Priority scheduling method based on large-model multi-agent concurrency and related device

Abstract

The invention belongs to the technical field of artificial intelligence, and provides a priority scheduling method based on large-model multi-agent concurrency and a related device. The method is applied to a large language model and an agent concurrence scheduling system, the agent concurrence scheduling system comprises a housekeeper agent and k exclusive task agents, and the agent concurrence scheduling system can schedule resources according to a priority scheduling algorithm by designing the priority scheduling algorithm; and the workflows corresponding to the plurality of exclusive task agents are executed concurrently, so that the user is allowed to continuously issue the tasks for multiple times without waiting for the completion of the previous task, and finally the time-sharing technology is realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a priority scheduling method and related devices based on large-model multi-agent concurrency. Background Art

[0002] As the digital transformation of society continues to deepen, the scale of computing power has exploded, greatly promoting the development of artificial intelligence technology. In particular, the rapid development of large artificial intelligence models represented by LLM (Large Language Model) this year has promoted the deep integration of artificial intelligence technology and application systems, and enabled application software to set off a climax in the development and application of large artificial intelligence models.

[0003] The technology of building application systems based on large language models refers to the use of pre-trained large language models as the core components of artificial intelligence systems, in conjunction with intelligent agents to realize reasoning about natural languages, and thus output reasoning conclusions. However, the current interaction mode between intelligent agents and users on the market is mostly serial, that is, after users issue tasks, they need to wait for the intelligent agent to complete before issuing new tasks. This mode limits the continuity and efficiency of user operations. There is an urgent need for a technology that can design priorities for different tasks according to the urgency of user needs, and automatically identify the priority of tasks to ensure that important tasks (such as video conferencing) can be processed immediately. Summary of the invention

[0004] In response to the defects in the prior art, the present invention provides a priority scheduling method and related devices based on large-model multi-agent concurrency, which allows users to issue tasks continuously for multiple times through time-sharing sharing technology, so that the agent concurrent scheduling system can execute multiple tasks concurrently without waiting for the previous task to be completed.

[0005] The priority scheduling method based on large model multi-agent concurrency is applied to large language models, agent concurrent scheduling system, the agent concurrent scheduling system includes a housekeeper agent and k dedicated task agents, and the large language model is deployed with a pre-built agent concurrent scheduling system. The method includes:

[0006] The housekeeper agent performs personalized optimization on the obtained k first task instructions respectively to obtain k second task instructions, and the k second task instructions correspond one-to-one to the k first task instructions;

[0007] The housekeeper agent sends each of the k second task instructions to the corresponding dedicated task agent;

[0008] Each of the k dedicated task agents performs enhancement processing on the received second task instructions according to the corresponding hierarchical memory module to obtain k third task instructions;

[0009] Each of the k dedicated task agents performs intent recognition and process reasoning decomposition on the corresponding third task instruction to generate workflows corresponding to the k dedicated task agents;

[0010] The agent concurrent scheduling system concurrently executes the workflows corresponding to k dedicated task agents according to the priority scheduling algorithm.

[0011] Furthermore, the priority scheduling algorithm also includes:

[0012] The priority scheduling algorithm is characterized by the following formula:

[0013] π agent =α·π pre +β·(L default -L remaining )

[0014] Among them, π agent represents the real-time priority of the workflow corresponding to the agent that executes the dedicated task, π pre is the scene priority of the predefined dedicated task agent, L default represents the default length of the workflow corresponding to the dedicated task agent, L remaining represents the remaining workflow length of the unexecuted dedicated task agent, α represents the priority weight of the predefined dedicated task agent, and β represents the influence weight of the remaining workflow length of the dedicated task agent on the real-time priority.

[0015] Furthermore, the priority weight of the predefined dedicated task agent is 0.7, and the influence weight of the remaining workflow length of the dedicated task agent on the real-time priority is 0.3.

[0016] Furthermore, the housekeeper agent performs personalized optimization on the obtained k first task instructions respectively to obtain k second task instructions, including:

[0017] The housekeeper agent performs task scenario analysis on each of the k first task instructions obtained, and determines optimization information corresponding to each of the k first task instructions;

[0018] The housekeeper agent optimizes the k first task instructions respectively according to the optimization information corresponding to the k first task instructions, and obtains k optimized task instructions, and the k optimized task instructions correspond to the k first task instructions one by one;

[0019] The housekeeper agent calls the knowledge graph to supplement the background context of each of the k optimization task instructions to obtain k second task instructions.

[0020] Furthermore, the knowledge graph is constructed in the following way:

[0021] The large language model qwen2 is fine-tuned using the LORA method to obtain a knowledge graph construction model;

[0022] The knowledge graph construction model constructs the system text based on the GraphRAG method to obtain the knowledge graph.

[0023] Furthermore, the hierarchical memory module includes a core memory area, a main memory area and a fuzzy memory area.

[0024] Furthermore, the method further comprises:

[0025] Each of the k dedicated task agents uses the ReAct method to perform process reasoning and decomposition on the corresponding third task instructions.

[0026] The priority scheduling device based on large model multi-agent concurrency includes an optimization unit, a sending unit, an enhancement unit, a generation unit and an execution unit, wherein:

[0027] An optimization unit, used for the housekeeper agent to perform personalized optimization on the obtained k first task instructions respectively, to obtain k second task instructions, where the k second task instructions correspond one to one with the k first task instructions;

[0028] A sending unit, used for the housekeeper agent to send each second task instruction of the k second task instructions to the corresponding dedicated task agent;

[0029] An enhancement unit, configured for each of the k dedicated task agents to enhance the received second task instructions according to the corresponding hierarchical memory module to obtain k third task instructions;

[0030] A generation unit, configured for each of the k dedicated task agents to perform intent recognition and process reasoning decomposition on the corresponding third task instruction, and generate workflows corresponding to the k dedicated task agents respectively;

[0031] An execution unit is used for the agent concurrent scheduling system to concurrently execute the workflows corresponding to the k dedicated task agents according to the priority scheduling algorithm.

[0032] A terminal comprises a processor, an input device, an output device and a memory, wherein the processor, input device, output device and memory are interconnected, wherein the memory is used to store a computer program, the computer program comprises program instructions, and the processor is configured to call the program instructions to execute a priority scheduling method based on large-model multi-agent concurrency as described in any one of the above items.

[0033] A computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor executes any of the above-mentioned priority scheduling methods based on large-model multi-agent concurrency.

[0034] Beneficial effects of the present invention: The present invention allows users to issue tasks multiple times in succession through time-sharing sharing technology, so that dedicated task agents can execute multiple tasks concurrently without waiting for the previous task to be completed, thereby improving the continuity and efficiency of user operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the specific embodiments of the present invention, the following will briefly introduce the drawings required for the specific embodiments or the prior art description. In all the drawings, each element or part is not necessarily drawn according to the actual scale.

[0036] Figure 1 A flowchart of a priority scheduling method based on large model multi-agent concurrency provided by an embodiment of the present invention;

[0037] Figure 2 A schematic diagram of the structure of a priority scheduling device based on large-model multi-agent concurrency provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0038] The following embodiments of the technical solution of the present invention are described in detail in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and are therefore only used as examples, and cannot be used to limit the protection scope of the present invention.

[0039] It should be noted that, unless otherwise specified, the technical terms or scientific terms used in this application should have the common meanings understood by those skilled in the art to which the present invention belongs.

[0040] In one embodiment, if Figure 1 As shown, a priority scheduling method based on large-model multi-agent concurrency is provided, which is applied to large-scale language models and an agent concurrent scheduling system. The agent concurrent scheduling system includes a butler agent and k exclusive task agents. The large-scale language model is deployed with a pre-built agent concurrent scheduling system, so that the user inputs the first task instruction on the front-end interface, and the butler agent can cooperate with the downstream exclusive task agent to complete the task issued by the user according to the input first task instruction.

[0041] Specifically, the priority scheduling method based on large model multi-agent concurrency includes:

[0042] 1. The housekeeper agent performs personalized optimization on the obtained k first task instructions respectively to obtain k second task instructions, and the k second task instructions correspond one-to-one to the k first task instructions;

[0043] Preferably, the housekeeper agent performs personalized optimization on the obtained k first task instructions respectively, including:

[0044] (1) The housekeeper agent performs task scenario analysis on each of the k first task instructions obtained, and determines the optimization information corresponding to each of the k first task instructions;

[0045] (2) The housekeeper agent optimizes the k first task instructions respectively according to the optimization information corresponding to the k first task instructions, and obtains k optimized task instructions, and the k optimized task instructions correspond to the k first task instructions one by one;

[0046] (3) The housekeeper agent calls the knowledge graph to supplement the background context of each of the k optimization task instructions to obtain k second task instructions.

[0047] The housekeeper agent receives k first task instructions, analyzes the task scenario of each first task instruction according to the user's requirements, and first determines the optimization information of each first task instruction. The optimization information includes rewriting, error correction, and disambiguation of the first task instruction. Error correction refers to restoring the first task instruction input by the user by mistake to an objective form through large model analysis, and disambiguation mainly eliminates reference disambiguation. When there is a dedicated task agent-office role agent, when the user needs to use the system text as the raw material for creative office work, the housekeeper agent needs to supplement the relevant context of the first task instruction, that is, fill in the task-related part of the local system document, analyze the task scenario, and determine the optimization information corresponding to the k first task instructions. Then the housekeeper agent optimizes each first task instruction according to the optimization information of each first task instruction to obtain k optimized task instructions. After that, the housekeeper agent calls the knowledge graph to supplement the background context of each optimized task instruction to obtain k second task instructions. Through the housekeeper agent's personalized optimization of the obtained k first task instructions, the communication cost can be reduced and the downstream dedicated task agents can be assisted to better complete the task.

[0048] Preferably, the knowledge graph is constructed in the following manner:

[0049] (1) Fine-tune the large language model qwen2 using the LORA method to obtain a knowledge graph construction model;

[0050] (2) The knowledge graph construction model constructs the system text based on the GraphRAG method to obtain the knowledge graph.

[0051] In order to reduce costs, the present invention chooses to fine-tune the large language model qwen2:7b as the base model for building knowledge graphs and queries. Use the fine-tuning script LoRA provided in the official project to fine-tune the large language model qwen2 to obtain the knowledge graph construction model, load the qwen2:7b model, and organize the new data set, dividing it into a training set and a validation set. Next, configure the fine-tuning parameters, including determining the learning rate (for example, from 1e-5 to 1e-3), the batch size (based on the available GPU memory size, such as 8 or 16), the number of iterations, and the evaluation frequency. Then, use the official fine-tuning script to load the knowledge graph construction model and apply LoRA to define the optimizer and loss function. In each training epoch, iterate the batches in the data set, perform forward propagation, calculate losses, back propagation, and weight updates. Regularly evaluate the performance of the knowledge graph construction model on the validation set, and adjust hyperparameters such as the learning rate and batch size based on feedback to optimize the knowledge graph construction model. The dataset focuses on instruction datasets, including NLP tasks such as entity recognition (NER), relation extraction (RE), event extraction (EE), and various Chinese and English datasets. Compared with the general GPT model, it has advantages in accuracy, completeness, and compliance with the specified output format, and reduces a lot of costs.

[0052] The GraphRAG method uses an external structured knowledge graph to enhance the contextual understanding of the large language model qwen2 and generate a more insightful response framework. Its goal is to retrieve the most relevant knowledge from the database and thus improve the quality of answers to downstream tasks.

[0053] 2. The housekeeper agent sends each of the k second task instructions to the corresponding dedicated task agent;

[0054] Each dedicated task agent has its own dedicated configuration file, which contains targeted system prompts. For example, the meeting agent (You are a meeting assistant, tasked with booking or instantly starting meetings. You are a meeting assistant, responsible for booking or starting meetings immediately). The setting of system instructions can fully explore the execution potential of the large model in special tasks. At the same time, the collected or self-made tool catalog is written into the dedicated task agent, and equipped with special tools, so that the dedicated task agent can better complete the tasks assigned by the user.

[0055] 3. Each of the k dedicated task agents performs enhancement processing on the received second task instructions according to the corresponding hierarchical memory module to obtain k third task instructions;

[0056] Preferably, the hierarchical memory module includes a core memory area, a main memory area and a fuzzy memory area.

[0057] The hierarchical memory module stores and retrieves memories with different access frequencies and importance levels in a hierarchical manner. Drawing on the design idea of ​​hierarchical cache in the operating system, the text memory is hierarchically partitioned into core memory area, main memory area and fuzzy memory area, and differentiated management strategies are adopted for different memory partitions.

[0058] Core memory area: This area has the smallest capacity but the fastest retrieval speed. It mainly stores memory items that are frequently accessed and important, usually involving the user's personal preferences and key information. The core memory area adopts a priority replacement strategy. When the capacity of this area reaches the upper limit, the memory items with the lowest priority will be replaced first.

[0059] Priority of memory entry π memory The calculation formula is as follows:

[0060] π memory =γ·f hit +δ·I importance

[0061] where f hit Indicates the access frequency, that is, the number of hits of the memory entry; I importance It represents the importance of information, and the big model rates the importance of memory items; γ represents the weight of access frequency; δ represents the weight of information importance.

[0062] Main memory area: This area has a large capacity and a relatively slow retrieval speed. It is mainly used to store conversation records between users and dedicated task agents. When the main memory area is saturated, the system will adopt the LRU (least recently used) or LFU (least frequently used) strategy for management. Memory entries that have not been used for a long time or are used less frequently will be transferred to the fuzzy memory area, while those memory entries with a hit count exceeding the threshold can be moved to the core memory area.

[0063] Fuzzy memory area: This area stores entries removed from the main memory area and is saved after further compression and summary. Since the fuzzy memory area has the largest capacity and a low hit rate, the data heat is low, so the FIFO (first in, first out) strategy is used for management to effectively reduce time complexity. This strategy achieves automatic forgetting of memory by eliminating the oldest entries.

[0064] During the interaction between the dedicated task agent and the user, when information outside the short-term memory (i.e., the context window) is involved, the system automatically triggers the memory retrieval mechanism. Following the principle of locality, the system first searches from the core memory area, which can quickly provide key or frequently used memory items; if it misses, it continues to search the main memory area to obtain the historical interaction records between the user and the agent; if it still misses, it enters the fuzzy memory area for retrieval. By calculating the cosine similarity between the memory item and the user input content, the system can evaluate the relevance of the memory item to the current task, thereby returning the memory information that best meets the needs.

[0065] The hierarchical memory model based on the operating system's hierarchical cache significantly improves the memory retrieval efficiency and adaptability of the dedicated task agent through the strategy of partition storage and hierarchical management. This hierarchical memory management method not only improves the efficiency of resource utilization, but also enhances the flexibility and scalability of the system when handling complex tasks.

[0066] 4. Each of the k dedicated task agents performs intent recognition and process reasoning decomposition on the corresponding third task instruction to generate workflows corresponding to the k dedicated task agents;

[0067] Preferably, each of the k dedicated task agents uses the ReAct method (ReAct) to perform process reasoning decomposition on the corresponding third task instruction.

[0068] ReAct uses the Chain of Thought to guide LLM (Large Language Model) to split complex problems, and perform reasoning and action step by step. It also introduces the observation link. After each execution (Action), it will first observe (Observation) the current status, and then proceed to the next step of reasoning (Reasoning). Developers guide LLM (Large Language Model) to reason step by step, and then determine which action to take based on the result of reasoning. The dedicated task agent (Agent) reasoning plans by itself to break down the task into a series of work steps (workflow). Each work step is completed by using one or more tools, and iterates at different links. Each step guides the final desired result.

[0069] The housekeeper agent manages many downstream dedicated task agents. When each dedicated task agent receives the corresponding second task instruction, it splits the task according to the existing tool resources, guides the LLM (Large Language Model) to perform reasoning, and then determines which action to take, that is, which tool to use, based on the result of reasoning. Then, based on the workflow of the ReAct paradigm, it further reasoned after combining the result observation and reflection, and automatically generated a complete workflow.

[0070] Preferably, each dedicated task agent will automatically generate a workflow based on the ReAct paradigm according to the target task assigned by the user and its own tool resources. The workflow length is limited to no more than 10, and errors are allowed during the execution of the workflow. After an error, the subsequent workflow will be reorganized, but the number of errors shall not exceed three times.

[0071] The downstream dedicated task agent understands the task and, based on the ReAct paradigm, automatically generates the steps required to complete the task according to the existing tool capabilities, executes the steps, and determines whether the task is successful. If successful, it directly provides feedback to the user. If unsuccessful, it switches to the template workflow to execute manually predefined steps.

[0072] 5. The agent concurrent scheduling system concurrently executes the workflows corresponding to k dedicated task agents according to the priority scheduling algorithm.

[0073] Preferably, the priority scheduling algorithm further includes:

[0074] The priority scheduling algorithm is characterized by the following formula:

[0075] π agent =α·π pre +β·(L default -L remaining )

[0076] Among them, π agent represents the real-time priority of the workflow corresponding to the agent that executes the dedicated task, π pre is the scene priority of the predefined dedicated task agent, L default represents the default length of the workflow corresponding to the dedicated task agent, L remaining represents the remaining workflow length of the unexecuted dedicated task agent, α represents the priority weight of the predefined dedicated task agent, and β represents the influence weight of the remaining workflow length of the dedicated task agent on the real-time priority.

[0077] Preferably, the priority weight of the predefined dedicated task agent is 0.7, and the influence weight of the remaining workflow length of the dedicated task agent on the real-time priority is 0.3.

[0078] In a general multi-agent system, each dedicated task agent usually executes linearly in time sequence, which will cause a dedicated task agent to continue to occupy resources before the task is completed, thereby hindering the issuance and execution of other tasks, and the user is therefore unable to submit new tasks during the execution process. In order to solve this problem, the present invention designs a time-sharing and shared multi-agent concurrent priority scheduling method, allowing users to continuously issue tasks without waiting. The core of this mechanism is to subdivide the continuous resource usage cycle into multiple discrete workflow time slices and allocate exclusive resource access time periods for different dedicated task agents. When the user submits a natural language instruction (the first task instruction), the housekeeper agent receives and analyzes the instruction and routes it to the corresponding dedicated task agent. The large language model automatically generates a workflow for the dedicated task agent according to the task requirements. For different dedicated task agents and different tasks, the workflow generated by the large model is also different. The present invention regards a single workflow step as the smallest unit of scheduling by the agent concurrent scheduling system, and uses the first-in-first-out (FIFO) and priority scheduling algorithms to switch workflows between dedicated task agents.

[0079] First-in-first-out scheduling algorithm: Based on the classic scheduling method in the operating system, the system allocates resources according to the arrival order of the workflow steps of the dedicated task agents.

[0080] Priority scheduling algorithm: Dynamically adjust the priority of the dedicated task agent based on the urgency of the task and the weighted average of the length of the generated unexecuted workflow. The urgency of the task is based on the predefined scenario priority of the dedicated task agent. The smaller the priority value, the higher the priority. The real-time priority of the workflow corresponding to the dedicated task agent is π agent The calculation formula is as follows:

[0081] π agent =α·π pre +β·(10-L remaining )

[0082] The default length of the workflow corresponding to the dedicated task agent is 10, π pre The scenario priority of the predefined dedicated task agent (such as "meeting agent" is 0, "application installation agent" is 1, and "document writing agent" is 2).

[0083] Through this scheduling mechanism, the execution of each dedicated task agent is transformed from linear order to staggered discrete execution, which significantly improves the throughput of the agent concurrent scheduling system, avoids a single dedicated task agent occupying resources for a long time, and thus reduces the average waiting time and turnaround time of the dedicated task agent, allowing users to obtain feedback information from the dedicated task agent earlier.

[0084] In one of the embodiments 1, the user issues a first task instruction: write a ppt of my annual work summary according to the system document. The housekeeper agent receives the first task instruction, analyzes whether the first task instruction needs to be corrected and corrected to eliminate ambiguity (the optimization information corresponding to the first task instruction). If the first task instruction is correct and does not need to rewrite the system instruction systemprompt, then the first task instruction is determined to be an optimized task instruction. According to the optimized task instruction, the housekeeper agent analyzes and concludes that a knowledge graph is needed. Among many system documents, the user's series of related work documents are found in the form of a global search through graphrag (graph retrieval enhancement), and the ppt material related to the annual summary is generated as the background context of the first task instruction, and the second task instruction (the first task instruction + the background context of the first task instruction) is obtained. The housekeeper agent routes the second task instruction to the downstream exclusive task agent - the ppt agent. After receiving the instruction, the ppt agent starts to start the hierarchical memory module to search for the user's preference memory about the ppt. The ppt agent then analyzes the theme, combines the user's style preference, and combines the ppt generation tool to automatically generate a workflow, and then executes the workflow to generate a ppt file.

[0085] In one of the embodiments 2, the user issues a first task instruction: install a document processing software. The housekeeper agent receives the first task instruction, analyzes that the user has not clearly specified the software, the instruction is ambiguous, and needs to perform reference disambiguation (optimization information corresponding to the first task instruction), obtains the optimized task instruction, and then chooses to call graphrag (graph retrieval enhancement) to retrieve the user's software usage preference. The user prefers to use wps software for document processing. The housekeeper agent modifies the user's first task instruction and fills the first task instruction with: install a document processing software, and recommends wps. Get the second task instruction. The new second task instruction is then routed to the downstream dedicated task agent-the software installation agent. The software installation agent generates a workflow by itself and then executes it.

[0086] In one of the embodiments 3, the user issues a first task instruction: schedule a meeting about the start of school for new students at 9 a.m. tomorrow (scheduled meeting). After receiving the first task instruction, the housekeeper agent is immediately routed to the exclusive task agent - the meeting agent. Then the user immediately issues the next first task instruction: immediately initiate a meeting about China Tourism Day (immediate meeting). The housekeeper agent is also immediately routed to the exclusive task agent - the meeting agent. At this time, two different meeting agents are started in different threads to complete two different tasks respectively. Each generates a workflow and then executes concurrently. Finally, the tasks of the scheduled meeting and the immediate meeting are basically completed at the same time.

[0087] In one of the embodiments 4, the user successively issues three first task instructions, namely: immediately initiate a meeting about the parent-teacher conference; install QQ; write a document about the Mid-Autumn Festival; after the housekeeper agent receives the three first task instructions, it concurrently starts three exclusive task agents. In daily use, video conferences usually need to be processed immediately, the urgency of software installation depends on the purpose of the software and the needs of the user, and document writing is a creative and time-intensive task. The three exclusive task agents automatically generate workflows based on their own tool resources, execute concurrently during the execution of the workflow, subdivide the continuous resource usage cycle into multiple discrete workflow time slices, and allocate exclusive resource access time periods to different exclusive task agents according to the priority scheduling algorithm. A single workflow step is regarded as the smallest unit of scheduling by the agent concurrent scheduling system, and a priority scheduling algorithm is used to switch workflows between exclusive task agents. The system finally achieves the effect of processing three tasks in parallel.

[0088] In one of the embodiments, a priority scheduling device based on large model multi-agent concurrency is also provided, including an optimization unit, a sending unit, an enhancement unit, a generation unit and an execution unit, wherein:

[0089] An optimization unit, used for the housekeeper agent to perform personalized optimization on the obtained k first task instructions respectively, to obtain k second task instructions, where the k second task instructions correspond one to one with the k first task instructions;

[0090] A sending unit, used for the housekeeper agent to send each second task instruction of the k second task instructions to the corresponding dedicated task agent;

[0091] An enhancement unit, configured for each of the k dedicated task agents to enhance the received second task instructions according to the corresponding hierarchical memory module to obtain k third task instructions;

[0092] A generation unit, configured for each of the k dedicated task agents to perform intent recognition and process reasoning decomposition on the corresponding third task instruction, and generate workflows corresponding to the k dedicated task agents respectively;

[0093] The execution unit is used in the agent concurrent scheduling system to concurrently execute the workflows corresponding to k dedicated task agents according to the priority scheduling algorithm.

[0094] A terminal includes a processor, an input device, an output device and a memory, wherein the processor, the input device, the output device and the memory are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute part or all of the steps of any priority scheduling method based on large model multi-agent concurrency as recorded in the above method embodiment.

[0095] A computer-readable storage medium stores a computer program, wherein the computer program includes program instructions. When the program instructions are executed by a processor, the processor executes part or all of the steps of any priority scheduling method based on large-model multi-agent concurrency as recorded in the above method embodiments.

[0096] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0097] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.

Claims

1. A priority scheduling method based on large model multi-agent concurrency, characterized by: Applied to a large language model, an agent concurrent scheduling system, the agent concurrent scheduling system includes a housekeeper agent and k dedicated task agents, the large language model is deployed with a pre-built agent concurrent scheduling system, the method includes: The housekeeper agent performs personalized optimization on the obtained k first task instructions respectively to obtain k second task instructions, and the k second task instructions correspond one to one to the k first task instructions; The housekeeper agent sends each of the k second task instructions to the corresponding dedicated task agent; Each of the k dedicated task agents performs enhancement processing on the received second task instructions according to the corresponding hierarchical memory module to obtain k third task instructions; Each of the k dedicated task agents performs intent recognition and process reasoning decomposition on the corresponding third task instruction to generate workflows corresponding to the k dedicated task agents respectively; The agent concurrent scheduling system concurrently executes the workflows corresponding to the k dedicated task agents respectively according to the priority scheduling algorithm.

2. The priority scheduling method based on large model multi-agent concurrency as claimed in claim 1 is characterized in that: The priority scheduling algorithm also includes: The priority scheduling algorithm is characterized by the method described in the following formula: p agent =a·p pre +β·(L default -L remaining ) Among them, π agent represents the real-time priority of the workflow corresponding to the agent that executes the dedicated task, π pre is the scene priority of the predefined dedicated task agent, L default represents the default length of the workflow corresponding to the dedicated task agent, L remaining represents the remaining workflow length of the unexecuted dedicated task agent, α represents the priority weight of the predefined dedicated task agent, and β represents the influence weight of the remaining workflow length of the dedicated task agent on the real-time priority.

3. The priority scheduling method based on large model multi-agent concurrency as claimed in claim 2 is characterized in that: The priority weight of the predefined dedicated task agent is 0.7, and the influence weight of the remaining workflow length of the dedicated task agent on the real-time priority is 0.

3.

4. The priority scheduling method based on large model multi-agent concurrency according to any one of claims 1 to 3, characterized in that: The housekeeper agent performs personalized optimization on the obtained k first task instructions respectively to obtain k second task instructions, including: The housekeeper agent performs task scenario analysis on each of the k first task instructions obtained, and determines optimization information corresponding to the k first task instructions respectively; The housekeeper agent optimizes the k first task instructions respectively according to the optimization information respectively corresponding to the k first task instructions to obtain k optimized task instructions, and the k optimized task instructions correspond to the k first task instructions one by one; The housekeeper agent calls the knowledge graph to supplement the background context of each of the k optimization task instructions to obtain k second task instructions.

5. The priority scheduling method based on large model multi-agent concurrency as claimed in claim 4 is characterized in that: The knowledge graph is constructed in the following way: The large language model qwen2 is fine-tuned using the LORA method to obtain a knowledge graph construction model; The knowledge graph construction model constructs the system text based on the GraphRAG method to obtain a knowledge graph.

6. The priority scheduling method based on large model multi-agent concurrency as claimed in claim 5 is characterized in that: The hierarchical memory module includes a core memory area, a main memory area and a fuzzy memory area.

7. The priority scheduling method based on large model multi-agent concurrency as claimed in claim 6 is characterized in that: Also includes: Each of the k dedicated task agents uses the ReAct method to perform process reasoning and decomposition on the corresponding third task instruction.

8. A priority scheduling device based on large model multi-agent concurrency, characterized in that: It includes an optimization unit, a sending unit, an enhancement unit, a generation unit and an execution unit, wherein: The optimization unit is used for the housekeeper agent to perform personalized optimization on the obtained k first task instructions respectively to obtain k second task instructions, and the k second task instructions correspond one-to-one to the k first task instructions; The sending unit is used for the housekeeper agent to send each of the k second task instructions to the corresponding dedicated task agent; The enhancement unit is used for each of the k dedicated task agents to enhance the received second task instructions according to the corresponding hierarchical memory module to obtain k third task instructions; The generating unit is used for each of the k dedicated task agents to perform intention recognition and process reasoning decomposition on the corresponding third task instruction, and generate workflows corresponding to the k dedicated task agents respectively; The execution unit is used in the agent concurrent scheduling system to concurrently execute the workflows corresponding to the k dedicated task agents according to the priority scheduling algorithm.

9. A terminal, characterized in that: The method comprises a processor, an input device, an output device and a memory, wherein the processor, the input device, the output device and the memory are interconnected, wherein the memory is used to store a computer program, the computer program comprises program instructions, and the processor is configured to call the program instructions to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor is caused to perform the method according to any one of claims 1 to 7.

Images