A multi-agent collaborative emergency response method and system for zirconium-based MOF synthesis accidents

Abstract

This invention relates to the field of emergency management technology and discloses a multi-agent collaborative emergency response method for zirconium-based MOF synthesis accidents, comprising: Step 1: collecting file data, extracting text content and storing it in a knowledge database; Step 2: converting the text content into semantic vectors and storing them in a vector database; Step 3: analyzing the text content, constructing a knowledge graph and storing it in a graph database; Step 4: the user agent receives query data input by the user, performs classification and keyword recognition to obtain the question type and first keyword; the knowledge retrieval agent obtains the fusion result based on the question type and first keyword; the verification agent filters the fusion result to obtain a compliant fusion result; the fusion decision agent inputs prompt words into a large model and outputs an emergency response plan. Simultaneously, a multi-agent collaborative emergency response system for zirconium-based MOF synthesis accidents is also disclosed.

Description

Technical Field

[0001] This application belongs to the field of emergency management technology, and more specifically, relates to a multi-agent collaborative emergency response method and system for zirconium-based MOF synthesis accidents. Background Technology

[0002] Zirconium-based metal-organic frameworks (Zr-MOFs), as a novel type of porous material, have broad application prospects in gas adsorption and separation, catalysis, and sensing. However, the synthesis process of Zr-MOFs involves several hazardous steps, including high-temperature and high-pressure reactions, the use of organic solvents, and metal salt treatment, posing the following safety risks: 1. Chemical hazards: The zirconium salts used in the synthesis process (such as ZrCl4) are highly corrosive, the organic ligands (such as terephthalic acid) are flammable, and the solvents (such as DMF and methanol) are flammable and explosive; 2. Process hazards: Hydrothermal / solventothermal synthesis requires high temperature and high pressure (autogenous pressure) conditions, posing an explosion risk; 3. Accident complexity: Accident types are diverse (leakage, fire, explosion, poisoning), and the causes are complex (equipment failure, operational errors, process loss of control). Emergency response requires comprehensive consideration of chemical properties, reaction mechanisms, equipment status, and other factors; 4. High difficulty in emergency decision-making: Traditional emergency plans are mostly static texts, which are difficult to cope with dynamically changing accident scenarios. Emergency personnel need to quickly retrieve a large amount of professional knowledge, resulting in high decision-making pressure and tight time constraints.

[0003] Currently, emergency response to chemical accidents mainly relies on paper-based emergency plans, expert consultation systems, and knowledge base systems. However, these methods are inefficient, have poor knowledge correlation, and make it difficult to generate personalized solutions.

[0004] Therefore, the technical problem addressed by this application is: how to improve the accuracy and efficiency of generating emergency response plans. Summary of the Invention

[0005] The main objective of this application is to provide a multi-agent collaborative emergency response method for zirconium-based MOF synthesis accidents. This method acquires various file data and constructs multiple databases. When generating an emergency response plan, it performs a mixed search on multiple databases to comprehensively obtain information related to the accident. Then, through multi-agent collaboration, it generates an emergency response plan, thereby improving accuracy and efficiency.

[0006] In addition, a multi-agent collaborative emergency response system for zirconium-based MOF synthesis accidents is also provided.

[0007] To achieve the above objectives, the technical solution adopted in this application is as follows:

[0008] A multi-agent collaborative emergency response method for zirconium-based MOF synthesis accidents includes the following steps:

[0009] Step 1: Collect accident-related document data, use parsing tools to parse the document data, extract the text content and store it in the knowledge database;

[0010] Step 2: Convert the text content into semantic vectors using a text embedding model and store them in a vector database;

[0011] Step 3: Analyze the text content using a pre-set model to obtain various entities and the relationships between different entities. Based on the entities and the relationships between different entities, construct a knowledge graph and store it in the graph database.

[0012] Step 4: Construct multiple intelligent agents, including a user agent intelligent agent, a knowledge retrieval intelligent agent, a rule verification intelligent agent, and a fusion decision-making intelligent agent;

[0013] The user agent receives query data input by the user, and classifies and identifies keywords in the query data to obtain the question type and the first keyword;

[0014] The knowledge retrieval agent performs vector retrieval from the vector database, keyword retrieval from the knowledge database, and knowledge graph retrieval from the graph database according to the question type and the first keyword, respectively, to obtain vector retrieval results, keyword retrieval results, and knowledge graph retrieval results. Then, the vector retrieval results, keyword retrieval results, and knowledge graph retrieval results are reordered by a reordering model to obtain a fusion result.

[0015] The verification agent filters the fusion results based on preset security rules, removes information from the fusion results that do not conform to the preset security rules, and obtains compliant fusion results;

[0016] The fusion decision-making agent constructs prompt words based on problem type, keywords, and compliance fusion results, inputs the prompt words into the large model, and outputs emergency response solutions.

[0017] Preferably, the accident-related document data includes standard document data, expert experience data, accident case data, literature research data, and emergency plan data; the preset model includes NLP model and deep learning model.

[0018] Preferably, step 2 includes the following sub-steps:

[0019] Step A1: Set multiple secondary keywords, identify the text content type using these keywords, and label it as the corresponding type;

[0020] Step A2: Based on the second keyword, a rule-based segmentation strategy is adopted to obtain multiple semantic blocks. The size of each semantic block is 500 to 2000 characters, and an overlap area of ​​100 to 200 characters is set between adjacent semantic blocks.

[0021] Step A3: Convert semantic blocks into semantic vectors using a text embedding model and store them in a vector database.

[0022] Preferably, step 3 includes the following sub-steps:

[0023] Step B1: Entity recognition is performed on the text content using an NLP model to obtain multiple entities, including accident name, equipment name, product name, emergency measures, and cause.

[0024] Step B2: Analyze the text content using a deep learning model, extract the relationships between multiple entities, and construct multiple triples based on the multiple entities and their relationships. The relationships between multiple entities include causal relationships, temporal relationships, and hierarchical relationships.

[0025] Step B3: Construct a knowledge graph based on multiple triples and store it in the graph database.

[0026] Preferably, the specific operation process of the knowledge retrieval intelligent agent includes the following sub-steps:

[0027] Step C1: Set the cosine similarity threshold, vectorize the question type and the first keyword using the text2vec-base-chinese model to obtain the query vector block, calculate the similarity between the query vector block and the semantic vector in the vector database according to the cosine similarity calculation formula, and obtain the semantic vector with a cosine similarity greater than the cosine similarity threshold as the vector retrieval result;

[0028] Step C2: Based on the first keyword, perform keyword retrieval on the text content in the knowledge database and obtain the text content containing the first keyword as the keyword retrieval result;

[0029] Step C3: Use the Cypher query language to search for entities associated with the first keyword in the knowledge graph database, and obtain the entity information associated with the first keyword as the knowledge graph retrieval result;

[0030] Step C4: Input the vector retrieval results, keyword retrieval results, and knowledge graph retrieval results into the re-ranking model to obtain the fusion results.

[0031] Preferably, the preset safety rules include industry safety standard rules, industry production operation rules, and industry production management rules.

[0032] Preferably, the plurality of intelligent agents further includes a management intelligent agent; after the fusion decision-making intelligent agent outputs an emergency response plan, the following steps are also included:

[0033] Step 5: The inspection agent scans the emergency response plan to check if there is any content that does not comply with the preset security rules. If the emergency response plan contains content that does not comply with the preset security rules, the corresponding content is marked and sent to the fusion decision agent. The fusion decision agent regenerates the emergency response plan based on the marked content. If the emergency response plan does not contain any content that does not comply with the preset security rules, proceed to the next step.

[0034] Step 6: The fusion decision-making agent performs logical consistency verification on the emergency response plan, and determines whether there are any logical defects in the emergency response plan. If there are logical defects in the emergency response plan, the corresponding logical defects are recorded, and the fusion decision-making agent regenerates the emergency response plan based on the corresponding logical defects; if there are no logical defects in the emergency response plan, proceed to the next step.

[0035] Step 7: The management agent scores the emergency response plan from multiple dimensions, obtaining the score value for each dimension. Then, a weighted summation method is used to calculate the final score value. When the final score value is greater than the preset score value, the final emergency response plan is output. When the final score value is less than the preset score value, modification suggestions are output and sent to the fusion decision-making agent. The fusion decision-making agent regenerates the emergency response plan based on the modification suggestions. The multiple dimensions include: coverage of standard basis, completeness of handling steps, emergency operation annotation rate, diversity of knowledge sources, and semantic relevance to reference materials.

[0036] Preferably, the method further includes step 8: receiving the user's satisfaction with the emergency response plan and the errors pointed out; updating the knowledge database, vector database, and graph database based on the pointed-out errors; and adjusting the weight of the corresponding first keyword in keyword retrieval based on the satisfaction level, and / or adjusting the value of the cosine similarity threshold.

[0037] It should be noted that:

[0038] NLP Models: NLP models refer to algorithms or systems used for Natural Language Processing (NLP) tasks. Their core goal is to enable computers to understand, interpret, generate, and process human language.

[0039] The text2vec-base-chinese model is a sentence embedding model optimized for Chinese scenarios. It can convert Chinese text of arbitrary length into fixed-dimensional (768-dimensional) semantic vectors to measure the semantic similarity between texts.

[0040] Cypher query language: Cypher is a dedicated query language for graph databases. It functions similarly to SQL in relational databases and is used for pattern matching and data processing of nodes, relations, and attributes in graph databases.

[0041] Furthermore, a multi-agent cooperative emergency response system for zirconium-based MOF synthesis accidents is also provided to implement the aforementioned multi-agent cooperative emergency response method for zirconium-based MOF synthesis accidents, comprising the following modules:

[0042] Knowledge Acquisition Module: Used to collect accident-related file data, and uses parsing tools to parse the file data, extract the text content and store it in the knowledge database;

[0043] The knowledge vectorization module is used to convert text content into semantic vectors through a text embedding model and store them in a vector database.

[0044] Knowledge graph construction module: It is used to analyze text content through preset models, obtain various entities and the relationships between different entities, construct knowledge graphs based on entities and the relationships between different entities, and store them in the graph database;

[0045] Multi-agent collaboration module: used to construct multiple agents, including user agent agent, knowledge retrieval agent, rule verification agent, and fusion decision agent;

[0046] The user agent receives query data input by the user, and classifies and identifies keywords in the query data to obtain the question type and the first keyword;

[0047] The knowledge retrieval agent performs vector retrieval from the vector database, keyword retrieval from the knowledge database, and knowledge graph retrieval from the graph database according to the question type and the first keyword, respectively, to obtain vector retrieval results, keyword retrieval results, and knowledge graph retrieval results. Then, the vector retrieval results, keyword retrieval results, and knowledge graph retrieval results are reordered by a reordering model to obtain a fusion result.

[0048] The verification agent filters the fusion results based on preset security rules, removes information from the fusion results that do not conform to the preset security rules, and obtains compliant fusion results;

[0049] The fusion decision-making agent constructs prompt words based on problem type, keywords, and compliance fusion results, inputs the prompt words into the large model, and outputs emergency response solutions.

[0050] One of the above-mentioned technical solutions in this application has at least one of the following advantages or beneficial effects:

[0051] The emergency response method of this application constructs a knowledge database, a vector database, and a graph database by collecting and processing various types of file data. Then, a hybrid retrieval method is used to obtain information related to the query data from the knowledge database, vector database, and graph database respectively. In this way, comprehensive information related to the user can be obtained. By re-sorting, the closest relevance can be quickly analyzed. Finally, the solution is verified by a verification agent, thereby improving the accuracy and efficiency of the solution. Attached Figure Description

[0052] The present application will be further described below with reference to the accompanying drawings and embodiments;

[0053] Figure 1 This is a flowchart of the multi-agent collaborative emergency response method in Example 1;

[0054] Figure 2 This is a block diagram of the multi-agent collaborative emergency response system in Example 2. Detailed Implementation

[0055] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0056] The following disclosure provides many different implementation methods or examples for different schemes of implementing this application.

[0057] Example 1

[0058] refer to Figure 1 A multi-agent collaborative emergency response method for zirconium-based MOF synthesis accidents includes the following steps:

[0059] Step 1: Collect accident-related document data, use parsing tools to parse the document data, extract the text content and store it in the knowledge database;

[0060] The data sources in this embodiment include various heterogeneous data such as publicly available standards and specifications, expert experience, accident cases, literature research, and emergency response plans. Document data primarily comes from local knowledge base directories. Standard document data includes publicly available industry and national standards (specifically in PDF format), expert experience data (such as expert summaries, rules of thumb, and guidelines, specifically in JSON or Word format), accident case data (such as corporate accident review reports and accident investigation materials, specifically in Word format), literature research data (such as research materials obtained from publicly available academic databases or locally collected literature, specifically in PDF format), and emergency response plan data (such as historically generated or accumulated emergency response plans, specifically in JSON format). Data parsing tools include PyMuPDF (fitz), the python-docx library, and a JSON parsing module. The PyMuPDF (fitz) library is used to parse PDF data, extracting content page by page while preserving layout information. The python-docx library is used to parse Word data, iterating through all paragraphs and extracting paragraph text. The json module of the Python standard library is used to parse JSON data to obtain text content information. The data obtained from parsing PDF, Word, and JSON formats are used as the text content.

[0061] Set up a directory structure for the knowledge database, including "Standard Documents", "Expert Experience", "Accident Cases", "Literature Research" and "Emergency Plans". Then, store the parsed text content in the corresponding directories for easy classification, management and retrieval.

[0062] It should be noted that:

[0063] PyMuPDF library: A library that can process PDF format information, find detailed information about each text character, matrix, and row, and also extract and visualize tables.

[0064] The python-docx library: python-docx is a Python library for creating, modifying, and parsing Microsoft Word documents. By using python-docx, you can easily create, edit, and save Word documents without installing Microsoft Word.

[0065] The json module in the Python standard library: The function (json.loads()) in the json module of the Python standard library is used to parse (decode) a JSON-formatted string into a Python object.

[0066] Step 2: Convert the text content into semantic vectors using a text embedding model and store them in a vector database;

[0067] Step 2 includes the following sub-steps:

[0068] Step A1: Set multiple secondary keywords, identify the text content type using these keywords, and label it as the corresponding type;

[0069] This method will preset multiple secondary keywords corresponding to the user's chemical production scenario, such as "leak", "explosion", "poisoning", etc., and perform type identification of text content based on the secondary keywords. For example, for an accident report about DMF leakage, if the text content contains "leak", the text content will be marked as "leak" type.

[0070] Step A2: Based on the second keyword, a rule-based segmentation strategy is adopted to obtain multiple semantic blocks. The size of each semantic block is 500 to 2000 characters, and an overlap area of ​​100 to 200 characters is set between adjacent semantic blocks.

[0071] In this embodiment, firstly, during the document segmentation stage, the system adopts a "accident keyword priority" segmentation strategy: the original text is first segmented based on the second keyword (such as "explosion, leakage, fire, poisoning, accident, case, combustion, corrosion, spill, evaporation, pollution, spontaneous combustion, runaway reaction"). If the second keyword is not matched, the entire text is treated as a single block. For long text blocks obtained after segmentation by the second keyword, a secondary segmentation is triggered when its length exceeds a threshold. Specifically, in the keyword segmentation function, when the block length exceeds max_chunk_len=2000 characters, it continues to be split according to the character length. In the main process, when the block length exceeds 2×chunk_size, a recursive character segmenter is used for refined segmentation. The current configuration parameters are chunk_size=500 and chunk_overlap=100, and segmentation is prioritized based on delimiters such as line breaks, Chinese periods, exclamation marks, question marks, English punctuation, and spaces, thereby controlling the block size while preserving contextual continuity.

[0072] Step A3: Convert semantic blocks into semantic vectors using a text embedding model and store them in a vector database.

[0073] The text embedding model is the text2vec-base-chinese model, which converts semantic blocks into fixed-length semantic vectors. The specific process is as follows: first, the semantic blocks are segmented and encoded, and then input into the pre-trained model to obtain the last hidden state representation. Then, the hidden state is averaged and pooled based on the attention mask to obtain fixed-dimensional sentence vectors. Subsequently, multiple semantic vectors, along with the corresponding semantic block content and source identifier, are written into the vector database.

[0074] The determination of the semantic block detail level adopts a rule mechanism: semantic segments are located based on accident keywords (such as explosion, leakage, fire, etc.); if the length of the semantic segment exceeds the preset threshold (about twice the block length parameter), a secondary segmentation is triggered; if the length does not exceed the threshold, the original block is kept.

[0075] Therefore, the level of detail determination in this embodiment is mainly based on a combination rule of text length (number of words) and keyword hit results, rather than a separately trained "level of detail scoring model".

[0076] Step 3: Analyze the text content using a pre-set model to obtain various entities and the relationships between different entities. Based on the entities and the relationships between different entities, construct a knowledge graph and store it in the graph database.

[0077] The preset models include NLP models and deep learning models.

[0078] Step 3 includes the following sub-steps:

[0079] Step B1: Entity recognition is performed on the text content using an NLP model to obtain multiple entities, including accident name, equipment name, product name, emergency measures, and cause.

[0080] The system performs entity recognition on text content using an NLP model. Specifically, the system first loads the zh_core_web_sm Chinese NLP model to perform word segmentation and named entity recognition on the text, extracting candidate entities. Then, it determines the type of the candidate entities according to the preset entity type keyword mapping rules. Entity types include accident, chemical, disposal, standard, equipment, personnel, location, and time. For example, for an accident report, the NLP model identifies the accident name (e.g., hydrothermal reactor explosion), chemical name (e.g., terephthalic acid), equipment name (e.g., high-pressure reactor), emergency measures (e.g., opening a fume hood), and cause (e.g., excessive pressure).

[0081] Step B2: Analyze the text content using a deep learning model, extract the relationships between multiple entities, and construct multiple triples based on the multiple entities and their relationships. The relationships between multiple entities include causal relationships, temporal relationships, and hierarchical relationships.

[0082] In this embodiment, the deep learning model is the spaCy Chinese model, which performs named entity recognition on the text content, extracting multiple entities such as chemicals, accidents, handling, equipment, and personnel. Then, it combines predefined relational vocabulary rules to extract relationships between multiple entities. These predefined relational vocabulary rules include: 1. Identifying causal relationships between entities through matching words such as "cause," "lead," and "cause"; 2. Identifying temporal relationships between entities by extracting co-occurrence relationships between time markers and event entities using regular expressions; 3. Identifying hierarchical relationships between entities by associating the accident subject with its attributes such as time, location, cause, handling, and consequences using structured field extraction rules for different document types. Finally, based on the extracted entities and relationships, multiple triples are constructed with subject-verb-object as the basic unit, such as: <hydrothermal reactor explosion - cause - excessive pressure> and <hydrothermal reactor explosion - emergency measures - power cut-off>.

[0083] It should be noted that:

[0084] The specific process for extracting the co-occurrence relationship between time markers and event entities using regular expressions is as follows: First, define a set of regular expression patterns for the time field, including formats such as "20XX year X month X day", "the accident occurred on XXXX year XX month XX day", and "time: XXXX year XX month XX day"; second, perform matching on each pattern in the text to extract the matched time strings; then, combine the event entity (usually the standardized result of the document title or event name) to construct a structured relationship record; finally, organize the "event entity - time field" together with the extracted fields such as location, cause, process, handling, and consequences into a structured triple and write it into the knowledge graph.

[0085] The specific process for structured field extraction rules for different document types is as follows: The system first determines the document type based on the parent directory name of the document path. If the path cannot be determined, it degenerates into matching the title and body keywords. Document types include accident cases, standard documents, expert knowledge, literature research, and emergency plans. Then, the corresponding field extraction rules are called according to the document type: accident cases extract time, location, cause, process, handling, and consequences; standard documents extract standard number, publication date, implementation date, and scope of application; expert knowledge extracts expert name, professional field, core viewpoints, and recommended measures; literature research extracts author, publication year, research method, main findings, and conclusions; and emergency plans extract plan name, issuing unit, response level, emergency measures, and person in charge. The extraction results are then assembled into a "subject-predicate-object" structure, combined with a general relation ternary, and deduplicated. Finally, the results are uniformly written into a knowledge graph to form a consistent structured knowledge representation for multiple document types.

[0086] Step B3: Construct a knowledge graph based on multiple triples and store it in the graph database.

[0087] In this embodiment, the graph database is the Neo4j graph database.

[0088] For example, the "DMF leak" incident node is connected to the "DMF" node through the "involved chemicals" relationship, while the "DMF" node is connected to nodes such as "flammable" and "toxic" through the "hazard" relationship. This graph structure enables the system to quickly perform complex reasoning and discover deep connections between knowledge through simple multi-hop queries (such as "find all protective measures related to DMF leak").

[0089] Step 4: Construct multiple intelligent agents, including a user agent intelligent agent, a knowledge retrieval intelligent agent, a rule verification intelligent agent, a fusion decision-making intelligent agent, and a management intelligent agent;

[0090] Suppose a user enters: "My reactor experiences a sudden pressure surge during the synthesis of UiO-66, posing a risk of explosion. What should I do?"

[0091] The user agent receives query data input by the user, and classifies and identifies keywords in the query data to obtain the question type and the first keyword, specifically:

[0092] The `classify_question()` function is called to classify user input into question types using keyword matching: if the input contains words such as "emergency," "handling," "how to," "measures," or "contingency plan," it is classified as a "disposal" question; if it contains words such as "reason," "why," or "cause," it is classified as a "reason" question; if it contains words such as "case," "accident," or "analysis," it is classified as a "case" question; otherwise, it is classified as a "general" question. Simultaneously, the `extract_field()` function is called to extract structured fields from the query data based on predefined regular expression patterns, identifying key elements such as the time, location, cause, handling method, and consequences of the accident, thus converting the user's natural language input into question types and primary keywords.

[0093] For example, if the user agent receives a message such as "My reactor is experiencing a sudden pressure surge during the synthesis of UiO-66, posing a risk of explosion, what should I do?", and identifies "what should I do", then the query is classified as a disposal type. The first keywords identified include "reactor", "synthesis of UiO-66", "pressure", and "explosion".

[0094] It should be noted that:

[0095] The classify_question() function is a question classification function, whose core function is to automatically identify the type of input question;

[0096] The extract_field() function is an information extraction function whose core function is to automatically extract specified key fields from text / strings.

[0097] The knowledge retrieval agent performs vector retrieval from the vector database, keyword retrieval from the knowledge database, and knowledge graph retrieval from the graph database according to the question type and the first keyword, respectively, to obtain vector retrieval results, keyword retrieval results, and knowledge graph retrieval results. Then, the vector retrieval results, keyword retrieval results, and knowledge graph retrieval results are reordered by a reordering model to obtain a fusion result.

[0098] Preferably, the specific operation process of the knowledge retrieval intelligent agent includes the following sub-steps:

[0099] Step C1: Set the cosine similarity threshold, vectorize the question type and the first keyword using the text2vec-base-chinese model to obtain the query vector block, calculate the similarity between the query vector block and the semantic vector in the vector database according to the cosine similarity calculation formula, and obtain the semantic vector with a cosine similarity greater than the cosine similarity threshold as the vector retrieval result;

[0100] In this embodiment, the text2vec-base-chinese model is used to semantically encode the question type and the first keyword. The question type and the first keyword are concatenated into query text and then input into the text2vec-base-chinese model. The text2vec-base-chinese model performs word segmentation and word vector calculation on the query text, and performs weighted mean pooling on all word vectors in the last hidden state according to the attention mask, compressing the variable-length text into a dense floating-point vector of fixed dimensions, thereby obtaining the query vector block.

[0101] Then, rank-bm25 is used to search the vector database based on the query vector block. The similarity between the query vector block and each semantic vector in the vector database is calculated using the cosine similarity calculation formula. The cosine similarity between the query vector block and each semantic vector is obtained. All semantic vectors with a cosine similarity greater than the cosine similarity threshold are extracted (in this embodiment, the cosine similarity threshold is 0.01). The semantic vectors with a cosine similarity greater than 0.01 are extracted and used as the vector retrieval result.

[0102] The formula for calculating cosine similarity is:

[0103]

[0104] Where q is the query vector block and d is the semantic vector in the vector database. When sim(q,d)>0.01, the semantic vector is retained as a vector retrieval result.

[0105] Step C2: Based on the first keyword, perform keyword retrieval on the text content in the knowledge database and obtain the text content containing the first keyword as the keyword retrieval result;

[0106] For example, if the first keyword is "reaction vessel", "synthesis of UiO-66", "pressure" and "explosion", then the relevant paragraphs containing the first keyword will be searched in the knowledge database and used as the keyword search results.

[0107] Step C3: Use the Cypher query language to search for entities associated with the first keyword in the knowledge graph database, and obtain the entity information associated with the first keyword as the knowledge graph retrieval result;

[0108] The specific process of using the Cypher query language to find entities associated with the first keyword in a knowledge graph database is as follows: First, the query data is processed by Chinese word segmentation, preferably using spaCy segmentation. If spaCy segmentation is unavailable, it degenerates to jieba segmentation, and then to regular expression segmentation based on punctuation and whitespace. Then, each segmentation result is used as a search term (the first keyword), and Cypher statements are executed to perform fuzzy matching on the Concept node. The matching condition is that the node name or node description contains the search term (the first keyword). The results of multiple keywords are merged and deduplicated by entity name to obtain the set of associated entities corresponding to the first keyword. Further relational queries are then performed to retrieve the rule nodes and document nodes associated with the entities. Structured relational queries (such as time relations and numerical relations) are added if necessary. Finally, the entity name, description, category, its association rules, and associated documents are summarized into knowledge graph retrieval results and returned. For example, based on "reactor" and "pressure," and according to the entities and their corresponding relationships, entity information such as the cause and measures for a sudden increase in reactor pressure is retrieved and used as knowledge graph retrieval results.

[0109] Step C4: Input the vector retrieval results, keyword retrieval results, and knowledge graph retrieval results into the re-ranking model to obtain the fusion results.

[0110] The re-ranking model is the bge-reranker re-ranking model. This model re-ranks vector results, keyword results, and knowledge graph results. Once loaded, the model is in inference mode (evaluation mode) and can automatically calculate relevance scores for any query-document pair, thereby performing the re-ranking.

[0111] Reordering criteria:

[0112] The re-ranking model uses the semantic relevance score between the query vector block and the search results (vector search results, keyword search results, or knowledge graph search results) as the ranking metric. Specifically, the query vector block and each candidate search result are input as a pair and fed into a cross-encoder model for joint encoding. The re-ranking model outputs a binary logistic regression score, which is then normalized using the Softmax function. The probability value of the positive class (positive relevance) is taken as the relevance score between the document and the query. A higher score indicates a stronger semantic relevance between the candidate search result and the query vector. Finally, all candidate search results are sorted from highest to lowest score, and the top 10 are selected as the final output.

[0113] The verification agent filters the fusion results based on preset security rules, removes information from the fusion results that do not conform to the preset security rules, and obtains compliant fusion results;

[0114] The preset safety rules include industry safety standard rules, industry production operation rules, and industry production management rules. More specifically, in this embodiment, the preset safety rules include: GB 15603—2022 "General Rules for Storage of Hazardous Chemicals in Warehouses", GB / T 29639—2020 "Guidelines for the Preparation of Emergency Response Plans for Production Safety Accidents in Production and Business Units", and GB13690—2009 "General Rules for Classification and Hazard Disclosure of Chemicals". Of course, corresponding safety rules can also be added according to the needs of the site.

[0115] For example, if the fusion results are "The reaction may be out of control, heating should be stopped immediately," "For solvothermal reactors, they should be transferred to a fume hood and cooled," and "The pressure is too high, open the reactor immediately to reduce the pressure," then the verification agent will analyze the content of the fusion results according to the preset safety rules. If the preset safety rules include "It is forbidden to open the reactor before the pressure is reduced," then the verification agent will remove "The pressure is too high, open the reactor immediately to reduce the pressure" and obtain compliant fusion results: "The reaction may be out of control, heating should be stopped immediately" and "For solvothermal reactors, they should be transferred to a fume hood and cooled."

[0116] The fusion decision-making agent constructs prompt words based on problem type, keywords, and compliance fusion results, inputs the prompt words into the large model, and outputs emergency response solutions.

[0117] The fusion decision-making agent guides a large model (in this embodiment, the large model is the DeepSeek large language model) to generate high-quality emergency response plans by designing prompt words. The specific design of the prompt words is as follows:

[0118] 1. Role Definition: The system message informs the model that "you are a senior chemical safety emergency expert".

[0119] 2. Context Injection: Organize all relevant information retrieved in the previous step (compliance fusion results) – such as the handling process in historical cases, specific clauses in security standards, and key operational points in expert experience – with clear headings and place them at the beginning of the user message.

[0120] 3. Specific task: After the context, clearly state the user's question, such as "Based on the above information and the on-site situation (sudden increase in pressure), please generate an emergency response plan."

[0121] 4. Output format constraints: It is explicitly required that the large model be output in JSON format, and the structure of JSON is specified in detail. For example, "Accident Analysis" must include subfields such as "Accident Type" and "Chemicals Involved", which ensures the standardization and parsability of the output.

[0122] 5. Strengthen safety constraints: At the end of the Prompt, reiterate the safety principles in an emphasized tone, such as "must comply with the Regulations on the Safety Management of Hazardous Chemicals", "prohibit the generation of recommendations that may lead to secondary accidents", and "prioritize personnel safety".

[0123] These carefully designed prompts effectively suppress the "illusion" of large models, guiding them to reason and generate within a controlled and professional framework, ultimately outputting a structured emergency response plan that is both safe and highly operable.

[0124] After generating the emergency response plan, a three-tier verification process is initiated to improve the security and reliability of the emergency response plan, as shown in steps 5 to 7:

[0125] Step 5: The inspection agent scans the emergency response plan to check if there is any content that does not comply with the preset security rules. If the emergency response plan contains content that does not comply with the preset security rules, the corresponding content is marked and sent to the fusion decision agent. The fusion decision agent regenerates the emergency response plan based on the marked content. If the emergency response plan does not contain any content that does not comply with the preset security rules, proceed to the next step.

[0126] This is the first layer of verification, rule verification. Specifically, the inspection agent performs a complete scan of the emergency response plan to check for any content that does not comply with the preset safety rules. For example, if the emergency response plan incorrectly suggests "using water to extinguish zirconium salt fires," while the preset safety rule includes "it is forbidden to use water to extinguish metal fires," then the inspection agent will mark "using water to extinguish zirconium salt fires" with the content: "Using water to extinguish zirconium salt fires" does not comply with the preset safety rule "it is forbidden to use water to extinguish metal fires." This marked content is then sent back to the fusion decision agent, which regenerates the emergency response plan based on the marked content, until the inspection agent no longer detects any content that does not comply with the preset safety rules.

[0127] Step 6: The fusion decision-making agent performs logical consistency verification on the emergency response plan, and determines whether there are any logical defects in the emergency response plan. If there are logical defects in the emergency response plan, the corresponding logical defects are recorded, and the fusion decision-making agent regenerates the emergency response plan based on the corresponding logical defects; if there are no logical defects in the emergency response plan, proceed to the next step.

[0128] This is the second layer of verification: logical consistency verification. Specifically, after the fusion decision-making agent receives a valid rule verification result, it calls the verifier to verify the internal logic of the emergency response plan. This verifies whether the emergency response plan has issues such as disordered order or operational contradictions. For example, checking the order of "emergency measures" in the emergency response plan might require ensuring that "personnel evacuation" precedes "site cleanup," or requiring both "enhanced ventilation" and "completely enclosed space," which constitutes an operational contradiction, as ventilation is impossible in a completely enclosed space. The verifier sends the type and content of the logical defects back to the fusion decision-making agent. The fusion decision-making agent then regenerates the emergency response plan based on the type and content of the logical defects until the verifier no longer detects any logical defects.

[0129] Step 7: The management agent scores the emergency response plan from multiple dimensions, obtaining the score value for each dimension. Then, a weighted summation method is used to calculate the final score value. When the final score value is greater than the preset score value, the final emergency response plan is output. When the final score value is less than the preset score value, modification suggestions are output and sent to the fusion decision-making agent. The fusion decision-making agent regenerates the emergency response plan based on the modification suggestions. The multiple dimensions include: coverage of standard basis, completeness of handling steps, emergency operation annotation rate, diversity of knowledge sources, and semantic relevance to reference materials.

[0130] This is the third layer of verification, a comprehensive scoring verification, specifically:

[0131] The management agent scores emergency response plans. The management and coordination agent comprehensively scores multiple verified candidate emergency response plans from the following five dimensions, with each dimension having a maximum score of 20 points, for a total score of 100 points. Only plans that score 80 points or higher are considered as the final output:

[0132] First dimension: Coverage of normative basis (20 points): The number and accuracy of national standards and industry normative clauses cited in the inspection plan. The more standard documents cited and the higher the degree of matching with the accident type, the higher the score.

[0133] The matching degree is specifically calculated using keyword matching rate. Let the set of keywords for accident types be... The standard clause keyword set is The matching degree formula is:

[0134]

[0135] Match score Where A is the total matching score, which is 10 points in this embodiment, and M is the matching degree. When the number of referenced standard documents is 1-2, the score is 4 points; when the number of referenced standard documents is 3-4, the score is 7 points; and when the number of referenced standard documents is ≥5, the score is 10 points. The matching degree score and the number of references score are calculated using a weighted summation method to obtain the standard coverage score, as follows:

[0136]

[0137] in, To standardize the scoring based on coverage values, Score based on the number of standard documents. To score the matching degree, The weighting of the score based on the number of standard documents. The weights for the matching score, + =1;

[0138] In this embodiment, = =0.5.

[0139] Second dimension: Completeness of handling steps (20 points): The assessment plan covers the complete process nodes of alarm assessment, personnel protection, leakage control, environmental protection and post-disposal. If a key stage is missing, points will be deducted accordingly. In this embodiment, if a key stage is missing, (4) points will be deducted. The deducted points will be subtracted from 20 to obtain the score for the completeness of handling steps.

[0140] Third dimension: Prohibited operation labeling rate (20 points): Check whether the plan clearly warns against erroneous operations that may cause secondary accidents (such as prohibiting water rinsing in the event of a zirconium oxychloride leak) and explains the reasons for the prohibition. The more comprehensive the labeling, the higher the score.

[0141] The score for whether a clear warning was given is 12 points, and the score for whether the reason for the contraindication was explained is 8 points. The score for the contraindication marking rate is calculated as follows: (Ratio of operations with clear warnings to all operations requiring clear warnings × corresponding score) + (Ratio of operations with explained reasons for contraindication to all operations requiring clear warnings × corresponding score). For example, if an emergency response plan requires 5 operations with clear warnings and 4 operations with clear warnings, and 3 of these operations explain the emergency reason, then the contraindication marking rate score is... .

[0142] Fourth Dimension: Diversity of Knowledge Sources (20 points): The number of information sources integrated into the emergency response plan, including knowledge graphs, standard document data, expert experience data, accident case data, literature research data, and emergency plan data. The richer and more diverse the sources, the higher the credibility of the plan and the higher the score.

[0143] In this embodiment, the number of information sources is 6. The score for knowledge source diversity is determined by the ratio of the number of information sources included in the emergency response plan to the total number of information sources. For example, if the emergency response plan includes 5 different information sources, then the score for knowledge source diversity is = .

[0144] Fifth Dimension: Semantic Relevance to References (20 points): The semantic similarity between the generated solution and the provided references is calculated. Cosine similarity is calculated after encoding using the text2vec-base-chinese model. Higher similarity indicates that the solution faithfully reflects the retrieved factual basis, resulting in a higher score. Semantic Relevance to References = Cosine Similarity × 20 (Score set for semantic relevance to references).

[0145] The five dimensions are weighted and summed to obtain the total score. When the total score is not less than 80 points, the management coordination agent returns the solution as the final emergency response solution to the user; otherwise, a regeneration process is triggered, and the solution is sent back to the fusion decision-making agent to regenerate the emergency response solution until the scoring requirements are met.

[0146] The formula for weighted summation is:

[0147]

[0148] in, This is the final score. To standardize the scoring based on coverage values, The completeness score for the handling procedures. The score is the percentage of prohibited operations labeled. As a score for the diversity of knowledge sources, The score represents the semantic relevance to the reference materials. , , , and These are the corresponding weights. + + + + =1.

[0149] In this embodiment, = = = = =0.2.

[0150] Step 8: Receive user satisfaction with the emergency response plan and the errors pointed out. Based on the errors pointed out, update the knowledge database, vector database, and graph database. Based on the satisfaction level, the knowledge retrieval agent adjusts the weight of the corresponding first keyword in keyword retrieval and / or adjusts the value of the cosine similarity threshold.

[0151] After receiving the emergency response plan, users can provide feedback, including their satisfaction level (1-10 points) and any errors they point out (e.g., users point out that the generated plan of "covering the DMF leak with sand" is not accurate enough, because for large DMF leaks, a dedicated absorbent should be used first).

[0152] Extract the keywords indicating errors. Taking the example of "The generated solution of 'covering DMF leaks with sand' is not accurate enough because for large DMF leaks, specialized adsorbents should be used first," extract the keywords "large DMF leaks," "sand," "inaccurate," and "adsorbent." Create a new entry in the knowledge database to store "For large DMF leaks, it is recommended to use specialized adsorbents instead of ordinary sand," and add it to the "Expert Experience" directory to save it in the knowledge database. Then, through steps 2 and 3, store this entry in the vector database and the graph database.

[0153] The handling of satisfaction levels involves adjusting the weight of the corresponding primary keyword in keyword retrieval and / or adjusting the cosine similarity threshold. For example, if the satisfaction level for queries related to "DMF leakage" remains consistently low, the retrieval strategy will be adjusted. This could involve increasing the weight of the keyword "DMF" in keyword retrieval or adjusting the recall of relevant documents during vector retrieval (by lowering the cosine similarity threshold, the number of retrieved semantic vector blocks will increase accordingly). This ensures that the newly added expert knowledge can be retrieved when a similar query is encountered again.

[0154] It should be noted that:

[0155] The text2vec-base-chinese text vectorization model (text2vec-base-chinese model) comes from the Hugging Face official tool library and is an officially trained model. In this embodiment, it is directly called.

[0156] The bge-reranker-base reranking model (bge-reranker reranking model) comes from the Hugging Face official tool library and is an officially trained model. In this embodiment, it is directly called.

[0157] The spaCy zh_core_web_sm natural language processing model (spaCy Chinese model) comes from the official spaCy tool library and is a pre-trained model. In this embodiment, it is directly called.

[0158] Cypher query language, derived from the graph database Neo4j, is a declarative graph query language, which is directly invoked in this embodiment.

[0159] NLP Models: spaCy Chinese Model + jieba Model

[0160] The jieba model is a pre-trained word segmenter model from the jieba project on GitHub.

[0161] rank-bm25 is a pre-trained algorithm from a GitHub project.

[0162] Example 2

[0163] refer to Figure 2 A multi-agent collaborative emergency response system for zirconium-based MOF synthesis accidents, used to implement the aforementioned multi-agent collaborative emergency response method for zirconium-based MOF synthesis accidents, includes the following modules:

[0164] Knowledge Acquisition Module: Used to collect accident-related file data, and uses parsing tools to parse the file data, extract the text content and store it in the knowledge database;

[0165] The knowledge vectorization module is used to convert text content into semantic vectors through a text embedding model and store them in a vector database.

[0166] Knowledge graph construction module: It is used to analyze text content through preset models, obtain various entities and the relationships between different entities, construct knowledge graphs based on entities and the relationships between different entities, and store them in the graph database;

[0167] Multi-agent collaboration module: used to construct multiple agents, including user agent agent, knowledge retrieval agent, rule verification agent, and fusion decision agent;

[0168] The user agent receives query data input by the user, and classifies and identifies keywords in the query data to obtain the question type and the first keyword;

[0169] The knowledge retrieval agent performs vector retrieval from the vector database, keyword retrieval from the knowledge database, and knowledge graph retrieval from the graph database according to the question type and the first keyword, respectively, to obtain vector retrieval results, keyword retrieval results, and knowledge graph retrieval results. Then, the vector retrieval results, keyword retrieval results, and knowledge graph retrieval results are reordered by a reordering model to obtain a fusion result.

[0170] The verification agent filters the fusion results based on preset security rules, removes information from the fusion results that do not conform to the preset security rules, and obtains compliant fusion results;

[0171] The fusion decision-making agent constructs prompt words based on problem type, keywords, and compliance fusion results, inputs the prompt words into the large model, and outputs emergency response solutions.

[0172] In this embodiment, the system adopts a layered architecture and is developed entirely using the Python language. At the data layer, the system integrates a knowledge database, a vector database, and a graph database. At the processing layer, the system integrates various parsing tools, such as the pdfplumber library for PDF parsing and the python-docx library for Word document parsing. At the intelligence layer, the system implements a multi-agent collaboration framework and a large model invocation module, which calls large models via APIs. At the application layer, the system builds a user-friendly web interface based on the Grado framework, supporting text input and file uploads, and providing functions such as session management and knowledge base statistics to facilitate user interaction with the system.

[0173] The specific operation process of the emergency response system is as follows: the knowledge acquisition module collects accident-related file data, uses parsing tools to parse the file data, extracts the text content and stores it in the knowledge database, and sends the text content to the knowledge vectorization module and the knowledge graph construction module;

[0174] After receiving the text content, the knowledge vectorization module converts the text content into semantic vectors through a text embedding model and stores them in the vector database;

[0175] After receiving text content, the knowledge graph construction module analyzes the text content through a preset model to obtain various entities and the relationships between different entities. Based on the entities and the relationships between different entities, it constructs a knowledge graph and stores it in the graph database.

[0176] The multi-agent collaboration module constructs multiple agents, including a user agent agent, a knowledge retrieval agent, a rule verification agent, and a fusion decision agent.

[0177] Users input query data through a web interface. The user agent receives the query data, classifies it, identifies keywords, obtains the question type and primary keyword, and then sends the question type and primary keyword to the knowledge retrieval agent.

[0178] The knowledge retrieval agent performs vector retrieval from the vector database, keyword retrieval from the knowledge database, and knowledge graph retrieval from the graph database based on the question type and the first keyword, respectively, and obtains vector retrieval results, keyword retrieval results, and knowledge graph retrieval results. Then, the vector retrieval results, keyword retrieval results, and knowledge graph retrieval results are reordered by the reordering model to obtain the fusion result, and then the fusion result is sent to the verification agent.

[0179] The verification agent filters the fusion results based on preset security rules, removing those that do not comply with the preset security rules, thus obtaining compliant fusion results, which are then sent to the fusion decision agent.

[0180] The fusion decision-making agent constructs prompt words based on problem type, keywords, and compliance fusion results. The prompt words are then input into the large model to output emergency response solutions.

[0181] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A multi-agent collaborative emergency response method for accidents during zirconium-based MOF synthesis, characterized in that, Includes the following steps: Step 1: Collect accident-related document data, use parsing tools to parse the document data, extract the text content and store it in the knowledge database; Step 2: Convert the text content into semantic vectors using a text embedding model and store them in a vector database; Step 3: Analyze the text content using a pre-set model to obtain various entities and the relationships between different entities. Based on the entities and the relationships between different entities, construct a knowledge graph and store it in the graph database. Step 4: Construct multiple intelligent agents, including a user agent intelligent agent, a knowledge retrieval intelligent agent, a rule verification intelligent agent, and a fusion decision-making intelligent agent; The user agent receives query data input by the user, and classifies and identifies keywords in the query data to obtain the question type and the first keyword; The knowledge retrieval agent performs vector retrieval from the vector database, keyword retrieval from the knowledge database, and knowledge graph retrieval from the graph database according to the question type and the first keyword, respectively, to obtain vector retrieval results, keyword retrieval results, and knowledge graph retrieval results. Then, the vector retrieval results, keyword retrieval results, and knowledge graph retrieval results are reordered by a reordering model to obtain a fusion result. The verification agent filters the fusion results based on preset security rules, removes information from the fusion results that do not conform to the preset security rules, and obtains compliant fusion results; The fusion decision-making agent constructs prompt words based on problem type, keywords, and compliance fusion results, inputs the prompt words into the large model, and outputs emergency response solutions.

2. The multi-agent cooperative emergency response method for zirconium-based MOF synthesis accidents according to claim 1, characterized in that, The accident-related document data includes standard document data, expert experience data, accident case data, literature research data, and emergency response plan data; the preset models include NLP models and deep learning models.

3. The multi-agent cooperative emergency response method for zirconium-based MOF synthesis accidents according to claim 2, characterized in that, Step 2 includes the following sub-steps: Step A1: Set multiple secondary keywords, identify the text content type using these keywords, and label it as the corresponding type; Step A2: Based on the second keyword, a rule-based segmentation strategy is adopted to obtain multiple semantic blocks. The size of each semantic block is 500 to 2000 characters, and an overlap area of ​​100 to 200 characters is set between adjacent semantic blocks. Step A3: Convert semantic blocks into semantic vectors using a text embedding model and store them in a vector database.

4. The multi-agent cooperative emergency response method for zirconium-based MOF synthesis accidents according to claim 2, characterized in that, Step 3 includes the following sub-steps: Step B1: Entity recognition is performed on the text content using an NLP model to obtain multiple entities, including accident name, equipment name, product name, emergency measures, and cause. Step B2: Analyze the text content using a deep learning model, extract the relationships between multiple entities, and construct multiple triples based on the multiple entities and their relationships. The relationships between multiple entities include causal relationships, temporal relationships, and hierarchical relationships. Step B3: Construct a knowledge graph based on multiple triples and store it in the graph database.

5. The multi-agent cooperative emergency response method for zirconium-based MOF synthesis accidents according to claim 2, characterized in that, The specific operation process of the knowledge retrieval intelligent agent includes the following sub-steps: Step C1: Set the cosine similarity threshold, vectorize the question type and the first keyword using the text2vec-base-chinese model to obtain the query vector block, calculate the similarity between the query vector block and the semantic vector in the vector database according to the cosine similarity calculation formula, and obtain the semantic vector with a cosine similarity greater than the cosine similarity threshold as the vector retrieval result; Step C2: Based on the first keyword, perform keyword retrieval on the text content in the knowledge database and obtain the text content containing the first keyword as the keyword retrieval result; Step C3: Use the Cypher query language to search for entities associated with the first keyword in the knowledge graph database, and obtain the entity information associated with the first keyword as the knowledge graph retrieval result; Step C4: Input the vector retrieval results, keyword retrieval results, and knowledge graph retrieval results into the re-ranking model to obtain the fusion results.

6. The multi-agent cooperative emergency response method for zirconium-based MOF synthesis accidents according to claim 1, characterized in that, The preset safety rules include industry safety standard rules, industry production operation rules, and industry production management rules.

7. The multi-agent cooperative emergency response method for zirconium-based MOF synthesis accidents according to claim 1, characterized in that, The multiple intelligent agents also include a management intelligent agent; after the fusion decision-making intelligent agent outputs an emergency response plan, the following steps are also included: Step 5: The inspection agent scans the emergency response plan to check if there is any content that does not comply with the preset security rules. If the emergency response plan contains content that does not comply with the preset security rules, the corresponding content is marked and sent to the fusion decision agent. The fusion decision agent regenerates the emergency response plan based on the marked content. If the emergency response plan does not contain any content that does not comply with the preset security rules, proceed to the next step. Step 6: The fusion decision-making agent performs logical consistency verification on the emergency response plan, and determines whether there are any logical defects in the emergency response plan. If there are logical defects in the emergency response plan, the corresponding logical defects are recorded, and the fusion decision-making agent regenerates the emergency response plan based on the corresponding logical defects; if there are no logical defects in the emergency response plan, proceed to the next step. Step 7: The management agent scores the emergency response plan from multiple dimensions, obtaining the score value for each dimension. Then, a weighted summation method is used to calculate the final score value. When the final score value is greater than the preset score value, the final emergency response plan is output. When the final score value is less than the preset score value, modification suggestions are output and sent to the fusion decision-making agent. The fusion decision-making agent regenerates the emergency response plan based on the modification suggestions. The multiple dimensions include: coverage of standard basis, completeness of handling steps, emergency operation annotation rate, diversity of knowledge sources, and semantic relevance to reference materials.

8. The multi-agent cooperative emergency response method for zirconium-based MOF synthesis accidents according to claim 5, characterized in that, It also includes step 8: receiving user satisfaction with the emergency response plan and the errors pointed out; updating the knowledge database, vector database, and graph database based on the pointed-out errors; and adjusting the weight of the corresponding first keyword in keyword retrieval based on the satisfaction level, and / or adjusting the value of the cosine similarity threshold.

9. A multi-agent collaborative emergency response system for zirconium-based MOF synthesis accidents, characterized in that, The multi-agent cooperative emergency response method for zirconium-based MOF synthesis accidents as described in any one of claims 1-8 includes the following modules: Knowledge Acquisition Module: Used to collect accident-related file data, and uses parsing tools to parse the file data, extract the text content and store it in the knowledge database; The knowledge vectorization module is used to convert text content into semantic vectors through a text embedding model and store them in a vector database. Knowledge graph construction module: It is used to analyze text content through preset models, obtain various entities and the relationships between different entities, construct knowledge graphs based on entities and the relationships between different entities, and store them in the graph database; Multi-agent collaboration module: used to construct multiple agents, including user agent agent, knowledge retrieval agent, rule verification agent, and fusion decision agent; The user agent receives query data input by the user, and classifies and identifies keywords in the query data to obtain the question type and the first keyword; The knowledge retrieval agent performs vector retrieval from the vector database, keyword retrieval from the knowledge database, and knowledge graph retrieval from the graph database according to the question type and the first keyword, respectively, to obtain vector retrieval results, keyword retrieval results, and knowledge graph retrieval results. Then, the vector retrieval results, keyword retrieval results, and knowledge graph retrieval results are reordered by a reordering model to obtain a fusion result. The verification agent filters the fusion results based on preset security rules, removes information from the fusion results that do not conform to the preset security rules, and obtains compliant fusion results; The fusion decision-making agent constructs prompt words based on problem type, keywords, and compliance fusion results, inputs the prompt words into the large model, and outputs emergency response solutions.

Images