal profile injection, and pointer handoff zero-copy.

[0014] The intelligent context caching system also includes a virtual file system (VFS) for: An address protocol is used to cache virtual file paths obtained from the global data base; wherein, the global data base includes the distributed object storage system MinIO, the graph database Neo4j, and the vector database Weaviate; Implement fine-grained access control for enterprise-level identity permissions; In response to user access commands, the virtual file path is sent to the Model Context Protocol (MCP) Tool interface.

[0015] The virtual file system (VFS) is also used for: A multi-agent cross-session compressed chain link is established using the graph database Neo4j; wherein, the graph database Neo4j supports range queries using the #[start-end] syntax; After the cross-session ends, a Process Improvement Plan (PIP) report is generated; the PIP report contains thematic summaries, lessons learned, and optimization recommendations, and the PIP report can be applied to the Agent configuration with one click.

[0016] The Context Engine also includes a Delta incremental cache submodule, used for: Cache newly added or changed data segments in the multimodal data returned by the target tool Tool.

[0017] On one hand, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the methods described above.

[0018] On the one hand, a storage medium is provided that stores computer program instructions thereon, which, when executed by a processor, implement any of the methods described above.

[0019] Compared with the prior art, the beneficial effects of this application are as follows: In this application, the intelligent context caching method is applied to an enterprise-level Agentic AIOps platform. Based on this, when performing intelligent context caching, firstly, a Tool Call can be used to enable the intelligent agent layer (Agentic Layer) to communicate with the context engine (Context Engine). The intelligent agent layer includes a Large Language Model (LLM), multiple agents, multiple tools, and multiple skills. Then, the context engine can be used to intelligently process the multimodal data returned by the target tool. This intelligent processing includes tool data input, two-dimensional decision score calculation, Persona ACL permission control, intelligent decision-making, Virtual File System (VFS) storage, multimodal profile injection, and pointer handoff zero-copy.

[0020] Based on this, in this application, since intelligent processing includes intelligent decision-making, compared to existing technologies, this application does not require preloading strategies, avoiding excessive waste and insufficient invalidation, and improving token efficiency. Furthermore, since intelligent processing also includes Virtual File System (VFS) storage and pointer handoff, compared to existing technologies, this application not only enables zero-intermediary collaboration between Tools, enhancing Tools, but is also highly suitable for enterprises. Moreover, since intelligent processing also includes Persona ACL permission control, compared to existing technologies, this application not only allows Agents / users with different permissions to only access data corresponding to their permissions, but also automatically anonymizes sensitive data, preventing data leakage. Additionally, since intelligent processing also includes multimodal profile injection, compared to existing technologies, this application not only supports persistent execution of complex Skills through profile injection, enhancing Skills, but LLM can also accurately pull subsets after "previewing," thereby achieving addressable retrieval. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of an application scenario provided by an embodiment of this application; Figure 2 An architecture diagram of an intelligent context caching system provided in this application embodiment; Figure 3 An architecture diagram of the Context Engine provided in the embodiments of this application; Figure 4 A flowchart illustrating the intelligent context caching method provided in this application embodiment; Figure 5 A schematic flowchart of intelligent processing provided in an embodiment of this application; Figure 6 A flowchart illustrating the intelligent decision caching provided in this application embodiment; Figure 7 A schematic diagram of a VFS context caching and access process provided in an embodiment of this application; Figure 8 A schematic diagram of a process for generating an image provided in an embodiment of this application; Figure 9 A flowchart illustrating the target-oriented dynamic compression provided in this application embodiment. Figure 10 This is a timing interaction diagram of a network automation agent provided in an embodiment of this application.

[0023] The diagram is labeled as follows: 10-Smart Context Cache Device, 101-Processor, 102-Memory, 103-I / O Interface, 104-Database. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0025] Currently, while intelligent operations and maintenance (AIOps) platforms employing a "data post-processing + large language model LLM parsing" model offer strong stability, interpretability, and ease of governance, they suffer from significant limitations such as passive processing, static context, and a lack of agentic capabilities. Therefore, with the evolution of agentic AI technology, portable skills and execution tools are gradually becoming core development directions. Skills utilize portable formats, supporting community sharing, dynamic installation, and natural language descriptions; agents can automatically discover and combine them based on tasks. Tools include shells, file operations, external APIs, and browsers. While agents can achieve autonomous action through tool calls, they face challenges such as context bloat and memory management. Furthermore, existing open-source local agents (such as OpenClaw) and typical memory systems often employ a "file-first" strategy, automatically writing key facts into Markdown logs and using hybrid retrieval (vectors + keywords) for persistence. Before each interaction, relevant memories are injected into the context through "active loading" to prevent LLM from forgetting to call the memory_search tool. However, this can easily lead to token waste and insufficient support for enterprise-level structured massive amounts of data.

[0026] Furthermore, when using enterprise-grade agentic AI to process complex and heterogeneous data, existing context management solutions clearly have the following drawbacks: (1) When using "direct injection and truncation" for context management, SQL results, log messages, etc. can be directly concatenated with prompt words or simply truncated to a fixed window. However, in massive data scenarios, this can easily lead to context overflow, increased illusion rate, and a surge in token consumption. (2) When using "general compression and persistent memory" for context management, some open-source agent frameworks adopt a "file-first" strategy, which relies on global summaries and pre-loaded completion information. This results in a contradiction between excessive loading and wasted tokens and insufficient loading and missing information. During compression, key details of fault assessment are easily lost. (3) When using "conventional retrieval enhancement to generate RAG" for context management, it combines SQLite and vector index long-term memory persistence, or Claude-like memory management. However, all of these lack enterprise-level intelligent decision-making mechanisms and cannot handle structured massive data, shared state among multiple agents, and business governance needs. It can be seen that the existing context management scheme has disadvantages such as non-intelligent decision-making, high compression loss, unaddressable retrieval, enterprise incompatibility and low token efficiency; (4) When using “context compression memory of production-grade Agentic framework” for context management, although it adopts a three-layer memory architecture and three-level progressive context compression, it still cannot meet the enterprise-level limitations such as the context management needs of heterogeneous massive data in multi-tenant enterprise AIOps.

[0027] Based on this, this application provides an intelligent context caching method applied to an enterprise-level Agentic AIOps platform. When performing intelligent context caching, firstly, a ToolCall can be used to enable communication between the Agentic Layer and the Context Engine. The Agentic Layer includes a Large Language Model (LLM), multiple Agents, multiple Tools, and multiple Skills. Then, the Context Engine can intelligently process the multimodal data returned by the target Tool. This intelligent processing includes tool data input, two-dimensional decision score calculation, Persona ACL permission control, intelligent decision-making, Virtual File System (VFS) storage, multimodal profile injection, and zero-copy pointer handoff. Therefore, in this application, since intelligent processing includes intelligent decision-making, compared to existing technologies, this application does not require pre-loading strategies, avoiding excessive waste and insufficient invalidation, and improving token efficiency. Furthermore, since intelligent processing also includes VFS storage and pointer handoff, compared to existing technologies, this application not only enables zero-intermediary collaboration between Tools, enhancing Tools, but is also highly suitable for enterprise environments. Furthermore, since intelligent processing also includes Persona ACL access control, compared to existing technologies, this application not only ensures that agents / users with different permissions can only access data corresponding to their permissions, but also automatically de-identifies sensitive data to prevent data leakage. Additionally, since intelligent processing also includes multimodal profile injection, compared to existing technologies, this application not only supports persistent execution of complex skills through profile injection, enhancing skills, but LLM can also accurately retrieve subsets after "previewing," thereby achieving addressable retrieval.

[0028] After introducing the design concept of the embodiments of this application, the following is a brief introduction to the application scenarios to which the technical solutions of the embodiments of this application can be applied. It should be noted that the application scenarios described below are only for illustrating the embodiments of this application and are not intended to limit the scope. In specific implementation, the technical solutions provided by the embodiments of this application can be flexibly applied according to actual needs.

[0029] like Figure 1 The diagram shown illustrates an application scenario provided by an embodiment of this application. This application scenario may include an intelligent context caching device 10.

[0030] The intelligent context cache device 10 can be used for intelligent context caching, and can be, for example, a personal computer (PC), server, or laptop. The intelligent context cache device 10 may include one or more processors 101, memory 102, I / O interfaces 103, and databases 104. Specifically, the processor 101 may be a central processing unit (CPU) or a digital processing unit, etc. The memory 102 may be volatile memory, such as random-access memory (RAM); the memory 102 may also be non-volatile memory, such as read-only memory, flash memory, hard disk drive (HDD), or solid-state drive (SSD); or the memory 102 may be any other medium capable of carrying or storing desired program code in the form of instructions or data structures that can be accessed by a computer, but is not limited thereto. The memory 102 may be a combination of the above-mentioned memories. The memory 102 can store some program instructions of the intelligent context caching method provided in this application embodiment. When these program instructions are executed by the processor 101, they can be used to implement the steps of the intelligent context caching method provided in this application embodiment, so as to solve the technical problems existing in the prior art, such as retrieval unaddressability, enterprise incompatibility, and low token efficiency. The database 104 can be used to store multimodal data and data such as context engine, decision results, JSON profiles, and virtual file system (VFS) involved in the solution provided in this application embodiment.

[0031] In this embodiment, the intelligent context caching device 10 can obtain context caching instructions through the I / O interface 103. Then, the processor 101 of the intelligent context caching device 10 will solve the technical problems existing in the prior art, such as retrieval unaddressability, enterprise incompatibility, and low token efficiency, according to the program instructions of the intelligent context caching method provided in this embodiment in the memory 102. In addition, multimodal data and context engine, decision results, JSON profiles, and virtual file system (VFS) data can be stored in the database 104.

[0032] like Figure 2 The diagram shown is an architecture diagram of an intelligent context caching system provided in an embodiment of this application. The system includes: an Agentic Layer and a Context Engine.

[0033] The Agentic Layer is used to communicate with the Context Engine via Tool Call. The Agentic Layer includes a Large Language Model (LLM), multiple Agents, multiple Tools, and multiple Skills.

[0034] The Context Engine is used for intelligent processing of multimodal data returned by the target tool. This intelligent processing includes tool data input, two-dimensional decision score calculation, Persona ACL permission control, intelligent decision-making, Virtual File System (VFS) storage, multimodal profile injection, and pointer handoff zero-copy. The intelligent decision-making dynamically determines the cache path based on the scale score (S) and semantic relevance score (Rel). The VFS underlying layer uses MinIO, Neo4j, and Weaviate global data infrastructure, and implements enterprise-level fine-grained permission control through the Persona Access Control List (PACL) layer. The multimodal profile injection generates a multimodal JSON profile containing four top-level fields: stats, structure, semantics, and business. The pointer handoff zero-copy enables zero-copy direct connection collaboration between tools via the vfs: / / address protocol.

[0035] In one possible implementation, such as Figure 2 As shown, the intelligent context caching system also includes a Virtual File System (VFS) for: An address protocol is used to cache virtual file paths obtained from the global data base; wherein, the global data base includes the distributed object storage system MinIO, the graph database Neo4j, and the vector database Weaviate.

[0036] Implement fine-grained access control for enterprise-level identity permissions.

[0037] In response to user access commands, the virtual file path is sent to the Model Context Protocol (MCP) Tool interface.

[0038] In one possible implementation, the virtual file system (VFS) is further used for: A multi-agent cross-session compressed chain link is established using the graph database Neo4j; wherein, the graph database Neo4j supports range queries using the #[start-end] syntax; After the cross-session ends, a Process Improvement Plan (PIP) report is generated; the PIP report contains thematic summaries, lessons learned, and optimization recommendations, and the PIP report can be applied to the Agent configuration with one click.

[0039] In one possible implementation, such as Figure 2 As shown, the intelligent context caching system also includes an Enterprise Data Fabric architecture, which can store raw multimodal data. This enterprise data architecture has over 1700 data sources, each of which can be represented using cloud icons. Therefore, this application can eliminate data silos and achieve unified access by building a unified data foundation, making it highly suitable for enterprises.

[0040] In one possible implementation, such as Figure 3 The diagram shown is an architecture diagram of the Context Engine provided in an embodiment of this application. The Context Engine further includes: The intelligent decision caching module is used to receive multimodal data returned by the target tool; wherein, the multimodal data includes log fragments, SQL datasets, alarm details and interface receipts.

[0041] The intelligent decision caching module is also used to calculate the scale score and semantic relevance score based on the multimodal data.

[0042] The intelligent decision caching module is also used to make intelligent decisions on the multimodal data based on the scale score and semantic relevance score to obtain decision results; wherein, the intelligent decision is used to determine whether to perform context caching, compression, segmentation and reuse.

[0043] The intelligent decision caching module is also used to perform context caching on the multimodal data based on the decision result.

[0044] The multimodal data profile generation module is used to perform layered processing on the multimodal data and generate JSON profiles.

[0045] The pointer retrieval interface module is used to send JSON profiles to the Agentic Layer using pointer handoff, in order to achieve Skill chain collaboration.

[0046] In addition, the intelligent decision caching module can also be used to support the dynamic expansion of Agent Skills. The multimodal data profiling generation module can also be used to support preview decisions during Skill execution.

[0047] In one possible implementation, such as Figure 3As shown, the Context Engine also includes: An addressable virtual file system module, which may consist of object storage, a graph database, and a vector index, is used to provide a unified address space and tool-to-tool handoff capabilities. A goal-oriented dynamic compression engine is used to support persistent context management for multi-round Skill interactions; and the system is integrated with existing AIOps data pipelines to achieve an architectural upgrade from post-processing to full-stack Agentic AIOps.

[0048] In one possible implementation, the intelligent decision caching module of the Context Engine may further include a Delta incremental caching submodule, which is used to cache newly added or changed data fragments in the multimodal data returned by the target tool. That is, it can cache only the diff part of the multimodal data and the same tool returned last time, so as to reduce the overhead of repeated full caching.

[0049] In one possible implementation, the JSON output structure of the multimodal data profiling generation module may also include four top-level fields: stats, structure, semantic, and business. Furthermore, in the multimodal data profiling generation module, the semantic relevance score of the business layer can be calculated using the following formula: relevance=cosine_similarity(aggregate_vec, Goal_vec) Wherein, cosine_similarity is the cosine similarity; aggregate_vec is the semantic aggregation vector of historical segments; and Goal_vec is the target vector of the current operation and maintenance task.

[0050] In one possible implementation, the graph database of the addressable virtual file system module can be used to establish multi-agent cross-session compressed chain links, with address range queries [start-end].

[0051] In one possible implementation, the Tool-to-Tool Handoff of the pointer retrieval interface module is used to generate a handoff_pointer after the output of Tool A is written to the VFS, and directly pass it to Tool B as an input parameter without the need for LLM to transfer data content.

[0052] In one possible implementation, the micro-summary generation of the goal-oriented dynamic compression engine uses a prompt template: "For the current goal {G}, extract key facts and causal relationships of no more than 200 tokens from the following historical fragments."

[0053] In one possible implementation, the intelligent context caching system may also include an industry profile template library, such as pre-set templates for networks (device distribution histograms), energy (time series quantiles), and finance (transaction entity extraction).

[0054] In one possible implementation, the intelligent context caching system may also include a business result observation interface, which is used to call a review of the LLM generation process improvement plan (PIP) report after the session ends. The report includes three parts: "topic summary, lessons learned, and optimization suggestions", and supports one-click application to the Agent configuration.

[0055] Of course, the methods provided in the embodiments of this application are not limited to... Figure 1 The application scenarios shown can also be used in other possible scenarios, and this application embodiment does not impose any limitations. Figure 1 The functions that the various devices in the application scenarios shown can achieve will be described in subsequent method embodiments, and will not be elaborated on here. Below, the methods of the embodiments of this application will be described in conjunction with the accompanying drawings.

[0056] like Figure 4 The diagram shown is a flowchart of an intelligent context caching method provided in this application embodiment. This method can be applied to enterprise-level Agentic AIOps platforms and, through... Figure 1 The method is executed by the intelligent context caching device 10. Specifically, the process is described below.

[0057] Step 401: Use a tool call to enable the Agentic Layer to communicate with the Context Engine.

[0058] The intelligent agent layer comprises a Large Language Model (LLM), multiple agents, multiple tools, and multiple skills. This intelligent agent layer uses the LLM as its core, coupled with multiple agents with clearly defined roles (e.g., fault analysis agent, work order processing agent, configuration verification agent, etc.) to collaboratively complete operation and maintenance tasks. Furthermore, because it integrates multiple reusable skills (encapsulating operation and maintenance workflows) and multiple tools (e.g., SQL queries, log analysis, API calls, etc.), it can also achieve autonomous decision-making, toolchain-style execution, and business closed-loop. Therefore, the intelligent agent layer of this application can be responsible for receiving user operation and maintenance requests, generating root cause analysis, outputting handling suggestions, and completing human-computer interaction.

[0059] In this application, the Agentic Layer serves as the user interaction entry point. It can communicate with the Context Engine using tooltips or tool calls, rather than through hard-coded interfaces, to achieve flexible linkage between intelligent decision-making and underlying capabilities.

[0060] Step 402: Use the Context Engine to intelligently process the multimodal data returned by the target tool.

[0061] The intelligent processing includes tool data input, two-dimensional decision score calculation, Persona ACL permission control, intelligent decision-making, virtual file system (VFS) storage, multimodal profile injection, and pointer handoff zero-copy.

[0062] Specifically, such as Figure 5 The diagram shown is a flowchart illustrating an intelligent processing method provided in an embodiment of this application.

[0063] Step 501: Use the Context Engine to receive multimodal data returned by the target tool.

[0064] The multimodal data includes log fragments, SQL datasets, alarm details, and interface receipts.

[0065] Step 502: Calculate the scale score and semantic relevance score based on the multimodal data.

[0066] Step 503: Make intelligent decisions on multimodal data based on the scale score and semantic relevance score to obtain the decision results.

[0067] The intelligent decision-making process is used to determine whether to perform context caching, compression, segmentation, and reuse.

[0068] Step 504: Based on the decision results, perform context caching on the multimodal data.

[0069] like Figure 6 The diagram shown is a flowchart of an intelligent decision caching method provided in an embodiment of this application.

[0070] Step 601: If the decision result is that the size score is less than the preset size score and the semantic relevance score is greater than the preset relevance score, then the direct injection method is used for context caching.

[0071] That is, small amounts of highly relevant data are cached directly.

[0072] Step 602: If the decision result is that the scale score is not less than the preset scale score, or the semantic relevance score is not greater than the preset relevance score, then the Virtual File System (VFS) is used for context caching, and a structured profile is generated.

[0073] That is, apart from small-volume, highly relevant data, the remaining data is cached and profiled using VFS. Furthermore, in this application, the Virtual File System (VFS) provides structured, unified storage and lightweight addressing to replace scattered local files, enabling unified data management and rapid retrieval.

[0074] Therefore, when using the Virtual File System (VFS) for context caching and access, such as Figure 7 The diagram shown is a flowchart illustrating a VFS context caching and access process provided in an embodiment of this application.

[0075] Step 701: Using the address protocol, cache the virtual file path obtained from the global data base to the virtual file system (VFS).

[0076] The comprehensive data foundation includes the distributed object storage system MinIO, the graph database (DataBase, DB) Neo4j, and the vector database Weaviate. MinIO can be used to store unstructured / semi-structured data such as raw tool results, logs, and binary files; Neo4j can be used to store multi-agent relationship graphs, business causal relationships, memory links, and entity relationships; and Weaviate can be used to store data embedding vectors, semantic indexes, and business knowledge vectors, supporting semantic retrieval.

[0077] Furthermore, in this application, the address protocol is a unified addressing protocol, which is the core of the addressable VFS. It does not need to transport the full amount of original data, but only exposes standardized virtual file paths (e.g., vfs: / / tool / sql / result / xxx). It uses indexes, IDs, or paths to replace full-text transmission, thereby laying the foundation for subsequent zero-copy transmission by realizing unified addressing and lightweight referencing of data.

[0078] Step 702: Perform fine-grained access control for enterprise-level identity permissions in the Persona ACL layer of the Virtual File System (VFS).

[0079] In other words, fine-grained access control can be implemented at the Persona Access Control List (PACL) layer, based on roles, business domains, and agent identities, ensuring that agents / users with different permissions can only access data corresponding to their permissions. Furthermore, sensitive data (such as passwords, IP addresses, and user information) can be automatically anonymized to prevent data leakage. Therefore, this not only meets enterprise-level security compliance requirements but also achieves secure data isolation and access control.

[0080] Step 703: In response to the user's access command, obtain the virtual file path from the virtual file system (VFS) through the MCP Tool interface.

[0081] Among them, the Model Context Protocol (MCP) Tool interface can provide a unified and standardized tool call interface for the Agentic Layer, allowing the Agent to directly access data through Tool Call without having to worry about the underlying storage details. This connects the upper-layer intelligent agent with the underlying data, enabling flexible and scalable tool call capabilities.

[0082] Based on this, in order to construct an enterprise-level secure, efficient, and low-token intelligent proxy data access closed loop, this application, when accessing the VFS, can respond to user access commands and obtain the virtual file path from the virtual file system (VFS) through the MCP Tool interface. It is evident that, since only address pointers / index subsets are passed throughout the process, without copying, persisting, or repeatedly loading the original large data, the data can be accessed directly at the storage layer through addressing, eliminating redundant copying, maximizing token consumption, completely avoiding context overflow, and significantly improving data access efficiency. This achieves lightweight context management that "does not move data, only passes pointers."

[0083] In addition, before generating the JSON profile, the Context Engine calls the identification and authentication module to determine the data visibility radius of the Agent in the Virtual File System (VFS) based on the current Agent token; and the pointer handoff only transmits the vfs: / / address pointer throughout the process, and the data does not move in the physical storage layer (MinIO / Neo4j / Weaviate), achieving true zero-copy collaboration.

[0084] Step 505: Perform layered processing on the multimodal data to generate a JSON profile.

[0085] Specifically, such as Figure 8 The diagram shown is a flowchart illustrating a process for generating an image according to an embodiment of this application.

[0086] Step 801: In the lightweight layer, use the Pandas data analysis library to perform statistics on the multimodal data and obtain the statistical results.

[0087] In other words, a lightweight layer can be used with the Pandas data analysis library to perform fast and low-cost statistics on multimodal data, extracting basic features such as min (minimum value), max (maximum value), and histogram (distribution histogram) to obtain a "global overview" of the multimodal data. Without loading the full amount of data, the scale, magnitude, and distribution of the data can be quickly determined, laying the foundation for subsequent stratified processing, while filtering out invalid data.

[0088] Step 802: At the structural layer, extract the schema metadata of the multimodal data.

[0089] In other words, the schema metadata of multimodal data can be automatically extracted at the structure layer, such as field names, data types, nesting depth, index relationships, constraints, and field meanings. This allows the system to "understand what the data looks like" without requiring a large language model (LLM) to parse complex nested raw structures, significantly reducing the cost of understanding and enabling unified processing of structured / semi-structured data (such as JSON, logs, and tables).

[0090] Step 803: At the semantic layer, convert the text content of the multimodal data into semantic vectors and identify the business themes of the multimodal data.

[0091] The business topics include faults, performance, configuration, and security.

[0092] In other words, at the semantic layer, semantic embedding can be used to convert the textual content of multimodal data into high-dimensional semantic vectors to capture the deep semantic features of the data. Furthermore, Latent Dirichlet Allocation (LDA) topic modeling can be used to automatically identify the business themes of the multimodal data, such as identifying whether the multimodal data belongs to a fault, performance, configuration, or security scenario. This allows the system to "understand what the data is saying," comprehending the core content of the data at the semantic level and supporting subsequent business matching.

[0093] Step 804: At the business layer, match the semantic vector with the enterprise business knowledge vector library to obtain the matching result.

[0094] In other words, at the business layer, semantic vectors generated by the semantic layer can be matched with the enterprise business knowledge vector library for similarity. This calculates the relevance score between multimodal data and the current operational task (e.g., troubleshooting, root cause analysis). Simultaneously, the timeliness of the multimodal data is labeled (e.g., whether it's a recent alarm or historical log). This process adds business tags to the multimodal data, filtering out high-value information strongly relevant to the current task and removing irrelevant noise, allowing the Large Language Model (LLM) to acquire only precise business-relevant content.

[0095] Step 805: Generate a structured JSON profile based on the statistical results, schema metadata, semantic vectors, business themes, and matching results.

[0096] That is, all the features of the four-layer processing can be integrated into a very small structured JSON file. Thus, by using JSON profiles to replace the full amount of original multimodal data, the data can be directly input into the large language model LLM, which can completely preserve the core information of the multimodal data and compress the token to the extreme. In this way, efficient context management can be achieved by "not looking at the original big data, but only at the lightweight profile", and completely avoiding context overflow.

[0097] As can be seen, the "portrait generation" in this application can process the massive, heterogeneous raw data returned by the tool through four progressive layers to finally generate a lightweight, structured, LLM-friendly JSON portrait, fundamentally solving the pain points of traditional solutions such as "full data flooding into the Prompt, resulting in context overflow, token waste, and high illusion rate".

[0098] Step 506: Use pointer handoff to send the JSON profile to the Agentic Layer.

[0099] That is, pointer handoff can be used to pass only the index pointer to the Agentic Layer without moving the full amount of original multimodal data, so as to completely avoid context token overflow and redundant loading, and achieve zero-copy efficient calls.

[0100] In one possible implementation, to reduce compression loss, this application, after making intelligent decisions on the multimodal data based on scale score and semantic relevance score, and obtaining the decision results, can also perform "goal-oriented dynamic compression" on the data, such as... Figure 9 The diagram shown is a flowchart illustrating a target-oriented dynamic compression method provided in an embodiment of this application.

[0101] Step 901: Retrieve the global history vector associated with the current task from the global vector memory.

[0102] That is, anchor the current task, retrieve all relevant historical materials, and provide input for subsequent refinement.

[0103] Step 902: Segment the global history vector to obtain multiple historical segments.

[0104] That is, it can be broken down into several independent small historical segments according to session rounds, time nodes, and independent events. In this way, "large and difficult-to-process long history" is broken down into small units that can be evaluated one by one, avoiding redundant loading of the whole segment and preparing for fine-grained screening.

[0105] Step 903: Calculate the semantic relevance score for each historical segment.

[0106] That is, for each split historical segment, a semantic relevance score (the degree of matching between the historical segment and the current operation and maintenance task) can be calculated separately. In this way, historical segments that are strongly related to the current task can be accurately selected, while irrelevant historical noise, such as records of other unrelated faults and expired session content, can be filtered out.

[0107] Step 904: For each historical segment whose semantic relevance score is greater than the preset relevance score, generate the corresponding micro-summary.

[0108] That is, for each selected highly relevant historical segment, a concise and lightweight micro-summary can be generated, which retains only the core conclusions, key fields and key events, and removes redundant original text descriptions. In this way, the token is compressed to the extreme, and the original long historical text is compressed into a core summary of a few lines, which greatly reduces the context overhead.

[0109] Step 905: Determine the weight of each micro-summary based on its semantic relevance score.

[0110] Specifically, the weight of each micro-summary can be determined based on the mapping relationship between semantic relevance score and weight according to a pre-set weight mapping table. As shown in Table 1, this is a mapping table of semantic relevance score and weight provided in an embodiment of this application, where weight = semantic relevance score.

[0111] Table 1

[0112] Step 906: Sort the micro-abstracts in descending order according to their weights.

[0113] Step 907: Generate a list of compressed links and backup index pointers based on the weight sorting results.

[0114] Specifically, based on the weight ranking results, multiple micro-summaries with weights greater than a preset weight threshold can be concatenated to generate a compressed link. That is, as shown in Table 1, multiple micro-summaries with weights greater than 0.3 can be concatenated into a smooth, traceable, and concise compressed link according to the descending order of their weights.

[0115] Based on the weighted sorting results, the VFS access pointers of at least one micro-summary with a weight not greater than a preset weight threshold are placed in the standby index pointer list. That is, as shown in Table 1, the VFS access pointers of at least one micro-summary with a weight not greater than 0.3 can be placed in the standby index pointer list.

[0116] Step 908: Obtain the compression history context based on the current task, compression link, and list of backup index pointers.

[0117] For example, the compressed context can be represented as "current task + first 2 micro-summaries concatenated in order + [pointer:vfs: / / history / alert / disk / 20240325_xxx]".

[0118] In addition, multi-agent graph database sharing can be implemented. The processed micro-summaries, weights, and chain relationships can be uniformly stored in the multi-agent shared graph database. Then, different agents can directly reuse this refined memory without repeated retrieval and summarization, which greatly avoids the waste of redundant computing power.

[0119] As can be seen, this application can integrate scattered compressed fragments into a coherent historical context that LLM can directly understand through "goal-oriented dynamic compression", thereby preserving key historical information while minimizing token consumption.

[0120] In one possible implementation, during multi-agent collaboration and governance, operations engineers can see all metrics in the Persona matrix, while managers only see the aggregated view. Furthermore, at the end of the session, a debriefing LLM can be automatically invoked to generate a Persona Improvement Plan (PIP) report. This report can include three parts: "topic summary, lessons learned, and optimization suggestions," and can be applied to agent configurations with a single click. Specific Implementation Example 1: like Figure 10 The diagram shown is a time-series interaction diagram of a network automation agent provided in this application embodiment, wherein the query_network_metrics() tool returns 18,742 lines of JSON (device ID, metric, timestamp).

[0122] First, the query_network_metrics() tool can return 18742 lines of JSON to the cache of the Virtual File System (VFS); Then, the Virtual File System (VFS) performs decision caching and profiling based on these 18,742 lines of JSON; the data can be cached in vfs: / / sess_001 / tool_metrics / 20260216_1423 / abc123.

[0123] Next, we can inject the image and pointer handoff into the large language model LLM.

[0124] Then, the Large Language Model (LLM) can perform pointer retrieval to obtain the retrieval results fetch_vfs(...).

[0125] Finally, by calling `fetch_vfs(...)` through the tool, the top-200 high-risk records are returned to the large language model LLM to achieve zero-copy. Through 10 repeated tests, the token count was reduced from 65k to 8k, and the accuracy improved from 72% to 96%.

[0126] The core logic of intelligent decision-making (decide_cache pseudocode) is as follows: If the scale score S < 5000 and the semantic relevance score Rel > 0.85, then direct injection (direct_inject(R)) is performed; otherwise, vfs_store(R) is executed and generate_portrait(R) is called to generate a four-layer JSON profile. After the original 18,742 lines of JSON (approximately 65,000 tokens) are injected with the profile, the LLM context contains only a JSON profile of approximately 350 tokens and a vfs: / / pointer; on-demand fetch only returns the Top-200 high-risk records (approximately 8,000 tokens), resulting in a token reduction of 87.7%; after 10 repeated tests, the accuracy improved from 72% to 96% (+24 percentage points). Specific Implementation Example 2: The existing data pipeline output of AIOps is directly connected to the middleware as the input of the enterprise-level Agentic AIOps platform tool in this application, so as to realize the architecture upgrade from post-processing to full-stack Agentic AIOps (achieving seamless upgrade).

[0128] In summary, this application has the following advantages: (1) Since intelligent processing involves intelligent decision-making, compared with the prior art, this application does not require preloading strategy, which avoids excessive waste and insufficient failure, and improves token efficiency.

[0129] (2) Since intelligent processing also includes virtual file system (VFS) storage and pointer handoff, compared with the prior art, this application can not only realize zero-transfer collaboration between tools and enhance tools, but also is very suitable for enterprises.

[0130] (3) Since intelligent processing also includes portrait injection, compared with the existing technology, this application can not only support the persistent execution of complex skills through portrait injection and enhance skills, but LLM can also accurately pull subsets after "previewing", thereby realizing retrieval addressability.

[0131] (4) Since goal-oriented dynamic compression is also performed, compared with the prior art, this application can not only compress the token to the extreme through a minimalist and lightweight micro-summary, reducing compression loss and context overhead, but also integrate the scattered compressed fragments into a coherent historical context that LLM can directly understand, thereby minimizing token consumption while preserving key historical information.

[0132] In some possible implementations, various aspects of the methods provided in this application can also be implemented as a program product comprising program code that, when run on a computer device, causes the computer device to perform the steps of the methods according to the various exemplary embodiments of this application described above. For example, the computer device may perform actions such as... Figures 4-9 The method performed in the illustrated embodiment.

[0133] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks. Alternatively, if the integrated units of this application are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to the prior art, can be embodied in the form of software products. These computer software products are stored in a storage medium and include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0134] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0135] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A smart context caching method, characterized in that, When applied to an enterprise-level Agentic AIOps platform, the method includes: Tool Call is used to enable the Agentic Layer to communicate with the Context Engine; wherein, the Agentic Layer includes a Large Language Model (LLM), multiple Agents, multiple Tools, and multiple Skills; The Context Engine is used to intelligently process the multimodal data returned by the target tool. The intelligent processing includes tool data input, two-dimensional decision score calculation, Persona ACL permission control, intelligent decision-making, virtual file system (VFS) storage, multimodal profile injection, and pointer handoff zero-copy.

2. The method as described in claim 1, characterized in that, The steps for intelligently processing the multimodal data returned by the target tool using a context engine include: The system employs a context engine to receive multimodal data returned by the target tool; this multimodal data includes log fragments, SQL datasets, alarm details, and interface receipts. Based on the multimodal data, calculate the scale score and semantic relevance score; Based on the scale score and semantic relevance score, intelligent decision-making is performed on the multimodal data to obtain decision results; wherein, the intelligent decision-making is used to determine whether to perform context caching, compression, segmentation and reuse; Based on the decision result, the multimodal data is cached in context; The multimodal data is processed in layers to generate a JSON profile; The JSON profile is sent to the Agentic Layer using a pointer handoff.

3. The method as described in claim 2, characterized in that, The step of performing context caching on the multimodal data based on the decision result includes: If the decision result is that the size score is less than the preset size score and the semantic relevance score is greater than the preset relevance score, then the direct injection method is used for context caching. If the decision result is that the scale score is not less than the preset scale score, or the semantic relevance score is not greater than the preset relevance score, then the Virtual File System (VFS) is used for context caching, and a structured profile is generated.

4. The method as described in claim 3, characterized in that, The step of using a Virtual File System (VFS) for context caching includes: Using an address protocol, virtual file paths obtained from the global data base are cached in the virtual file system VFS; wherein, the global data base includes the distributed object storage system MinIO, the graph database Neo4j, and the vector database Weaviate; Fine-grained access control is implemented for enterprise-level identity permissions in the Persona ACL layer of the Virtual File System (VFS). In response to user access commands, the virtual file path is obtained from the Virtual File System (VFS) through the Model Context Protocol (MCP) Tool interface.

5. The method as described in claim 2, characterized in that, The step of performing layered processing on the multimodal data to generate a JSON profile includes: In the lightweight layer, the Pandas data analysis library is used to perform statistical analysis on the multimodal data and obtain statistical results. At the structural layer, the schema metadata of the multimodal data is extracted; At the semantic layer, the text content of the multimodal data is converted into semantic vectors, and the business themes of the multimodal data are identified; wherein, the business themes include faults, performance, configuration, and security. At the business layer, the semantic vector is matched with the enterprise business knowledge vector library to obtain the matching result; Based on statistical results, schema metadata, semantic vectors, business themes, and matching results, a structured JSON profile is generated.

6. The method as described in claim 2, characterized in that, After making intelligent decisions based on the multimodal data according to the scale score and semantic relevance score, and obtaining the decision results, the method further includes: Retrieve the global history vector associated with the current task from the global vector memory; The global history vector is segmented to obtain multiple historical segments; Calculate the semantic relevance score for each historical segment separately; For each historical segment whose semantic relevance score is greater than the preset relevance score, a corresponding micro-summary is generated. The weight of each micro-summary is determined based on its semantic relevance score. Based on the weight of each micro-summary, sort the micro-summaries in descending order; Based on the weighted sorting results, generate a list of compressed links and backup index pointers; Obtain the compression history context based on the current task, compression link, and list of backup index pointers.

7. The method as described in claim 6, characterized in that, The step of generating a list of compressed links and backup index pointers based on the weight sorting results includes: Based on the weight ranking results, multiple micro-summaries with weights greater than a preset weight threshold are concatenated to generate compressed links. Based on the weight sorting results, the VFS access pointers of at least one micro-digest with a weight not greater than a preset weight threshold are placed into the list of standby index pointers.

8. The method as described in claim 6, characterized in that, The semantic relevance score is calculated using the following formula: relevance=cosine_similarity(aggregate_vec, Goal_vec) Where relevance is the semantic relevance score; cosine_similarity is the cosine similarity; aggregate_vec is the semantic aggregation vector of historical segments; and Goal_vec is the target vector of the current operation and maintenance task.

9. An intelligent context caching system, characterized in that, The intelligent context caching system, applied in enterprise-level Agentic AIOps platforms, includes: The Agentic Layer is used to communicate with the ContextEngine via Tool Call; the Agentic Layer includes a Large Language Model (LLM), multiple Agents, multiple Tools, and multiple Skills. The Context Engine is used to intelligently process the multimodal data returned by the target tool; the intelligent processing includes tool data input, two-dimensional decision score calculation, Persona ACL permission control, intelligent decision-making, virtual file system (VFS) storage, multimodal profile injection, and pointer handoff zero-copy.

10. The intelligent context caching system as described in claim 9, characterized in that, The intelligent context caching system also includes a virtual file system (VFS) for: An address protocol is used to cache virtual file paths obtained from the global data base; wherein, the global data base includes the distributed object storage system MinIO, the graph database Neo4j, and the vector database Weaviate; Implement fine-grained access control for enterprise-level identity permissions; In response to user access commands, the virtual file path is sent to the Model Context Protocol (MCP) Tool interface.

11. The intelligent context caching system as described in claim 10, characterized in that, The virtual file system (VFS) is also used for: A multi-agent cross-session compressed chain link is established using the graph database Neo4j; wherein, the graph database Neo4j supports range queries using the #[start-end] syntax; After the cross-session ends, a Process Improvement Plan (PIP) report is generated; the PIP report contains thematic summaries, lessons learned, and optimization recommendations, and the PIP report can be applied to the Agent configuration with one click.

12. The intelligent context caching system as described in claim 9, characterized in that, The ContextEngine also includes a Delta incremental cache submodule, used for: Cache newly added or changed data segments in the multimodal data returned by the target tool Tool.

13. An electronic device, characterized in that, The device includes: Memory, used to store program instructions; A processor is configured to invoke program instructions stored in the memory and execute the method described in any one of claims 1-8 according to the obtained program instructions.

14. A storage medium, characterized in that, The storage medium stores computer-executable instructions for causing a computer to perform the method described in any one of claims 1-8.

Images