Deep search method and device, equipment and medium

Abstract

The invention discloses a deep search method and device, equipment and a medium, and relates to the technical field of artificial intelligence, and the method comprises the steps: employing a planning agent to decompose a currently input preprocessing problem into a plurality of sub-problems; screening and analyzing from an external knowledge base by adopting a search agent according to the characteristics of each sub-problem, and gradually extracting corresponding effective information; and integrating the extracted effective information to generate an answer corresponding to the pre-processing question. In this way, preprocessing problems are decomposed through the planning agent and the searching agent, screening and analysis are carried out from an external knowledge base to gradually extract corresponding effective information, then the extracted effective information is integrated, and therefore the problems can be managed more easily; in addition, it can be ensured that each sub-question can be fully concerned and accurately answered, the information retrieval efficiency is improved, and scattered information resources are effectively integrated, so that accurate and timely replies can be generated, and the reply quality and accuracy are ensured.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a deep search method, device, equipment and medium. Background Art

[0002] In today's era of information explosion, handling complex problems is a major challenge in the field of artificial intelligence. Traditional information retrieval and processing methods have several shortcomings. For example, the Retrieval-Augmented Generation (RAG) approach cannot ensure the timeliness of information retrieval and has limited ability to handle complex problems. Network agents lack operational flexibility and are difficult to deploy. Reinforcement learning and behavioral cloning methods require extensive training time, rely on human behavioral data, and have low processing efficiency. Large language models have made significant progress in reasoning and language understanding, but they still face challenges in providing accurate information responses.

[0003] Therefore, how to make timely and accurate responses to user's question requests in a short period of time is a technical problem that needs to be solved urgently by those skilled in the art. Summary of the Invention

[0004] The purpose of the present invention is to provide a deep search method, device, equipment and medium that can make problems easier to manage, improve the efficiency of information retrieval, and effectively integrate scattered information resources to generate accurate and timely responses, thereby ensuring the quality and accuracy of the responses.

[0005] In order to solve the above technical problems, the present invention provides a depth search method, comprising:

[0006] Use a planning agent to decompose the current input preprocessing problem into multiple sub-problems;

[0007] Based on the characteristics of each sub-problem, a search agent is used to screen and analyze the external knowledge base, gradually extracting the corresponding effective information;

[0008] The extracted valid information is integrated to generate the answer corresponding to the preprocessing question.

[0009] In the first aspect, in the above-mentioned deep search method provided by the present invention, a planning agent is used to decompose the current input preprocessing problem into multiple sub-problems, including:

[0010] Use planning agents to define decomposition functions;

[0011] Analyzing the semantic content, contextual information, and related domain knowledge of the currently input preprocessing question through the decomposition function to identify multiple sub-questions;

[0012] According to each identified sub-problem, a corresponding sub-problem set is constructed to decompose the preprocessing problem into multiple sub-problems, thereby achieving a mapping from the preprocessing problem to the sub-problem set.

[0013] On the other hand, in the above-mentioned depth search method provided by the present invention, while identifying multiple sub-questions, it also includes:

[0014] Based on planning hints, targeted planning is performed on each sub-problem to generate planning sub-problems corresponding to different key aspects of the preprocessing problem; the planning hints include decomposition guidance information for different domain requests;

[0015] Divide each planning sub-problem according to the preset threshold to obtain sub-problems of different complexity;

[0016] Correspondingly, based on the identified sub-problems, a corresponding set of sub-problems is constructed, including:

[0017] According to the sub-problems of different complexity obtained by division, a corresponding set of sub-problems is constructed.

[0018] On the other hand, in the above-mentioned depth search method provided by the present invention, after constructing the corresponding sub-problem set, it also includes:

[0019] When any subproblem in the subproblem set is a complex category subproblem, an adjacency list expansion strategy is adopted to expand the complex category subproblem through a hierarchical traversal method until the complexity of all subproblems drops below the preset threshold.

[0020] On the other hand, in the above-mentioned deep search method provided by the present invention, based on the characteristics of each sub-problem, a search agent is used to screen and analyze from an external knowledge base, gradually extracting corresponding valid information, including:

[0021] Use search agents to define search functions;

[0022] Screening and analyzing the external knowledge base through the search function to gradually extract corresponding valid information;

[0023] In the process of screening and analysis, the search prompts are parsed according to the characteristics of each sub-question to generate targeted search sub-questions;

[0024] Based on the search sub-question, selecting a corresponding tool from a set of equipped tools;

[0025] Perform retrieval of information from external knowledge bases using tools of choice.

[0026] On the other hand, in the above-mentioned deep search method provided by the present invention, the process of performing information retrieval includes:

[0027] Conduct preliminary screening of the retrieved materials based on relevant information including titles, abstracts and summaries provided by the external knowledge base;

[0028] Conduct semantic understanding and keyword matching on the full text of the initially screened materials to analyze the content of the initially screened materials and extract effective information.

[0029] On the other hand, in the above-mentioned deep search method provided by the present invention, the extracted valid information is integrated to generate the answer corresponding to the preprocessing question, including:

[0030] Define the integration function;

[0031] According to the content and context of the valid information, the valid information is comprehensively analyzed through the integration function, the analysis results are organized in a logical order, and converted into natural language expressions to obtain answers corresponding to the preprocessing questions.

[0032] In order to solve the above technical problems, the present invention further provides a depth search device, comprising:

[0033] Planning agent, used to decompose the current input preprocessing problem into multiple sub-problems;

[0034] The search agent is used to filter and analyze the external knowledge base based on the characteristics of each sub-problem, and gradually extract the corresponding effective information;

[0035] The information integration module is used to integrate the extracted valid information and generate the answer corresponding to the preprocessing question.

[0036] In order to solve the above technical problems, the present invention further provides an electronic device, comprising:

[0037] memory for storing computer programs;

[0038] A processor is configured to implement the steps of the above-mentioned depth search method when executing the computer program.

[0039] In order to solve the above technical problems, the present invention further provides a non-volatile storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned depth search method are implemented.

[0040] It can be seen from the above technical solution that the present invention provides a deep search method, which includes: using a planning agent to decompose the currently input preprocessing problem into multiple sub-problems; based on the characteristics of each sub-problem, using a search agent to screen and analyze from an external knowledge base, and gradually extract corresponding valid information; integrating the extracted valid information to generate an answer corresponding to the preprocessing problem.

[0041] The beneficial effect of the present invention is that the above-mentioned deep search method provided by the present invention has two core components: a planning agent and a search agent. First, the planning agent can decompose the pre-processing problem currently input into multiple sub-problems. Through this subdivision, not only can the problem be made easier to manage, but it can also ensure that each sub-problem can receive sufficient attention and accurate answers. Next, the search agent uses an external knowledge base to screen and analyze, and gradually extracts the corresponding effective information to solve these sub-problems and improve the efficiency of information retrieval. Finally, the extracted effective information is integrated to generate the answer corresponding to the pre-processing problem. This effectively integrates scattered information resources, thereby generating accurate and timely replies, and ensuring the quality and accuracy of the replies. The overall approach not only provides an efficient and reliable solution for solving complex user queries, but also provides new ideas and technical foundations for the development of future intelligent customer service and automated transaction systems, and can be applied to scenarios requiring high precision and real-time response.

[0042] In addition, the present invention also provides a corresponding deep search device, electronic device and non-volatile storage medium for the deep search method, which have the same or corresponding technical features as the above-mentioned deep search method and have the same effects as above. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0044] Figure 1 A flowchart of a depth search method provided by an embodiment of the present invention;

[0045] Figure 2 A schematic diagram of the framework of a planning agent provided by an embodiment of the present invention;

[0046] Figure 3 A schematic diagram of the framework of the adjacency list expansion provided by an embodiment of the present invention;

[0047] Figure 4 A schematic diagram of a search agent framework provided by an embodiment of the present invention;

[0048] Figure 5 A schematic diagram of the structure of a depth search device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0050] It should be noted that, in the description of the present invention, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. The terms "first," "second," etc., in the present invention are used to distinguish similar objects, and are not used to describe a particular order or precedence.

[0051] In order to enable those skilled in the art to better understand the solutions of the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0052] In conjunction with the specific application environment architecture or specific hardware architecture on which the execution of the deep search method depends, the specific application environment architecture or specific hardware architecture is described here.

[0053] An embodiment of the present invention provides a depth search method, and the method is described in detail in conjunction with the execution flow of the depth search method. Figure 1 A flowchart of a depth search method provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the method includes:

[0054] S101. Use a planning agent to decompose the current input preprocessing problem into multiple sub-problems.

[0055] It should be noted that this paper designs a DeepAgent framework, which consists of two core components: a planning agent and a search agent. This framework leverages these two core components to achieve efficient information processing and response generation.

[0056] The planning agent is an intelligent system that uses planning to make autonomous decisions. It performs a key preprocessing task within the DeepAgent framework. Its core function is to accurately decompose complex user input into multiple, relatively simple sub-questions. This step is crucial for subsequent processing, enhancing the manageability of the problem and the accuracy of the answer.

[0057] S102. Based on the characteristics of each sub-problem, a search agent is used to screen and analyze from an external knowledge base, and corresponding valid information is gradually extracted.

[0058] In practice, the search agent is an intelligent system that uses a set of search algorithms and strategies to find a target or solution that meets specific criteria within a given search space. It plays a key role in information acquisition and screening within the DeepAgent framework. After the planning agent breaks down a complex problem into a set of subproblems, the search agent can filter and analyze the vast external knowledge base K, accurately extracting relevant and effective information to solve these subproblems, thereby improving information retrieval efficiency and response quality.

[0059] S103: Integrate the extracted valid information to generate answers corresponding to the preprocessing questions.

[0060] In practice, step S103 integrates the dispersed, valid information extracted in step S102 to generate the final answer U that can be used to respond to the user. This approach is particularly applicable to fields requiring rapid response, such as financial markets and judicial services. Whether in the rapidly changing financial markets, facing the diverse demands of judicial services, or other fields, the DeepAgent framework can provide efficient support. Furthermore, the framework is flexible and scalable, capable of adapting to diverse application scenarios and rapidly adjusting and optimizing its performance.

[0061] The above-mentioned deep search method provided by the embodiment of the present invention has two core components: a planning agent and a search agent. First, the planning agent can decompose the pre-processing problem currently input into multiple sub-problems. Through this subdivision, not only can the problem be easier to manage, but also it can be ensured that each sub-problem can receive sufficient attention and accurate answers. Next, the search agent uses the external knowledge base to screen and analyze, and gradually extracts the corresponding effective information to solve these sub-problems and improve the efficiency of information retrieval. Finally, the extracted effective information is integrated to generate the answer corresponding to the pre-processing problem. This effectively integrates scattered information resources, thereby generating accurate and timely responses and ensuring the quality and accuracy of the responses. The overall approach not only provides an efficient and reliable solution for solving complex user queries, but also provides new ideas and technical foundations for the development of future intelligent customer service and automated transaction systems. It can be applied to scenarios that require high precision and real-time response.

[0062] Furthermore, in a specific implementation, in the above-mentioned deep search method provided in an embodiment of the present invention, step S101 uses a planning agent to decompose the currently input preprocessing problem into multiple sub-problems, which may specifically include: using a planning agent to define a decomposition function; analyzing the semantic content, contextual information and related domain knowledge of the currently input preprocessing problem through the decomposition function to identify multiple sub-problems; constructing a corresponding sub-problem set based on each identified sub-problem to decompose the preprocessing problem into multiple sub-problems and realize the mapping from the preprocessing problem to the sub-problem set.

[0063] In the implementation, assuming that the problem to be processed (i.e., complex request) input by the user is R, the goal of the planning agent is to decompose it into a set of sub-problems:

[0064] (1)

[0065] To achieve this goal, this paper defines a decomposition function D(R) that maps a complex request R to a set of subproblems S, namely, S = D(R). D(R) goes beyond simple rule matching and deeply analyzes the semantic content, context, and relevant domain knowledge of the request R.

[0066] Taking the financial trading field as an example, if the user request is "Under the current market situation, what are the risks and returns of investing in technology stocks, and what investment opportunities are worth paying attention to?", the D(R) function will identify key sub-problems such as "current market situation analysis", "technology stock return analysis", "technology stock risk assessment", and "technology stock investment opportunity recommendation" based on professional knowledge in the financial field, thereby constructing the corresponding sub-problem set S.

[0067] Furthermore, in specific implementation, in the above steps, while identifying multiple sub-problems, it can also include: based on planning prompts, targeted planning is carried out for each sub-problem to generate planning sub-problems corresponding to different key aspects of the preprocessing problem; the planning prompt contains decomposition guidance information for requests in different fields; each planning sub-problem is divided according to a preset threshold to obtain sub-problems of different complexity.

[0068] Correspondingly, a corresponding sub-problem set is constructed based on each identified sub-problem, which may specifically include: constructing a corresponding sub-problem set based on sub-problems of different complexity obtained by division.

[0069] Figure 2 This is a schematic diagram of the framework of the planning agent provided by the embodiment of the present invention. Figure 2As shown in the figure, when performing the decomposition task, the planning agent uses planning prompts to guide the decomposition process. Planning prompts contain decomposition guidance information for requests in different fields. For example, in the financial field, it clearly instructs the agent to analyze the request from the perspectives of market trends, investment advice, risk assessment, etc., helping the agent to more specifically extract sub-questions. Based on the planning prompts, the planning agent generates a series of planning sub-questions that precisely correspond to different key aspects of the original request, laying the foundation for subsequent in-depth processing of each sub-question.

[0070] During the planning process, "Hard" and "Easy" thresholds play a crucial role in optimizing resource allocation. These thresholds are based on the complexity of the subproblems. Complex subproblems (such as financial risk assessments involving multi-factor analysis) are classified as "Hard," while simpler subproblems (such as single data queries) are classified as "Easy." The system allocates more computing resources and time to "Hard" subproblems, while "Easy" subproblems are processed more quickly, improving overall system efficiency.

[0071] Furthermore, in specific implementation, in the above steps, after constructing the corresponding sub-problem set, it can also include: when the sub-problem in any sub-problem set is a complex category sub-problem, an adjacency list expansion strategy is used to expand the complex category sub-problem through a hierarchical traversal method until the complexity of all sub-problems is reduced to below a preset threshold.

[0072] In practice, this invention introduces an adjacency list expansion strategy when handling complex subproblems. Specifically, when faced with complex subproblems, the large model uses prompts to decompose these complex problems into multiple simpler subproblems. If a subproblem is marked as "Hard," indicating that it still has a high degree of complexity, it will be further decomposed. This process will continue until the difficulty of all subproblems drops below a preset threshold.

[0073] During the adjacency list expansion process, the decomposed sub-problems are expanded using a hierarchical traversal method, which means that the decomposition of the second-level sub-problems will not begin until all sub-problems in the first level have been decomposed. Figure 3 Schematic diagram of the framework of the adjacency list expansion provided by the embodiment of the present invention. Figure 3 As shown in Figure 2, subproblem D is processed before subproblem E. This approach ensures that the subproblems in the third level will not be repeated with those in the second level, because the generation process of the third level subproblems will take into account the situations of all the second level subproblems.

[0074] This approach not only improves the efficiency and accuracy of problem decomposition but also ensures the logical coherence and independence between sub-problems at different levels. In this way, even extremely complex problems can be systematically and methodically broken down and solved, significantly improving the overall problem-solving capabilities and effectiveness.

[0075] Furthermore, in specific implementation, in the above-mentioned deep search method provided in the embodiment of the present invention, step S102 uses a search agent to screen and analyze from an external knowledge base based on the characteristics of each sub-problem, and gradually extracts corresponding valid information. Specifically, it may include: using a search agent to define a search function; screening and analyzing from an external knowledge base through a search function, and gradually extracting corresponding valid information; wherein, in the process of screening and analyzing, according to the characteristics of each sub-problem, parsing the search prompt to generate targeted search sub-questions; based on the search sub-questions, selecting a corresponding tool from the equipped tool set; and using the selected tool to perform information retrieval from the external knowledge base.

[0076] In implementation, the present invention uses a search agent to define a search function F(S, K), through which valid information I is obtained, that is, I=F(S, K).

[0077] During actual operation, the search agent first parses the search prompt based on the characteristics of the sub-questions. For example, in the market trend analysis sub-question, the search prompt guides the agent to focus on finding the latest market data and analysis reports. The search agent then generates a series of more targeted search sub-questions based on these instructions, such as "Stock price trends in the past month" and "Performance of major industry sectors." These search sub-questions refine the market trend analysis requirements from different dimensions, helping to comprehensively and deeply mine relevant information within the knowledge base.

[0078] Search agents are equipped with a rich variety of tools to perform information retrieval tasks, and the choice of tools depends on the specific application scenario and the nature of the sub-problem. Figure 4 Schematic diagram of the framework of the search agent provided by the embodiment of the present invention. Figure 4 As shown, the present invention can select appropriate tools from a rich tool set based on the specific application scenarios and properties of the sub-problems. When dealing with sub-problems involving web page information, a web crawler tool can be enabled to obtain the latest information on the web page; when it is necessary to query professional field knowledge, such as financial expertise, the knowledge base query interface can be used to accurately locate relevant professional documents; for the retrieval of structured data, database search engines will play an important role. For example, when dealing with sub-problems in the financial field, if you want to obtain the latest stock trading data, you can use a database search engine to quickly query the relevant financial database; if you need to understand industry news, you can use a web crawler to obtain relevant web page information.

[0079] Furthermore, in specific implementation, in the above steps, in the process of executing information retrieval, it can specifically include: preliminary screening of the retrieved materials based on relevant information including titles, abstracts and summaries provided by the external knowledge base; semantic understanding and keyword matching of the full text of the preliminary screened materials to analyze the content of the preliminary screened materials and extract valid information.

[0080] In practice, during the information retrieval process, to improve efficiency and reduce unnecessary computational overhead, the search agent will first perform a preliminary screening of the data based on information such as the title, abstract, and summary. This information acts as a "condensed essence" of the data, helping the agent quickly determine which data is likely to be highly relevant to the sub-question. For example, when searching for information on the development of artificial intelligence technology, if the title of an article includes "Latest AI Breakthroughs" and the abstract mentions improvements to key technical indicators, then this article is more likely to be the required information. In this way, the agent can quickly filter out the most valuable data subset from a large amount of potential information.

[0081] However, relying solely on titles, abstracts, and summaries may not guarantee the accuracy and completeness of information. Therefore, for the initially selected materials, the search agent further reads the full text. During this reading process, the agent analyzes the content using semantic understanding and keyword matching. This allows the agent to extract the most valuable information from the material, ensuring that the information obtained is both accurate and complete, laying a solid foundation for generating high-quality answers.

[0082] After screening and extraction by the search agent, a valid information set is obtained:

[0083] (2)

[0084] Among them, i1, i2, ..., i m is the valid information in the valid information set I, and m is the maximum number of valid information in the valid information set I.

[0085] Furthermore, in specific implementation, in the above-mentioned deep search method provided in an embodiment of the present invention, step S103 integrates the extracted valid information to generate an answer corresponding to the preprocessing question, which may specifically include: defining an integration function; comprehensively analyzing the valid information through the integration function according to the content and context of the valid information, organizing the analysis results in a logical order, and converting them into natural language expressions to obtain the answer corresponding to the preprocessing question.

[0086] In implementation, the present invention defines an integration function G(I), through which valid information is integrated into the final response, that is, U=G(I).

[0087] It's important to note that the G(I) function is a complex mapping process. It doesn't simply pile up information. Instead, it uses natural language generation technology and logical reasoning based on the content and context of valid information to generate coherent, accurate, and user-friendly responses. In financial applications, if a user asks, "What are the current prospects for investing in new energy stocks?", the search agent might extract valid information including market growth data for the new energy industry, policy support, the financial status of related companies, and technological innovation progress. The G(I) function then comprehensively analyzes this information, first explaining the overall market growth trends of the new energy industry, then interpreting its driving effect on industry development in light of policy support, then analyzing the strengths and challenges of related companies. Finally, based on this analysis, it provides a comprehensive assessment and recommendation on the prospects for investing in new energy stocks.

[0088] Furthermore, in order to comprehensively and accurately evaluate the performance of the DeepAgent framework, the present invention collected experimental datasets in the financial and judicial fields.

[0089] In the financial field, 500 major global financial consulting questions have been collected, including stock price trends, trading volume, market capitalization, etc.; macroeconomic indicator data, such as gross domestic product (GDP) growth rate, inflation rate, interest rate, etc.; and financial statement data of thousands of listed companies, including key information such as balance sheets, income statements, and cash flow statements.

[0090] In the judicial field, the present invention has carefully compiled a dataset containing more than 1,000 legal consultation questions, aiming to cover various major legal issues and case analysis. This dataset includes but is not limited to the following aspects: (1) Criminal legal consultation: covers consultation questions on common criminal cases, such as criminal cases, and the application and interpretation of criminal law provisions; (2) Civil dispute resolution: collects consultations on contract disputes, property rights disputes, marriage and family cases, etc., reflecting the specific application of the Civil Code and its related judicial interpretations; (3) Commercial legal affairs: includes consultation content on various aspects such as company establishment and change, shareholder rights protection, intellectual property protection, etc., showing common legal problems and solutions in commercial activities; (4) Administrative legal affairs: includes answers to questions on the legality of government actions, interpretation of administrative penalty decisions, administrative reconsideration procedures, etc., reflecting the actual operation and implementation standards of administrative laws and regulations.

[0091] The present invention can adopt a combination of automatic scoring and manual evaluation, and use a cross-validation scheme to ensure the reliability and accuracy of the evaluation results.

[0092] For automatic scoring, the GPT-4o model can be used to perform a preliminary scoring of responses generated by the DeepAgent framework. This pre-trained model identifies and quantifies quality features of the text, such as logical consistency, information accuracy, and relevance. The GPT-4o model then assigns a score to the response by setting a series of scoring metrics and weights. For example, logical consistency may be weighted 30%, information accuracy 50%, and relevance 20%, ultimately yielding a composite score.

[0093] Manual evaluation can be performed by a team of experts comprised of financial researchers, experienced legal practitioners (such as judges and lawyers), and natural language processing researchers. Team members score the responses based on a standardized scoring system, taking into account the accuracy, completeness, practicality, and fluency of the responses. For responses in the financial sector, the focus is on the rationality of the investment advice and the accuracy of the risk assessment. For responses in the legal sector, the emphasis is on the correctness of the legal analysis, the appropriateness of the case, and the practicality of the legal advice.

[0094] Cross-validation is to divide the data set into several parts and use one part as the test set in turn to test the consistency and stability of the scoring system.

[0095] The scoring criteria is to score the quality of each generated result from 1 to 5, with 3 points being a pass. The specific criteria are as follows:

[0096] 1 (Very Poor): The generated results are seriously inconsistent with user needs, contain a large amount of erroneous information, or are completely irrelevant to the question. For example, in the financial field, the investment advice provided is inconsistent with market conditions; in the legal field, the legal opinions provided are significantly biased or irrelevant legal provisions are provided.

[0097] 2 (Poor): While some relevant information is provided, the overall quality is low, with numerous errors or inaccuracies that may mislead users. For example, key risk factors may be omitted in the financial risk assessment, or the judicial analysis may lack understanding of the case context or improper application of legal provisions.

[0098] 3 points (Fair): The response generally meets the requirements, but there are minor issues with details, such as insufficient information or depth. For example, the analysis of financial market trends only mentions the main trends without discussing potential influencing factors; the response to the judicial consultation addresses basic issues but lacks in-depth analysis or personalized advice.

[0099] 4 (Good): High-quality results, with only minor deficiencies in a few non-critical areas, and overall excellent performance. For example, the investment strategy recommendations accurately analyze market conditions and provide reasonable advice, although the interpretation of some emerging market trends is less in-depth. The legal advice recommendations not only comprehensively and accurately address issues, but also express themselves fluently and naturally, though there is room for improvement in approachability.

[0100] 5 (Excellent): The generated results are nearly perfect, not only fully meeting user expectations but also exceeding them in value or providing novel insights. In the financial sector, investment advice accurately captures market dynamics, fully considers various risks and opportunities, and even proposes innovative investment ideas. In the legal sector, the client's issues are efficiently resolved, potential needs are proactively explored, and thoughtful advice is provided, all delivered with engaging and professional language.

[0101] Through automatic scoring, manual evaluation, and cross-validation, based on the detailed scoring criteria described above, we can comprehensively, objectively, and accurately evaluate the quality of responses generated by the DeepAgent framework, providing a strong basis for research improvement and optimization.

[0102] This paper tests DeepAgent and selects GPT-4o as the test model. The purpose is to further explore the performance of this framework in responding to complex requests in the financial and judicial fields, meeting high timeliness requirements and accurately retrieving and integrating information. The following is a detailed discussion of the experimental results.

[0103] The experiment evaluated the quality of responses. We collected time-sensitive questions from the financial and legal fields, had the model generate responses, and evaluated them through both manual evaluation and automatic scoring. The scoring scale ranged from 1 to 5, with 3 being considered a pass. The results are shown in Tables 1 and 2 below:

[0104] Table 1. DeepAgent evaluation results in the financial field

[0105]

[0106] Table 2. DeepAgent evaluation results in the judicial field

[0107]

[0108] From the above data, we can see that DeepAgent has performed well in both the financial and judicial fields.

[0109] First, in terms of human evaluation, DeepAgent achieved higher average scores in both areas compared to several other methods (search engine, ReActSearch, MindSearch), indicating that the answers it generates are not only more in line with the expectations of human reviewers, but also have obvious advantages in understanding and handling complex problems.

[0110] Specifically, in the financial sector, DeepAgent achieved an 82% pass rate for manual evaluations, meaning that over four-fifths of responses were deemed qualified across all test cases. This rate is approximately 4 percentage points higher than that of MindSearch, the next-best performer. This significant improvement in pass rates is particularly significant, assuming the difficulty and background knowledge required for each domain are the same. This demonstrates DeepAgent's superior ability to process highly specialized and timely financial information.

[0111] Secondly, in the legal domain, DeepAgent achieved a pass rate of 81% in human evaluations, only slightly lower than its performance in the financial domain, but still significantly higher than other comparison models. This demonstrates that DeepAgent can effectively understand and respond to complex legal queries, reflecting its high performance in legal information analysis and application.

[0112] DeepAgent also ranked highly in automated scoring. In the judicial field, DeepAgent's automated scoring pass rate was 83%, 15 percentage points higher than the lowest search engine. Applying this percentage increase to actual numbers, for 800 judicial questions, DeepAgent could have provided approximately 120 additional qualified answers compared to a search engine, significantly improving the efficiency and accuracy of the decision support system.

[0113] Further analysis reveals that while all models generally outperformed the financial sector in the legal sector, DeepAgent saw the smallest increase, potentially suggesting that the model maintains consistent high performance across different domains. This consistency is particularly important for companies that need to provide services across multiple domains, as it means that investing in DeepAgent can bring broad benefits, not just in a specific area.

[0114] In summary, DeepAgent's outstanding performance in both manual evaluation and automated scoring demonstrates its ability to effectively retrieve and integrate information while meeting stringent timeliness and accuracy requirements. Whether facing complex financial inquiries or legal requests, DeepAgent delivers high-quality answers, making it a highly promising tool for scenarios requiring rapid and accurate responses to diverse and complex requests.

[0115] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus the necessary general hardware platform, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method.

[0116] An embodiment of the present application also provides a depth search device. Figure 5 This is a schematic diagram of the structure of the depth search device provided by an embodiment of the present invention. This embodiment is based on the perspective of functional modules, such as Figure 5 As shown, the device includes:

[0117] A planning agent 10 is used to decompose the preprocessing problem of the current input into multiple sub-problems;

[0118] The search agent 11 is used to screen and analyze the external knowledge base according to the characteristics of each sub-problem and gradually extract the corresponding effective information;

[0119] The information integration module 12 is used to integrate the extracted valid information and generate answers corresponding to the preprocessing questions.

[0120] In the above-mentioned deep search device provided by the embodiment of the present invention, the currently input pre-processing problem can be decomposed into multiple sub-problems through the planning agent and the search agent, and the external knowledge base can be used for screening and analysis to gradually extract the corresponding valid information. Finally, the extracted valid information is integrated to generate the answer corresponding to the pre-processing problem, which not only makes the problem easier to manage, but also ensures that each sub-problem can receive sufficient attention and accurate answers, thereby improving the efficiency of information retrieval. This effectively integrates scattered information resources, thereby generating accurate and timely replies, ensuring the quality and accuracy of the replies. The overall device not only provides an efficient and reliable solution for solving complex user queries, but also provides new ideas and technical foundations for the development of future intelligent customer service and automated transaction systems, and can be applied to scenarios that require high precision and real-time response.

[0121] Since the embodiments of the apparatus part correspond to the embodiments of the method part, the embodiments of the apparatus part refer to the description of the embodiments of the method part and are not described here in detail. The embodiments of the apparatus part have the same beneficial effects as the depth search method mentioned above.

[0122] Furthermore, in a specific implementation, in the above-mentioned deep search device provided in an embodiment of the present invention, the planning agent 10 can be specifically used to define a decomposition function; the semantic content, contextual information and related domain knowledge of the currently input preprocessing problem are analyzed through the decomposition function to identify multiple sub-problems; based on the identified sub-problems, a corresponding sub-problem set is constructed to decompose the preprocessing problem into multiple sub-problems, thereby realizing the mapping from the preprocessing problem to the sub-problem set.

[0123] Furthermore, in specific implementation, in the above-mentioned deep search device provided in the embodiment of the present invention, the search agent 11 can be specifically used to define a search function; through the search function, screening and analysis are performed from the external knowledge base, and corresponding valid information is gradually extracted; wherein, in the process of screening and analysis, the search prompts are parsed according to the characteristics of each sub-question, and targeted search sub-questions are generated; based on the search sub-questions, the corresponding tools are selected from the equipped tool set; and the selected tools are used to perform information retrieval from the external knowledge base.

[0124] Furthermore, in specific implementation, in the above-mentioned deep search device provided in the embodiment of the present invention, the information integration module 12 can be specifically used to define an integration function; according to the content and context of the effective information, the effective information is comprehensively analyzed through the integration function, and the analysis results are organized in a logical order and converted into natural language expressions to pre-process the answers corresponding to the questions.

[0125] An embodiment of the present application further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any of the above-mentioned depth search method embodiments.

[0126] An embodiment of the present invention further provides a non-volatile storage medium, in which a computer program is stored, wherein the computer program is configured to execute the steps of any of the above-mentioned depth search method embodiments when running.

[0127] In an exemplary embodiment, the non-volatile storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk, or an optical disk.

[0128] An embodiment of the present invention further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps of any of the above-mentioned depth search method embodiments are implemented.

[0129] An embodiment of the present invention further provides another computer program product, including a non-volatile computer storage medium, wherein the non-volatile computer storage medium stores a computer program, and when the computer program is executed by a processor, the steps of any of the above-mentioned depth search method embodiments are implemented.

[0130] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0131] The above is a detailed introduction to a deep search method, device, equipment and medium provided by the present invention. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, the present invention can also be improved and modified in a number of ways, and these improvements and modifications also fall within the scope of protection of the present invention.

Claims

1. A depth search method, characterized in that: include: Use a planning agent to decompose the current input preprocessing problem into multiple sub-problems; Based on the characteristics of each sub-problem, a search agent is used to screen and analyze the external knowledge base, gradually extracting the corresponding effective information; The extracted valid information is integrated to generate the answer corresponding to the preprocessing question.

2. The depth search method according to claim 1, characterized in that: A planning agent is used to decompose the current input preprocessing problem into multiple sub-problems, including: Use planning agents to define decomposition functions; Analyzing the semantic content, contextual information, and related domain knowledge of the currently input preprocessing question through the decomposition function to identify multiple sub-questions; According to each identified sub-problem, a corresponding sub-problem set is constructed to decompose the preprocessing problem into multiple sub-problems, thereby achieving a mapping from the preprocessing problem to the sub-problem set.

3. The depth search method according to claim 2, characterized in that: While identifying multiple sub-problems, these also include: Based on planning hints, targeted planning is performed on each sub-problem to generate planning sub-problems corresponding to different key aspects of the preprocessing problem; the planning hints include decomposition guidance information for different domain requests; Divide each planning sub-problem according to the preset threshold to obtain sub-problems of different complexity; Correspondingly, based on the identified sub-problems, a corresponding set of sub-problems is constructed, including: According to the sub-problems of different complexity obtained by division, a corresponding set of sub-problems is constructed.

4. The depth search method according to claim 3, characterized in that: After constructing the corresponding set of sub-problems, it also includes: When any subproblem in the subproblem set is a complex category subproblem, an adjacency list expansion strategy is adopted to expand the complex category subproblem through a hierarchical traversal method until the complexity of all subproblems drops below the preset threshold.

5. The depth search method according to claim 1, characterized in that: Based on the characteristics of each sub-problem, a search agent is used to screen and analyze the external knowledge base, gradually extracting the corresponding effective information, including: Use search agents to define search functions; Screening and analyzing the external knowledge base through the search function to gradually extract corresponding valid information; In the process of screening and analysis, the search prompts are parsed according to the characteristics of each sub-question to generate targeted search sub-questions; Based on the search sub-question, selecting a corresponding tool from a set of equipped tools; Perform retrieval of information from external knowledge bases using tools of choice.

6. The depth search method according to claim 5, characterized in that: The process of performing information retrieval includes: Conduct preliminary screening of the retrieved materials based on relevant information including titles, abstracts and summaries provided by the external knowledge base; Conduct semantic understanding and keyword matching on the full text of the initially screened materials to analyze the content of the initially screened materials and extract effective information.

7. The depth search method according to claim 1, characterized in that: Integrate the extracted valid information to generate the answer corresponding to the preprocessing question, including: Define the integration function; According to the content and context of the valid information, the valid information is comprehensively analyzed through the integration function, the analysis results are organized in a logical order, and converted into natural language expressions to obtain answers corresponding to the preprocessing questions.

8. A depth search device, characterized in that: include: Planning agent, used to decompose the current input preprocessing problem into multiple sub-problems; The search agent is used to filter and analyze the external knowledge base based on the characteristics of each sub-problem, and gradually extract the corresponding effective information; The information integration module is used to integrate the extracted valid information and generate the answer corresponding to the preprocessing question.

9. An electronic device, characterized in that: include: memory for storing computer programs; A processor, configured to implement the steps of the depth search method according to any one of claims 1 to 7 when executing the computer program.

10. A non-volatile storage medium, characterized in that: The non-volatile storage medium stores a computer program, which, when executed by a processor, implements the steps of the depth search method according to any one of claims 1 to 7.

Images