Project review method and device, electronic equipment and storage medium

Abstract

The invention provides a project review method, and the method comprises the steps: carrying out the matching of review key points and review guidance information corresponding to a to-be-reviewed project according to the content of the to-be-reviewed project; based on the project review key point information and the review guidance information, a target review instruction set of the to-be-reviewed project is generated through the multiple first-class intelligent agents, and the target review instruction set comprises at least one target review instruction; based on the target review instruction set and the large language model, constructing at least one second-class agent, each second-class agent corresponding to at least one target review instruction; and based on the at least one second-class agent, reviewing the to-be-reviewed item to obtain a reviewing result of the to-be-reviewed item. According to the invention, the to-be-reviewed item can be reviewed automatically without depending on manual review, and the accuracy of the review result can be improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a project review method, device, electronic equipment and storage medium. Background Art

[0002] In the field of project management, project review is an important link to ensure that the project establishment plan is scientific, reasonable and feasible. Traditional project review methods often rely on manual review, which is not only time-consuming and labor-intensive, but also easily affected by the subjective factors of the reviewers and the reviewers' understanding of the project, resulting in inaccurate project review results. Therefore, a new project review solution is urgently needed to improve the accuracy of project review. Summary of the invention

[0003] The embodiment of the present invention provides a project review method, aiming to provide a new project review scheme to improve the accuracy of project review. A first-type intelligent agent generates a target review instruction for a project to be reviewed based on project review key information and review guidance information, and a second-type intelligent agent is constructed by combining the target review instruction with a large language model. The second-type intelligent agent is used to automatically review the project to be reviewed, which does not rely on manual review and can improve the accuracy of the review result.

[0004] In a first aspect, an embodiment of the present invention provides a project review method, the method comprising the following steps:

[0005] According to the content of the project to be reviewed, match the review points and review guidance information corresponding to the project to be reviewed;

[0006] Based on the project review key point information and the review guidance information, generating a target review instruction set for the project to be reviewed through a plurality of first-type agents, wherein the target review instruction set includes at least one target review instruction;

[0007] Based on the target review instruction set and the large language model, construct at least one second-type intelligent agent, each second-type intelligent agent corresponds to at least one of the target review instructions;

[0008] Based on at least one of the second-type intelligent agents, the project to be reviewed is reviewed to obtain a review result of the project to be reviewed.

[0009] Optionally, before constructing the target review instruction set of the to-be-reviewed project by the first type of agent based on the project review key points information and the review guidance information, the method further includes:

[0010] Based on the project review key points information, determine the number of project review key points;

[0011] Based on the number of project review points, determine the number of first-type intelligent agents to be constructed;

[0012] Construct a plurality of agents corresponding to the number of the first type of agents to obtain a plurality of the first type of agents.

[0013] Optionally, constructing at least one second type of intelligent agent based on the target review instruction set and the large language model includes:

[0014] Based on each of the target review instructions in the review instruction set, the large language model is encapsulated to obtain a corresponding second-category intelligent agent.

[0015] Optionally, the reviewing the project to be reviewed based on at least one of the second-type intelligent agents to obtain the review result of the project to be reviewed includes:

[0016] If there are multiple second-type agents, then arranging the multiple second-type agents according to the context order of the items to be reviewed to obtain a workflow of the second-type agents;

[0017] The project proposal of the project to be reviewed is reviewed based on the workflow to obtain the review result of the project to be reviewed.

[0018] Optionally, after reviewing the project to be reviewed based on at least one of the second-type intelligent agents to obtain a review result of the project to be reviewed, the method further includes:

[0019] Obtain historical review data;

[0020] Based on the third type of agents, the review result is compared with the historical review data, and whether the first type of agents needs to be updated according to the comparison result;

[0021] If the first type of agent needs to be updated, the first type of agent is updated based on the comparison result.

[0022] Optionally, updating the first type of agent based on the comparison result includes:

[0023] Based on the comparison result, the first-category agent corresponding to the review result that failed the comparison is determined as the first-category agent to be updated, and the review result that failed the comparison and the corresponding historical review data are determined as update data corresponding to the first-category agent to be updated, and each first-category agent to be updated corresponds to one update data;

[0024] Based on the update data, the first type of agent to be updated is updated.

[0025] Optionally, updating the first type of agent to be updated based on the update data includes:

[0026] Calculating the error value between the review result and the historical review data;

[0027] The first type of agent to be updated is updated with the minimum error value as the optimization goal.

[0028] In a second aspect, an embodiment of the present invention further provides a project review device, the project review device comprising:

[0029] A first acquisition module is used to match the review points and review guidance information corresponding to the project to be reviewed according to the content of the project to be reviewed;

[0030] A generating module, configured to generate a target review instruction set for the project to be reviewed through a plurality of first-type agents based on the project review key point information and the review guidance information, wherein the target review instruction set includes at least one target review instruction;

[0031] A first construction module is used to construct at least one second-class intelligent agent based on the target review instruction set and the large language model, each second-class intelligent agent corresponds to at least one target review instruction;

[0032] The review module is used to review the project proposal of the project to be reviewed based on at least one of the second-type intelligent agents to obtain the review result of the project to be reviewed.

[0033] In a third aspect, an embodiment of the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the project review method provided in the embodiment of the present invention when executing the computer program.

[0034] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps in the project review method provided in the embodiment of the invention are implemented.

[0035] In an embodiment of the present invention, according to the content of the project to be reviewed, the review points and review guidance information corresponding to the project to be reviewed are matched; based on the project review points information and review guidance information, a target review instruction set of the project to be reviewed is generated by multiple first-class agents, and the target review instruction set includes at least one target review instruction; based on the target review instruction set and the large language model, at least one second-class agent is constructed, and each second-class agent corresponds to at least one target review instruction; based on the at least one second-class agent, the project to be reviewed is reviewed to obtain the review result of the project to be reviewed. The target review instruction of the project to be reviewed is generated by the first-class agent according to the project review points information and review guidance information, and the second-class agent is constructed by combining the target review instruction and the large language model. The second-class agent is used to automatically review the project to be reviewed, which does not rely on manual review and can improve the accuracy of the review result. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. 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 paying creative work.

[0037] Figure 1 is a flow chart of a project review method provided by an embodiment of the present invention;

[0038] Figure 2 is a flow chart of another project review method provided by an embodiment of the present invention;

[0039] Figure 3 It is a structural schematic diagram of a project review device provided by an embodiment of the present invention;

[0040] Figure 4 It is a structural schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] like Figure 1 As shown, Figure 1 1 is a flowchart of a project review method provided by an embodiment of the present invention, wherein the project review method comprises the following steps:

[0043] 101. According to the content of the project to be reviewed, match the review points and review guidance information corresponding to the project to be reviewed.

[0044] In an embodiment of the present invention, the project review method can be applied to a project review platform, and the project review platform can be constructed based on a server or a distributed server. The project review platform may include a user-side interface, a large language model (LLM) or a calling interface of a large language model, wherein the user-side interface can upload relevant information of the project to be reviewed, and the relevant information of the project to be reviewed may include project review key points information, review guidance information and project establishment plan. The large language model may be a large language model such as the GPT series, the BERT series, etc.

[0045] The above-mentioned project to be reviewed may be a government project or a business project, etc., and the embodiment of the present invention does not limit the specific review project.

[0046] For a project to be reviewed, the user can upload the content of the project to be reviewed through the user terminal. The project review platform receives the content of the project to be reviewed through the user terminal interface and can match the project review key points information and review guidance information based on the content.

[0047] The above-mentioned project review points may be important review points in the project to be reviewed, may be specified by users, professionals or technical experts, or may be obtained by automatic text recognition of project documents of the project to be reviewed.

[0048] The project documents of the above-mentioned projects to be reviewed are mainly indicative documents that guide the implementation of the projects to be reviewed, such as "Guidelines for the Preparation of Project Establishment Plans", "Detailed Rules for Project Establishment and Approval", etc. Specifically, the text structure of the project documents can be extracted through natural language processing technology. The above-mentioned text structure can include text structures divided according to chapters, sections, pages, etc., and can also include text structures divided according to semantic content. The project review points corresponding to the projects to be reviewed are extracted based on the text structure.

[0049] The above-mentioned review guidance information may be the guidance information in the project to be reviewed, and may be proposed by professionals or technical experts for the project to be reviewed. The above-mentioned review guidance information may be the guidance information proposed by professionals or technical experts for the key points of project review.

[0050] 102. Based on the project review key information and review guidance information, a target review instruction set for the project to be reviewed is generated by multiple first-category intelligent agents.

[0051] In an embodiment of the present invention, after obtaining the project review key information and review guidance information, the target review instruction of the project to be reviewed can be generated based on the first agent. The first type of agent can be used to generate the Prompt instruction (target review instruction) of the large language model. The Prompt instruction can be understood as the input of the large language model, which is used to prompt the large language model to output the corresponding answer.

[0052] Considering functional atomization, the workflow of project review is more flexible. Multiple first-class agents can be set up, and each first-class agent is used to generate a target review instruction for a type of project review points. The same first-class agent can be reused for the same type of project review points.

[0053] The above-mentioned first-class intelligent agent can be an intelligent agent formed based on a combination of a large language model and a Prompt instruction. The Prompt instruction in the first-class intelligent agent is used to guide the large language model to output the corresponding target review instruction. Different Prompt instructions can be combined with the large language model to construct different first-class intelligent agents, and output different target review instructions.

[0054] Specifically, the project review key points information and review guidance information can be input into the first type of intelligent agent, and the project review key points information and review guidance information are generated and processed by the first type of intelligent agent to generate review instructions corresponding to the project review key points information. The number of the above-mentioned target review instructions can be one or more, and the project review key points information can include one or more project review key points, and each project review key point can correspond to one target review instruction.

[0055] The number of the first type of agents mentioned above may be related to the number of project review points. Specifically, the number of the first type of agents may be equal to the number of project review points, and one project review point corresponds to one preset agent. One project to be reviewed may correspond to one target review instruction set, and one target review instruction set includes at least one target review instruction.

[0056] The above-mentioned project review points may include name, demand, plan, deadline, budget, team, risk, emergency, benefit, management, etc. The above-mentioned target review instructions may include name review instructions, demand review instructions, plan review instructions, deadline review instructions, budget pre-review instructions, team review instructions, risk review instructions, emergency review instructions, benefit review instructions, management review instructions, etc.

[0057] 103. Construct at least one second-class intelligent agent based on the target review instruction set and the large language model.

[0058] In an embodiment of the present invention, after obtaining the target review instruction set, each target review instruction in the target review instruction set is combined with the large language model to form a corresponding second-type intelligent agent, which is only responsible for the review of the corresponding project review points in the project to be reviewed.

[0059] It can be understood that the input of a second type of intelligent agent is the content of the project to be reviewed, or the content of the project to be reviewed that corresponds to the key points of the project review.

[0060] 104. Based on at least one second-type intelligent agent, the project to be reviewed is reviewed to obtain a review result of the project to be reviewed.

[0061] In an embodiment of the present invention, after obtaining the second type of intelligent agent, the content of the project to be reviewed is input into the second type of intelligent agent, and the second type of intelligent agent automatically reviews the project to be reviewed, and outputs the review result of the project to be reviewed. In one embodiment, the review result is output as a suggestion to the reviewer for reference.

[0062] In a possible embodiment, the content corresponding to the project review points in the content of the project to be reviewed can be separated out and respectively input into the second type of intelligent agents corresponding to the project review points to obtain the review results corresponding to each project review point, and the review results corresponding to all project review points can be used as the review results of the project to be reviewed.

[0063] In an embodiment of the present invention, according to the content of the project to be reviewed, the review points and review guidance information corresponding to the project to be reviewed are matched; based on the project review points information and review guidance information, a target review instruction set of the project to be reviewed is generated by multiple first-class agents, and the target review instruction set includes at least one target review instruction; based on the target review instruction set and the large language model, at least one second-class agent is constructed, and each second-class agent corresponds to at least one target review instruction; based on the at least one second-class agent, the project to be reviewed is reviewed to obtain the review result of the project to be reviewed. The target review instruction of the project to be reviewed is generated by the first-class agent according to the project review points information and review guidance information, and the second-class agent is constructed by combining the target review instruction and the large language model. The second-class agent is used to automatically review the project to be reviewed, which does not rely on manual review and can improve the accuracy of the review result.

[0064] It is understandable that in the specific implementation of this application, related data such as project data, evaluation standard data, evaluation opinion data, etc. are involved. When the embodiments in this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data, as well as the training, deployment and calling of large language models, must comply with relevant laws, regulations and standards of relevant countries and regions.

[0065] Optionally, before the step of constructing a target review instruction set for the project to be reviewed through the first type of intelligent agent based on the project review key point information and the review guidance information, the number of project review key points can be determined based on the project review key point information; the number of first type of intelligent agents to be constructed can be determined based on the number of project review key points; and multiple intelligent agents corresponding to the number of first type of intelligent agents to be constructed can be constructed to obtain multiple first type intelligent agents.

[0066] In an embodiment of the present invention, the above-mentioned first-class intelligent agent can be understood as an instruction generator or an instruction editor. After the project review key information and review guidance information are input into the first-class intelligent agent, the corresponding project review instructions are generated by the first-class intelligent agent.

[0067] The above-mentioned project review key points information may include at least one project review key point. The number of project review key points included in the project review key points information can be counted, and the number of project review key points can be used as the number of intelligent entities to be constructed. The corresponding number of intelligent entities can be constructed through the number of intelligent entity construction. These intelligent entities are the preset intelligent entities of the large language model.

[0068] The above-mentioned intelligent agent can be an instruction generator, and the above-mentioned instruction generator can be an instruction generator based on a recurrent neural network RNN ​​or a long short-term memory network LSTM, which can generate corresponding review instructions according to project review key information and review guidance information.

[0069] The above-mentioned intelligent agent may also be an instruction editor. The above-mentioned instruction editor may provide an instruction editing page to the user. The user edits the instruction through the instruction editing page to obtain the corresponding review instruction.

[0070] In a possible embodiment, the above-mentioned first-category intelligent agent can be an intelligent agent formed based on a combination of a large language model and a Prompt instruction. The Prompt instruction in the first-category intelligent agent is used to guide the large language model to output corresponding target review instructions. Different Prompt instructions can be combined with the large language model to construct different first-category intelligent agents, which output different target review instructions.

[0071] After obtaining the number of project review points, a first type of intelligent agent with the same number of project review points can be constructed for the large language model, and then the target review instructions corresponding to the project to be reviewed can be separately constructed through the preset intelligent agent, thereby improving the accuracy of each target review instruction.

[0072] In a possible embodiment, the target review instructions for the output of the first type of intelligent agent may include roles-skills-workflows-examples, forming a COT (Chain Of Thought) thinking chain + BROKE framework (a prompt engineering framework) to guide the output of the second type of intelligent agent. The BROKE framework includes Background: providing sufficient background information so that the large language model can understand the context of the problem. Role: setting a specific role so that the large language model can generate a response based on the role. Objectives: clarify the task objectives so that the large language model knows clearly what needs to be achieved. Key Results: define key and measurable results so that the large language model knows how to measure the completion of the goals. Evolve: test the results through experiments and adjustments, and optimize as needed. In the target review instructions, a clear and typical few-shot can also be constructed, and example specifications are given in the target review instructions. In the target review instructions, a hierarchical prompt format can also be used, for example, using markdown (a lightweight markup language) format as a generation template, and the hierarchical prompt format can reduce the difficulty of model understanding. In the target review instructions, jargon translation can also be carried out, such as words such as "xinchuang", and the explanation of special terms can be injected to understand the context.

[0073] In a possible embodiment, the output requirements of the first type of agent may be:

[0074] Data interface definition: fixed JSON data structure, consistent interfaces between upstream and downstream intelligent entities, thus ensuring collaboration.

[0075] Prompt word classification: distinguish between tool-type and business-type, so as to facilitate management and locate problems.

[0076] Multiple rounds of robustness testing: The test phase runs the Prompt instruction multiple times to observe and optimize performance.

[0077] Optionally, in the step of constructing at least one second-category intelligent agent based on the target review instruction set and the large language model, the large language model can be encapsulated based on each target review instruction in the review instruction set to obtain a corresponding second-category intelligent agent.

[0078] In the embodiment of the present invention, after obtaining the target review instruction set, each target review instruction in the target review instruction set can be combined with the large language model to construct the corresponding second-class agent. It should be noted that the above-mentioned target review instruction is equivalent to the Prompt instruction of the large language model, and the large language model in the second-class agent can be the same as or different from the large language model in the first-class agent. The first-class agent is used for the generation and processing of the review instruction, and the second-class agent is used for the review processing of the project to be reviewed.

[0079] Optionally, in the step of reviewing the project to be reviewed based on at least one second-category intelligent agent to obtain the review result of the project to be reviewed, if there are multiple second-category intelligent agents, the multiple second-category intelligent agents can be arranged in the context order of the project to be reviewed to obtain the workflow of the second-category intelligent agents; based on the workflow, the project establishment plan of the project to be reviewed is reviewed to obtain the review result of the project to be reviewed.

[0080] In the embodiment of the present invention, if the number of the second type of intelligent agent is one, the second type of intelligent agent is directly used to review the project to be reviewed to obtain the review result of the project to be reviewed.

[0081] If there are multiple second-category agent data, the contextual information of each project review point in the content of the project to be reviewed can be considered to construct a workflow, and the project to be reviewed can be reviewed through the workflow to obtain the review results of the project to be reviewed.

[0082] The second-category agent is determined according to the number of target review instructions. Each target review instruction is combined with a large language model to obtain a second-category agent. In this way, multiple second-category agents can be subdivided, the tasks of each second-category agent can be simplified, and the functional definition can be vertically defined for easy management. For high-demand second-category agents, their generated content is reused as context background to other second-category agents, thereby achieving the reuse of second-category agents, which can also be understood as the reuse of target review instructions. For project review points that span chapters and multiple structures, the corresponding multiple second-category agents can be combined into a workflow, and the review output of the review project can be performed through the workflow.

[0083] Optionally, after the step of reviewing the items to be reviewed based on at least one second-category intelligent agent and obtaining the review results of the items to be reviewed, historical review data can also be obtained; based on the third-category intelligent agent, the review results are compared with the historical review data, and based on the comparison results, it is determined whether the first-category intelligent agent needs to be updated; if the first-category intelligent agent needs to be updated, the first-category intelligent agent is updated based on the comparison results.

[0084] In an embodiment of the present invention, the above-mentioned historical review data may be data that experts believe is correctly reviewed, and the above-mentioned third type of intelligent agent may be a discriminator, which is used to determine whether the review result is the same or similar to the historical review data, and further determine whether the target review instruction is correct.

[0085] In a possible embodiment, the third type of agent may be an agent based on a combination of a large language model and a Prompt instruction, and the Prompt instruction in the third type of agent is used to guide whether the review result output by the large language model is the same or similar to the historical review data.

[0086] It is understandable that when the review results of the project to be reviewed are significantly different from the historical review data, it can be considered that the target review instructions are inaccurate, resulting in poor performance of the generated second-type agent. Therefore, the first-type agent can be updated so that the first-type agent generates more accurate target review instructions, and then combined with the large language model to form a second-type agent with better performance. When the review results of the project to be reviewed are the same or similar to the historical review data, there is no need to update the first-type agent.

[0087] When the first-class intelligent agent needs to be updated, it can be manually updated or automatically updated. Manual update is to analyze the comparison results through experts, and adjust the input prompt instructions of the first-class intelligent agent according to the comparison results to form a new first-class intelligent agent. Automatic update is to train and adjust the parameters in the first-class intelligent agent through a training data set containing positive samples and negative samples to obtain a new first-class intelligent agent. The update of the above-mentioned first-class intelligent agent can also be achieved by using a reinforcement learning strategy. A larger reward value is set for a successful comparison, and a smaller or negative reward value is set for a failed comparison. After each review, the parameters in the first-class intelligent agent are adjusted with the maximum reward value as the goal. Through reinforcement learning, the first-class intelligent agent can be updated in real time.

[0088] When the first type of intelligence needs to be updated, the review results that are different from the historical review data can be determined in the comparison results, and the corresponding second type of intelligent agent can be determined through the review results, and the corresponding first type of intelligent agent can be determined through the corresponding second type of intelligent agent. At the same time, the data that failed to be compared are screened out from the comparison results as negative samples, and the data that successfully compared are used as positive samples, so as to obtain a training data set, and the training data set is used to train and adjust the corresponding first type of intelligent agent to obtain an updated first type of intelligent agent, and then the updated first type of intelligent agent is used to generate a new target review instruction, and the new target review instruction is combined with the large language model to construct a new second type of intelligent agent, and the new second type of intelligent agent is used to review the items to be reviewed.

[0089] Optionally, in the step of updating the first-category intelligent agent based on the comparison result, the first-category intelligent agent corresponding to the review result that failed the comparison can be determined as the first-category intelligent agent to be updated based on the comparison result, and the review result that failed the comparison and the corresponding historical review data can be determined as the update data corresponding to the first-category intelligent agent to be updated, and each first-category intelligent agent to be updated corresponds to one update data; based on the update data, the first-category intelligent agent to be updated is updated.

[0090] In the embodiment of the present invention, the comparison result includes a comparison failure and a comparison success. The output of each second-class intelligent agent can correspond to a comparison failure or a comparison success. When the output of a second-class intelligent agent is corresponded to a comparison failure, it means that the target review instruction corresponding to the second-class intelligent agent is incorrect, that is, the target review instruction generated by the corresponding first-class intelligent agent is incorrect, and the first-class intelligent agent is determined as the first-class intelligent agent to be updated. The output (review result) of the second-class intelligent agent and the corresponding historical review data are determined as the update data corresponding to the first-class intelligent agent to be updated. The function of the update data is to update the first-class intelligent agent so that the first-class intelligent agent generates a new target review instruction, and the new target review instruction is combined with the large language model to form a new second-class intelligent agent. The new second-class intelligent agent reviews the project to be reviewed, and the output of the new second-class intelligent agent is the same or similar to the corresponding historical review data.

[0091] The above-mentioned update may be an adjustment to the Prompt instruction of the first type of intelligence, or an adjustment to the parameters of the first type of intelligence.

[0092] Optionally, in the step of updating the first type of intelligent agent to be updated based on the updated data, the error value between the review result and the historical review data can be calculated; and the first type of intelligent agent to be updated is updated with the minimum error value as the optimization goal.

[0093] In the embodiment of the present invention, an error function may be used to calculate the error value between the review result and the historical review data. The error function may be a cross entropy error function or an average error function.

[0094] The above error value is used to indicate the difference between the review result and the historical review data. With the minimum error value as the optimization goal, the parameters or prompt instructions of the first agent to be updated are updated and adjusted. In this way, the difference between the updated review result and the historical review data can be reduced, and a review result close to the correct one can be obtained.

[0095] like Figure 2 As shown, Figure 2 is a flowchart of another project review method provided by an embodiment of the present invention. Figure 2In the stage of building the intelligent agent, the project review points are extracted through the "Guidelines for the Preparation of Project Establishment Plans" and the "Detailed Rules for the Approval of Government Information Project Establishment". The review guidance information is provided by technical experts, and the target review instructions are generated based on the first type of intelligent agent in combination with the project review points and review guidance information. The second type of intelligent agent model is formed by combining the target review instructions with the large language model. The second type of intelligent agent model is used to review the review project to obtain the review results. The historical review data is obtained, and the review results are judged based on the historical review data to determine whether they meet the standards. If they meet the standards, the standard requirements are output; if they do not meet the standard requirements, the questions and standard requirements are given. The business experts are combined to review the given questions and standard requirements. The review opinions are output. The corresponding target review instructions are judged to be correct based on the review opinions. If they are correct, the target review instructions are determined as valid review instructions; if they are incorrect, the corresponding first type of intelligent agent is updated, so that the target review instructions output by the first type of intelligent agent can be corrected and the review can continue. In the agent use stage, based on the project review key information and review guidance information, multiple first-class agents can generate a target review instruction set for the project to be reviewed; based on the target review instruction set and the large language model, at least one second-class agent can be constructed; based on at least one second-class agent, the project to be reviewed is reviewed to obtain the review result of the project to be reviewed. The second agent reviews the project plan to be reviewed to determine whether the project to be reviewed meets the standard. If it meets the standard, the standard requirements are output; if it does not meet the standard requirements, questions and standard requirements are given, and the questions and standard requirements are output to the user, so that the user can modify the project to be reviewed according to the questions and standard requirements.

[0096] like Figure 3 As shown, an embodiment of the present invention provides a project review device, which includes:

[0097] The first acquisition module 301 is used to match the review points and review guidance information corresponding to the project to be reviewed according to the content of the project to be reviewed;

[0098] A generating module 302, configured to generate a target review instruction set for the project to be reviewed through a plurality of first-type agents based on the project review key information and the review guidance information, wherein the target review instruction set includes at least one target review instruction;

[0099] A first construction module 303 is used to construct at least one second-class intelligent agent based on the target review instruction set and the large language model, each second-class intelligent agent corresponds to at least one target review instruction;

[0100] The review module 304 is used to review the project to be reviewed based on at least one of the second-type intelligent agents to obtain a review result of the project to be reviewed.

[0101] Optionally, the device further comprises:

[0102] A first processing module, configured to determine the number of project review key points based on the project review key point information;

[0103] A second processing module is used to determine the number of first-type intelligent agents to be constructed based on the number of project review points;

[0104] The second construction module is used to construct multiple agents corresponding to the construction number of the first type of agents, so as to obtain multiple first type of agents.

[0105] Optionally, the first construction module 303 is further used to encapsulate the large language model based on each target review instruction in the review instruction set to obtain a corresponding second-type intelligent agent.

[0106] Optionally, the review module 304 is also used to, if there are multiple second-category intelligent agents, arrange the multiple second-category intelligent agents in the context order of the project to be reviewed to obtain the workflow of the second-category intelligent agents; based on the workflow, review the project establishment plan of the project to be reviewed to obtain the review result of the project to be reviewed.

[0107] Optionally, the device further comprises:

[0108] The second acquisition module is used to acquire historical review data;

[0109] A comparison module, for comparing the review result with the historical review data based on the third type of agent, and determining whether the first type of agent needs to be updated according to the comparison result;

[0110] An updating module is used to update the first type of agents based on the comparison result if the first type of agents need to be updated.

[0111] Optionally, the update module is also used to determine the first-category intelligent agent corresponding to the review result that failed the comparison as the first-category intelligent agent to be updated based on the comparison result, and determine the review result that failed the comparison and the corresponding historical review data as update data corresponding to the first-category intelligent agent to be updated, each first-category intelligent agent to be updated corresponds to one update data; based on the update data, the first-category intelligent agent to be updated is updated.

[0112] Optionally, the updating module is further used to calculate the error value between the review result and the historical review data; and to update the first type of intelligent agent to be updated with the minimum error value as the optimization goal.

[0113] like Figure 4 As shown, an embodiment of the present invention further provides an electronic device, including a processor, and the processor can execute any one of the above-mentioned project review methods.

[0114] Specifically, it includes a processor 401 and a memory 402, and a computer program for executing the project review method stored in the memory 402 and capable of running on the processor 401, wherein:

[0115] The processor 401 runs the computer program of the project review method stored in the memory 402 to perform the following steps:

[0116] According to the content of the project to be reviewed, match the review points and review guidance information corresponding to the project to be reviewed;

[0117] Based on the project review key point information and the review guidance information, generating a target review instruction set for the project to be reviewed through a plurality of first-type agents, wherein the target review instruction set includes at least one target review instruction;

[0118] Based on the target review instruction set and the large language model, construct at least one second-type intelligent agent, each second-type intelligent agent corresponds to at least one of the target review instructions;

[0119] Based on at least one of the second-type intelligent agents, the project to be reviewed is reviewed to obtain a review result of the project to be reviewed.

[0120] Optionally, before constructing the target review instruction set of the to-be-reviewed project by the first type of agent based on the project review key points information and the review guidance information, the method executed by the processor 401 further includes:

[0121] Based on the project review key points information, determine the number of project review key points;

[0122] Based on the number of project review points, determine the number of first-type intelligent agents to be constructed;

[0123] Construct a plurality of agents corresponding to the number of the first type of agents to obtain a plurality of the first type of agents.

[0124] Optionally, the step of constructing at least one second-type agent based on the target review instruction set and the large language model executed by the processor 401 includes:

[0125] Based on each of the target review instructions in the review instruction set, the large language model is encapsulated to obtain a corresponding second-category intelligent agent.

[0126] Optionally, the processor 401 performs the review of the to-be-reviewed project based on at least one of the second-type intelligent agents to obtain the review result of the to-be-reviewed project, including:

[0127] If there are multiple second-type agents, then arranging the multiple second-type agents according to the context order of the items to be reviewed to obtain a workflow of the second-type agents;

[0128] The project proposal of the project to be reviewed is reviewed based on the workflow to obtain the review result of the project to be reviewed.

[0129] Optionally, after reviewing the project to be reviewed based on at least one of the second-type agents to obtain a review result of the project to be reviewed, the method executed by the processor 401 further includes:

[0130] Obtain historical review data;

[0131] Based on the third type of agents, the review result is compared with the historical review data, and whether the first type of agents needs to be updated according to the comparison result;

[0132] If the first type of agent needs to be updated, the first type of agent is updated based on the comparison result.

[0133] Optionally, the updating of the first type of agent based on the comparison result performed by the processor 401 includes:

[0134] Based on the comparison result, the first-category agent corresponding to the review result that failed the comparison is determined as the first-category agent to be updated, and the review result that failed the comparison and the corresponding historical review data are determined as update data corresponding to the first-category agent to be updated, and each first-category agent to be updated corresponds to one update data;

[0135] Based on the update data, the first type of agent to be updated is updated.

[0136] Optionally, the updating of the first type of agents to be updated based on the update data performed by the processor 401 includes:

[0137] Calculating the error value between the review result and the historical review data;

[0138] The first type of agent to be updated is updated with the minimum error value as the optimization goal.

[0139] An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the computer program implements the various processes of the project review method or the application-side project review method provided in the embodiment of the present invention, and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0140] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0141] The above disclosure is only the preferred embodiment of the present invention, which certainly cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

1. A project review method, characterized in that: The method comprises the following steps: According to the content of the project to be reviewed, match the review points and review guidance information corresponding to the project to be reviewed; Based on the project review key point information and the review guidance information, generating a target review instruction set for the project to be reviewed through a plurality of first-type agents, wherein the target review instruction set includes at least one target review instruction; Based on the target review instruction set and the large language model, construct at least one second-type intelligent agent, each second-type intelligent agent corresponds to at least one of the target review instructions; Based on at least one of the second-type intelligent agents, the project to be reviewed is reviewed to obtain a review result of the project to be reviewed.

2. The project review method according to claim 1, characterized in that: Before constructing the target review instruction set of the to-be-reviewed project by the first type of agent based on the project review key points information and the review guidance information, the method further includes: Based on the project review key points information, determine the number of project review key points; Based on the number of project review points, determine the number of first-type intelligent agents to be constructed; Construct a plurality of agents corresponding to the number of the first type of agents to obtain a plurality of the first type of agents.

3. The project review method according to claim 1, characterized in that: The step of constructing at least one second type of intelligent agent based on the target review instruction set and the large language model comprises: Based on each of the target review instructions in the review instruction set, the large language model is encapsulated to obtain a corresponding second-category intelligent agent.

4. The project review method according to claim 1, characterized in that: The step of reviewing the project to be reviewed based on at least one of the second-type intelligent agents to obtain a review result of the project to be reviewed includes: If there are multiple second-type agents, then arranging the multiple second-type agents according to the context order of the items to be reviewed to obtain a workflow of the second-type agents; The project proposal of the project to be reviewed is reviewed based on the workflow to obtain the review result of the project to be reviewed.

5. The project review method according to claim 1, characterized in that: After the project to be reviewed is reviewed based on at least one of the second-type intelligent agents to obtain a review result of the project to be reviewed, the method further includes: Obtain historical review data; Based on the third type of agents, the review result is compared with the historical review data, and whether the first type of agents needs to be updated according to the comparison result; If the first type of agent needs to be updated, the first type of agent is updated based on the comparison result.

6. The project review method according to claim 5, characterized in that: The updating of the first type of agents based on the comparison result comprises: Based on the comparison result, the first-category agent corresponding to the review result that failed the comparison is determined as the first-category agent to be updated, and the review result that failed the comparison and the corresponding historical review data are determined as update data corresponding to the first-category agent to be updated, and each first-category agent to be updated corresponds to one update data; Based on the update data, the first type of agent to be updated is updated.

7. The project review method according to claim 6, characterized in that: The updating of the first type of agents to be updated based on the update data includes: Calculating the error value between the review result and the historical review data; The first type of agent to be updated is updated with the minimum error value as the optimization goal.

8. A project review device, characterized in that: The project review device comprises: A first acquisition module is used to match the review points and review guidance information corresponding to the project to be reviewed according to the content of the project to be reviewed; A generating module, configured to generate a target review instruction set for the project to be reviewed through a plurality of first-type agents based on the project review key point information and the review guidance information, wherein the target review instruction set includes at least one target review instruction; A first construction module is used to construct at least one second-class intelligent agent based on the target review instruction set and the large language model, each second-class intelligent agent corresponds to at least one target review instruction; The review module is used to review the items to be reviewed based on at least one of the second-type intelligent agents to obtain review results of the items to be reviewed.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps in the project review method as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the project review method according to any one of claims 1 to 7 are implemented.

Images