LLM and LangChain-based event management system optimization method

Abstract

The invention relates to an optimization method of an emergency management system, in particular to an optimization method of an event management system based on LLM and LangChain. According to the method, rapid processing of information such as various intelligence, laws and regulations and cases is accelerated by utilizing an LLM technology, intelligent enabling is performed on an event management system, a field LLM application is rapidly constructed by utilizing a LangChain framework to improve automation of an event disposal process, interaction with a user is performed through a natural language mode, an event entity relationship is associated, and the event disposal efficiency is improved. The method is convenient for commanders to comprehensively understand event venation, event situations and event related backgrounds, and can assist the commanders to quickly make correct reflection and timely adjustment. According to the invention, the LangChain framework is adopted to quickly construct the event management system, the LLM is accessed to carry out private deployment, the strong information understanding and transaction reasoning capabilities are utilized to deal with sudden problems in a command combat scene, and the event handling efficiency is improved.

Description

Technical Field

[0001] The present invention relates to an optimization method for an emergency management system, in particular to an optimization method for an emergency management system based on LLM and LangChain. Background Art

[0002] A large number of intelligent combat equipment and new combat styles have emerged on the modern battlefield, which has improved the overall battlefield perception capability and gradually evolved into an integrated and complex combat system. It has become very difficult to rely on manual processing of massive information. Commanders need to master a lot of knowledge including combat equipment, combat tactics, laws and regulations, battlefield intelligence, social public opinion, etc. Therefore, there is an urgent need for artificial intelligence technology to assist commanders in planning, decision-making and business processing.

[0003] At present, artificial intelligence is divided into dedicated artificial intelligence and general artificial intelligence. Dedicated artificial intelligence is good at handling specific tasks in a single scenario. Most intelligent applications currently belong to this category. For example, "AlphaGO" launched by DeepMind can defeat the human world chess champion in chess, but it is unrealistic to ask it to generate a summary for a text. For large systems with complex business, a large number of models need to be integrated, which is too costly and difficult to integrate. However, general artificial intelligence pays more attention to flexibility and versatility, and is better at handling complex and comprehensive tasks, especially in complex scenarios such as command and combat. Emergencies are changeable and the development of events is difficult to predict. It is necessary to integrate multiple aspects of intelligence for comprehensive analysis. General artificial intelligence can play its strengths in reasoning, planning, problem solving, abstract thinking, understanding complex ideas, rapid learning and learning from experience in affairs, and assist commanders in making correct event handling decisions under tense situations.

[0004] General artificial intelligence mainly refers to LLM (Large Language Model). The most representative application is "ChatGPT" launched by OpenAI at the end of 2022. It has become popular on the global Internet with its powerful language organization, text processing, logical reasoning and other capabilities, and has become a phenomenal application in the field of artificial intelligence. LLM uses the Transformer model at the bottom layer, uses the self-attention mechanism to improve the ability to understand context, and obtains the ability to reason, plan, solve problems, think abstractly, understand complex ideas, learn quickly and learn from experience in transactions through large-scale training (usually with more than one billion parameters). Summary of the invention

[0005] In order to improve the efficiency of the emergency management system in the command and combat scenario, the present invention provides an event management system optimization method based on LLM and LangChain, which uses LLM technology to accelerate the rapid processing of various intelligence, laws and regulations, cases and other information, and intelligently empowers the event management system. The LangChain framework is used to quickly build a field LLM application to improve the automation of the event handling process, and interact with users in a natural language manner, and associate event entity relationships, so as to facilitate the commander to have a comprehensive understanding of the event context, event situation and event-related background. In an emergency, the commander can quickly make correct reflections and immediate adjustments based on historical cases stored in the LLM vector library, thereby improving the command and combat capabilities. The present invention uses the LangChain framework to quickly build an event management system, which can be connected to ChatGLM, Qwen, LocalGPT, Ollama and other LLMs for private deployment, and uses its powerful information understanding and transaction reasoning capabilities to deal with sudden problems in the command and combat scenario, and improve the efficiency of event handling.

[0006] The present invention is implemented by adopting the following technical scheme: an event management system optimization method based on LLM and LangChain, the event management system adopts a three-layer system architecture design of application layer, framework layer and model layer, the application layer is designed around the command and combat emergency event handling business logic, and the emergency event handling process is divided into four stages, including event data management, event situation grasp, event handling suggestion, and event handling organization and coordination; the framework layer adopts the large language application development framework LangChain for business series connection and logic splicing, optimizes each stage of the emergency event handling process facing the application layer, provides intelligent auxiliary decision support for the commander's decision-making under emergency conditions, and ensures that the emergency situation can be responded to immediately and correctly under tense situations; the model layer is connected to the open source LLM, responsible for forming the expected event situation summary through text understanding and knowledge reasoning of the input natural language information, can call the external interface to query the event hot spot information and the related case support system to generate the event handling plan, and can interact with the external system to automatically issue command orders, combat documents, message warnings, etc. for inter-seat organization and coordination.

[0007] In the above-mentioned event management system optimization method based on LLM and LangChain, the event data management module obtains intelligence data from the intelligence system and business system to provide data support for building the field LLM; the data source of the event situation grasping module is reported by the commanders of each seat, and the content summary is formed through LLM; the event handling suggestion module uses LLM to abstract the event handling experience from historical cases, form a handling plan, and weigh it according to the comparison of forces, material support, and social public opinion factors, provide handling suggestions and reasons, and help commanders make handling decisions; the event handling organization and coordination module responds to emergencies according to the event handling plan formulated by the commander, conducts action monitoring, force deployment, and counter-planning, forms task work orders and distributes them according to the grouping, supports data flow between various elements and seats, and realizes online organization and coordination of the entire process of emergency handling.

[0008] The above-mentioned event management system optimization method based on LLM and LangChain, the event data management module specifically includes two parts: data introduction and loading and content word vector storage; when data is introduced and loaded, the LangChain data loader component is used to introduce and load the data in the intelligence system and the business system to realize the data service request and data extraction functions. The intelligence data and business data introduced by the data loader need to be stored in the database to provide knowledge support for historical case query, legal and regulatory query, and disposal plan recommendation during business execution. The data process includes three stages: document data attribute extraction, word vector embedding training, and data storage: ① The introduced data uses the Doctran component to extract document attributes for accurate query and related retrieval of relevant information; ② The content of the document is embedded with word vector training, semantically associated with the context and distinguished between different contexts to form a high-dimensional vector format that can be understood by the computer; ③ The Chroma vector database is used to store the trained word vector content.

[0009] The above-mentioned event management system optimization method based on LLM and LangChain adds a BufferMemory component window filtering mechanism to smooth the peaks and fill the valleys of data flows in order to prevent congestion of the data connection channel during data connection loading.

[0010] The above-mentioned event management system optimization method based on LLM and LangChain, the event situation grasping module includes two parts: document metadata tagging, knowledgeable label category generation and content comprehensive summary generation. The document metadata label tagging documents historical cases of emergencies, constructs structured metadata labels of events to form a knowledgeable label category system in the command and operation field, and stores it in the Chroma vector database through a text embedding model; the situation reports submitted by each seat are cut into file blocks through StuffDocumentChain, and the relevant information is queried from the file blocks according to the event metadata through the indexer, and filled into the corresponding summary prompt word template, and finally input into LLM to generate content comprehensive summaries of multiple situation reports.

[0011] In the above-mentioned event management system optimization method based on LLM and LangChain, the event handling suggestion module is triggered by the user, and the user queries the business system access database through the SQLDatabaseChain component to realize relevant real-time information query, and calls the custom method through the MultiFunctionAgent component to query the existing combat knowledge query system; using the thinking chain technology, the complex problem Q involved in solving the emergency is decomposed into a series of simple problems Q={q 1 ,q 2 ,q 3 ,…,q n}, forming a problem decomposition template to provide reference ideas for LLM; the system fills the thinking chain into the prompt word template for generating the disposal plan, and uses the example selector to automatically match the appropriate thinking chain, so that LLM can generate the disposal plan according to the same thinking; at the same time, the system provides a custom output template to generate the disposal plan according to the specified format, and then inputs the result into LLM again and enters the scoring prompt word template, and finally analyzes and compares the disposal plan through the output analyzer to generate disposal suggestions.

[0012] In the above-mentioned event management system optimization method based on LLM and LangChain, the event handling organization and coordination module relies on the Callback mechanism provided by LangChain to preset the method and timing of task execution. The custom tool can set the specific content of task execution, and the execution agent can set the conditions for task execution. The system can execute multiple tasks in series in an orderly manner according to the business logic to form a complete workflow; when the emergency handling enters the plan implementation stage, the user triggers the task dispatching action, and the system will add the task to the workflow according to the handling steps in the handling plan. The unstructured handling plan is converted into a formatted command through LLM translation, and its content includes elements such as the command sender, the command receiver, the command type, the command content, and the urgency; the system completes the construction of the graph of event entities and the relationship between entities through the dialogue knowledge graph memory component, and completes the filling of various elements of the formatted command command; after the system monitors the generation of the command command, it calls the interface through the Callback mechanism to complete the drafting of the combat document, and finally the commander manually checks, modifies, and sends it, realizing the task work order dispatching of people in the loop.

[0013] The present invention is novel in two aspects: the selection of intelligent models and the optimization method of business processing: ① The selection of intelligent models is novel, and the most advanced LLM with broad development prospects is selected as the underlying support. Existing event management systems are mainly reported and summarized in the form of written reports, most of which are based on keywords, rules, and structured data for matching and extraction of event information to assist commanders in understanding the situation of emergencies. This method is limited by rules and formats and has poor effects. It often requires manual participation in the review and modification of situation reports, and is inefficient. The present invention adopts LLM, which has advantages in content understanding, is not restricted by rules, structures, etc., and can automatically understand and summarize various situation reports, so it is more efficient.

[0014] ② The optimization method of business processing is novel. While ensuring business scalability and system integration, LLM is used to improve the business processing efficiency of emergency handling. The existing event management system uses a dedicated artificial intelligence model for specific functional requirements. One model intelligently solves a certain type of specific problem. A large number of models are required for complex systems and multiple types of scenarios, resulting in high integration costs, and it is difficult to achieve logical connection and contextual association between models. The present invention uses the LangChain framework to connect business logic, which can quickly build and expand system functions, thereby reducing system integration costs. At the same time, the LLM used is a general intelligent model that uses a unified model for scheduling business understanding, which is convenient for storing business context content, and is conducive to understanding the relationship between event context and event subject, and is more efficient in business processing.

[0015] The present invention is creative in two aspects: 1. In terms of the improvement of field applications, LLM is creatively applied to the command and combat field to solve the pain points of commanders in handling emergencies, and provides cutting-edge exploration for the application and popularization of LLM, an emerging technology; 2. In terms of the application of advanced technologies, innovative methods such as massive data connection and fusion, multi-text summary extraction, handling solution reasoning based on thinking chain, and workflow construction based on graph memory are proposed to realize the automation and intelligent transformation of the event handling process, thereby improving the work efficiency of the emergency management system.

[0016] The present invention uses the current mainstream LangChain framework to optimize the command and combat emergency incident handling business, supports docking with various types of LLMs, supports private deployment and can be put into practical use, can flexibly build business processes to quickly form LLM-based field applications, and uses the generalization and transfer learning capabilities of LLM to support low-cost business expansion. Compared with traditional rule-based and probability-based machine learning methods, this method has higher accuracy and availability in semantic understanding, content generation, and logical reasoning. It can improve the speed of commanders processing massive amounts of information in actual use and quickly grasp the battlefield situation. It can build a combat command knowledge base to provide commanders with reasonable disposal plans and suggestions, realize the generation and distribution of combat documents, and improve organizational coordination efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is the system architecture diagram.

[0018] Figure 2 This is the structure diagram of the event data management module.

[0019] Figure 3 Master the module structure diagram for event situation.

[0020] Figure 4 Suggest a module structure diagram for event handling.

[0021] Figure 5 Diagram of the organizational coordination module for incident handling. DETAILED DESCRIPTION

[0022] System Architecture

[0023] The present invention adopts a three-layer system architecture design of application layer, framework layer and model layer. The application layer is designed around the business logic of command and operation emergency handling, and the emergency handling process is divided into four stages, including event data management, event situation grasp, event handling suggestions, and event handling organization and coordination. The design is optimized for each stage to provide intelligent support in the analysis and judgment of emergencies, reasoning and decision-making, etc. The framework layer adopts the six major components of the large language application development framework LangChain for business series connection and logic splicing, optimizes each stage of the emergency handling process facing the application layer, and provides intelligent auxiliary decision support for commanders in emergency situations. To ensure that in tense situations, it can respond to emergencies immediately and correctly. For the model layer, it improves the model effect through prompt word templates, tool agents, document summaries, word vector embedding and other technologies; the model layer connects to mainstream open source LLMs such as ChatGLM, Qwen, LocalGPT, Ollama, etc., and is mainly responsible for converting the input natural language information into the expected event summary through text understanding and knowledge reasoning. It can call external interfaces to query event hotspot information and related case support systems to generate event handling plans, and can interact with external systems to automatically issue command orders, combat documents, message warnings, etc. for inter-seat organization and coordination. The system architecture of the present invention is as follows Figure 1 shown.

[0024] Event Data Management

[0025] In order to ensure that the command and operation can take immediate response and adequate countermeasures to emergencies, it is necessary to make advance judgments and prepare plans based on massive intelligence information. Therefore, the present invention designs an event data management module for the emergency management system, which can obtain a large amount of intelligence data from the intelligence system and business system to provide data support for building domain LLM. Figure 2 As shown, it includes two parts: data link loading and content word vector storage.

[0026] 1) Data loading The intelligence system contains a large amount of strategic intelligence data and battlefield perception data, including meteorology and hydrology, geographical environment, air and sea intelligence, and social opinion. At the same time, the business system contains a large amount of field business data and field general data, including combat documents, key map drawing, information retrieval query, personnel equipment, etc. The emergency management system of the present invention uses the LangChain data loader component to connect and load the data in the intelligence system and the business system, realizes data service request and data extraction functions, and supports data formats such as Json, CSV, PDF, and URL. In order to prevent congestion in the data connection channel, the BufferMemory component window filtering mechanism is added to cut the peak and fill the valley of the data flow.

[0027] 2) Content word vector storage The intelligence data and business data introduced by the data loader need to be stored in the database to provide data query support for historical case queries, legal and regulatory queries, and disposal plan recommendations during business execution. The data process includes three stages: document data attribute extraction, word vector embedding training, and data storage: ① The introduced content uses the Doctran component to extract document attributes for accurate query and related retrieval of relevant materials. ② The document content is embedded with word vectors for semantic association and distinction between different contexts to form a high-dimensional vector format that can be understood by computers. ③ The Chroma vector database is used to store the trained word vector content.

[0028] Event situation

[0029] Emergency response requires commanders to be able to integrate information from all aspects to form an overall understanding and judgment of the incident. By understanding the background, motivation, capabilities, etc. of the subject of the incident, they can form a complete response plan for the context and development of the incident. The incident situation grasping module mainly provides commanders with a quick understanding of information to form usable knowledge, and integrates it with existing knowledge to accelerate commanders' research and judgment. The data source of the incident situation grasping module is mainly reported by commanders at various seats. Originally, the commander-in-chief summarized and refined the situation reports received from various directions. Now, this function of quickly forming content summaries can be achieved through the powerful natural language understanding capabilities of LLM. The incident situation grasping module is as follows: Figure 3 shown.

[0030] The event situation grasping module designed by the present invention mainly includes two parts: document metadata tagging, knowledgeable label category generation and content comprehensive summary generation. The document metadata label tagging performs document tagging on historical cases of emergencies, constructs structured metadata labels of events to form a knowledgeable label category system in the command and combat field, including event type, event level, event content, event background, event entity and other contents. It is stored in the Chroma vector database through the text embedding model. The situation reports submitted by each seat are cut into file blocks through StuffDocumentChain, and the relevant information is queried from the file blocks according to the event metadata through the indexer, and filled into the corresponding summary prompt word template, and finally input into the LLM to generate a comprehensive summary of the content of multiple situation reports.

[0031] Incident handling suggestions

[0032] After understanding the situation of the incident, the commander needs to give several reasonable disposal plans in combination with the current situation, and through consultation with the commanders of each seat, decide on the best disposal plan to implement. In this process, the formation of disposal plans often requires commanders to have rich experience in handling emergencies. Through in-depth study and assumptions of historical cases, laws and regulations, tactics and methods, they can predict the costs of various disposal plans and the impact on the development of the situation. The present invention uses the ability of LLM to learn quickly and learn from experience, which can abstract the experience of incident handling from historical cases, quickly form disposal plans, and weigh factors such as troop strength comparison, material support, and social opinion, provide disposal suggestions and reasons, and help commanders make disposal decisions quickly.

[0033] The event handling suggestion module is triggered by the user. The user can query the business system database through the SQLDatabaseChain component to realize relevant real-time information query, or call the custom method through the MultiFunctionAgent component to query the existing combat knowledge query system, including historical cases, laws and regulations, tactics and strategies, etc. The present invention adopts the thinking chain technology to decompose the complex problem Q into a series of simple problems Q={q 1 ,q 2 ,q 3 ,…,q n}, forming a problem decomposition template to provide reference ideas for LLM. The system fills the thought chain into the prompt word template for generating the disposal plan, and uses the example selector to automatically match the appropriate thought chain, so that LLM can generate the disposal plan according to the same thought. At the same time, the system provides a custom output template to generate the disposal plan according to the specified format, and then inputs the result into LLM again and inputs the scoring prompt word template, and finally analyzes and compares the disposal plan through the output analyzer to generate disposal suggestions.

[0034] Incident handling organization and coordination

[0035] The incident handling organization and coordination responds to emergencies according to the incident handling plan formulated by the commander, carries out tasks such as action monitoring, force deployment, countermeasure planning, etc., forms task work orders and distributes them according to the group, supports data flow between various elements and seats, and realizes online organization and coordination of the entire process of emergency handling. Figure 5 shown.

[0036] The present invention relies on the Callback mechanism provided by LangChain to preset the method and timing of task execution. The custom tool can set the specific content of task execution, the execution agent can set the conditions for task execution, and the system can execute multiple tasks in series and orderly according to the business logic to form a complete workflow. When the emergency disposal enters the implementation stage of the plan, the user triggers the task dispatching action, and the system will add the task to the workflow according to the disposal steps in the disposal plan. The unstructured disposal plan is converted into a formatted command through LLM translation, and its content includes elements such as the command sender, the command receiver, the command type, the command content, and the degree of urgency. The system completes the construction of the graph of the relationship between event entities and entities through the dialogue knowledge map memory component, and completes the filling of various elements of the formatted command command. After the system monitors the generation of the command command, it calls the interface through the Callback mechanism to complete the drafting of the combat document, and finally the commander manually checks, modifies, and sends it, realizing the dispatch of the task work order with people in the loop.

[0037] The key point of the present invention is to closely combine the natural language understanding, content generation, learning and reasoning capabilities of LLM with the command and combat incident handling business, give full play to the universal and easy generalization advantages of LLM to deal with various emergency scenarios, be able to abstract and learn from historical cases and easily migrate to future scenarios, and improve business development capabilities. At the same time, the present invention conducts an in-depth analysis of the command and combat incident handling business, and solves the business pain points of commanders through methods such as massive data connection and fusion, multi-text summary extraction, handling solution reasoning based on thinking chain, and workflow construction based on graph memory, so as to realize the automation and intelligent transformation of the event handling process and improve the work efficiency of the emergency management system.

[0038] The present invention mainly proposes four optimization methods for business processing based on LLM at each stage of the emergency handling process: ① For event data management, various formats of data in intelligence systems and business systems are converted into recognizable and usable data through methods such as connection, loading, segmentation, word vector embedding, persistence, etc., to support LLM to learn domain knowledge; ② For event situation grasping, metadata tags are used to construct label categories for existing domain knowledge, and content summary technology is used to understand the event situation reported by each seat to form a comprehensive situation report, so as to improve the efficiency of commanders in grasping the emergency situation; ③ For event handling suggestions, wrappers and agent designs are used to add "hands and feet" to the system to realize data query access to hot information, historical cases, etc., and a thinking chain method is used to provide prompt references for LLM to generate handling plans, so as to assist commanders in quickly providing handling plans and suggestions; ④ For event handling organization and coordination, the callback mechanism and custom tools provided by LangChain are used to build workflows, format commands are generated for specific implementation steps of the handling plan, and commanders are assisted to fill in key information such as task executors, locations, events, etc., and finally a work order in the format of combat documents is generated to circulate between various seats and levels.

Claims

1. An event management system optimization method based on LLM and LangChain, characterized in that: The incident management system adopts a three-layer system architecture design of application layer, framework layer and model layer. The application layer is designed around the business logic of command and operation emergency incident handling, and the emergency incident handling process is divided into four stages, including incident data management, incident situation grasp, incident handling suggestions, and incident handling organization and coordination; The framework layer uses the large language application development framework LangChain for business concatenation and logic splicing, and optimizes each stage of the emergency handling process for the application layer, providing intelligent auxiliary decision-making support for commanders in emergency situations, ensuring that they can respond to emergencies immediately and correctly in tense situations; The model layer is connected to the open source LLM, which is responsible for forming the expected event situation summary through text understanding and knowledge reasoning of the input natural language information. It can call external interfaces to query event hotspot information and related case support systems to generate event handling plans, and can interact with external systems to automatically issue command orders, combat documents, and message warnings to organize and coordinate between seats.

2. The event management system optimization method based on LLM and LangChain according to claim 1, characterized in that: The event data management module obtains intelligence data from the intelligence system and business system to provide data support for building domain LLM; the data source of the event situation grasping module is reported by the commanders of each seat, and the content summary is formed through LLM; The incident handling suggestion module uses LLM to abstract incident handling experience from historical cases, form a handling plan, and weigh factors such as troop strength, material support, and social opinion to provide handling suggestions and reasons to help commanders make handling decisions; The incident handling organization and coordination module conducts emergency responses according to the incident handling plan formulated by the commander, carries out action monitoring, troop deployment, and counter-measure planning, forms task work orders and distributes them according to groups, supports data flow between various elements and seats, and realizes online organization and coordination of the entire process of emergency handling.

3. The event management system optimization method based on LLM and LangChain according to claim 2 is characterized in that: The event data management module specifically includes two parts: data loading and content word vector storage. When loading data, the LangChain data loader component is used to load data from the intelligence system and business system to realize data service request and data extraction functions. The intelligence data and business data loaded by the data loader need to be stored in the database to provide knowledge support for historical case query, legal and regulatory query, and disposal plan recommendation during business execution. Its data process includes three stages: document data attribute extraction, word vector embedding training, and data storage: ① The Doctran component is used to extract document attributes from the loaded data to facilitate accurate query and related retrieval of relevant information; ② The word vector embedding training is performed on the content of the document to semantically associate with the context and distinguish different contexts to form a high-dimensional vector format that can be understood by the computer; ③ The Chroma vector database is used to store the trained word vector content.

4. The event management system optimization method based on LLM and LangChain according to claim 3 is characterized in that: To prevent congestion in the data connection channel during data connection loading, a BufferMemory component window filtering mechanism is added to smooth the data flow.

5. The event management system optimization method based on LLM and LangChain according to claim 3 is characterized in that: The event situation grasping module includes two parts: document metadata tagging, knowledgeable label category generation and content comprehensive summary generation. Document metadata tagging tags historical cases of emergencies, builds structured metadata tags for events to form a knowledgeable label category system in the command and operation field, and stores it in the Chroma vector database through a text embedding model. The situation reports submitted by each seat are cut into file blocks through StuffDocumentChain, and the indexer queries relevant information from the file blocks according to the event metadata, and fills it into the corresponding summary prompt word template, and finally inputs it into the LLM to generate content comprehensive summaries of multiple situation reports.

6. The event management system optimization method based on LLM and LangChain according to claim 5 is characterized in that: The event handling suggestion module is triggered by the user. The user queries the database connected to the business system through the SQLDatabaseChain component to realize relevant real-time information query, and calls the custom method through the MultiFunctionAgent component to query the existing combat knowledge query system. Using the thinking chain technology, the complex problem Q involved in solving the emergency is decomposed into a series of simple problems Q={q1,q2,q3,…,q n }, forming a problem decomposition template to provide reference ideas for LLM; the system fills the thinking chain into the prompt word template for generating the disposal plan, and uses the example selector to automatically match the appropriate thinking chain, so that LLM can generate the disposal plan according to the same thinking; at the same time, the system provides a custom output template to generate the disposal plan according to the specified format, and then inputs the result into LLM again and enters the scoring prompt word template, and finally analyzes and compares the disposal plan through the output analyzer to generate disposal suggestions.

7. The event management system optimization method based on LLM and LangChain according to claim 6 is characterized in that: The event handling organization and coordination module relies on the Callback mechanism provided by LangChain to preset the method and timing of task execution, customize the tool to set the specific content of task execution, and execute the agent to set the conditions for task execution. The system connects multiple tasks in series and executes them in an orderly manner according to the business logic to form a complete workflow; when the emergency event handling enters the plan implementation stage, the user triggers the task dispatch action, and the system will add the task to the workflow according to the handling steps in the handling plan. The unstructured handling plan is converted into a formatted command through LLM translation, and its content includes the command sender, command receiver, command type, command content, and urgency; The system uses the conversation knowledge graph memory component to complete the graph construction of event entities and the relationship between entities, and completes the filling of various elements of the formatted command. After the system monitors the generation of the command, it calls the interface through the Callback mechanism to complete the drafting of the combat document, which is finally manually checked, modified, and sent by the commander to realize the dispatch of task work orders with people in the loop.

Images