Information integration large model interaction method based on RAG framework

Abstract

The invention relates to an information integration large model interaction method based on an RAG framework. The method comprises the steps that work logs uploaded by multiple groups of workers are received, and a work log database is established; constructing a multi-level agent; generating a first-level system summary according to the work log database in combination with the cue word corresponding to the first-level agent, constructing a first-level database according to the first-level system summary, generating each level of system summary according to the superior database in combination with the cue word corresponding to each level of agent, and constructing each level of database according to each level of system summary until the last level of database; and according to the corresponding agent level, performing data retrieval in a database corresponding to the agent level by adopting an RAG framework, and generating a summary report of the management personnel by adopting a large model in combination with a cue word corresponding to the agent level. According to the invention, through cooperative work of a plurality of agent models and combination of an artificial intelligence technology, each agent generates a customized summary according to demands of managers of different levels, so that the summary efficiency and precision are greatly improved.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of artificial intelligence technology, and in particular, to an interaction method for an information integration large model based on the RAG framework. Background Art

[0002] Artificial intelligence (AI) technology, especially large language models, has made remarkable progress in recent years. Through deep learning and training with massive amounts of data, AI models have demonstrated powerful capabilities in the field of natural language processing (NLP), being able to understand, generate, and process complex text data. Large models such as GPT-3 and BERT, with their excellent text generation and understanding capabilities, have been widely applied in multiple industries, showing advantages over traditional technologies in various aspects from automatic translation to intelligent customer service and text creation.

[0003] In the process of traditional work summary, there is a lot of manual participation, and it often takes a large amount of time and effort to collect, organize, and analyze relevant data. This not only affects work efficiency, but also due to the different needs and concerns of each administrator, manual summaries are often lacking in personalization and difficult to precisely meet the needs of managers at different levels. In addition, most existing work summary systems have not fully utilized AI technology, resulting in insufficient accuracy and pertinence of the summary content and unable to effectively improve the efficiency of management decision-making.

[0004] Therefore, it is necessary to improve one or more problems existing in the above-mentioned related technical solutions.

[0005] It should be noted that this part aims to provide background or context for the technical solution of the present invention stated in the claims. The description herein is not admitted to be prior art just because it is included in this part. Summary of the Invention

[0006] The purpose of the present invention is to provide an interaction method for an information integration large model based on the RAG framework, thereby at least to a certain extent solving one or more problems caused by the limitations and defects of the related technologies.

[0007] The present invention provides an interaction method for an information integration large model based on the RAG framework, including:

[0008] Receiving work logs uploaded by multiple groups of staff and establishing a work log database;

[0009] Constructing multi-level agents, where each level of agent corresponds to the relative permissions of different managers, and generating corresponding prompt words for each level of agent;

[0010] Generate a first-level system summary based on the work log database in combination with the prompt words corresponding to the first-level agent, and construct a first-level database according to the first-level system summary. Each level of system summary is generated based on the upper-level database in combination with the prompt words corresponding to each level of agent, and each level of database is constructed according to the system summary of each level until the last-level database;

[0011] Match the corresponding agent level according to the relative permissions of the management personnel, perform data retrieval in the database corresponding to the agent level using the RAG framework, and generate a summary report for the management personnel using a large model in combination with the prompt words corresponding to the agent level.

[0012] Optionally, the steps of matching the corresponding agent level according to the relative permissions of the management personnel, performing data retrieval in the database corresponding to the agent level using the RAG framework, and generating a summary report for the management personnel using a large model in combination with the prompt words corresponding to the agent level include:

[0013] Transmit the prompt words as input to the large model, and construct a framework for the work summary report through the natural language processing and generation capabilities of the large model.

[0014] Optionally, the steps of transmitting the prompt words as input to the large model and constructing a framework for the work summary report through the natural language processing and generation capabilities of the large model include:

[0015] The large model is fine-tuned using LoRA with a small dataset.

[0016] Optionally, the steps of transmitting the prompt words as input to the large model and constructing a framework for the work summary report through the natural language processing and generation capabilities of the large model include:

[0017] Determine whether the management personnel are satisfied with the framework of the work summary report. When it is determined that they are satisfied, output the corresponding work summary report. When it is determined that they are not satisfied, determine whether it is necessary to check the upper-level database.

[0018] Optionally, the steps of determining whether it is necessary to check the upper-level database when it is determined that they are not satisfied include:

[0019] When it is determined that they are not satisfied, generate prompt words for intent recognition, and determine whether it is necessary to check the upper-level database through the large model in combination with the prompt words for intent recognition.

[0020] Optionally, the steps of determining whether it is necessary to check the upper-level database when it is determined that they are not satisfied include:

[0021] When it is determined that it is necessary to check the upper-level database, according to the new requirements, start a multi-level data retrieval mechanism and initiate a query request to the upper-level database through a preset query path.

[0022] Optionally, the step of, when it is determined that it is necessary to query the superior database, starting a multi-level data retrieval mechanism according to the new requirements and sending a query request to the superior database through a preset query path includes:

[0023] When sending a query request to the superior database, a temporary proxy is generated based on the current proxy level of the manager, a corresponding prompt word is generated according to the temporary proxy, and a work summary report is generated again in combination with the retrieval data of the superior database.

[0024] Optionally, the step of, when sending a query request to the superior database, generating a temporary proxy based on the current proxy level of the manager, generating a corresponding prompt word according to the temporary proxy, and generating a work summary report again in combination with the retrieval data of the superior database includes:

[0025] Combining the keywords in the manager's feedback and the prompt words generated by the temporary proxy, the retrieval data is transmitted to the large model, and a work report that meets the personalized needs of the manager is generated through the processing and analysis of the large model.

[0026] Optionally, the large model uses the ChatGLM3-6B model.

[0027] The technical solution provided by the present invention may include the following beneficial effects:

[0028] In the present invention, through the collaborative work of multiple proxy models and the combination of artificial intelligence technology, each proxy generates a customized summary according to the needs of managers at different levels, and the management level can obtain accurate and personalized work summaries, greatly improving the summary efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The drawings herein are incorporated into the specification and constitute a part of this specification, showing embodiments consistent with the present invention, and are used together with the specification to explain the principles of the present invention. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts.

[0030] Figure 1 A flowchart showing the information integration large model interaction method based on the RAG framework in an exemplary embodiment of the present invention;

[0031] Figure 2 A schematic diagram showing the text description of the manager in an exemplary embodiment of the present invention;

[0032] Figure 3 A schematic diagram showing a partial work log data set in an exemplary embodiment of the present invention;

[0033] Figure 4 Schematic diagram showing the system report generation prompt words in an exemplary embodiment of the present invention;

[0034] Figure 5 Schematic diagram showing the judgment working process of the information integration large model interaction method based on the RAG framework in an exemplary embodiment of the present invention;

[0035] Figure 6 Schematic diagram showing the system report generation in an exemplary embodiment of the present invention;

[0036] Figure 7 Schematic diagram showing the database retrieval in an exemplary embodiment of the present invention;

[0037] Figure 8 Schematic diagram showing the preliminary personalized report generation prompt words in an exemplary embodiment of the present invention;

[0038] Figure 9 Schematic diagram showing the intention recognition in an exemplary embodiment of the present invention;

[0039] Figure 10 Schematic diagram showing the process of querying the superior database in an exemplary embodiment of the present invention;

[0040] Figure 11 Schematic diagram showing the temporary agent generation prompt words in an exemplary embodiment of the present invention;

[0041] Figure 12 Schematic diagram showing the personalized report generation prompt words in an exemplary embodiment of the present invention;

[0042] Figure 13 Schematic diagram showing the opinion feedback in an exemplary embodiment of the present invention. Detailed implementation manners

[0043] Example embodiments will now be described more fully with reference to the accompanying drawings. However, the example embodiments can be implemented in various forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this invention will be more complete and comprehensive, and will fully convey the concept of the example embodiments to those skilled in the art. The features, structures, or characteristics described may be combined in any suitable manner in one or more embodiments.

[0044] In addition, the drawings are only schematic illustrations of the embodiments of the present invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and thus their repeated description will be omitted. Some of the block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0045] The present invention provides an information integration large model interaction method based on the RAG framework, referring to Figure 1 as shown in

[0046] Step S100: Receive work logs uploaded by multiple groups of staff and establish a work log database.

[0047] Step S200: Construct multi-level agents, where each level of agent corresponds to the relative authority of different managers, and generate corresponding prompt words for each level of agent.

[0048] Step S300: Generate a first-level system summary based on the work log database in combination with the prompt words corresponding to the first-level agent, and construct a first-level database based on the first-level system summary. Each level of system summary is generated based on the superior database in combination with the prompt words corresponding to each level of agent, and each level of database is constructed based on the system summary of each level until the last-level database.

[0049] Step S400: Match the corresponding agent level according to the relative authority of the manager, perform data retrieval in the database corresponding to the agent level using the RAG framework, and generate the summary report of the manager using the large model in combination with the prompt words corresponding to the agent level.

[0050] It should be understood that RAG is the abbreviation of "Retrieval-Augmented Generation", that is, the retrieval enhanced generation framework. The RAG framework aims to combine information retrieval technology with a Natural Language Generation (NLG) model to improve the performance and effect of a generative artificial intelligence system. Its core idea is to utilize an external knowledge source for retrieval when generating text, and integrate the retrieved relevant information into the generation process, so that the generated text is more accurate, rich, and relevant.

[0051] It should also be understood that by constructing a work log dataset for power operations and personalized design of text agents for multi-level management personnel, the accuracy and pertinence of work summary content have been improved. This method generates different report generation prompt words for each management agent, combines historical data and task requirements to ensure the efficient completion of summary reports. When the management personnel's feedback requirements change, the system can extract the keywords in the feedback, generate new personalized prompt words and adjust the work summary content. If the management personnel have higher requirements for the detail level of the report, the system can retrieve more detailed data by querying the superior database and generate a more in-depth summary. In addition, the system also integrates the ChatGLM3-6B and sentence-BERT (Bidirectional Encoder Representations from Transformers) models, and uses cosine similarity to match relevant data, thereby improving the accuracy of work summary and the efficiency of management decision-making.

[0052] It should also be understood that a work log dataset has been established. The present invention constructs a work log dataset for power operations, which is the data basis for the summary tasks of the present invention.

[0053] It should also be understood that text agents and system agents for multi-level management personnel are designed. The present invention designs separate personalized prompt words for management personnel in different positions, and makes personalized designs in terms of detail level, project preferences, etc. The generated prompt words are more targeted, and thus the accuracy of the summary content can be guaranteed. For example Figure 2 is an example diagram of text agents for multi-level management personnel. Considering that managers may put forward different temporary requirements at any time, which will affect the generation of multi-level reports, we have not only designed separate personalized agents, but also designed system prompt words for generating multi-level databases (system prompt words are fixed prompt words that are not affected by the feedback of management personnel). A general and relatively comprehensive work report is generated according to the system prompt words.

[0054] It should also be understood that the present invention generates a system work summary according to the system agent, and the operation steps are as follows: The system work summary is summarized according to time. For example, the secondary report is summarized every three days.

[0055] Data retrieval: Since, for the integrity of work log information, we have processed the data in chunks according to the work logs of each person per day, the retrieval of the system work summary can be directly carried out according to the metadata label information of the work logs. The retrieved data is combined with the pre-set system agent to generate new prompt words, and then the API of the large model is called for response to generate the summary for these three days.

[0056] It should also be understood that if the management is not satisfied with the system summary, the present invention will generate new prompt words based on the feedback of the management and in combination with the personalized text agent.

[0057] Keyword extraction: The large model extracts keywords from the management's feedback and combines them with the original fixed agent of the management to generate a new temporary agent.

[0058] Prompt word generation: Retrieve the work logs of the corresponding staff according to the keywords of the temporary agent. Then, in combination with the system agent, generate prompt words and input them into the called large model. In order to find all the data most similar to the keywords, generate new prompt words in combination with the temporary agent.

[0059] Work summary generation: Finally, pass the prompt words to the large model to generate a work summary report that meets the personalized needs of the management.

[0060] It should also be understood that if the management has higher requirements for the level of detail of the report, the present invention will recursively query the superior database to retrieve more detailed data to meet the needs of the management.

[0061] Perform intention recognition on the newly generated temporary agent to determine whether it is necessary to recursively query the superior database. If it is determined that a recursive query is required, query the superior database during data retrieval so that more detailed and specific work log data can be obtained. Thus, generate a detailed work report for a certain task.

[0062] It should also be understood that, for example, in step S100, a high-quality work log dataset of power system staff is constructed.

[0063] It should also be understood that, for example, step S100 constructs a high-quality work log dataset of power system staff.

[0064] Adopt a work log structure that is continuous in time and logical before and after, making the dataset closer to the actual work situation and task execution process. The specific construction idea and content format are as follows:

[0065] Construction idea: The dataset is based on a work team consisting of N groups, with X staff members in each group. To ensure the coherence and rationality of tasks, a main task is arranged for each group every Y days. This task is split into multiple subtasks according to the division of labor of each staff member and completed jointly by X people, and each person reports their working hours for three days. In this way, the task arrangements and work logs in the dataset can reflect the collaboration and task allocation processes in actual work, while ensuring the continuity of work content and task progress.

[0066] Content format:

[0067] The work log of each staff member contains the following key information:

[0068] Personality traits: The basic personality and work habits of the staff (such as being careful, proactive, etc.), which help the model to match and analyze tasks according to the characteristics of the personnel.

[0069] Time: Record the specific working dates to ensure the time series of the log, which is convenient for subsequent analysis and summary.

[0070] Group number: Indicate the group to which the staff member belongs, which is convenient for task allocation and tracking of team collaboration.

[0071] Name: Record the name of each staff member to help distinguish individual work records.

[0072] Work content: Record the work tasks and activities performed by the staff member on specific dates.

[0073] Problems encountered: Record the specific problems that occurred during the work, which helps with subsequent feedback on work summary and problem-solving solutions.

[0074] Specific solutions: Describe the solutions and measures to solve the problems, as well as the specific steps taken.

[0075] Feedback after implementation: Record the feedback on the effect after implementing the solution, and evaluate the task completion situation and the effectiveness of the solution.

[0076] As Figure 3 shown, these contents will be stored in the JSON file format, which is convenient for structured processing and subsequent data analysis. To ensure the integrity of the data and facilitate efficient retrieval, we will store the work logs in chunks according to the daily work records of each person after vectorization. In this way, the work logs of each person can be clearly segmented in time, ensuring the continuity, integrity, and efficiency of the data. The storage structure of the dataset can intuitively show how this refined work log dataset provides a strong foundation for subsequent report generation.

[0077] It is also necessary to understand that a system agent is constructed and a text agent for multi-level management personnel is designed. In order to generate a summary report that can meet the personalized needs of management personnel, the present invention designs an independent text agent for each management personnel, which serves as the basis for generating new prompt words. These text agents can not only clearly convey the specific requirements of management personnel for the report, but also cover their personalized needs in aspects such as the level of detail and key points of the report content. For example, some management personnel may require the report to provide a higher-level overview, or to report in detail on certain key aspects of the power system, project progress, etc., for the convenience of decision-making and supervision. Each text agent needs to clearly clarify these requirements to ensure that the report generated by the large model can accurately meet the expectations and actual work needs of management personnel. As Figure 4 shown in the example of some text agents, it demonstrates how through personalized design, the specific requirements of each management personnel can be reflected through the generated new prompt words, and then report content that meets personalized needs can be generated.

[0078] It is also necessary to understand that a multi-level database is constructed according to the dataset and the system agent to generate a comprehensive system report. The present invention constructs a multi-level database that is not affected by the feedback of management personnel and the personalized report generation process when generating reports, ensuring the stability and consistency of the system report as the basic data. The system report is designed as the core component of the database, and its summary content requirements are relatively comprehensive to ensure the provision of basic and comprehensive work summary information. According to the construction idea of the work log dataset of power system staff, the generation frequency of the system report is closely related to information such as time and group. For example, a secondary report will generate a summary every three days for each group to ensure timely reflection of the work progress and task completion of each group. In this way, the system can provide comprehensive and reliable data support for subsequent personalized summaries and decision-making, avoiding distortion or omission of the system report due to the influence of individual feedback.

[0079] It is also important to understand that we have designed a process framework for work log summaries and proposed a high-efficiency interactive framework for a multi-agent work summary large model, which has improved the efficiency of work summaries. We have designed text agents (different agent text descriptions are designed for managers at different levels, so that work summaries that meet the personal requirements of different managers can be made for different managers). Temporary dialogues generate characteristic summaries. For temporary requests, the agents we designed are divided into two parts: fixed agents and temporary agents. Fixed agents will be dynamically added or deleted based on the number of searches. We check the superior knowledge base step by step and build a multi-level database based on the degree of generalization of the summary. When superior managers need a more detailed work summary, we will search the superior database step by step until the user's requirements are met. We have built a detailed work log of the staff as the basic data set, and based on the RAG framework, we have improved the knowledge limitations and illusion problems of the large model and enhanced the interactive performance of the large model.

[0080] It is also necessary to understand that to improve the efficiency of management decision-making: through the multi-agent hierarchical summarization method based on the RAG framework, personalized work summaries can be generated quickly and accurately, reducing the time required for managers to obtain information, thereby improving the efficiency and timeliness of management decisions.

[0081] Enhance the accuracy and pertinence of work summaries: By designing personalized prompts for managers in different positions and making dynamic adjustments based on manager feedback, we can ensure that the content of work summaries accurately meets the needs of managers and avoid problems with incomplete or biased information.

[0082] Improve task analysis and agent efficiency: Use large models to intelligently match and analyze agent background information (such as name, age, professional field, and task completion status), making task allocation and execution more efficient and able to adapt to changing needs in different work scenarios.

[0083] Responding to dynamic management needs: This system incorporates a flexible feedback mechanism. When managers raise temporary requirements or revisions, the system can quickly adjust the work summary and generate a new, personalized report. This flexibility ensures that work summaries are not only efficient but also accurately reflect the latest management needs.

[0084] Intelligent data processing and information integration: By integrating the ChatGLM3-6B large model and the sentence embedding representation model Sentence-BERT, the present invention can efficiently perform data retrieval and similarity analysis, accurately recall relevant data from massive work logs, avoid the tedious manual screening, and improve the intelligence level of work summaries.

[0085] By adopting the above-mentioned information integration large model interaction method based on the RAG framework, through the collaborative work of multiple proxy models and the combination of artificial intelligence technology, each proxy generates customized summaries according to the needs of managers at different levels. The management can obtain accurate and personalized work summaries, greatly improving the summary efficiency and accuracy.

[0086] Next, with reference to Figures 1 to 13 each step of the above-mentioned information integration large model interaction method based on the RAG framework in this exemplary embodiment will be described in more detail.

[0087] In some embodiments, with reference to Figure 5 and Figure 6 as shown, step S400 includes:

[0088] Step S500: Transmit the prompt as input to the large model, and construct a framework for the work summary report through the natural language processing and generation capabilities of the large model.

[0089] It should be understood that in order to achieve higher accuracy and professionalism in the database retrieval process, the present invention embeds a sentence embedding representation model: Sentence-BERT in the large model. Sentence-BERT is a natural language processing model based on BERT, which is an improvement of the pre-trained BERT network. It uses the Siamese Network and Triplet Network structures to output semantically valuable sentence embeddings (SentenceEmbedding), which can be compared using cosine similarity to extract text data whose similarity meets the threshold. As Figure 7 shown, for sentence embedding, first perform embedding on the task name to generate a fixed-length vector, and then perform average pooling operation. Calculate the average value of all vectors obtained by passing the sentence through the model, and use the average value as the sentence vector of each sentence. Finally, calculate the cosine similarity to obtain sentences whose similarity is greater than the set threshold, and recall the vector block where the sentence is located. The cosine similarity calculation formula is as follows:

[0090]

[0091] where A and B are two sentences, and Embedding is the vector generated by Sentence-BERT. The higher the similarity score, the stronger the semantic similarity between task descriptions. With the help of the embedding representation of Sentence-BERT and the cosine similarity calculation, we can efficiently retrieve and capture the vector blocks we need in the database, improving the response speed of the overall task.

[0092] In some embodiments, step S500 includes:

[0093] The large model is fine-tuned using a small dataset through LoRA.

[0094] It should be understood that in order to ensure the accuracy of intent recognition, the present invention uses a small dataset to fine-tune the large model through LoRA (Low-Rank Adaptation). Through this fine-tuning method, the model can perform more accurate keyword extraction in a specific field, thereby improving the retrieval accuracy of task-related data. The fine-tuned large model can identify and extract truly critical keywords, which will be used to generate more accurate prompt words, and then optimize the effects of data retrieval and report generation. Some examples of training datasets are shown as Figure 8 shown, demonstrating how the fine-tuned large model can determine from a specific context to ensure that the system can provide a work summary that meets the requirements.

[0095] Therefore, the designed prompt words contain three key parts: the judgment purpose, the judgment content, and the judgment basis. As Figure 9 shown, the judgment purpose is to clarify whether a detailed data backcheck is required for this operation, such as whether it involves task details or requires refined information; the judgment content is to analyze the specific requirements feedback by the management personnel, such as whether a higher-level summary or a detailed analysis is required; the judgment basis is to determine whether to retrieve more data in depth based on the known context information and work logs in the system, combined with the feedback from the management personnel. Through the comprehensive judgment of these three parts, the system can flexibly adjust whether to perform a backcheck in different scenarios, avoiding unnecessary data noise from interfering with the content of the final generated document, thereby ensuring the accuracy and pertinence of the report.

[0096] In some embodiments, referring to Figure 5 and Figure 10 shown, step S500 includes:

[0097] Step S600: Determine whether the management personnel are satisfied with the framework of the work summary report. When it is determined that they are satisfied, output the corresponding work summary report. When it is determined that they are not satisfied, determine whether it is necessary to backcheck the superior database.

[0098] It should be understood that as Figure 13 shown, if the management personnel are not satisfied with the system summary, the present invention will generate new prompt words based on the feedback from the management personnel and in combination with the personalized text agent. Perform intent recognition on the newly generated temporary agent to determine whether it is necessary to backcheck the superior database. If it is determined that a backcheck is required, retrieve the superior database during data retrieval so as to obtain more detailed and specific work log data. Thus, a detailed work report for a certain task is generated.

[0099] In some embodiments, referring to Figure 10As shown, step S600 includes:

[0100] Step S700: When it is determined that the result is not satisfactory, a prompt word for intention recognition is generated, and the large model is combined with the prompt word for intention recognition to determine whether it is necessary to check the upper-level database.

[0101] It's important to understand that when managers express dissatisfaction and offer their own requests and opinions, we first identify their intent. This invention incorporates specialized prompts for intent identification, allowing us to refer back to the higher-level database for relevant requests during ad hoc conversations. When managers request more general documents, the system doesn't need to refer back, as the retrieved detailed data might be noisy, impacting the quality of the final report. Therefore, the system must make a judgment based on the manager's feedback and determine whether a referral to the higher-level database is necessary.

[0102] In some embodiments, reference Figure 10 As shown, step S700 includes:

[0103] Step S800: When it is determined that a query to the superior database is required, a multi-level data retrieval mechanism is started according to the new requirement, and a query request is initiated to the superior database through a preset query path.

[0104] It is important to understand that when the system determines through intent recognition that it needs to review the superior database, the system will activate a multi-level data retrieval mechanism based on these new requirements and initiate a query request to the superior database through a preset query path. To ensure the accuracy and relevance of the information, the system will accurately match it based on the data structure in the superior database. Usually, the superior database contains more comprehensive and historical data, so the system needs to perform more detailed retrieval operations to ensure that it can obtain information that is highly relevant to the current task or problem. Avoid redundant information interference and ensure that the most relevant records are quickly located. During the retrieval process, the system will combine the temporary agents that have been generated to determine whether it is necessary to retrieve specific categories of information based on more detailed requirements, such as specific historical events, the background of the execution of key tasks, or cross-departmental collaboration details. The system will also automatically analyze the information integrity of the superior database during the review, filter out the parts that are most relevant to the current report content, and extract them for subsequent work summary generation.

[0105] In some embodiments, reference Figure 10 As shown, step S800 includes:

[0106] Step S900: When a query request is initiated to the superior database, a temporary agent is generated based on the current agent level of the manager, and corresponding prompt words are generated according to the temporary agent, and a work summary report is generated again in combination with the search data of the superior database.

[0107] It should be understood that, based on the fixed agent of the manager himself / herself, specific prompt words are designed according to the keywords and feedback of the manager, and these prompt words are transmitted to the fine-tuned large model. The large model regenerates a temporary agent according to these prompt words, so that the summary content can more accurately reflect the personalized needs of the manager. This process can flexibly respond to the manager's immediate feedback on the report content, level of detail or other aspects, ensuring that each generated report is highly targeted and customized. As Figure 11 shown, the prompt words generated by the temporary agent can be adjusted according to different task requirements and the manager's feedback, thereby further improving the accuracy and effectiveness of the report.

[0108] In some embodiments, referring to Figure 10 shown, step S900 includes:

[0109] Combining the keywords in the manager's feedback and the prompt words generated by the temporary agent, the retrieved data is transmitted to the large model, and a work report that meets the personalized needs of the manager is generated through the processing and analysis of the large model.

[0110] It should be understood that, as Figure 12 shown, the retrieved data is combined with the keywords in the manager's feedback and the temporary agent to generate new prompt words, and these prompt words are transmitted to the large model. Through the processing and analysis of the large model, a work report that meets the personalized needs of the manager is generated. This process can not only conduct a detailed analysis for specific tasks, but also adjust the content and format of the summary according to the manager's requirements, thereby ensuring that the report content is more accurate, comprehensive and meets the actual needs, and ultimately providing more valuable support for management decisions.

[0111] In some embodiments, the large model adopts the ChatGLM3-6B model. It should be understood that the ChatGLM3-6B model will make full use of its natural language processing and generation capabilities to automatically construct the framework of the work summary report based on the provided data and prompts. The model will not only summarize the progress of the current work, problems encountered, measures taken, and next steps based on the retrieved data, but also flexibly adjust the language style and content detail of the report according to the personalized requirements of the management personnel. For example, if the management personnel require a more detailed background description, the system will provide a more comprehensive task background and work details based on the reasoning ability of the model; if highlighting performance indicators is required, the system will strengthen the presentation of content related to key performance indicators (KPIs). The system will generate a preliminary work summary report automatically by combining the prompts generated according to the personalized needs of the management personnel with the relevant data retrieved from the work log through calling a large model (such as ChatGLM3-6B). In this process, the system will first adjust the specific content of the generated prompts according to the levels and needs of different management personnel, including the detail level of the report, the key areas of focus, and the key tasks or projects of concern. After being carefully designed, the prompts will be passed as input to the large model to help the model understand the needs of the management personnel and guide it to generate more accurate report content. This process is completely automated without manual intervention, and the processing speed of the large model is extremely fast, and it can generate a work summary report that meets personalized needs in a short time, greatly improving the efficiency and accuracy of report generation. Through this intelligent method, the management can quickly obtain high-quality work summaries, ensuring that the decision-making basis is timely, comprehensive, and accurate.

[0112] In the description of this specification, the descriptions referring to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine different embodiments or examples described in this specification.

[0113] After considering the specification and practicing the invention disclosed herein, those skilled in the art will readily conceive of other embodiments of the present invention. The present invention is intended to cover any variations, uses, or adaptations of the present invention, which follow the general principles of the present invention and include common general knowledge or conventional technical means in the technical field not disclosed in the present invention. The specification and embodiments are only regarded as exemplary, and the true scope and spirit of the present invention are pointed out by the appended claims.

Claims

1. An information integration large model interaction method based on the RAG framework, characterized in that, Including: Receiving work logs uploaded by multiple groups of staff and establishing a work log database; Constructing multi-level agents, where each level of agent corresponds to the relative permissions of different managers, and generating corresponding prompt words for each level of agent; Generating a first-level system summary based on the work log database in combination with the prompt words corresponding to the first-level agent, and constructing a first-level database based on the first-level system summary. Each level of system summary is generated based on the superior database in combination with the prompt words corresponding to each level of agent, and each level of database is constructed based on the system summary of each level until the last-level database; Matching the corresponding agent level according to the relative permissions of the manager, performing data retrieval in the database corresponding to the agent level using the RAG framework, and generating the summary report of the manager using a large model in combination with the prompt words corresponding to the agent level.

2. The large model interaction method according to claim 1, wherein The steps of matching the corresponding agent level according to the relative permissions of the manager, performing data retrieval in the database corresponding to the agent level using the RAG framework, and generating the summary report of the manager using a large model in combination with the prompt words corresponding to the agent level include: Transmitting the prompt words as input to the large model, and constructing a framework for the work summary report through the natural language processing and generation capabilities of the large model.

3. The large model interaction method according to claim 2, wherein The steps of transmitting the prompt words as input to the large model and constructing a framework for the work summary report through the natural language processing and generation capabilities of the large model include: The large model is fine-tuned by LoRA using a small dataset.

4. The large model interaction method according to claim 3, characterized in that The steps of transmitting the prompt words as input to the large model and constructing a framework for the work summary report through the natural language processing and generation capabilities of the large model include: Determining whether the manager is satisfied with the framework of the work summary report. When it is determined that the manager is satisfied, the corresponding work summary report is output. When it is determined that the manager is not satisfied, it is judged whether it is necessary to check the superior database.

5. The large model interaction method according to claim 4, characterized in that, The steps of judging whether it is necessary to check the superior database when it is determined that the manager is not satisfied include: Generating prompt words for intention recognition when it is determined that the manager is not satisfied, and judging whether it is necessary to check the superior database through the large model in combination with the prompt words for intention recognition.

6. The large model interaction method according to claim 5, wherein The steps of judging whether it is necessary to check the superior database when it is determined that the manager is not satisfied include: When it is judged that it is necessary to check the superior database, according to the new requirements, starting a multi-level data retrieval mechanism and initiating a query request to the superior database through a preset query path.

7. The large model interaction method according to claim 6, wherein The steps of when it is judged that it is necessary to check the superior database, according to the new requirements, starting a multi-level data retrieval mechanism and initiating a query request to the superior database through a preset query path include: When initiating a query request to the superior database, generating a temporary agent based on the current agent level of the manager, generating corresponding prompt words according to the temporary agent, and generating a work summary report again in combination with the retrieved data from the superior database.

8. The large model interaction method according to claim 7, wherein The steps of when initiating a query request to the superior database, generating a temporary agent based on the current agent level of the manager, generating corresponding prompt words according to the temporary agent, and generating a work summary report again in combination with the retrieved data from the superior database include: Combine the keywords in the manager's feedback and the prompt words generated by the temporary agent, transmit the retrieved data to the large model, and generate a work report that meets the personalized needs of the manager through the processing and analysis of the large model.

9. The large model interaction method according to any one of claims 1-8, characterized in that, The large model uses the ChatGLM3-6B model.

Images