Scientific research question and answer reasoning all-in-one machine based on multi-agent cooperation and control method

Abstract

The invention provides a scientific research question answering and reasoning all-in-one machine based on multi-agent cooperation and a control method, and relates to the technical field of artificial intelligence and multi-agent reasoning, and the all-in-one machine comprises a first agent which is used for carrying out the semantic understanding of a scientific research problem inputted by a user, and obtaining a semantic understanding result; performing semantic extension on the semantic understanding result according to a configured extension rule to obtain a first query request corresponding to the scientific research problem; the second intelligent agent is used for screening response information associated with the query request sent by the first intelligent agent from a pre-constructed scientific research knowledge base and sending the corresponding response information to the first intelligent agent; and the third agent is used for evaluating whether the target response information, the corresponding semantic understanding result, the first query request and the second query request meet the configured evaluation rule or not to obtain an evaluation result. The scientific research problem of the user can be accurately, comprehensively and reliably responded.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence and multi-agent reasoning technology, and more specifically, to a multi-agent collaborative scientific research question-answering and reasoning all-in-one machine and control method. Background Art

[0002] Existing question-answering platforms or large question-answering models have the following main problems: First, data sparsity and lack of professionalism: existing technologies usually rely on large-scale pre-training corpus, but the data coverage of professional fields (such as microbial natural products, drug development, biochemistry, etc.) is still insufficient, resulting in the model's limited understanding of professional terminology, complex mechanisms or marginal knowledge, resulting in limited professional reasoning ability.

[0003] Second, scientific question answering requires a deep understanding of specialized terminology and specific concepts. Current models often exhibit bias or even semantic misunderstandings when faced with rare or highly specialized expressions. For complex questions in the biochemistry field, a single round of search often fails to cover the full range of knowledge required. Professional problem solving typically requires multi-level, multi-dimensional knowledge association and reasoning.

[0004] Third, the information type is single and cross-modal fusion is insufficient: most existing question-answering systems mainly use text as input and output, and it is difficult to efficiently process various data forms such as structural data, experimental maps, spectral maps or gene sequences; for scientific research problems, especially when involving multi-level and multi-dimensional information, single-modal information is often not enough to provide comprehensive and accurate answers, and the system still needs to make breakthroughs in integrating multimodal data.

[0005] Fourth, the task process is not sufficiently split, and logical reasoning and causal relationship identification are insufficient: When faced with problems with multiple steps or hierarchical structures in the field of scientific research, existing technologies often find it difficult to effectively split complex tasks into manageable sub-tasks, resulting in an overall solution process that is not rigorous enough; scientific research problems usually require detailed reasoning and causal relationship analysis in details, and the lack of logical reasoning capabilities of existing systems may lead to errors in links or omissions of key steps in the solution process.

[0006] Fifth, there is a high risk of hallucinations and erroneous information, and insufficient real-time feedback and dynamic correction: Due to the lack of an effective self-assessment mechanism, existing technologies are prone to produce "hallucination" outputs that are inconsistent with actual biological or chemical facts in the absence of auxiliary professional tools. In addition, there is a lack of sufficient output monitoring mechanisms, and it is impossible to issue alarms in a timely manner when errors are discovered, and to continuously update and correct relevant information through feedback mechanisms, thereby reducing the risk of erroneous information being misused. Summary of the Invention

[0007] The purpose of the embodiments of the present application is to provide a scientific research question-answering and reasoning all-in-one machine and control method based on multi-agent collaboration, which solves the above-mentioned problems existing in the prior art, can understand the task of natural language and complete complex multimodal scientific research question reasoning and answering through dialogue.

[0008] In a first aspect, the present invention provides a scientific research question-answering and reasoning all-in-one machine based on multi-agent collaboration, the all-in-one machine comprising: The first agent is used to perform semantic understanding on the scientific research question input by the user and obtain a semantic understanding result; according to the configured expansion rule, the semantic understanding result is semantically expanded to obtain a first query request corresponding to the scientific research question, and the first query request is sent to the second agent; and, the first response information sent by the second agent in response to the first query request is received; if the richness of the first response information does not meet the configured richness threshold, a second query request is generated according to the first response information and the semantic understanding result, and the second query request is sent to the second agent; and, the second response information sent by the second agent in response to the second query request is received; if the lexical richness of the second response information meets the configured lexical richness threshold, the second response information is used as the target response information of the scientific research question; an evaluation request is generated, and the evaluation request is sent to a third agent; and, the evaluation result sent by the third agent in response to the evaluation request is received; wherein the evaluation request includes: the semantic understanding result, the first query request, the second query request and the target response information; if the evaluation result meets the configured evaluation rule, the target response information is output; The second agent is configured to filter response information associated with the query request sent by the first agent from a pre-built scientific research knowledge base, and send the corresponding response information to the first agent; The third intelligent agent is used to evaluate whether the target response information and the corresponding semantic understanding results, the first query request and the second query request meet the configured evaluation rules to obtain an evaluation result.

[0009] In an optional embodiment, the all-in-one machine further comprises: a knowledge base construction module and an interactive display device; The interactive display device is used for the user to input the scientific research question and display the target response information output by the first intelligent agent; The knowledge base construction module is specifically used to: Acquiring multi-source scientific research data, including: individual compounds, comprehensive properties of different compounds, and scientific research literature containing any compound; Building a first database based on the comprehensive characteristics of each compound and different compounds; generating a summary vector of each scientific research document based on the abstracts of the scientific research documents containing each compound; and constructing a second database based on the summary vectors of each scientific research document and each scientific research document; Extracting images associated with any compound and image descriptions of the corresponding images from various scientific research documents to construct a third database; Using each compound as a unique identifier, we associate the comprehensive characteristics of each compound, the abstract vector of each scientific research paper, and each image description to construct a three-dimensional knowledge graph. A scientific research knowledge base is generated based on the first database, the second database, the third database, the three-dimensional knowledge graph and a pre-configured tool library.

[0010] In an optional embodiment, the first agent is further configured to: Obtaining the form of the scientific research question input by the user; wherein the form of the scientific research question includes: text form, image form, and mixed text and image form; From the configuration table of different configuration forms and different preprocessing means, match the target preprocessing means corresponding to the form; The target preprocessing means is used to preprocess the scientific research problem to obtain the preprocessed scientific research problem.

[0011] In an optional embodiment, the first agent is specifically configured to: Performing semantic understanding on the pre-processed scientific research question to obtain a semantic understanding result; wherein the semantic understanding result includes: the entity to be queried, the keyword contained in the scientific research question, the first question intention corresponding to the scientific research question, and the semantic relationship between the entity to be queried and the question intention; Splitting the first question intention according to the semantic relationship to obtain multiple second question intentions; Generate multiple first query questions according to the multiple second question intentions and the entity to be queried; Filtering a list of association relationships associated with the entity to be queried and the keyword from the three-dimensional knowledge graph; Generate an extended list according to the association relationship list and the entity to be queried; Expanding the plurality of first query questions according to the expansion list to obtain a plurality of second query questions; According to the configured chain deformation rule, the plurality of first query questions and the plurality of second query questions are deformed and adjusted to obtain a plurality of third query questions; Filtering the plurality of first query questions, the plurality of second query questions, and the plurality of third query questions according to the semantic understanding result to obtain at least one target query question; Based on the obtained target query question, a first query request corresponding to the scientific research question is generated.

[0012] In an optional embodiment, the first agent is further configured to: Counting the number of words contained in the received response information; The richness of the response information is determined based on the number of words counted.

[0013] In an optional embodiment, the second agent is specifically configured to: Filtering, from the three-dimensional knowledge graph, summary vectors, compounds, comprehensive characteristics, and image descriptions associated with the query request sent by the first agent; Using the tool library to obtain associated data associated with the query request; Determine the scientific research document corresponding to the abstract vector as the target scientific research document; Searching the target scientific research literature for a text block associated with the query request; Calculate the cosine similarity between each text block; Determine the citation weight of each text block based on the citation relationship of the target scientific research literature corresponding to each text box; sorting the text blocks according to the cosine similarities between the text blocks and the corresponding citation weights; According to the configured first screening rule, a target text block is screened from the sorted text blocks; screening the compound, the comprehensive characteristic, the associated data, and the image according to the configured second screening rule to obtain a target compound, a target comprehensive characteristic, a target associated data, and a target image; Response information associated with the query request is generated based on the target text block, the target compound, the target comprehensive characteristic, the target associated data, and the target image.

[0014] In an optional embodiment, the third agent is further configured to: If the evaluation result does not satisfy the configured evaluation rule, generating a configured response deviation prompt; and sending the response deviation prompt to the first agent; The first intelligent agent is further configured to: In response to the received response deviation prompt, a new query request is generated based on the response deviation prompt and the corresponding query request and response information.

[0015] In a second aspect, the present invention provides a scientific research question-answering reasoning control method based on multi-agent collaboration, the device comprising: Obtain scientific research questions input by users; Controlling the first agent to perform semantic understanding on the scientific research question to obtain a semantic understanding result; performing semantic expansion on the semantic understanding result according to the configured expansion rules to obtain a first query request corresponding to the scientific research question, and sending the first query request to the second agent; Controlling the second agent to select first response information associated with the first query request from a pre-built scientific research knowledge base, and sending the first response information to the first agent; Controlling the first agent to receive the first response information; if the richness of the first response information does not meet a configured richness threshold, generating a second query request based on the first response information and the semantic understanding result, and sending the second query request to the second agent; Controlling the second agent to select second response information associated with the second query request from a pre-built scientific research knowledge base, and sending the second response information to the first agent; Controlling the first agent to receive the second response information; if the lexical richness of the second response information meets the configured lexical richness threshold, using the second response information as the target response information for the scientific research question; generating an evaluation request, and sending the evaluation request to the third agent; wherein the evaluation request includes: the semantic understanding result, the first query request, the second query request, and the target response information; Controlling a third agent to evaluate whether the target response information and the corresponding semantic understanding result, the first query request, and the second query request satisfy configured evaluation rules, and obtaining an evaluation result; Control the first agent to receive the evaluation result; if the evaluation result meets the configured evaluation rules, output the target response information.

[0016] In a third aspect, the present invention provides an electronic device, comprising a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus; Memory for storing computer programs; The processor is configured to implement any of the methods described in the foregoing embodiments when executing the program stored in the memory.

[0017] In a fourth aspect, the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method described in any one of the aforementioned embodiments is implemented.

[0018] This application establishes a professional literature vector library in a specific field and uses retrieval augmentation generation (RAG) technology. When answering scientific research questions, the model retrieves vector representations of literature in related fields to obtain accurate and traceable information support, thereby improving the professionalism and credibility of the answers.

[0019] This application ensures semantic consistency between question input and responses by building a multi-agent framework. The first agent is responsible for semantically parsing and professionally expressing the user's input question, while the second agent is responsible for generating answers based on retrieved domain knowledge. This multi-agent collaborative mechanism ensures semantic consistency from question input to generated response, effectively avoiding information bias and improving the professional quality of answers.

[0020] This application designed a biochemistry evaluation benchmark dataset covering five key research scenarios in the natural products field. To minimize manual intervention, the results were evaluated for plausibility using both experts and large-scale language models. Experimental results showed that this application significantly outperformed other tested models and, in some cases, demonstrated superior answer diversity, highlighting its potential as an intelligent assistant for natural product research.

[0021] This application creates a paradigm for building external databases, expands the application capabilities of large-scale language models in biology and chemistry, thereby improving data processing efficiency and scientific accuracy; proposes a multi-agent framework to promote interdisciplinary integration between biology and chemistry, and support collaborative cooperation between the two scientific fields, especially for application areas such as natural product research; establishes a sound set of evaluation criteria to evaluate the performance of the all-in-one machine, thereby verifying the effectiveness of the entire framework. The evaluation criteria of this application can be migrated to any relevant agent evaluation task. Through a multi-agent collaborative framework, this application gives the model the ability to self-evaluate, enabling it to determine whether the current knowledge reserve is sufficient to solve the problem, and trigger a more in-depth analysis process when necessary.

[0022] This application can adaptively allocate computing resources based on the complexity of the problem and conduct multiple rounds of progressive searches on key questions to ensure the comprehensiveness and accuracy of knowledge coverage. This application establishes a strict information traceability process to ensure that all output content can be traced back to the authoritative source, effectively reducing the risk of hallucinations and improving the credibility of the scientific research support system.

[0023] Based on the data specific to the field of biochemistry, this application constructs vector databases in various forms to ensure the accuracy of retrieval and provide multimodal output; each LLM undergoes different processing to become an agent that performs its own function; this application uses thinking chain technology to chain thinking and expand problems, which can better output the response information expected by users, and the response information is more comprehensive.

[0024] The first intelligent agent of this application uses the combination of thought chain technology and reinforcement learning to understand and reason about the scientific research questions input by the user; the second intelligent agent searches or calls the corresponding scientific research tools using code based on the results of the understanding and reasoning of the first intelligent agent, and returns the search results to the first intelligent agent to generate a new analysis report and return it to the third intelligent agent; the third intelligent agent is used to trace and evaluate the search results, so as to ensure accurate and comprehensive answers to the user's scientific research questions, and improve the factual reliability and practical application value of the answer content. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0026] Figure 1 An architectural diagram of a multi-agent collaborative scientific research question-answering and reasoning machine provided in an embodiment of the present application; Figure 2 A schematic diagram of the structure of a scientific research question-answering and reasoning all-in-one machine based on multi-agent collaboration provided in an embodiment of the present application; Figure 3 A flowchart of a scientific research question-answering reasoning control method based on multi-agent collaboration provided in an embodiment of the present application. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0028] like Figure 1 As shown, the scientific research question-answering and reasoning all-in-one machine based on multi-agent collaboration provided in the embodiment of the present application is applied to scientific research question-answering in professional fields (such as microbial natural products, drug development, biology, chemistry, etc.); Figure 2As shown, the scientific research question-answering and reasoning all-in-one machine based on multi-agent collaboration may include: an interactive display device, a first agent for question-answering, reading comprehension and reasoning, a second agent for retrieval and calling, a third agent for tracing and evaluation, a scientific research knowledge base and a knowledge base construction module; the first agent is respectively connected to the interactive display device, the second agent and the third agent; the second agent is also connected to the third agent and the scientific research knowledge base; the scientific research knowledge base is also connected to the knowledge base construction module; Specifically: An interactive display device, used for a user to input a scientific research question and to display target response information output by the first intelligent agent; A knowledge base construction module is used to construct a scientific research knowledge base including a first database, a second database, a third database, a three-dimensional knowledge graph, and a pre-configured tool library; A scientific research knowledge base, which is used to store corresponding scientific research data, three-dimensional knowledge graphs or various tool call interfaces through the first database, the second database, the third database, the three-dimensional knowledge graph and the tool library; The first intelligent agent is used to perform semantic understanding on the scientific research questions input by the user and obtain semantic understanding results; perform semantic expansion on the semantic understanding results according to the configured expansion rules to obtain a first query request corresponding to the scientific research questions, and send the first query request to the second intelligent agent; and receive the first response information sent by the second intelligent agent in response to the first query request; if the richness of the first response information does not meet the configured richness threshold, generate a second query request based on the first response information and the semantic understanding results, and send the second query request to the second intelligent agent; and receive the second response information sent by the second intelligent agent in response to the second query request; if the lexical richness of the second response information meets the configured lexical richness threshold, use the second response information as the target response information of the scientific research question; generate an evaluation request, and send the evaluation request to the third intelligent agent; and receive the evaluation result sent by the third intelligent agent in response to the evaluation request; if the evaluation result meets the configured evaluation rules, send the target response information to the display module; The second agent is configured to filter response information associated with the query request sent by the first agent from a pre-built scientific research knowledge base and send the corresponding response information to the first agent; The third agent is used to evaluate whether the target response information and the corresponding semantic understanding results, the first query request and the second query request meet the configured evaluation rules to obtain an evaluation result; The first, second, and third agents can be hosted in a processor or server, or in a terminal with strong computing power. The processor can be two NVIDIA A6000 48GB processors. The server can be a physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The terminal can be a user equipment (UE) such as a mobile phone, smartphone, laptop, digital broadcast receiver, personal digital assistant (PDA), tablet computer (PAD), handheld device, in-vehicle device, wearable device, computing device, or other processing device connected to a wireless modem, mobile station (MS), mobile terminal, etc. The terminal and server can be connected directly or indirectly via wired or wireless communication methods, which is not limited in this application.

[0029] In the embodiment of the present application, the knowledge base construction module is specifically used to: Acquire multi-source scientific research data; construct a first database based on the comprehensive characteristics of each compound and different compounds; generate a summary vector of each scientific research document based on the abstract of the scientific research document containing each compound; construct a second database based on the summary vector of each scientific research document and each scientific research document; extract images associated with any compound and image descriptions of the corresponding images from each scientific research document to construct a third database; use each compound as a unique identifier, associate the comprehensive characteristics of each compound, the summary vector of each scientific research document and each image description to construct a three-dimensional knowledge graph; generate a scientific research knowledge base based on the first database, the second database, the third database, the three-dimensional knowledge graph and the pre-configured tool library.

[0030] In the embodiments of the present application, the scientific research data include: comprehensive characteristics of each compound, different compounds, and scientific research literature containing any compound; specifically, the compounds include: microbial secondary metabolites and chemical compounds; chemical compounds include: type II polyketide compounds; the comprehensive characteristics of microbial secondary metabolites include: gene cluster data, chemical structure (SMILES), biological activity, gene sequence (FASTA) and functional domain prediction results; gene cluster data include: gene cluster ID, species origin, gene composition and functional annotation.

[0031] In the examples of the present application, the comprehensive properties of each compound and different compounds are derived from structured data in MIBiG (Microbial Secondary Metabolites Database), PubChem (Chemical Compound Database) and other professional databases, as well as literature mining (Web of Science) text data and expert annotation information.

[0032] In the embodiment of the present application, based on the comprehensive characteristics of each compound and different compounds, a first database is constructed, including: Normalize the multi-source scientific research data; generate the article table articles based on the processed scientific research documents; generate the product table products based on the compound name, chemical structure and biological activity of each processed type II polyketide compound; generate the gene cluster table BGCs based on the gene cluster ID, species origin and functional annotation of each processed compound; generate the gene table genes based on the gene sequence and functional domain prediction results of each processed compound; generate the document-product association table article_product through many-to-many relationship mapping between the synthesis / characterization association between the processed documents and each compound; establish the biosynthetic pathway mapping between gene clusters and their synthetic products to generate the gene cluster-product association table BGC_product; use the product table as the hub to associate chemical research (articles), gene function (genes) and synthetic biology data (BGCs) in scientific research documents; use the document table, product table, gene cluster table and gene table as entity tables; generate the first database based on the association data between the entity table and the gene cluster-product association table, the document-product association table and the product table.

[0033] In an embodiment of the present application, the literature table contains metadata and abstracts of various scientific research documents; each compound entity is stored in each entity table in the first database; and each entity table in the first database is provided with a foreign key constraint (such as product_id) to ensure the cross-reference consistency between chemical entities and biological entities; the present application establishes a "gene cluster-compound" biosynthesis relationship through the BGC_product table, linking microbiology and natural product chemistry data; and cross-links the synthesis method (chemistry) in the literature with the product activity data (pharmacology) through the article_product table, thereby realizing the integration of multidisciplinary scientific research data.

[0034] In an embodiment of the present application, the normalization processing includes: normalizing the names of each compound; structuring the gene cluster data; extracting the metadata of the scientific research literature through regular expressions, and storing the metadata and the experimental description content in the scientific research literature separately; dividing each scientific research literature into blocks to obtain multiple text blocks of each scientific research literature; the metadata includes: author, title, journal, year and digital object unique identifier DOI.

[0035] In the embodiments of the present application, since the compound names in PubChem (chemical compound database), InChIKey, and other professional databases may be different, it is necessary to normalize and map the names of the compounds to eliminate naming differences.

[0036] In the embodiment of the present application, the structured processing of gene cluster data includes: extracting the gene composition, functional annotation and other fields of BGC (biosynthetic gene cluster) from MIBiG, and establishing an association with the product.

[0037] In the embodiment of the present application, based on the abstract of the scientific research literature containing each compound, an abstract vector of each scientific research literature is generated, including: The domain-adapted embedding model embedding-3-small is used to vectorize the abstracts of each scientific research document to obtain the abstract vectors of each scientific research document.

[0038] In an embodiment of the present application, each scientific document is segmented, including: using pdfminer to extract the plain text of each scientific document and remove headers, footers, and / or watermark noise; adopting a hybrid segmentation strategy (fixed length + overlapping windows) to segment each scientific document to balance contextual coherence and retrieval granularity; and using the domain-adapted embedding model embedding-3-small to vectorize the segmented text blocks to preserve semantic features. This application improves the accuracy of fragment retrieval for long documents (such as reviews) by balancing the semantic integrity of the text (fixed segmentation) with the need for contextual overlap (sliding windows).

[0039] In the embodiments of the present application, the image description of each image is obtained by manually annotating the experimental flow charts and chemical structure diagrams in the scientific research literature, and storing the image URL and standardized description.

[0040] In the embodiment of the present application, the third database also includes: structural formula images of each compound generated in real time by tool calling (such as chemical drawing software) to supplement missing data.

[0041] In an embodiment of the present application, the tool library includes: functional annotation tools, PubChem interface tools and multiple sequence alignment tools, bioinformatics tools, scientific research mapping statistical tools and computational chemistry tools; among them, the functional annotation tool is used to align the gene sequence with the type II polyketide gene database, and screen the conserved domains (such as KR, ACP) by Bit-score threshold; the PubChem interface tool is used to retrieve physical and chemical property data according to the product name and supplement the chemical context; the multiple sequence alignment tool is used to identify the evolutionary relationship of gene clusters through the MAFFT algorithm to support the generation of synthetic biology modification suggestions; bioinformatics tools may include BLAST, DIAMOND, SPADES and HISAT2, etc., scientific research mapping statistical tools may include: R and python, etc.; computational chemistry tools may include: Gromacs and Gaussian.

[0042] In an embodiment of the present application, a three-dimensional knowledge graph is used to implement cross-references of relational data (structured attributes, i.e., compounds and comprehensive properties in the first database), vector data (semantics, i.e., summary vectors), and image data (visual).

[0043] In this embodiment of the present application, the first agent is a mentor; the second agent is an executor; and the third agent is an evaluator. T rounds of interaction occur between the mentor and the executor. In each round, the mentor submits a query request to the executor. The executor processes the query and retrieves relevant contextual information from various knowledge bases; this information may include text, images, references, and other supporting materials to provide assistance for the current query. Based on the user's original intent and the results returned by the executor in previous rounds, the mentor formulates the next query. After completing t rounds of interaction, the mentor summarizes all questions and answers generated during the interaction process to form a final response (i.e., answer information), which is returned to the user. Evaluation process: The evaluator interacts with the executor throughout the entire process of retrieving and generating answers (including before, during, and after) to improve the quality of the answers. Error reporting: If the Evaluator finds anomalies during the Executor's search process, these issues will be fed back to the Mentor.

[0044] In the embodiment of the present application, the first agent is further configured to: Obtain the form of the scientific research question input by the user; match the target preprocessing means corresponding to the form from the configuration table of different forms and different preprocessing means; use the target preprocessing means to preprocess the scientific research question to obtain the preprocessed scientific research question.

[0045] In the embodiments of the present application, the forms of scientific research questions include: text form, image form and mixed text and image form.

[0046] In the embodiment of the present application, the target preprocessing means corresponding to the text form include: using internal natural language processing (NLP) technology to perform word segmentation, syntactic analysis and semantic extraction on scientific research questions to obtain keywords, entities, contextual relationships and question intentions corresponding to the scientific research questions; based on keywords and semantic units, identify the fields and problem backgrounds that the scientific research questions may involve.

[0047] In an embodiment of the present application, the target preprocessing means corresponding to the image form include: performing visual feature extraction and local pattern recognition on the image, obtaining scene information, object labels, color information and other descriptive features from the image; the generated image labels and descriptions will be integrated into subsequent analysis in text form.

[0048] In the embodiment of the present application, the first agent is further configured to: Perform semantic understanding on the pre-processed scientific research questions to obtain semantic understanding results; split the first question intention according to the semantic relationship to obtain multiple second question intentions; generate multiple first query questions according to the multiple second question intentions and the entity to be queried; filter the association relationship list associated with the entity to be queried and the keyword from the three-dimensional knowledge graph; generate an extension list according to the association relationship list and the entity to be queried; expand the multiple first query questions according to the extension list to obtain multiple second query questions; deform and adjust the multiple first query questions and the multiple second query questions according to the configured chain deformation rules to obtain multiple third query questions; filter the multiple first query questions, the multiple second query questions and the multiple third query questions according to the semantic understanding results to obtain at least one target query question; generate a first query request corresponding to the scientific research question according to the obtained target query question.

[0049] In an embodiment of the present application, the semantic understanding result includes: the entity to be queried, keywords contained in the scientific research question, the first question intention corresponding to the scientific research question, and the semantic relationship between the entity to be queried and the question intention.

[0050] In the embodiment of the present application, the first question intention corresponding to the scientific research question includes the research direction involved in the scientific research question (such as biosynthesis, chemical modification and structural optimization, function, application and classification).

[0051] Based on the semantic relationship, the first question intention is split to obtain multiple second question intentions, including: For scientific research questions in text form, the first question intention is split into multiple fine-grained questions (i.e., the second question intention) according to logical or semantic relationships, such as "keyword definition", "principle exploration", "data relationship", etc. For scientific research questions in the form of images, semantic extraction is performed on the image to obtain the first question intention; based on the image semantic extraction results, follow-up questions about the image content, namely the second question intention, are generated; for example, "What is the relationship between the objects in the image?", "What is the relevance of this scene to the question?", etc.

[0052] For scientific research problems in the form of mixed images and text, the descriptions generated by the images are integrated with the text information, taking into account both visual cues and language details when understanding the problem as a whole, to ensure a more comprehensive information decomposition.

[0053] In an embodiment of the present application, after obtaining the user's question intention, a query request will not be directly generated based on the user's question intention. Instead, the question intention will be further expanded and deformed to obtain multiple questions to make the response information sufficiently rich and complete.

[0054] For example, let's assume the input research question is: "How to improve rice's disease resistance through gene editing technology?" First, we perform word segmentation and grammatical analysis on the question to extract key entities (such as "gene editing technology," "rice," and "disease resistance") and semantic relationships (such as "improve," "through"). Based on these semantic relationships, we obtain the semantic understanding result: the user's core intent is to understand how gene editing technology can be used on rice to enhance its disease resistance. According to the semantic understanding results, the first question intention is split into multiple second question intentions: Figure 1 :What is the mechanism of action of gene editing technology? Figure 2 :Which genes in rice are related to disease resistance? Figure 3 :How to apply gene editing technology to rice genetic modification; Combined with the entities to be queried ("gene editing technology", "rice", "disease resistance"), multiple first query questions are generated: Question 1: What is the working principle of gene editing technology; Question 2: What are the genes in rice related to disease resistance; Question 3: How to use gene editing technology to modify rice genes.

[0055] From the three-dimensional knowledge graph, a list of relationships associated with "gene editing technology", "rice", and "disease resistance" is screened: Relationship 1: gene editing technology → CRISPR-Cas9 system; Relationship 2: rice gene → OsNPR1 (related to disease resistance); Relationship 3: disease resistance → immune signaling pathway; Based on the association list and the entity to be queried, an extension list is generated: Extension 1: Specific operation process of the CRISPR-Cas9 system; Extension 2: Function and regulatory mechanism of the OsNPR1 gene; Extension 3: Role of immune signaling pathways in plant disease resistance; According to the expanded list, the first query question is expanded to obtain multiple second query questions: Question 4: How does the CRISPR-Cas9 system achieve gene editing? Question 5: The function of the OsNPR1 gene and its role in rice disease resistance? Question 6: How does the immune signaling pathway affect the disease resistance of plants?

[0056] According to the configured chain deformation rules, the first query question and the second query question are deformed and adjusted to obtain multiple third query questions: deformation 1: How to use the CRISPR-Cas9 system to edit the OsNPR1 gene in rice; deformation 2: Whether the mutation of the OsNPR1 gene will affect the disease resistance of rice; deformation 3: Whether the activation of the immune signaling pathway can enhance the disease resistance of rice.

[0057] Based on the semantic understanding results, the first query question, the second query question, and the third query question are filtered to finally obtain the target query question: Target Question 1: How to use the CRISPR-Cas9 system to edit the OsNPR1 gene in rice to enhance disease resistance; Target Question 2: The function of the OsNPR1 gene and its specific role in rice disease resistance.

[0058] Based on the target query question, generate the first query request corresponding to the scientific research question: Query request: Please provide relevant literature and experimental data on using the CRISPR-Cas9 system to edit the rice OsNPR1 gene to enhance disease resistance.

[0059] In an embodiment of the present application, the first agent is also used to: count the number of words contained in the received response information; determine the richness of the response information based on the counted number of words; when the richness of the response information meets the preset richness threshold, the first agent determines that the response information can be output; if not, the first agent needs to generate a new query request again based on the response information and the scientific research question to obtain a richer second response.

[0060] In another embodiment of the present application, a method for determining the richness of response information includes: Extract all key information points from the response; key information points include entities, attributes, and relationships; count the number of key information points to obtain a first scoring result; Classify the response information by type (such as definition, explanation, example, application, etc.); count the number of information of each type; score according to the diversity and number of information types to obtain a second scoring result; Count the vocabulary in the response information and determine the weight of each word based on its relevance to the query entity, keywords, and question intent; Calculate the third evaluation result based on the frequency, quantity, distribution in the sentence and weight of each word; The first scoring result, the second scoring result, and the third scoring result are weighted and summed to obtain the richness of the response information.

[0061] Specifically, the weights of the first scoring result, the second scoring result, and the third scoring result may be obtained by manual configuration.

[0062] In an embodiment of the present application, reinforcement learning is used to pre-train the first agent so that the first agent can move from single-step reasoning to multi-step logic construction; positive feedback is given when the first agent successfully decomposes a biochemical problem and guides the second agent to complete the task; when the first agent decomposes a natural product research problem, it not only gives direct guidance but also records the complete thinking process; the first agent automatically searches for background knowledge related to natural products and organizes information fragments in a logical order.

[0063] In the embodiment of the present application, the generation of the second query request includes: Obtain the richness of the first answer information; based on the first answer information, extract the question intentions without answer information, i.e., unanswered question intentions, from the question intentions corresponding to the first query request and the scientific research question to obtain a third question intention; extract multiple entities from the first query request and the first answer information; extract a new association relationship list associated with the multiple entities; based on the first answer information, split the third question intention into multiple fourth question intentions; based on the configured expansion rules, expand the new association relationship list and the fourth question intention to obtain a fifth question intention; based on the third question intention and the extracted entities, generate multiple fourth query questions; based on the expansion list, expand the fourth query question to generate a fifth query question; based on the configured chain deformation rule, deform and adjust the fourth query question and the fifth query question to generate a sixth query question; based on the semantic understanding results, filter the generated fourth query question, fifth query question, and sixth query question to remove redundant, repeated, or irrelevant questions to obtain a new target query question; based on the new target query question, generate a second query request.

[0064] In the embodiment of the present application, the query questions are generated by the tutor using a hierarchical progressive reasoning method, that is, starting from the overall question and gradually refining it into specific sub-questions, and each level generates corresponding reasoning steps: Preliminary summary: Based on the preprocessing results, the core intention and context of the user's question are summarized. Intermediate reasoning link: For each subdivision, intermediate reasoning conclusions are generated and key information is extracted to form multiple intermediate reasoning units. Final synthesis: All intermediate results are integrated to form a logically coherent answer, which is verified and corrected according to historical conversation records.

[0065] In the embodiment of the present application, the tutor, i.e., the first intelligent agent, not only subdivides the current problem, but also generates new follow-up questions (GENERATE NEXT QUESTION) based on the existing analysis in order to further explore the details of the problem or confirm the accuracy of the information, thereby ensuring the continuity of the reasoning chain, continuous feedback and correction in interaction with the user, and the gradual improvement of knowledge logic.

[0066] In an embodiment of the present application, the generation of the first query request includes two steps: generating follow-up questions based on historical dialogues and reasoning status to promote in-depth discussion of the problem; and constructing detailed chain thinking prompts to break down complex problems into multiple sub-problems.

[0067] In an embodiment of the present application, the generation of the first response information includes four steps: generating follow-up questions based on historical dialogues and reasoning status to promote in-depth discussion of the problem; constructing detailed chain thinking prompts to break down complex problems into multiple sub-problems; providing feedback to the user on the current work progress in the intermediate stage, and requesting additional information to further refine and accurately meet the needs; integrating the outputs of each sub-problem to generate a final comprehensive scientific report or answer document.

[0068] In an embodiment of the present application, the first intelligent agent is also used to: respond to the received response deviation prompt, generate a new query request based on the response deviation prompt and the corresponding query request and response information; the specific steps are similar to the generation process of the second query request and will not be elaborated here.

[0069] In an embodiment of the present application, the third agent will evaluate the query request generated by the first agent and the response information of the second agent to evaluate whether the query request and the response information are related to the scientific research question input by the user. When there is no correlation, the third agent will generate a deviation prompt and send it to the first agent and the second agent to prompt the first agent to regenerate the query request so that the second agent can retrieve the response information based on the new query request again; at the same time, when the scientific research question input by the user is any question in the configured refusal to answer question library, the third agent is used to feedback the refusal to answer prompt corresponding to the corresponding question to the second agent and the first agent to refuse to answer the corresponding question; the refusal to answer question library is user-defined and contains multiple questions and refusal to answer prompts for corresponding questions.

[0070] In the embodiment of the present application, the second agent is specifically configured to: From the three-dimensional knowledge graph, screen the summary vectors, compounds, comprehensive characteristics and image descriptions associated with the query request sent by the first intelligent agent; use the tool library to obtain the associated data associated with the query request; determine the scientific research literature corresponding to the summary vector as the target scientific research literature; search for the text blocks associated with the query request from the target scientific research literature; calculate the cosine similarity between each text block; determine the citation weight corresponding to each text block according to the citation relationship of the target scientific research literature corresponding to each text box; sort each text block according to the cosine similarity and the corresponding citation weight between each text block; according to the configured first screening rule, screen the target text block from the sorted text blocks; according to the configured second screening rule, screen the compounds, comprehensive characteristics, associated data and images respectively to obtain the target compound, target comprehensive characteristics, target associated data and target image; based on the target text block, target compound, target comprehensive characteristics, target associated data and target image, generate response information associated with the query request.

[0071] In an embodiment of the present application, the second intelligent agent adopts hierarchical retrieval to quickly locate relevant articles through abstract vector matching of scientific research documents; searches for text blocks in the hit documents (i.e., target scientific research documents), and combines cosine similarity calculation with citation relationship weight optimization to sort the results; associates the DOI of the original text with each text block in the vector database (i.e., the second database) to achieve a two-way jump with the articles table in the relational database.

[0072] In an embodiment of the present application, screening the image description associated with the query request sent by the first agent includes: using the BM25 algorithm to match image title keywords (such as "polyketide synthase") to avoid the problem of insufficient sensitivity of traditional semantic embedding to professional terms; the retrieval results are embedded in the LLM response through the URL to achieve mixed output of pictures and text.

[0073] For example, when a user searches for "clorobiocin biosynthesis": The vector database (the second database) returns relevant literature fragments. The relationship database (the first database) extracts the corresponding product (clorobiocin). The BGC table is linked to obtain the synthetic gene cluster. The tool library calls ClustalW to align homologous gene sequences. The search process automatically connects multiple dimensions of chemical, genetic, and bioinformatics data.

[0074] In an embodiment of the present application, during the screening process, low-relevance text blocks (cosine similarity <0.7) are excluded at the second database level; the final ranking is generated by integrating the BM25 score, number of citations, and expert annotation weights, and multidisciplinary cross-evidence is returned first; a hybrid block strategy is used to balance the semantic integrity of the text (fixed block) and the context overlap requirements (sliding window) to improve the accuracy of fragment retrieval of long documents (such as reviews); hierarchical vector retrieval is adopted to reduce computational overhead through two-level screening of "abstract → full text" while retaining fine-grained association capabilities.

[0075] In the embodiment of the present application, the evaluation request includes: semantic understanding results, the first query request, the second query request and target response information; In an embodiment of the present application, the third agent is further configured to: generate a configured response deviation prompt if the evaluation result does not meet the configured evaluation rules; and send the response deviation prompt to the first agent.

[0076] In an embodiment of the present application, the evaluation rules include: quality rules, association rules and reference rules; among them, the quality rules are used to ensure the quality of the response information; the association rules are used to ensure that the response information is associated with the semantic understanding results and the corresponding query requests; the reference rules are used to ensure that the response information must have a reference source, thereby eliminating hallucination information.

[0077] In this embodiment of the present application, the evaluation of the third agent specifically includes: The configured quality rules are used to reject evaluation requests. If the query question in the evaluation request and the question intent in the semantic understanding fall into a pre-configured rejection list, or exceed the pre-configured answer range, a prompt indicating that the question is out of range is generated. The all-in-one machine or a human will pre-screen the corresponding question or question intent and decide whether to allow the answer to be given, so as to avoid the system providing inappropriate or risky answers to questions. Irrelevant information detection is performed on the evaluation request using configured association rules. The response information is semantically parsed and matched with the corresponding query question and question intent, filtering out unnecessary content. If all content is judged as irrelevant, a feedback mechanism is triggered, indicating that the response information is irrelevant to the corresponding question and generating a response deviation prompt. Tracking citations for evaluation requests using configured citation rules: Evaluating whether the citations contained in the response information conform to the predetermined citation format and logic to ensure the accuracy of the citation information in the response information and avoid improper use of citations (such as citing content that is not actually used or not citing used content); by checking whether each supporting information block correctly defines the citation category, and returning guidance prompts for problematic answers, requiring the executor to regenerate answers that meet the citation requirements based on the feedback.

[0078] In this embodiment of the present application, if the evaluation finds that the response information does not meet the evaluation rules, the evaluator will specifically point out the problem, such as failure to reject the answer, excessive irrelevant information, or incorrect citations. The feedback content will be conveyed to the tutor in the form of guidance prompts, clearly indicating which part of the content does not meet the expected standards. Based on the evaluation feedback, the tutor (the first agent) may need to adjust the original question or the current chained thinking prompt (CoT), making changes in the generation of the next search or question refinement. If the evaluator eliminates a large amount of irrelevant information, the tutor may regenerate more precise search keywords and query parameters, thereby improving the relevance of subsequent search results.

[0079] In one embodiment of the present application, when the third agent determines that the response information does not meet the evaluation rules, it may directly request the executor to initiate a new information retrieval, or issue a prompt to the user to request additional or clarify questions in order to form a more accurate context.

[0080] In an embodiment of the present application, the third intelligent agent, namely the evaluator, pre-builds a biochemistry professional evaluation environment and performs real-time evaluation of the output of the executor; the third intelligent agent evaluates the scientific accuracy of the natural product information based on the configured scoring criteria; adjusts the evaluation criteria based on external feedback to improve the ability to judge scientific accuracy; and optimizes the expression of the biochemistry content generated by the executor to ensure professionalism and comprehensibility.

[0081] like Figure 1As shown in the figure, in the scientific research question-answering reasoning machine based on multi-agent collaboration, there are t rounds of interaction between the tutor and the executor. In each round, the tutor submits a query request to the executor, which processes the query and retrieves relevant contextual information from various knowledge bases. This information may include text, images, references, and other supporting materials to provide assistance for the current query. Based on the user's original intention and the results returned by the executor in the previous rounds, the tutor will formulate the next query. After completing t rounds of interaction, the tutor summarizes the questions and answers generated during all interactions to form a final response and return it to the user. Evaluation process: The evaluator interacts with the executor throughout the entire process of retrieving and generating answers (including before, during, and after) to improve the quality of the answers. Error reporting: If the evaluator finds any anomalies during the executor's retrieval process, they will feedback these issues to the tutor.

[0082] In the embodiment of the present application, the first intelligent agent is the mentor: in order to solve complex problems, users may need to interact with large language models multiple times to point out misunderstandings in their answers or shortcomings in temporary answers until the model provides the expected answer, this system introduces a mentor as an intermediary between the user and the system executor. After the user submits the input, the mentor uses the chain-of-thought (CoT) technology to analyze and process the user query step by step with the executor to guide the executor to retrieve sufficient information, and then comprehensively generate a complete response and present it to the user (see Figure 1 In addition, the tutor plays an important role in providing feedback to the user, guiding them to refine their query or elaborate on their question after identifying anomalies found by the evaluator (see Figure 1 The tutor's function is similar to that of the brain, capable of completing tasks such as thinking, discovery, and error correction. Its dynamic capabilities are not limited to simple query processing, but can also generate insights beyond the scope of the original problem.

[0083] In this embodiment of the present application, the second agent is the executor: To address the problem of hallucinated information in domain-specific queries, the implementers of ChatT2 adopted a Retrieval-Augmented Generation (RAG) framework to retrieve relevant information from external databases and combine it with the contextual learning capabilities of large language models to generate traceable answers. Figure 1As shown in Figure 1, the actor is responsible for understanding the input delivered by the tutor and making corresponding action decisions based on historical memory. When responding to the tutor's query, the actor retrieves additional information based on the tutor's input and historical conversation records by querying a database or calling a tool. It then integrates this information from various sources into a structured text format (such as XML). This text serves as the contextual input for the large language model to generate the final answer. The final supporting context consists of four parts: <text>(text), <figure>(image) and <tool>(tool).

[0084] In the process of retrieving information from relational and vector databases, queries are converted into SQL statements and text embeddings. SQL statements are executed to retrieve table information related to the query, which is expressed as< / tool> < / figure> < / text> (sheet), s v v s <text>For image database retrieval, the query is matched with the image title by the BM2.5 score, and the image with the high score is identified and its URL is saved as the image information. <figure>Label rendering.

[0085] During the tool repository retrieval process, the executor parses the key query information into the input of pre-packaged execution scripts. The execution results of these scripts are regarded as tool repository information and encapsulated in <tool>The collaboration of these four different databases ensures that implementers have access to the maximum amount of traceable supporting materials, allowing them to effectively respond to queries from their supervisors. The detailed database construction method is described in the Methods section.

[0086] In this embodiment of the present application, the third agent is the evaluator: While the executor takes various measures to ensure that ChatT2 can provide accurate answers to domain-specific questions, another key goal is to ensure that ChatT2 can refuse to answer content that is beyond its capabilities, thereby improving the credibility of the results, which is particularly important for knowledge-intensive tasks such as question answering. To achieve this goal, this application designs an additional intelligent agent, the evaluator. The evaluator interacts with the executor and performs three key tasks, namely, refusing to answer, filtering irrelevant information, and tracking citations, in the following three stages: (1) the pre-processing stage when the user query is submitted; (2) the intermediate processing stage when the executor retrieves context information; and (3) the post-processing stage when the executor generates the final context. This process ensures the credibility of the final results and prevents the inclusion of erroneous or fictitious information. For queries that ChatT2 cannot answer or the executor fails to retrieve relevant information, the evaluator provides feedback by refusing to answer; for answerable queries, the evaluator ensures that the executor includes the correct citation source in the final answer, thereby improving the factual reliability and practical application value of the answer, which is particularly beneficial for researchers when analyzing the results.

[0087] The preferred embodiments of the present application are described below in conjunction with the drawings in the specification. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application and are not used to limit the present application. In addition, the embodiments and features in the embodiments of the present application can be combined with each other if there is no conflict.

[0088] Figure 3 This is a flow chart of a scientific research question-answering reasoning control method based on multi-agent collaboration provided in an embodiment of the present application. Figure 3 As shown, the method may include: Step S310: obtain the scientific research question input by the user; control the first intelligent agent to perform semantic understanding on the scientific research question to obtain a semantic understanding result; according to the configured extension rules, perform semantic expansion on the semantic understanding result to obtain a first query request corresponding to the scientific research question, and send the first query request to the second intelligent agent.

[0089] Step S320: Control the second agent to filter the first response information associated with the first query request from the pre-built scientific research knowledge base, and send the first response information to the first agent; control the first agent to receive the first response information.

[0090] Step S330: If the richness of the first response information does not meet the configured richness threshold, a second query request is generated based on the first response information and the semantic understanding result, and the second query request is sent to the second agent.

[0091] Step S340: Control the second agent to filter the second response information associated with the second query request from the pre-built scientific research knowledge base, and send the second response information to the first agent; control the first agent to receive the second response information.

[0092] Step S350: If the vocabulary richness of the second response information meets the configured vocabulary richness threshold, the second response information is used as the target response information for the scientific research question; an evaluation request is generated, and the evaluation request is sent to the third agent.

[0093] Step S360: Control the third agent to evaluate whether the target response information and the corresponding semantic understanding results, the first query request and the second query request meet the configured evaluation rules to obtain the evaluation results.

[0094] Step S370: Control the first agent to receive the evaluation result; if the evaluation result meets the configured evaluation rules, output the target response information.

[0095] In an embodiment of the present application, the evaluation request includes: a semantic understanding result, a first query request, a second query request, and target response information.

[0096] For example, user: "Please predict the possible antibacterial activity of compound Y"; 1. Mentor’s Thinking Chain: 1.1 Analyze the pharmacophore of compound Y; 1.2 Determine the similar structure database to be searched; 1.3 Design a framework for activity prediction methodology; 2. Executor performs: 2.1 Search for similar compounds according to the instructor's guidance; 2.2 Application of QSAR model for prediction; 2.3 Generate preliminary activity prediction report; 3. Evaluator's comments: 3.1 Check the rationality of the prediction (whether it conforms to the structure-activity relationship law); 3.2 Compare with data of known similar compounds; 3.3 Propose correction suggestions and give scores; 4. Feedback Loop: 4.1 Adjust the executor’s prediction strategy based on the score; 4.2 The instructor adjusts the direction of the thinking chain based on the feedback; 4.3 System integration of final prediction results.

[0097] In another embodiment provided in the present application, a computer-readable storage medium is also provided, which stores instructions. When the computer-readable storage medium is run on a computer, it enables the computer to execute the scientific research question-answering reasoning control method based on multi-agent collaboration described in any of the above embodiments.

[0098] In another embodiment provided in the present application, a computer program product containing instructions is also provided. When the computer is run on a computer, the computer executes the scientific research question-answering reasoning control method based on multi-agent collaboration described in any of the above embodiments.

[0099] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the embodiments of the present application may take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware. Furthermore, the embodiments of the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0100] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0101] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0102] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0103] Although preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they become aware of the basic creative concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present invention.

[0104] Obviously, those skilled in the art can make various changes and modifications to the embodiments of the present application without departing from the spirit and scope of the embodiments of the present application. Thus, if these modifications and variations of the embodiments of the present application fall within the scope of the claims and their equivalents, the embodiments of the present application are also intended to include these modifications and variations.< / tool> < / figure> < / text> Labels are presented; at the same time, relevant literature sources are collected to form a document set denoted as I. The text embedding part is used to perform hierarchical semantic similarity matching based on cosine similarity. First, a set of semantically similar document sets is determined and denoted as I. Then, the query embedding vector is used to match the semantic similarity of the document fragments obtained from I∪I, thereby extracting the main text of the paper related to the query. This information is expressed as

Claims

1. A scientific research question-answering and reasoning machine based on multi-agent collaboration, characterized by: The all-in-one machine includes: The first agent is used to perform semantic understanding on the scientific research question input by the user and obtain a semantic understanding result; according to the configured expansion rule, the semantic understanding result is semantically expanded to obtain a first query request corresponding to the scientific research question, and the first query request is sent to the second agent; and, the first response information sent by the second agent in response to the first query request is received; if the richness of the first response information does not meet the configured richness threshold, a second query request is generated according to the first response information and the semantic understanding result, and the second query request is sent to the second agent; and, the second response information sent by the second agent in response to the second query request is received; if the lexical richness of the second response information meets the configured lexical richness threshold, the second response information is used as the target response information of the scientific research question; an evaluation request is generated, and the evaluation request is sent to a third agent; and, the evaluation result sent by the third agent in response to the evaluation request is received; wherein the evaluation request includes: the semantic understanding result, the first query request, the second query request and the target response information; if the evaluation result meets the configured evaluation rule, the target response information is output; The second agent is configured to filter response information associated with the query request sent by the first agent from a pre-built scientific research knowledge base, and send the corresponding response information to the first agent; The third intelligent agent is used to evaluate whether the target response information and the corresponding semantic understanding results, the first query request and the second query request meet the configured evaluation rules to obtain an evaluation result.

2. The all-in-one machine according to claim 1, wherein: The all-in-one machine also includes: a knowledge base building module and an interactive display device; The interactive display device is used for the user to input the scientific research question and display the target response information output by the first intelligent agent; The knowledge base construction module is specifically used to: Acquiring multi-source scientific research data, including: individual compounds, comprehensive properties of different compounds, and scientific research literature containing any compound; Building a first database based on the comprehensive characteristics of each compound and different compounds; generating a summary vector of each scientific research document based on the abstracts of the scientific research documents containing each compound; and constructing a second database based on the summary vectors of each scientific research document and each scientific research document; Extracting images associated with any compound and image descriptions of the corresponding images from various scientific research documents to construct a third database; Using each compound as a unique identifier, we associate the comprehensive characteristics of each compound, the abstract vector of each scientific research paper, and each image description to construct a three-dimensional knowledge graph; A scientific research knowledge base is generated based on the first database, the second database, the third database, the three-dimensional knowledge graph and a pre-configured tool library.

3. The all-in-one machine according to claim 2, wherein: The first intelligent agent is further configured to: Obtaining the form of the scientific research question input by the user; wherein the form of the scientific research question includes: text form, image form, and mixed text and image form; From the configuration table of different configuration forms and different preprocessing means, match the target preprocessing means corresponding to the form; The target preprocessing means is used to preprocess the scientific research problem to obtain the preprocessed scientific research problem.

4. The all-in-one machine according to claim 3, wherein: The first intelligent agent is specifically used to: Performing semantic understanding on the pre-processed scientific research question to obtain a semantic understanding result; wherein the semantic understanding result includes: the entity to be queried, the keyword contained in the scientific research question, the first question intention corresponding to the scientific research question, and the semantic relationship between the entity to be queried and the question intention; Splitting the first question intention according to the semantic relationship to obtain multiple second question intentions; Generate multiple first query questions according to the multiple second question intentions and the entity to be queried; Filtering a list of association relationships associated with the entity to be queried and the keyword from the three-dimensional knowledge graph; Generate an extended list according to the association relationship list and the entity to be queried; Expanding the plurality of first query questions according to the expansion list to obtain a plurality of second query questions; According to the configured chain deformation rule, the plurality of first query questions and the plurality of second query questions are deformed and adjusted to obtain a plurality of third query questions; filtering the plurality of first query questions, the plurality of second query questions, and the plurality of third query questions according to the semantic understanding result to obtain at least one target query question; Based on the obtained target query question, a first query request corresponding to the scientific research question is generated.

5. The all-in-one machine according to claim 3, wherein: The first intelligent agent is further configured to: Counting the number of words contained in the received response information; The richness of the response information is determined based on the number of words counted.

6. The all-in-one machine according to claim 4, wherein: The second intelligent agent is specifically used to: Filtering, from the three-dimensional knowledge graph, summary vectors, compounds, comprehensive characteristics, and image descriptions associated with the query request sent by the first agent; Using the tool library to obtain associated data associated with the query request; Determine the scientific research document corresponding to the abstract vector as the target scientific research document; Searching the target scientific research literature for a text block associated with the query request; Calculate the cosine similarity between each text block; Determine the citation weight of each text block based on the citation relationship of the target scientific research literature corresponding to each text box; sorting the text blocks according to the cosine similarities between the text blocks and the corresponding citation weights; According to the configured first screening rule, a target text block is screened from the sorted text blocks; screening the compound, the comprehensive characteristic, the associated data, and the image according to the configured second screening rule to obtain a target compound, a target comprehensive characteristic, a target associated data, and a target image; Response information associated with the query request is generated based on the target text block, the target compound, the target comprehensive characteristic, the target associated data, and the target image.

7. The all-in-one machine according to claim 1, wherein: The third agent is further configured to: If the evaluation result does not satisfy the configured evaluation rule, generating a configured response deviation prompt; and sending the response deviation prompt to the first agent; The first intelligent agent is further configured to: In response to the received response deviation prompt, a new query request is generated based on the response deviation prompt and the corresponding query request and response information.

8. A scientific research question-answering reasoning control method based on multi-agent collaboration, characterized in that: The method comprises: Obtain scientific research questions input by users; Controlling the first agent to perform semantic understanding on the scientific research question to obtain a semantic understanding result; performing semantic expansion on the semantic understanding result according to the configured expansion rules to obtain a first query request corresponding to the scientific research question, and sending the first query request to the second agent; Controlling the second agent to select first response information associated with the first query request from a pre-built scientific research knowledge base, and sending the first response information to the first agent; Controlling the first agent to receive the first response information; if the richness of the first response information does not meet a configured richness threshold, generating a second query request based on the first response information and the semantic understanding result, and sending the second query request to the second agent; Controlling the second agent to select second response information associated with the second query request from a pre-built scientific research knowledge base, and sending the second response information to the first agent; Controlling the first agent to receive the second response information; if the lexical richness of the second response information meets the configured lexical richness threshold, using the second response information as the target response information for the scientific research question; generating an evaluation request, and sending the evaluation request to the third agent; wherein the evaluation request includes: the semantic understanding result, the first query request, the second query request, and the target response information; Controlling a third agent to evaluate whether the target response information and the corresponding semantic understanding result, the first query request, and the second query request satisfy configured evaluation rules, and obtaining an evaluation result; Control the first agent to receive the evaluation result; if the evaluation result meets the configured evaluation rules, output the target response information.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to claim 8 is implemented.

Images