Deep research report generation method and system based on large model, DeepResearch and MCP

Abstract

The invention provides a deep research report generation method and system based on a large model, DeepResearch and MCP in the technical field of data intelligent analysis. The method comprises the following steps: S1, obtaining a research report declaration demand; s2, querying a tool list through the MCP client; s3, semantic understanding and preliminary retrieval are carried out on the research and report application demand through the large model, and a preliminary retrieval result is obtained; s4, content abstracts are extracted from the preliminary retrieval results in an intensive reading mode, and the preliminary retrieval results and the content abstracts are sliced and stored in a vector database; s5, performing topic clustering to obtain a core topic, and splitting the core topic into depth analysis angles; s6, executing depth retrieval based on the depth analysis angle to obtain a depth retrieval result; and S7, retrieving the information blocks from the vector database based on the deep retrieval results, and performing content iteration based on the deep retrieval results and the information blocks to generate research reports. The method has the advantages that the efficiency and the reliability of research report generation and the system expansibility are greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data intelligent analysis, and particularly discloses a deep research report generation method and system based on a large model, DeepResearch and MCP. BACKGROUND

[0002] Under the impetus of the rapid development of artificial intelligence technology, large language models (LLM, often referred to as "large models") have gradually become one of the key technologies supporting multiple industries, especially research and analysis tasks. In this context, the automated generation of high-quality, deep-insight research reports (hereinafter referred to as "research reports") is considered an important direction with great application potential. However, existing technical solutions still face multiple architectural-level technical challenges and performance bottlenecks in achieving this goal.

[0003] Currently, typical research report generation processes mostly adopt a collaborative work mode of "manual guidance + LLM assistance", and the analysis process still highly depends on the manual intervention of researchers. Researchers need to perform multiple rounds of information retrieval, filtering, reading, analysis, and content integration operations, resulting in low overall efficiency and being easily restricted by the subjective experience and cognitive boundaries of the operators. Although there are currently several LLM-based automated research report generation tools, there are still obvious limitations in actual deployment. First, LLM itself has a "knowledge timeliness barrier", that is, its training data has a fixed time cutoff point, and it cannot obtain and integrate real-time updated information, resulting in lagging behind the latest dynamics. Second, the "hallucination" problem commonly existing in LLM is particularly prominent in research scenarios, and the model may generate seemingly reasonable but actually inaccurate data, conclusions or references, seriously affecting the reliability and professionalism of the research report.

[0004] From the system architecture level, early solutions mostly adopt a structure of serial connection of retrieval modules and generation modules, which has a significant "insufficient retrieval-generation collaboration" problem, that is, there is a lack of deep semantic integration between retrieval results and generated content, which easily leads to information disconnection. In addition, the long context processing capability is another key bottleneck. When the system needs to comprehensively process a large number of retrieval documents, it is often difficult to effectively integrate all the information due to exceeding the model context window limit, or faces high computing cost due to excessive Token consumption, which restricts the practicality and economy of the system. At the same time, existing systems generally lack flexible and scalable architecture design, and have not formed a unified tool access standard, making it difficult to conveniently integrate and schedule diversified external resources (such as professional databases, real-time search engines, analysis tools, etc.), resulting in the need for repeated investment in highly customized integration and development work for each application.

[0005] In recent years, DeepResearch has emerged as a new AI research paradigm, focusing on enabling LLMs to autonomously plan and execute multi-step research tasks through agent workflows, covering problem decomposition, intelligent retrieval, information refinement, and report synthesis. Meanwhile, the Model Context Protocol (MCP) aims to establish a standardized interface for AI systems, providing a modular and standardized open protocol for the integration of LLMs with external tools and data sources, significantly enhancing the system's expandability and practicality.

[0006] However, the existing technology has not achieved the efficient and deep integration between the autonomous research capabilities represented by DeepResearch and the standardized tool ecosystem provided by the MCP protocol. Specifically, the current large model-based deep research report generation system still faces the following five key technical problems: 1. Information fragmentation and insufficient integration capabilities: Lack of effective deep fusion and semantic association analysis capabilities for multi-source and multi-modal information, resulting in simple listing of information, insufficient logical coherence and depth of insight, and obvious "retrieval-generation" semantic fragmentation.

[0007] 2. Limited real-time performance and knowledge update lag: Limited by the timeliness of LLM training data, without effective real-time information retrieval mechanisms, it is difficult to incorporate the latest industry dynamics, resulting in research reports that cannot accurately reflect rapidly changing market conditions.

[0008] 3. Low efficiency and high cost for processing long texts: When processing multiple documents and long context content, all original information needs to be input into the model, resulting in significant Token consumption, rising inference costs, and potential memory and computational power bottlenecks.

[0009] 4. Insufficient system flexibility and expandability: Lack of unified tool access and scheduling specifications, making it difficult to dynamically and flexibly integrate and coordinate various external tools (such as professional databases, search engines, analysis interfaces, etc.), resulting in poor system expandability and heavy development and maintenance burden.

[0010] 5. Lack of information authenticity protection and illusion suppression mechanisms: Without effective verification and credibility filtering of information sources when incorporating external retrieval information, it may introduce inaccurate content, leading to factual errors or "illusions" in model generation, including fabricated data, incorrect conclusions, or invalid references, seriously affecting the professionalism and reliability of research reports.

[0011] Therefore, how to provide a large model-based, DeepResearch, and MCP deep research report generation method and system to improve the efficiency, reliability, and system expandability of research report generation has become a technical problem that needs to be solved. SUMMARY

[0012] The technical problem solved by the present application is to provide a deep research report generation method and system based on a large model, DeepResearch and MCP, to improve the efficiency, reliability and system scalability of research report generation.

[0013] In a first aspect, the present application provides a deep research report generation method based on a large model, DeepResearch and MCP, comprising the following steps: Step S1, the DeepResearch service is constructed as an MCP client, an Agent integrating a large model is constructed, and a research report submission requirement based on natural language input is obtained through a user interface; Step S2, the Agent queries the registered MCP server through the MCP client to discover AI tools supported by each MCP server, and constructs a tool list based on each AI tool; Step S3, the Agent performs deep semantic understanding on the research report submission requirement through a large model, splits the research report submission requirement into a plurality of keywords, calls corresponding AI tools from the tool list based on the keywords through the MCP client to perform a preliminary retrieval operation, and obtains a preliminary retrieval result; Step S4, the Agent inputs the preliminary retrieval result into a large model for close reading to extract a content abstract, cuts the preliminary retrieval result and the content abstract into information blocks, and stores the information blocks into a vector database after vectorization; Step S5, the Agent performs theme clustering analysis on the research report submission requirement and the preliminary retrieval result through a large model, obtains a core issue of the research report submission requirement, and splits the core issue into a plurality of deep analysis angles; Step S6, the Agent calls corresponding AI tools to perform a deep retrieval operation based on each deep analysis angle, obtains a deep retrieval result, and stores the deep retrieval result into a cache; Step S7, the Agent retrieves the most relevant information block from the vector database through RAG technology based on the deep retrieval result of each deep analysis angle, and iterates the content based on each deep retrieval result and information block to generate a research report.

[0014] Further, in step S2, the AI tools at least include a keyword extraction tool, a search tool, a theme splitting tool and a data analysis tool.

[0015] Further, the step S3 is specifically: The Agent splits the research report declaration demand into a plurality of keywords through deep semantic understanding of the research report declaration demand by a large model, sets an iterative retrieval plan based on each keyword, and calls corresponding AI tools from a tool list to perform a preliminary retrieval operation based on the iterative retrieval plan through the MCP client to obtain a preliminary retrieval result. The MCP client caches and preloads the AI tools based on the calling frequency and the calling success rate.

[0016] Further, the step S6 is specifically: The Agent calls corresponding AI tools to perform a deep retrieval operation based on each deep analysis angle, obtains a deep retrieval result, filters the deep retrieval result corresponding to each deep analysis angle based on a preset filtering rule, and stores each deep retrieval result in a cache for subsequent direct matching calling based on similarity to bypass the MCP client.

[0017] Further, the step S7 is specifically: The Agent retrieves the most relevant information block from the vector database through the RAG technology based on the deep retrieval result of the first deep analysis angle, inputs the deep retrieval result and the information block into a large model to obtain corresponding chapter content; The chapter content is used as a prompt word for the next deep analysis angle, and corresponding chapter content is obtained through retrieval and reasoning combined with the prompt word until all chapter content is generated; Each chapter content is structured and embedded into a preset HTML template to obtain a research report.

[0018] In a second aspect, the present application provides a deep research report generation system based on a large model, DeepResearch, and MCP, comprising the following modules: A research report declaration demand input module is used to build a DeepResearch service as an MCP client, integrate an Agent of a large model, and obtain a research report declaration demand based on natural language input through a user interface; An AI tool discovery module is used for the Agent to query a registered MCP server through an MCP client to discover AI tools supported by each MCP server, and build a tool list based on each AI tool; A preliminary retrieval module is used for the Agent to split a research report declaration demand into a plurality of keywords through deep semantic understanding of the research report declaration demand by a large model, set an iterative retrieval plan based on each keyword, and call corresponding AI tools from a tool list to perform a preliminary retrieval operation based on the iterative retrieval plan through the MCP client to obtain a preliminary retrieval result. an information block vectorization storage module, configured to input, by the agent, the preliminary search result into a large model for in-depth reading to extract a content abstract, cut the preliminary search result and the content abstract into information blocks, and store the information blocks in a vector database after vectorization; a core issue splitting module, configured to perform, by the agent, theme clustering analysis on the research report filing demand and the preliminary search result based on a large model to obtain a core issue of the research report filing demand, and split the core issue into a plurality of depth analysis angles; a depth search module, configured to perform, by the agent, depth search operations based on each depth analysis angle by calling corresponding AI tools to obtain depth search results, and store the depth search results in a cache; a research report generation module, configured to retrieve, by the agent, the most relevant information blocks from the vector database based on the depth search results of each depth analysis angle by using a RAG technology, and perform content iteration based on each depth search result and information block to generate a research report.

[0019] Further, in the AI tool discovery module, the AI tools at least include a keyword extraction tool, a search tool, a theme splitting tool and a data analysis tool.

[0020] Further, the preliminary search module is specifically configured to: perform, by the agent, depth semantic understanding on the research report filing demand based on a large model, split the research report filing demand into a plurality of keywords, set an iterative search plan based on each keyword, call corresponding AI tools from a tool list based on the MCP client to perform preliminary search operations based on the iterative search plan, and obtain preliminary search results; the MCP client caches and preloads the AI tools based on calling frequency and calling success rate.

[0021] Further, the depth search module is specifically configured to: perform, by the agent, depth search operations based on each depth analysis angle by calling corresponding AI tools to obtain depth search results, and store each depth search result in a cache after screening the depth search results corresponding to each depth analysis angle based on a preset screening rule, for subsequent direct matching calling based on similarity to bypass the MCP client.

[0022] Further, the research report generation module is specifically configured to: retrieve, by the agent, the most relevant information blocks from the vector database based on the depth search results of the first depth analysis angle by using a RAG technology, input the depth search results and the information blocks into a large model to obtain corresponding chapter contents; The chapter content is used as a prompt word for the next in-depth analysis angle. The corresponding chapter content is obtained by searching and reasoning in combination with the prompt words until all chapter content is generated. The content of each chapter is structured and embedded into a preset HTML template to obtain the research report.

[0023] The advantages of this invention are: 1. By building the DeepResearch service as an MCP client and integrating a large-scale model agent, the agent obtains research report submission requirements based on natural language input through the user interface. The agent queries the registered MCP servers through the MCP client to discover the AI ​​tools supported by each MCP server and constructs a tool list based on these AI tools. The agent performs deep semantic understanding of the research report submission requirements through the large-scale model, breaking them down into several keywords. Based on these keywords, the agent calls the corresponding AI tools from the tool list through the MCP client to perform preliminary search operations, obtaining preliminary search results. These preliminary search results are then input into the large-scale model for detailed reading to extract content summaries. The preliminary search results and content summaries are segmented into information blocks, which are then vectorized and stored in a vector database. The agent performs topic clustering analysis on the research report submission requirements and preliminary search results through the large-scale model to obtain the core issues of the research report submission requirements. These core issues are then broken down into multiple in-depth analysis angles, and based on each in-depth analysis angle, the corresponding AI tools are called. The I tool performs deep retrieval operations, obtains deep retrieval results, and stores these results in a cache. The Agent, based on the deep retrieval results from various analysis angles, retrieves the most relevant information blocks from a vector database using RAG technology. It then iterates the content based on these deep retrieval results and information blocks to generate a research report. This involves a standardized tool integration framework based on the MCP protocol, dynamically scheduling AI tools such as keyword extraction, search, and topic segmentation to achieve efficient retrieval and filtering of real-time information, effectively overcoming the knowledge lag inherent in large models. Utilizing RAG technology to vectorize and retrieve massive amounts of information, only the most relevant information blocks are input into the large model for content iteration, significantly reducing the computational cost and contextual pressure of long text processing. Simultaneously, through deep semantic understanding, topic clustering analysis, and a multi-round retrieval-generation closed loop, external verification information is deeply coupled with the model generation process, systematically suppressing illusions. Ultimately, while ensuring the logical coherence and depth of insight in the research report, the efficiency, reliability, and system scalability of the report generation are greatly improved.

[0024] 2. By building the DeepResearch service as an MCP client and setting up an agent to integrate large models, flexible integration of various AI tools can be achieved. This modular design based on the MCP protocol allows the system to dynamically discover and call AI tools (such as keyword extraction tools, search tools, etc.) supported by registered MCP servers, thereby improving the system's scalability and interoperability. Enterprises can easily add or update AI tools without modifying the core code, reducing maintenance costs and adapting to ever-changing research needs.

[0025] 3. The agent utilizes a large model to perform deep semantic understanding of the user's natural language input research report application request, and breaks it down into keywords and core issues. This approach ensures that the system can accurately capture user intent and perform topic clustering analysis, thereby decomposing complex requirements into multiple in-depth analysis perspectives. The advantage is that it improves the accuracy and relevance of research report generation, avoids the limitations of traditional keyword matching, reflects the progress of artificial intelligence in the field of natural language processing, and helps to generate research report content that is more in line with user needs.

[0026] 4. A phased retrieval strategy is adopted, first conducting a preliminary retrieval, then performing a deep retrieval based on core topics, and combining caching and preloading mechanisms; for example, the MCP client caches AI tools based on call frequency and success rate, and stores deep retrieval results in the cache for direct matching and invocation. This significantly reduces network latency and redundant calculations, improves retrieval efficiency and system response speed, and reduces resource consumption, which has obvious practical value for research scenarios that process large-scale data.

[0027] 5. By retrieving relevant information blocks from the vector database using RAG technology and combining them with a large model for content iteration, research reports based on the latest data can be generated. Finally, the chapter content is structured and embedded into a preset HTML template to generate a standardized research report. This method ensures the richness and relevance of the generated content, avoids the problem of outdated or disconnected information, and at the same time, the output of structured HTML documents improves the readability and professionalism of the research report, meeting the high standards required for commercial research reports.

[0028] 6. The research report submission requirements are obtained through the user interface using natural language input. Users do not need to have professional technical knowledge to operate it. This lowers the threshold for use and enables non-professional users (such as analysts or decision-makers) to quickly generate in-depth research reports, thus expanding the application scope of the technology.

[0029] 7. By building the DeepResearch service as an MCP client, it can dynamically discover, call, and manage all external tools that conform to the MCP protocol (such as keyword extraction tools, search tools, topic splitting tools, and data analysis tools). This solves the fragmentation problem of tool integration, realizes standardized integration of "develop once, call everywhere", and greatly enhances the flexibility and scalability of the system.

[0030] 8. By leveraging DeepResearch's MCP-based large-scale model capabilities, the initial complex problem (research report submission requirements) is broken down into multiple iterative steps to form research sub-plans (keywords). In-depth analysis angles (such as technology, market, competition, risk, and trends) are independently planned. Information is obtained by cyclically calling different AI toolsets through the MCP client, forming a loop of in-depth analysis to ensure the comprehensiveness and depth of the research report. Multi-angle analysis prepares for the next in-depth search.

[0031] 9. A research report iterative generation method is proposed, in which a part of the content is generated by decomposing the deep problem, and then the deep search results are combined as a prompt to pass to the next part. The output of the first part is combined with a specific prompt as the input of the second part. This generation process is matched with corresponding similarity searches, so that the output of each part is logical and progressive, avoiding duplicate or contradictory content generated by multi-path large model generation, and summarizing the output at the same time.

[0032] 10. By caching frequently called AI tools, a more innovative approach is to introduce success rate as a weight. This involves analyzing historical task data and predictively preloading or setting high-priority caches for combinations of frequently called AI tools with high success rates (e.g., querying housing prices in a city). When a new task request hits the cache, the result can be returned directly, completely bypassing large model inference and AI tool calls from the MCP client. This achieves "zero token" consumption and millisecond-level response in specific scenarios, greatly improving efficiency and economy.

[0033] 11. By calling AI tools through the MCP client, the tool itself can serve as a "branch application." Its structured output (such as charts and statistical data) works in tandem with the text generation capabilities of the large model. The large model is responsible for the framework narrative and insights, while the professional content generated by the branch tool is seamlessly embedded through the MCP protocol, ultimately synthesizing a high-quality research report with rich data visualization, breaking the traditional linear process of "retrieval-generation."

[0034] 12. Significantly Enhanced Research Report Quality and Depth: Through in-depth iterative problem decomposition, multi-angle analysis, and evidence-based generation, it produces professional-grade research reports with stronger insights, more rigorous structure, and more comprehensive content, far surpassing the effects of simple question-and-answer or summary generation. Significantly Reduced Token Consumption and Computational Costs: Innovative dynamic RAG filtering and compression only transmits high-value information, filtering out unapproved personal website information and avoiding the wasteful practice of sending entire long original documents into large models. Simultaneously, the application of the MCP caching mechanism reduces a significant amount of token consumption, directly lowering the cost of calling AI tools and the demand for computing resources. Flexibility and Ecosystem Compatibility: Based on the MCP standard protocol, DeepResearch can seamlessly integrate any new tool or data source that conforms to the protocol without modifying the core code, allowing the system to easily adapt to the research needs of different professional fields and exhibiting extremely high scalability. Effectively ensuring the timeliness and authenticity of information: Utilizing the real-time search and segmentation tools integrated into MCP, research reports can cover the latest information, breaking down the time barriers of model knowledge. Simultaneously, the pre-screening mechanism by humans and the RAG retrieval mechanism greatly suppress the illusion of large-scale model-generated reports, significantly enhancing the credibility and commercial value of the reports. Improving research automation efficiency and user experience: Freeing researchers from tedious information gathering, reading, and preliminary integration work, allowing them to focus more on higher-level strategic thinking and decision-making; the fully automated process can run 24 hours a day, greatly shortening the report production cycle and improving work efficiency. Attached Figure Description

[0035] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0036] Fig. 1 This is a flowchart of a deep research report generation method based on large models, DeepResearch, and MCP according to the present invention.

[0037] Fig. 2 This is a schematic diagram of the structure of a deep research report generation system based on large models, DeepResearch, and MCP according to the present invention.

[0038] Fig. 3 This is a flowchart illustrating the present invention. Detailed Implementation

[0039] The overall approach of the technical solution in this application is as follows: By integrating a standardized tool framework based on the MCP protocol, AI tools such as keyword extraction, search, and topic segmentation are dynamically scheduled to achieve efficient retrieval and filtering of real-time information, effectively overcoming the knowledge lag of large models; RAG technology is used to vectorize and store massive amounts of information for retrieval, and only the most relevant information blocks are input into the large model for iterative content generation, significantly reducing the computational cost and contextual pressure of long text processing; at the same time, through deep semantic understanding, topic clustering analysis, and multi-round retrieval-generation closed loop, external verification information is deeply coupled with the model generation process, systematically suppressing the illusion phenomenon, and ultimately improving the efficiency, reliability, and system scalability of research report generation while ensuring the logical coherence and insight depth of the research report.

[0040] Please refer to Figs. 1 to 3 As shown, a preferred embodiment of the present invention, a method for generating deep research reports based on large models, DeepResearch, and MCP, includes the following steps: Step S1: Build the DeepResearch service as an MCP client, integrate the agent of large models (such as Qwen), and obtain research report application requirements based on natural language input through the user interface (such as "analyze the real estate development of a first-tier city in a certain country in 2024" or input "generate an in-depth research report on the artificial intelligence industry in 2025"). MCP, or Model Context Protocol, is a standardized communication protocol used in the practical application of large models to connect them with external data sources, tools, and services. From a technical perspective, MCP defines a unified interface specification that is flexible, scalable, and allows AI models to access external resources such as databases, file systems, API interfaces, and real-time data streams in a standardized way, thereby expanding the model's context window and functional boundaries. The protocol adopts a JSON-RPC-style request-response mechanism, supports synchronous and asynchronous operations, and provides core functions such as resource discovery, permission management, and error handling. This enables AI models to dynamically acquire and process information beyond the scope of their training data, achieving more intelligent and real-time interactive capabilities.

[0041] Step S2: The Agent queries the registered MCP servers through the MCP client to discover the AI ​​tools supported by each MCP server, and builds a tool list based on each AI tool. Step S3: The Agent performs deep semantic understanding of the research report application requirements through a large model, breaks down the research report application requirements into several keywords, and calls the corresponding AI tools from the tool list through the MCP client to perform preliminary search operations based on the keywords, and obtains preliminary search results. Step S4: The Agent inputs the preliminary search results into the large model for detailed reading to extract content summaries, cuts the preliminary search results and content summaries into information blocks, and vectorizes the information blocks and stores them in the vector database. This involves extracting content summaries to remove redundancy, compressing the initial search results, and using only the core content with the highest information density and most relevant information (rather than the full text) as context. This significantly reduces token consumption and computational costs, and improves generation speed. Step S5: The Agent performs topic clustering analysis on the research report application requirements and preliminary search results using a large model to obtain the core issues of the research report application requirements, and breaks down the core issues into multiple in-depth analysis angles (such as: analysis of real estate policy factors, the impact of the economic environment on real estate, and population flow analysis). Step S6: The Agent calls the corresponding AI tool to perform a deep search operation based on each of the deep analysis angles, obtains the deep search results, and stores the deep search results in the cache; Step S7: Based on the deep search results from each deep analysis perspective, the Agent retrieves the most relevant information blocks from the vector database using RAG technology, and iterates the content based on each deep search result and information block to generate a research report.

[0042] The core innovation of this invention lies in the creative integration of DeepResearch's proxy workflow, MCP's standardized tool ecosystem, and advanced RAG filtering technology with the content moderation mechanism to form a synergistic and efficient system.

[0043] In step S2, the AI ​​tools include at least a keyword extraction tool, a search tool (such as Sogou search), a topic segmentation tool, and a data analysis tool.

[0044] Step S3 specifically involves: The agent uses a large model to perform deep semantic understanding of research report submission requirements, breaking them down into several keywords (i.e., key research sub-questions, such as "real estate policy," "economic environment," "population flow," and "infrastructure construction"). Based on each keyword, an iterative retrieval plan is set, and based on the iterative retrieval plan, the agent uses the MCP client to call the corresponding AI tools from the tool list to perform preliminary retrieval operations and obtain preliminary retrieval results; for example, calling the search engine MCP tool to perform a search for "{keyword}," or calling the Sogou search MCP tool to retrieve related articles. The MCP client caches and preloads AI tools based on call frequency and call success rate.

[0045] The next time a user asks a similar question, a score weight is calculated based on the similarity. For questions with higher weights, AI tools can be applied directly without bypassing the MCP client, thereby reducing the need for MCP client calls and token consumption.

[0046] Step S6 specifically involves: The Agent calls the corresponding AI tool to perform deep search operations based on each of the deep analysis angles, obtains deep search results, filters the deep search results corresponding to each of the deep analysis angles based on preset filtering rules, and stores each of the deep search results in the cache for subsequent direct matching calls based on similarity to bypass the MCP client.

[0047] In practice, the screening rules can use authoritative articles published by official sources as a reference. For content published on personal websites, the authenticity of the content needs to be verified and tags added. The authenticity is debatable, and the large model will take this into consideration during the generation process.

[0048] Step S7 specifically involves: Based on the deep search results from the first deep analysis perspective, the Agent retrieves the most relevant information blocks from the vector database using RAG technology, and inputs the deep search results and information blocks into the large model to obtain the corresponding chapter content. The chapter content is used as a prompt word for the next in-depth analysis angle. The corresponding chapter content is obtained by searching and reasoning in combination with the prompt words until all chapter content is generated. The content of each chapter is structured and embedded into a preset HTML template to obtain the research report.

[0049] A preferred embodiment of the deep research report generation system based on large models, DeepResearch, and MCP of the present invention includes the following modules: The research report application requirement input module is used to build the DeepResearch service into an MCP client, integrate the agent of large models (such as Qwen), and obtain research report application requirements based on natural language input through the user interface (such as "analyze the real estate development of a first-tier city in a certain country in 2024" or input "generate an in-depth research report on the artificial intelligence industry in 2025"). MCP, or Model Context Protocol, is a standardized communication protocol used in the practical application of large models to connect them with external data sources, tools, and services. From a technical perspective, MCP defines a unified interface specification that is flexible, scalable, and allows AI models to access external resources such as databases, file systems, API interfaces, and real-time data streams in a standardized way, thereby expanding the model's context window and functional boundaries. The protocol adopts a JSON-RPC-style request-response mechanism, supports synchronous and asynchronous operations, and provides core functions such as resource discovery, permission management, and error handling. This enables AI models to dynamically acquire and process information beyond the scope of their training data, achieving more intelligent and real-time interactive capabilities.

[0050] The AI ​​tool discovery module is used by the Agent to query the registered MCP servers through the MCP client to discover the AI ​​tools supported by each MCP server and build a tool list based on each AI tool. The preliminary retrieval module is used by the Agent to perform deep semantic understanding of the research report application requirements through a large model, break down the research report application requirements into several keywords, and call the corresponding AI tools from the tool list through the MCP client to perform preliminary retrieval operations based on the keywords, so as to obtain preliminary retrieval results. The information block vectorization storage module is used by the Agent to input the preliminary search results into a large model for detailed reading to extract content summaries, cut the preliminary search results and content summaries into information blocks, and store the information blocks in a vector database after vectorization. This involves extracting content summaries to remove redundancy, compressing the initial search results, and using only the core content with the highest information density and most relevant information (rather than the full text) as context. This significantly reduces token consumption and computational costs, and improves generation speed. The core topic decomposition module is used by the Agent to perform topic clustering analysis on the research report application requirements and preliminary search results through a large model, to obtain the core topics of the research report application requirements, and to decompose the core topics into multiple in-depth analysis angles (such as: analysis of real estate policy factors, the impact of the economic environment on real estate, and population flow analysis). The deep search module is used by the Agent to call the corresponding AI tool to perform deep search operations based on each of the deep analysis angles, obtain deep search results, and store the deep search results in the cache; The research report generation module is used by the Agent to retrieve the most relevant information blocks from the vector database based on the deep search results from various deep analysis perspectives, and to perform content iteration based on the deep search results and information blocks to generate a research report.

[0051] The core innovation of this invention lies in the creative integration of DeepResearch's proxy workflow, MCP's standardized tool ecosystem, and advanced RAG filtering technology with the content moderation mechanism to form a synergistic and efficient system.

[0052] The AI ​​tool discovery module includes at least a keyword extraction tool, a search tool (such as Sogou search), a topic segmentation tool, and a data analysis tool.

[0053] The preliminary retrieval module is specifically used for: The agent uses a large model to perform deep semantic understanding of research report submission requirements, breaking them down into several keywords (i.e., key research sub-questions, such as "real estate policy," "economic environment," "population flow," and "infrastructure construction"). Based on each keyword, an iterative retrieval plan is set, and based on the iterative retrieval plan, the agent uses the MCP client to call the corresponding AI tools from the tool list to perform preliminary retrieval operations and obtain preliminary retrieval results; for example, calling the search engine MCP tool to perform a search for "{keyword}," or calling the Sogou search MCP tool to retrieve related articles. The MCP client caches and preloads AI tools based on call frequency and call success rate.

[0054] The next time a user asks a similar question, a score weight is calculated based on the similarity. For questions with higher weights, AI tools can be applied directly without bypassing the MCP client, thereby reducing the need for MCP client calls and token consumption.

[0055] The deep search module is specifically used for: The Agent calls the corresponding AI tool to perform deep search operations based on each of the deep analysis angles, obtains deep search results, filters the deep search results corresponding to each of the deep analysis angles based on preset filtering rules, and stores each of the deep search results in the cache for subsequent direct matching calls based on similarity to bypass the MCP client.

[0056] In practice, the screening rules can use authoritative articles published by official sources as a reference. For content published on personal websites, the authenticity of the content needs to be verified and tags added. The authenticity is debatable, and the large model will take this into consideration during the generation process.

[0057] The research report generation module is specifically used for: Based on the deep search results from the first deep analysis perspective, the Agent retrieves the most relevant information blocks from the vector database using RAG technology, and inputs the deep search results and information blocks into the large model to obtain the corresponding chapter content. The chapter content is used as a prompt word for the next in-depth analysis angle. The corresponding chapter content is obtained by searching and reasoning in combination with the prompt words until all chapter content is generated. The content of each chapter is structured and embedded into a preset HTML template to obtain the research report.

[0058] In summary, the advantages of this invention are: 1. By building the DeepResearch service as an MCP client and integrating a large-scale model agent, the agent obtains research report submission requirements based on natural language input through the user interface. The agent queries the registered MCP servers through the MCP client to discover the AI ​​tools supported by each MCP server and constructs a tool list based on these AI tools. The agent performs deep semantic understanding of the research report submission requirements through the large-scale model, breaking them down into several keywords. Based on these keywords, the agent calls the corresponding AI tools from the tool list through the MCP client to perform preliminary search operations, obtaining preliminary search results. These preliminary search results are then input into the large-scale model for detailed reading to extract content summaries. The preliminary search results and content summaries are segmented into information blocks, which are then vectorized and stored in a vector database. The agent performs topic clustering analysis on the research report submission requirements and preliminary search results through the large-scale model to obtain the core issues of the research report submission requirements. These core issues are then broken down into multiple in-depth analysis angles, and based on each in-depth analysis angle, the corresponding AI tools are called. The I tool performs deep retrieval operations, obtains deep retrieval results, and stores these results in a cache. The Agent, based on the deep retrieval results from various analysis angles, retrieves the most relevant information blocks from a vector database using RAG technology. It then iterates the content based on these deep retrieval results and information blocks to generate a research report. This involves a standardized tool integration framework based on the MCP protocol, dynamically scheduling AI tools such as keyword extraction, search, and topic segmentation to achieve efficient retrieval and filtering of real-time information, effectively overcoming the knowledge lag inherent in large models. Utilizing RAG technology to vectorize and retrieve massive amounts of information, only the most relevant information blocks are input into the large model for content iteration, significantly reducing the computational cost and contextual pressure of long text processing. Simultaneously, through deep semantic understanding, topic clustering analysis, and a multi-round retrieval-generation closed loop, external verification information is deeply coupled with the model generation process, systematically suppressing illusions. Ultimately, while ensuring the logical coherence and depth of insight in the research report, the efficiency, reliability, and system scalability of the report generation are greatly improved.

[0059] 2. By building the DeepResearch service as an MCP client and setting up an agent to integrate large models, flexible integration of various AI tools can be achieved. This modular design based on the MCP protocol allows the system to dynamically discover and call AI tools (such as keyword extraction tools, search tools, etc.) supported by registered MCP servers, thereby improving the system's scalability and interoperability. Enterprises can easily add or update AI tools without modifying the core code, reducing maintenance costs and adapting to ever-changing research needs.

[0060] 3. The agent utilizes a large model to perform deep semantic understanding of the user's natural language input research report application request, and breaks it down into keywords and core issues. This approach ensures that the system can accurately capture user intent and perform topic clustering analysis, thereby decomposing complex requirements into multiple in-depth analysis perspectives. The advantage is that it improves the accuracy and relevance of research report generation, avoids the limitations of traditional keyword matching, reflects the progress of artificial intelligence in the field of natural language processing, and helps to generate research report content that is more in line with user needs.

[0061] 4. A phased retrieval strategy is adopted, first conducting a preliminary retrieval, then performing a deep retrieval based on core topics, and combining caching and preloading mechanisms; for example, the MCP client caches AI tools based on call frequency and success rate, and stores deep retrieval results in the cache for direct matching and invocation. This significantly reduces network latency and redundant calculations, improves retrieval efficiency and system response speed, and reduces resource consumption, which has obvious practical value for research scenarios that process large-scale data.

[0062] 5. By retrieving relevant information blocks from the vector database using RAG technology and combining them with a large model for content iteration, research reports based on the latest data can be generated. Finally, the chapter content is structured and embedded into a preset HTML template to generate a standardized research report. This method ensures the richness and relevance of the generated content, avoids the problem of outdated or disconnected information, and at the same time, the output of structured HTML documents improves the readability and professionalism of the research report, meeting the high standards required for commercial research reports.

[0063] 6. The research report submission requirements are obtained through the user interface using natural language input. Users do not need to have professional technical knowledge to operate it. This lowers the threshold for use and enables non-professional users (such as analysts or decision-makers) to quickly generate in-depth research reports, thus expanding the application scope of the technology.

[0064] 7. By building the DeepResearch service as an MCP client, it can dynamically discover, call, and manage all external tools that conform to the MCP protocol (such as keyword extraction tools, search tools, topic splitting tools, and data analysis tools). This solves the fragmentation problem of tool integration, realizes standardized integration of "develop once, call everywhere", and greatly enhances the flexibility and scalability of the system.

[0065] 8. By leveraging DeepResearch's MCP-based large-scale model capabilities, the initial complex problem (research report submission requirements) is broken down into multiple iterative steps to form research sub-plans (keywords). In-depth analysis angles (such as technology, market, competition, risk, and trends) are independently planned. Information is obtained by cyclically calling different AI toolsets through the MCP client, forming a loop of in-depth analysis to ensure the comprehensiveness and depth of the research report. Multi-angle analysis prepares for the next in-depth search.

[0066] 9. A research report iterative generation method is proposed, in which a part of the content is generated by decomposing the deep problem, and then the deep search results are combined as a prompt to pass to the next part. The output of the first part is combined with a specific prompt as the input of the second part. This generation process is matched with corresponding similarity searches, so that the output of each part is logical and progressive, avoiding duplicate or contradictory content generated by multi-path large model generation, and summarizing the output at the same time.

[0067] 10. By caching frequently called AI tools, a more innovative approach is to introduce success rate as a weight. This involves analyzing historical task data and predictively preloading or setting high-priority caches for combinations of frequently called AI tools with high success rates (e.g., querying housing prices in a city). When a new task request hits the cache, the result can be returned directly, completely bypassing large model inference and AI tool calls from the MCP client. This achieves "zero token" consumption and millisecond-level response in specific scenarios, greatly improving efficiency and economy.

[0068] 11. By calling AI tools through the MCP client, the tool itself can serve as a "branch application." Its structured output (such as charts and statistical data) works in tandem with the text generation capabilities of the large model. The large model is responsible for the framework narrative and insights, while the professional content generated by the branch tool is seamlessly embedded through the MCP protocol, ultimately synthesizing a high-quality research report with rich data visualization, breaking the traditional linear process of "retrieval-generation."

[0069] 12. Significantly Enhanced Research Report Quality and Depth: Through in-depth iterative problem decomposition, multi-angle analysis, and evidence-based generation, it produces professional-grade research reports with stronger insights, more rigorous structure, and more comprehensive content, far surpassing the effects of simple question-and-answer or summary generation. Significantly Reduced Token Consumption and Computational Costs: Innovative dynamic RAG filtering and compression only transmits high-value information, filtering out unapproved personal website information and avoiding the wasteful practice of sending entire long original documents into large models. Simultaneously, the application of the MCP caching mechanism reduces a significant amount of token consumption, directly lowering the cost of calling AI tools and the demand for computing resources. Flexibility and Ecosystem Compatibility: Based on the MCP standard protocol, DeepResearch can seamlessly integrate any new tool or data source that conforms to the protocol without modifying the core code, allowing the system to easily adapt to the research needs of different professional fields and exhibiting extremely high scalability. Effectively ensuring the timeliness and authenticity of information: Utilizing the real-time search and segmentation tools integrated into MCP, research reports can cover the latest information, breaking down the time barriers of model knowledge. Simultaneously, the pre-screening mechanism by humans and the RAG retrieval mechanism greatly suppress the illusion of large-scale model-generated reports, significantly enhancing the credibility and commercial value of the reports. Improving research automation efficiency and user experience: Freeing researchers from tedious information gathering, reading, and preliminary integration work, allowing them to focus more on higher-level strategic thinking and decision-making; the fully automated process can run 24 hours a day, greatly shortening the report production cycle and improving work efficiency.

[0070] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A deep research report generation method based on a large model, DeepResearch, and MCP, characterized by: The method comprises the following steps: Step S1, the DeepResearch service is constructed as an MCP client, and an agent integrating a large model is constructed to obtain a research report submission demand based on natural language input through a user interface; Step S2, the agent queries a registered MCP server through the MCP client to discover AI tools supported by each MCP server, and constructs a tool list based on each AI tool; Step S3, the agent performs deep semantic understanding on the research report submission demand through the large model, splits the research report submission demand into a plurality of keywords, calls corresponding AI tools from the tool list based on the keywords to perform a preliminary retrieval operation through the MCP client, and obtains preliminary retrieval results; Step S4, the agent inputs the preliminary retrieval results into the large model for close reading to extract content abstracts, cuts the preliminary retrieval results and the content abstracts into information blocks, and stores the information blocks into a vector database after vectorization; Step S5, the agent performs theme clustering analysis on the research report submission demand and the preliminary retrieval results through the large model, obtains a core issue of the research report submission demand, and splits the core issue into a plurality of deep analysis angles; Step S6, the agent calls corresponding AI tools to perform a deep retrieval operation based on each deep analysis angle, obtains deep retrieval results, and stores the deep retrieval results into a cache; Step S7, the agent retrieves the most relevant information blocks from the vector database through RAG technology based on the deep retrieval results of each deep analysis angle, and performs content iteration based on each deep retrieval result and the information blocks to generate a research report.

2. The deep research report generation method based on a large model, DeepResearch, and MCP of claim 1, wherein: In the step S2, the AI tools at least include a keyword extraction tool, a search tool, a theme splitting tool, and a data analysis tool.

3. The deep research report generation method based on a large model, DeepResearch, and MCP of claim 1, wherein: In the step S3, the agent performs deep semantic understanding on the research report submission demand through the large model, splits the research report submission demand into a plurality of keywords, sets an iterative retrieval plan based on each keyword, calls corresponding AI tools from the tool list based on the iterative retrieval plan to perform a preliminary retrieval operation through the MCP client, and obtains preliminary retrieval results. The MCP client caches and preloads the AI tools based on a calling frequency and a calling success rate. In the step S6, the agent calls corresponding AI tools to perform a deep retrieval operation based on each deep analysis angle, obtains deep retrieval results, filters the deep retrieval results corresponding to each deep analysis angle based on a preset filtering rule, and stores each deep retrieval result into the cache for subsequent direct matching calling based on similarity to bypass the MCP client.

4. The deep research report generation method based on a large model, DeepResearch, and MCP of claim 1, wherein: In the step S7, the agent retrieves the most relevant information blocks from the vector database through RAG technology based on the deep retrieval results of the first deep analysis angle, inputs the deep retrieval results and the information blocks into the large model to obtain corresponding chapter contents. ​ 5. The method according to claim 1, wherein the method is characterized in that: ​ ​ The chapter content is taken as a prompt word of a next depth analysis angle, and corresponding chapter content is obtained by searching and reasoning in combination with the prompt word until all chapter contents are generated; The chapter content is structured and embedded into a preset HTML template to obtain the research report.

6. A deep research report generation system based on a large model, DeepResearch, and MCP, characterized by: The method comprises the following modules: A research report declaration requirement input module is configured to construct a DeepResearch service as an MCP client, integrate an Agent of a large model, and obtain a research report declaration requirement based on natural language input through a user interface. An AI tool discovery module is configured to query a registered MCP server through the MCP client to discover AI tools supported by each MCP server, and construct a tool list based on the AI tools. A preliminary search module is configured to perform deep semantic understanding on the research report declaration requirement through the large model, split the research report declaration requirement into a plurality of keywords, call corresponding AI tools from the tool list based on the keywords through the MCP client to perform a preliminary search operation, and obtain a preliminary search result. An information block vectorization storage module is configured to input the preliminary search result into the large model to extract a content abstract, split the preliminary search result and the content abstract into information blocks, and store the information blocks in a vector database after vectorization. A core issue splitting module is configured to perform theme clustering analysis on the research report declaration requirement and the preliminary search result through the large model, obtain a core issue of the research report declaration requirement, and split the core issue into a plurality of depth analysis angles. A deep search module is configured to call corresponding AI tools based on each depth analysis angle to perform a deep search operation, obtain a deep search result, and store the deep search result in a cache. A research report generation module is configured to retrieve the most relevant information blocks from the vector database through RAG technology based on the deep search result of each depth analysis angle, and perform content iteration based on each deep search result and information block to generate a research report.

7. The deep research report generation system based on a large model, DeepResearch, and MCP of claim 6, wherein: The AI tool discovery module comprises at least a keyword extraction tool, a search tool, a theme splitting tool, and a data analysis tool.

8. The deep research report generation system based on a large model, DeepResearch, and MCP of claim 6, wherein: The preliminary search module is specifically configured to: The Agent performs deep semantic understanding on the research report declaration requirement through the large model, splits the research report declaration requirement into a plurality of keywords, sets an iterative search plan based on each keyword, calls corresponding AI tools from the tool list based on the iterative search plan through the MCP client to perform a preliminary search operation, and obtains a preliminary search result. The MCP client caches and preloads the AI tools based on a calling frequency and a calling success rate.

9. The deep research report generation system based on a large model, DeepResearch, and MCP of claim 6, wherein: The deep search module is specifically configured to: The Agent calls corresponding AI tools to perform deep search operations based on each of the deep analysis angles, obtains deep search results, filters the deep search results corresponding to each of the deep analysis angles based on a preset filtering rule, and stores each of the deep search results in a cache for subsequent direct matching and calling based on similarity to bypass the MCP client.

10. The deep research report generation system based on a large model, DeepResearch, and MCP of claim 6, wherein: The research report generation module is specifically used for: The Agent retrieves the most relevant information block from the vector database through RAG technology based on the deep search result of the first deep analysis angle, inputs the deep search result and the information block into a large model to obtain corresponding chapter content; The chapter content is used as a prompt word for the next deep analysis angle, and corresponding chapter content is obtained through retrieval and reasoning combined with the prompt word until all chapter content is generated; Each of the chapter contents is structured and embedded into a preset HTML template to obtain a research report.

Images