Auxiliary work system based on AI large model development

Abstract

The invention belongs to the technical field of information networks, and particularly relates to a work auxiliary system based on AI large model development, which comprises a client, a technology stack and a server, can improve the work efficiency, solves the development pain points of unreasonable construction period arrangement, excessive repeated mechanical labor and the like in production and life, and improves the work efficiency. According to the intelligent core office assistant, the requirements of the public for a reasonable and efficient working mode are met, office information input by a user can be automatically subjected to priority arrangement through a built-in AI, all tasks can be executed according to a scientific and reasonable sequence, in addition, the intelligent core office assistant can also provide personalized task reminding and automatic file popping service, and the intelligent core office assistant has the advantages of being simple in structure and convenient to use. While efficient work is achieved, humanity is achieved, and good use experience of a user is guaranteed. Through linkage of the terminal part and the software part, team-based and public management is realized, and a simple, rapid and comfortable office environment is provided for users, so that the optimal office experience is realized.

Description

Technical Field

[0001] The present invention relates to the field of information network technology, and in particular to a work assistance system developed based on an AI big model. Background Art

[0002] Work assistance systems are software solutions that use advanced AI technologies, especially large pre-trained language models (such as GPT, BERT, etc.), to enhance and optimize human workflows. Such systems are designed to help users complete tasks, solve problems, or make decisions more efficiently by providing intelligent support and services;

[0003] However, the existing work assistance systems still have the following technical problems when used:

[0004] When in use, the existing auxiliary system is unable to effectively solve the problems of unreasonable schedule scheduling and excessive repetitive mechanical labor in production life, cannot mobilize network resources and local resources authorized by users, cannot perform in-depth analysis and preprocessing of tasks, cannot sort and execute multiple tasks in a reasonable sequence, and cannot provide personalized task reminders and automatic file pop-up services, causing great inconvenience to use.

[0005] To this end, a work assistance system developed based on an AI big model is proposed to solve the above-mentioned problems. Summary of the invention

[0006] The purpose of the present invention is to provide a work assistance system developed based on an AI big model to solve the problems raised in the above background technology.

[0007] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a work assistance system developed based on an AI big model, including a client, a technology stack and a server;

[0008] The client includes a user input module, an AI priority evaluation module, a task scheduling module, a personalized reminder module, a time management module, a file pop-up notification module and a user task processing module; the user input module allows users to input tasks, meetings, priorities and any related office information; the AI ​​priority evaluation module is a built-in AI system that will evaluate the priority of tasks based on user input and behavior patterns; the task scheduling module is an AI system that will schedule tasks based on the evaluation results and determine the execution order and schedule of tasks; the personalized reminder module is an AI system that will send personalized reminders at the appropriate time based on user preferences and past reaction time; the time management module will track the user's time to ensure that tasks are reminded and processed at the appropriate time; the file pop-up notification module will pop up files and tasks that need to be processed by the user at the appropriate time; the user task processing module processes tasks and files through a software interface;

[0009] The technology stack includes an architecture module, an asynchronous processing and responsive programming module, a dependency injection module, a persistent storage module, a test and continuous integration / continuous deployment module; the architecture module includes MVVM, which implements data-driven UI rendering and keeps the business logic and interface logic decoupled by adopting the MVVM architecture; the asynchronous processing and responsive programming module includes RxJava, which simplifies the logic of asynchronous operations and event processing by using responsive data streams such as RxJava's Observable and Flowable. Asynchronous processing refers to executing some time-consuming or resource-dependent operations without blocking the current thread. Responsive programming is a programming paradigm based on data flow and change propagation, which processes asynchronous events and data in a declarative manner without worrying about specific implementation details and synchronization issues; the dependency injection module includes DaggerHilt, which is an annotation-based dependency injection framework. Dagger Hilt can automatically manage the life cycle and dependencies of objects, reducing the burden of manual configuration; the persistent storage module includes Room, which acts as an abstract layer and provides a clear database access interface. It also generates necessary code through annotation processors, simplifying database operations; the testing and continuous integration / continuous deployment modules include JUnit, Espresso, and Jenkins, which ensure code quality through automated testing and continuous integration / continuous deployment, while accelerating the development and release process.

[0010] Preferably, in the framework module, MVVM completely decouples the view and the model, and implements the interaction between the view and the view model through data binding and command mode.

[0011] Preferably, the MVVM architecture is composed of the following parts: model: refers to a domain model representing the real state content, or refers to a data access layer representing the content. The model is independent of the specific platform or framework, defines business logic and data operations, and is usually implemented through interfaces or abstract classes; view: refers to the structure, layout and appearance seen by the user on the screen. The view is related to the specific platform or framework, is responsible for rendering the user interface, and responds to user input events. The view is usually a XAML file, which references the properties and commands defined in the view model through data binding; view model: refers to the abstraction of the view that exposes public properties and commands. The view model is a value converter responsible for exposing data objects from the model so that objects can be easily managed and presented. The view model is also responsible for processing user input events and calling business logic and data operations in the model. The view model usually implements the INotifyPropertyChanged interface to notify the view to update when properties change.

[0012] Preferably, in the asynchronous processing and responsive programming modules, asynchronous processing refers to executing some time-consuming or resource-dependent operations, such as network requests, file reading and writing, and database operations, without blocking the current thread; responsive programming is a programming paradigm based on data flow and change propagation, which processes asynchronous events and data in a declarative manner without worrying about specific implementation details and synchronization issues.

[0013] Preferably, the RxJava is a Java library based on the Reactive Extensions specification, providing a series of classes and methods for combining and consuming asynchronous operations and events using responsive data streams.

[0014] Preferably, the core concepts of RxJava include Observable, Observer, Operator, Scheduler, Flowable, Single, Maybe and Completable; the Observable represents an observable data stream that can emit zero or more data, errors or completion signals to the observer; the Observer represents an observer that can subscribe to an Observable and process the data, errors or completion signals emitted by it accordingly; the Operator represents an operator that transforms, filters, and combines the data emitted by the Observable and returns a new Observable; the Scheduler represents a scheduler that specifies The threads in which Observable and Observer run, thereby achieving asynchronous processing and concurrency control; the Flowable represents an Observable that supports the back pressure mechanism, which controls the speed of the data flow according to the processing power of the observer to avoid memory overflow or data loss; the Single represents an Observable that only emits a single data or error signal, which is suitable for scenarios with only one result, such as network requests; the Maybe represents an Observable that emits a single data, error or completion signal, which is suitable for scenarios where there may be no results, such as database queries; the Completable represents an Observable that only emits an error or completion signal, which is suitable for scenarios that do not care about the result but only care about whether the task is completed, such as file deletion.

[0015] Preferably, the Dagger Hilt provides a set of standard Android Dagger components, which are automatically generated according to the class path without manual creation and maintenance; provides a set of annotations to easily bind Dagger modules to corresponding components without manually writing component interfaces and injection methods; provides a way to use another annotation in the Android framework class to automatically inherit a base class, thereby implementing dependency injection without manually calling Dagger components; provides some special types of components and scopes to better manage the life cycle and dependencies of objects at the application level, activity level, fragment level, etc.; provides some testing tools to use the components generated by Hilt in tests, and provides methods to add or replace test bindings.

[0016] Preferably, the Room is a Jetpack architecture component in Android that provides an abstraction layer for data persistence and operations on a SQLite database.

[0017] Preferably, the testing and continuous integration / continuous deployment modules improve the quality and reliability of the code while shortening the development and release cycles. Testing refers to various types of inspections on the code to ensure that it meets the expected functions and performance. Continuous integration refers to frequently merging the developer's code into the main branch and performing automated building and testing. Continuous deployment refers to automatically deploying the tested code to the production environment.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] This application can improve work efficiency, solve development pain points such as unreasonable schedule scheduling and excessive repetitive mechanical labor in production and life, and meet the public's demand for a reasonable and efficient work mode. The product has built-in AI, which can automatically prioritize the office information input by the user, and strive to ensure that all tasks can be executed in a scientific and reasonable order. In addition, Zhihe Office Assistant can also provide personalized task reminders and automatic file pop-up services, while achieving efficient work, it does not lose humanity, and ensures a good user experience.

[0020] By linking the terminal part with the software part, team-based and corporate management can be achieved, providing users with a simple, fast and comfortable office environment, thus achieving the best office experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a task processing flow chart;

[0022] Figure 2 For technical implementation Figure 1 ;

[0023] Figure 3 For technical implementation Figure 2 ;

[0024] Figure 4 For technical implementation Figure 3 . DETAILED DESCRIPTION

[0025] 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.

[0026] In the description of the present invention, it is necessary to understand that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship are based on the orientation or position relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0027] Embodiment 1:

[0028] See also Figure 1-4 , the present invention provides a technical solution:

[0029] A work assistance system developed based on the AI ​​big model, including the client, technology stack and server;

[0030] The client includes a user input module, an AI priority evaluation module, a task scheduling module, a personalized reminder module, a time management module, a file pop-up notification module, and a user task processing module; the user input module allows users to input tasks, meetings, priorities, and any related office information; the AI ​​priority evaluation module is a built-in AI system that will evaluate the priority of tasks based on user input and behavior patterns; the task scheduling module is an AI system that will schedule tasks based on the evaluation results and determine the execution order and schedule of tasks; the personalized reminder module is an AI system that will send personalized reminders at the appropriate time based on user preferences and past reaction time; the time management module will track user time to ensure that tasks are reminded and processed at the appropriate time; the file pop-up notification module will pop up files and tasks that need to be processed by the user at the appropriate time; the user task processing module processes tasks and files through the software interface;

[0031] The technology stack includes architecture module, asynchronous processing and responsive programming module, dependency injection module, persistent storage module, testing and continuous integration / continuous deployment module; the architecture module includes MVVM (Model-View-ViewModel), which implements data-driven UI rendering by adopting MVVM architecture and keeps business logic and interface logic decoupled; the asynchronous processing and responsive programming module includes RxJava, which simplifies the logic of asynchronous operations and event processing by using responsive data streams such as RxJava's Observable and Flowable. Asynchronous processing refers to executing some time-consuming or resource-dependent operations without blocking the current thread. Responsive programming is a programming paradigm based on data flow and change propagation. It handles asynchronous events and data in a declarative way without worrying about specific implementation details and synchronization issues; the dependency injection module includes Dagger Hilt, which is an annotation-based dependency injection framework. Dagger Hilt can automatically manage the life cycle and dependencies of objects, reducing the burden of manual configuration; the persistent storage module includes Room, which acts as an abstraction layer. Room provides a clear database access interface and generates necessary code through annotation processors, simplifying database operations; the testing and continuous integration / continuous deployment modules include JUnit, Espresso, and Jenkins, which ensure code quality through automated testing and continuous integration / continuous deployment, while accelerating the development and release process.

[0032] In the framework module, MVVM completely decouples the view and model, and implements the interaction between the view and the view model through data binding and command pattern.

[0033] The MVVM architecture is composed of the following parts: Model: refers to the domain model (object-oriented) that represents the real state content, or refers to the data access layer (data-centric) that represents the content. The model is independent of the specific platform or framework, defines the business logic and data operations, and is usually implemented through interfaces or abstract classes; View: refers to the structure, layout, and appearance that the user sees on the screen. The view is related to the specific platform or framework, and is responsible for rendering the user interface and responding to user input events. The view is usually a XAML file that references the properties and commands defined in the view model through data binding; View model: refers to the abstraction of the view that exposes public properties and commands. The view model is a value converter that is responsible for exposing (converting) data objects from the model so that objects can be easily managed and presented. The view model is also responsible for processing user input events and calling business logic and data operations in the model. The view model usually implements the INotifyPropertyChanged interface to notify the view to update when properties change.

[0034] In the asynchronous processing and responsive programming modules, asynchronous processing refers to executing some time-consuming or resource-dependent operations, such as network requests, file reading and writing, and database operations, without blocking the current thread; responsive programming is a programming paradigm based on data flow and change propagation, which handles asynchronous events and data in a declarative manner without worrying about specific implementation details and synchronization issues.

[0035] RxJava is a Java library based on the Reactive Extensions specification, which provides a series of classes and methods for composing and consuming asynchronous operations and events using responsive data streams.

[0036] The core concepts of RxJava include Observable, Observer, Operator, Scheduler, Flowable, Single, Maybe and Completable; Observable represents an observable data stream that can emit zero or more data, errors or completion signals to observers; Observer represents an observer that can subscribe to an Observable and process the data, errors or completion signals emitted by it; Operator represents an operator that transforms, filters, and combines the data emitted by Observable and returns a new Observable; Scheduler represents a scheduler that specifies the observables. Observable and Observer run threads, thereby realizing asynchronous processing and concurrency control; Flowable represents an Observable that supports the back pressure mechanism, which controls the speed of data flow according to the processing power of the observer to avoid memory overflow or data loss; Single represents an Observable that only emits a single data or error signal, which is suitable for scenarios with only one result, such as network requests; Maybe represents an Observable that emits a single data, error or completion signal, which is suitable for scenarios where there may be no results, such as database queries; Completable represents an Observable that only emits error or completion signals, which is suitable for scenarios that do not care about the results but only care about whether the task is completed, such as file deletion.

[0037] Dagger Hilt provides a set of standard Android Dagger components, which are automatically generated according to the class path, without manual creation and maintenance; it provides a set of annotations to easily bind Dagger modules to corresponding components without manually writing component interfaces and injection methods; it provides a way to use another annotation in the Android framework class to automatically inherit a base class to achieve dependency injection without manually calling Dagger components; it provides some special types of components and scopes to better manage the life cycle and dependencies of objects at the application level, activity level, fragment level, etc.; it provides some testing tools to use Hilt-generated components in tests, and provides methods to add or replace test bindings.

[0038] Room is a Jetpack architecture component in Android that provides an abstraction layer for data persistence and operations on SQLite databases. The main advantages of Room are as follows: SQL queries are verified at compile time to avoid runtime errors; repetitive and error-prone boilerplate code is simplified through annotations, such as defining data entities, data access objects (DAOs), and database classes; a smooth database migration path is provided, and version control and data backup are supported. It supports integration with RxJava2, RxJava3, Guava and other libraries to achieve responsive programming and asynchronous operations, and supports integration with Paging 3 libraries to achieve paging data loading.

[0039] The testing and continuous integration / continuous deployment modules improve the quality and reliability of the code while shortening the development and release cycles. Testing refers to various types of inspections on the code to ensure that it meets the expected functionality and performance. Continuous integration refers to frequently merging the developer's code into the main branch and performing automated building and testing. Continuous deployment refers to automatically deploying the tested code to the production environment.

[0040] JUnit, Espresso and Jenkins are some commonly used testing and continuous integration / continuous deployment tools that can be used for Java and Android development. Here is a brief introduction to them:

[0041] JUnit is a Java unit testing framework that can perform fast, concise, and repeatable testing on Java code. JUnit can be integrated with IDEs (such as Eclipse) or build tools (such as Maven) to facilitate developers to write and run test cases.

[0042] Espresso is an Android UI testing framework that can simulate and verify the user interface of Android applications. Espresso can be integrated with Android Studio or Gradle to facilitate developers to write and run UI tests.

[0043] Jenkins is an open source continuous integration / continuous deployment tool that provides a user-friendly interface. It can be integrated with version control systems such as Git and SVN, as well as build tools such as Maven, Ant, and Gradle. Jenkins can monitor changes in code repositories, trigger automated build, test, and deployment tasks, and provide visual reports and notifications.

[0044] This work assistance system can realize intelligent task preprocessing and introduces an intelligent task preprocessing function based on a large language model. This function can mobilize network resources and local resources authorized by users to perform in-depth analysis and preprocessing of tasks, saving users a lot of time and energy.

[0045] 1. Powerful functions of task preprocessing:

[0046] Intelligent analysis: Through the large language model, the user's task request can be accurately analyzed to clarify the user's needs and intentions. Whether the user is asking about the weather, scheduling a meeting, or searching for information, the large language model can quickly understand and respond.

[0047] Resource mobilization: After parsing the task, the large language model will intelligently mobilize network resources and local resources authorized by the user, such as databases, file libraries, API interfaces, etc., to obtain the information and data required to complete the task.

[0048] Deep preprocessing: After obtaining relevant information, the large language model will perform deep preprocessing on the task. For example, for weather query tasks, the large language model will call the weather API in real time to obtain the latest weather information, and combine the user's historical data and behavior habits to provide users with personalized clothing and travel suggestions. For meeting arrangement tasks, the large language model will automatically filter out the documents required for the meeting and remind users to attend the meeting at the appropriate time.

[0049] Intelligent recommendation: In addition to basic task processing, the large language model can also recommend relevant information and resources to users based on their preferences and needs. For example, when checking the weather, the large language model will recommend suitable outdoor activities and tourist attractions; when arranging a meeting, the large language model will recommend suitable conference rooms and participants.

[0050] 2. Advantages of task preprocessing:

[0051] Efficiency: Through intelligent analysis and preprocessing of large language models, tasks assigned by users can be completed quickly, greatly saving users’ time and energy.

[0052] Personalization: The large language model can combine the user's historical data and behavioral habits to provide users with personalized task processing solutions and suggestions to meet their personalized needs;

[0053] Intelligence: The large language model has powerful learning and reasoning capabilities, and can continuously learn and optimize the task processing process to improve the intelligence and accuracy of task processing.

[0054] This system consists of two parts:

[0055] The first part is the software part: the software can use the built-in large language model to arrange the office information input by the user according to the processing priority. Provide personalized and intelligent task reminders at the appropriate time, and preliminarily organize the files that the user needs to process or the manuscripts that need to be edited, and feedback the results to the user. While providing users with preliminary ideas for task processing, it eliminates the need for coordinating a large number of tasks, and can directly provide users with dual support in efficiency and quality. As the user's usage time and frequency increase, exclusive data sets for different users can be generated, and these data sets can be used to train the built-in large language model to make the software's services more in line with personal needs. In this way, self-functional upgrades with low barriers and simple operations can be achieved.

[0056] The second part is the terminal: its main function is to display the work content and provide users with a convenient task processing platform. The built-in software part of the terminal can manage multiple devices in a unified team from one point. According to the actual needs of different teams, the terminal can further derive specialized functions such as conference collaborative editing of files, multi-device linkage, joint display of conference content, and helping users to interact with each other in real time based on the native version. This product is committed to helping users clarify tasks, plan tasks, and complete tasks, pursuing the simplest team collaboration experience and providing the most timely and comfortable convenience.

[0057] Zhihe Office Assistant is an inevitable product of modern fast-paced work. There is no shortage of "alarm clock" and "memo" type software and hardware products on the market, but most of them lack practicality and only stay at the reminder level, which does not provide practical assistance for completing tasks. Zhihe Office Assistant is different from other popular auxiliary office products on the market. In addition to arranging and reminding various pending tasks, the product can pop up the documents and materials needed for office work on the desktop as needed, so that users can proceed to the next step, providing the most direct and convenient service. In terms of traditional reminder functions, Zhihe Office Assistant can also decide the reminder method to be adopted according to different types of work and preferences of user groups, such as traditional alarm clock form, virtual character animation reminder, voice assistant reminder and other forms, and adjust the reminder intensity according to the task urgency assessed by the built-in large language model, striving to achieve a perfect assistance effect close to that of a real secretary. Terminal devices break the limitations of space and equipment, allowing Zhihe Office Assistant to not only be limited to mobile phones and computers, so that users can get started anytime and anywhere, and work can be carried out anywhere and at will. Through further demand surveys and actual user experience, terminal devices will specialize functions and develop special service versions for specific audiences such as companies and schools. With low production costs, they will increase coverage in production and life, making product services more ubiquitous, in line with demand, and solving pain points.

[0058] The software part of this product is installed in the user's mobile device and terminal device. Through the keywords entered by the user, such as "8 o'clock in the morning, conference room, year-end summary", with the help of built-in AI, the atomic events to be done are analyzed, "remind the user to go to the conference room for a meeting at 8 o'clock in the morning", "open the files required for the meeting on the user's mobile phone, computer and other devices in advance". After clarifying the task, the software will pre-process the task through the built-in AI. For example: before performing statistical tasks, edit the basic tables to be used; when a summary speech is required, pre-edit the speech according to the previous usage records or the files entered by the user. In this way, it plays an assisting role similar to that of a secretary, helping users avoid a lot of unnecessary mechanical operations and saving time and energy. When the user enters a large number of tasks with incomplete time information or unclear completion order, the software will arrange the multiple to-do tasks entered according to conventional cognition, and feedback the final results to the user, thereby helping the user to organize and complete complicated tasks. When the preset task execution time is reached, the software will make personalized and intelligent to-do reminders according to the importance of the task and the set reminder intensity. For example: when giving a reminder to wake up, a more intense method is used. If no timely and effective feedback is received, a second or even more repeated reminder is given. When giving a reminder of pending tasks, a scrolling barrage on the display screen or interaction with virtual characters can be used to eliminate the user's anxiety and aversion to pending tasks and optimize the software's user experience.

[0059] The hardware part is a tablet-like device with built-in software. Depending on the tasks to be completed and the user's needs, the hardware is divided into standard and specialized versions.

[0060] The standard version of the hardware is mainly aimed at individual users. The product can achieve three-point linkage with computers and mobile phones through built-in software, breaking through device limitations, being independent of mobile and PC terminals, and conducting work anytime and anywhere.

[0061] The specialized version is divided into two specialized directions: school and enterprise:

[0062] 1. The school version focuses more on a work mode that radiates from one point to all sides. The teacher can send learning goals and learning tasks to the student, and display the task progress of each student so that the teacher can better grasp the progress of the class. In addition to receiving information from the teacher, the student can also provide feedback to the teacher. This is not only conducive to the timely resolution of knowledge doubts, but also through the built-in AI in the software, it can complete the summary of high-frequency doubts and key points of knowledge difficulties, helping to carry out teaching tasks more effectively.

[0063] 2. The enterprise version focuses more on interactive collaboration between users. In a meeting, content sharing and collaborative editing can be carried out through the linkage of hardware devices within a certain range, changing the point-to-point communication mode to a point-to-point communication mode. While increasing the participation rate of participants, it further improves the work efficiency of the team and helps enterprises to conduct office work more conveniently and efficiently.

[0064] The hardware part can be further enhanced through cloud services. Cloud services can reduce the computing power requirements of the device. The required massive calculations are performed through the cloud, and the results are finally transmitted to the software and displayed on the hardware device. This can not only speed up the processing speed of the software for various tasks, but also reduce the configuration requirements of the hardware, thereby reducing the hardware cost, and more specialized hardware functions, making it more powerful.

[0065] The software uses built-in AI. Users can input to-do items, and AI will automatically arrange them according to the priority of things, put more important and urgent things at the front, plan many affairs reasonably, and make personalized task reminders at the right time, providing a general framework for solving problems. The AI ​​framework can fundamentally solve the psychological "head" of the public that "everything is difficult at the beginning", decompose the things to be done, and increase the positive feedback of enthusiasm. It fundamentally solves more than 90% of the procrastination problems of contemporary young people. Compared with other similar products, the hardware system has low cost and better effect. The product uses more scientific models and algorithms. As a new force that has never appeared in the market, it has good market prospects and market competitiveness.

[0066] Large models refer to deep learning models with billions or even tens of billions of parameters. These models have large capacity, massive parameters, and powerful computing power. They have gone through the development from early single-language pre-training models to multi-language pre-training models, and now multimodal pre-training models, and can cope with a variety of tasks. Large language models are deep learning algorithms that are trained using large-scale data sets to recognize, summarize, translate, predict, and generate content. Large language models are mainly based on a deep learning architecture called Transformer, which can learn context and meaning by capturing relationships in sequence data, integrate the required content, and output the information required by users. Google's BERT first demonstrated the powerful potential of pre-trained models, while models such as OpenAI's GPT series and Anthropic's Claude continue to promote the development of language model technology. These increasingly large-scale models have tens of billions or even hundreds of billions of parameters, which can capture the complex semantic relationships of language and achieve language interaction at the human level. The biggest innovation of the Transformer model is the introduction of the attention mechanism, which enables computers to comprehensively examine the entire sentence at once and identify which words are related to each other and which words are just incidental. This mechanism greatly improves the model's ability to learn long-distance dependencies and overcomes the limitations of traditional recurrent neural networks (RNNs) and convolutional neural networks (CNNs) when processing long sequence data. Before the advent of Transformer, natural language processing usually used RNNs or CNNs to model semantic information. However, the sequence structure of RNNs makes it difficult for information from earlier moments to be transmitted to later moments, similar to the original message in the game of passing the message to the last person, when it may have changed or disappeared, and the local perception of CNNs limits the capture of global semantic information, just like observing the entire comic strip through a small window, you may be able to see a few grids at a time, but you can't see the whole story at a glance. Therefore, RNNs and CNNs often find it difficult to fully learn long-distance dependencies between words when processing long sequences. Transformers can observe the entire sentence, understand the connection between words, and capture the complete story without passing information one by one or peeking through a window.

[0067] Transformer's attention mechanism breaks through the inherent limitations of RNN and CNN in processing long sequences. The Transformer model uses attention to weigh the importance of each word relative to other words to help understand the meaning of the entire sentence, even if the keywords are far apart, so that the language model can be pre-trained with rich language knowledge on a large corpus. This modular and scalable model structure also makes it easy to expand the model size and expressiveness by increasing the number of modules, providing a feasible way to achieve ultra-large parameter amounts. Transformer solves the difficulties of traditional models in processing long sequences and provides an infinitely scalable structure, laying the foundation for the implementation of large model technology. Large models represent a new pre-training-fine-tuning paradigm, the core of which is to first pre-train a model with extremely large parameters using a large dataset, and then fine-tune the model to adapt to specific tasks. This is in contrast to traditional single-task training and marks a major change in methodology. The most notable feature of large models is the doubling of the number of parameters, from the millions of early models to the billions or even tens of billions of current models, achieving an order of magnitude breakthrough compared to the past. The introduction of the Transformer architecture ushered in a new era of natural language processing model design. It introduced the self-attention mechanism and parallel computing ideas, greatly improving the model's ability to handle long-distance dependencies and laying the foundation for the subsequent development of large models. It is precisely because of the success of the Transformer architecture that researchers realized the importance of model architecture design when dealing with complex tasks and large-scale data. This understanding has inspired researchers to further expand the number of model parameters. Although methods of increasing the number of parameters have been tried before, due to limitations such as the model's own memory capacity at the time, the improvement of the model after increasing the parameters was not obvious.

[0068] An open source large language model is a deep neural network model that is pre-trained on a large corpus and can be used for various tasks of natural language understanding and generation. Currently, there are many mainstream open source large language models, such as:

[0069] 1. LLaMA2: A series of large language models open sourced by Meta, including models with 7B, 13B, 33B, and 65B parameter sizes, supporting multiple languages ​​and multiple tasks, with strong general and generative capabilities.

[0070] 2. ChatGLM2: A large language model open sourced by Tsinghua University that supports Chinese-English bilingual question-answering and dialogue. It is optimized for Chinese, has a 6B parameter scale, and can generate fluent, coherent, and logical text.

[0071] RLHF (Reinforcement Learning from Human Feedback) is an important technology used in training large language models. It optimizes the language model by combining human feedback and reinforcement learning, so that it can better generate text replies that meet human preferences. The training process of RLHF can be broken down into the following three steps:

[0072] 1. Pre-trained Language Model (LM): First, a language model is trained using the classic pre-training objective. This model can be a smaller version of GPT-3 or other large Transformer models. Based on the pre-training, additional text or conditions can be used for fine-tuning to further improve the performance of the model;

[0073] 2. Train the Reward Model (RM): In this step, the data required to train the reward model is generated using a pre-trained language model and human preference information is introduced. The reward model receives a sequence of texts and returns a scalar reward that measures human preference. The reward model can be a fine-tuned language model or a model trained from scratch. Training data can be obtained by manually ranking the generated answers;

[0074] 3. Fine-tune the language model using reinforcement learning: The last step is to fine-tune the language model using reinforcement learning. Since fine-tuning the parameters of the entire model is costly, a policy gradient reinforcement learning algorithm (such as PPO) is usually used to fine-tune some or all of the parameters of the initial language model. This can improve the performance of the model and the quality of the generated text.

[0075] Feedback plays a vital role in this process. It can be a score for the generated text or a ranking for the generated text. When training the reward model, human feedback can help the model learn human preferences, so as to better generate text replies that meet human preferences. When fine-tuning the language model, human feedback can help the model better understand human intentions, so as to generate more accurate and reasonable text replies. Therefore, human feedback plays a vital role in RLHF, which can help the model better understand human intentions, so as to generate more accurate and reasonable text replies.

[0076] Reasoning for large language models is a very difficult problem because it requires processing a large number of parameters and computations while ensuring the quality and efficiency of the generated results. Currently, there are some methods that can be used to speed up reasoning for large language models:

[0077] 1. Use low-precision or mixed-precision calculations. This method can reduce memory usage and calculation time and improve GPU utilization. For example, use 16-bit floating point numbers or Bfloat16 format instead of 32-bit floating point numbers, such as DeepSpeed, Megatron-LM, NVIDIA FasterTransformer, etc.

[0078] 2. Use model parallelism or pipeline parallelism. This method can split a large model into multiple parts and distribute them on multiple GPUs or nodes for parallel computing. This can overcome the memory limitations of a single device while improving throughput and scalability, such as DeepSpeed, Megatron-LM, NVIDIA FasterTransformer, vLLM, etc.

[0079] 3. Use cache optimization or memory management technology. This approach can reduce repeated calculations and memory waste, and improve memory efficiency and throughput. For example, use KV cache or PagedAttention technology to store and manage key-value pairs in attention mechanisms, such as NVIDIA FasterTransformer, vLLM, LLM Accelerator, etc.

[0080] 4. Use reference text or speculative sampling techniques. This method can use existing text information to speed up the generation process and avoid unnecessary calculations and searches. For example, use reference text similar to the target text as input, or generate multiple candidate results in advance and select the best result based on subsequent tokens, such as LLM Accelerator, StreamingLLM, Speculative Sampling, etc.

[0081] Humans have a remarkable ability to continuously absorb information, make decisions, take actions, observe changes, and make the next decision. Our entire life is an endless cycle of observation, thinking, and action. Researchers hope to apply this concept to large language models (LLMs) so that they can continuously make new decisions and gradually approach the ability to solve complex problems.

[0082] LLM agents are intelligent agents based on large language models that can perceive the environment, make decisions, and take actions. LLMs are artificial intelligence models that can process natural language. They can learn from massive amounts of network knowledge and demonstrate amazing capabilities. They can demonstrate capabilities beyond text generation, including having conversations, completing tasks, reasoning, and exhibiting a certain degree of autonomous behavior.

[0083] The technology of LLM agents can be divided into three main components: brain, perception, and action. The brain refers to the LLM itself, which is responsible for receiving input, generating output, storing memory, making decisions, etc. Perception refers to how LLM agents interact with the external environment, including obtaining information, understanding context, identifying goals, etc. Action refers to how LLM agents perform tasks, including selecting appropriate tools, adopting effective strategies, evaluating results, etc.

[0084] One of the defining characteristics of an AI agent is its ability to execute plans. To do this, it needs to be able to decompose tasks and improve through self-reflection. Task decomposition refers to the ability of an agent to break down complex tasks into smaller, manageable sub-goals. To this end, AI developers often use Chain of Thought (a series of intermediate natural language reasoning steps leading to a final output) and Tree of Thought (each thought is a coherent sequence of language that serves as an intermediate step in solving a problem) techniques. Once combined, these methods prompt the model to "think step by step" and break down difficult tasks into simpler steps. Some agents rely on external classical planners to perform long-term planning, using the Planning Domain Definition Language (PDDL) method to convert the problem into a format that the external planner can understand and generate a plan.

[0085] Break down complex problems into smaller subtasks and hypotheses and try Figure 1 We can prove or disprove them step by step to get closer to solving bigger problems. We can simulate this behavior with large language models by using the concept of intelligent agents. By providing the agents with the right tools, they can decide autonomously what action to take next.

[0086] If artificial agents do become smart enough to mimic human ingenuity and research skills, they could make huge advances in these areas. Humans would no longer be the bottleneck in inventing new things and exploring the mysteries of the world.

[0087] Memory is essential for AI agents, enabling the storage and retrieval of information required for various tasks. The memory process in AI is similar to human sensory perception, which enables agents to retain sensory impressions, including visual, auditory, and tactile, in the initial stage. Short-term memory (STM) or working memory acts as an active workspace for currently used information and is essential for performing complex cognitive tasks such as learning and reasoning. In this regard, the short-term memory of the LLM plays a role in contextual learning, but is limited by the limited context window length of the model. Long-term memory (LTM) is a repository of information that can be retained for a long time, ranging from days to decades. It has an almost unlimited storage capacity and includes two subtypes: explicit / declarative memory and implicit / procedural memory. Explicit / declarative memory stores facts and events that can be consciously recalled, including episodic memory and semantic memory. Implicit / procedural memory, on the other hand, operates unconsciously and includes skills and routines that are performed automatically, such as riding a bicycle or typing on a keyboard. Through these different types of memory, AI agents are able to handle various tasks and situations more intelligently.

[0088] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention; therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is limited by the attached claims rather than the above description. Therefore, it is intended to include all changes within the meaning and scope of the equivalent elements of the claims in the present invention, and any figure marks in the claims should not be regarded as limiting the claims involved.

[0089] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A work assistance system developed based on an AI big model, characterized in that: Including client, technology stack and server; The client includes a user input module, an AI priority evaluation module, a task scheduling module, a personalized reminder module, a time management module, a file pop-up notification module and a user task processing module; the user input module allows users to input tasks, meetings, priorities and any related office information; the AI ​​priority evaluation module is a built-in AI system that will evaluate the priority of tasks based on user input and behavior patterns; the task scheduling module is an AI system that will schedule tasks based on the evaluation results and determine the execution order and schedule of tasks; the personalized reminder module is an AI system that will send personalized reminders at the appropriate time based on user preferences and past reaction time; the time management module will track the user's time to ensure that tasks are reminded and processed at the appropriate time; the file pop-up notification module will pop up files and tasks that need to be processed by the user at the appropriate time; the user task processing module processes tasks and files through a software interface; The technology stack includes an architecture module, an asynchronous processing and responsive programming module, a dependency injection module, a persistent storage module, a test and continuous integration / continuous deployment module; the architecture module includes MVVM, which implements data-driven UI rendering and keeps the business logic and interface logic decoupled by adopting the MVVM architecture; the asynchronous processing and responsive programming module includes RxJava, which simplifies the logic of asynchronous operations and event processing by using responsive data streams such as RxJava's Observable and Flowable. Asynchronous processing refers to executing some time-consuming or resource-dependent operations without blocking the current thread. Responsive programming is a programming paradigm based on data flow and change propagation, which processes asynchronous events and data in a declarative manner without worrying about specific implementation details and synchronization issues; the dependency injection module includes DaggerHilt, which is an annotation-based dependency injection framework. Dagger Hilt can automatically manage the life cycle and dependencies of objects, reducing the burden of manual configuration; the persistent storage module includes Room, which acts as an abstract layer and provides a clear database access interface. It also generates necessary code through annotation processors, simplifying database operations; the testing and continuous integration / continuous deployment modules include JUnit, Espresso, and Jenkins, which ensure code quality through automated testing and continuous integration / continuous deployment, while accelerating the development and release process.

2. According to claim 1, a work assistance system developed based on an AI large model is characterized in that: In the framework module, MVVM completely decouples the view and the model, and implements the interaction between the view and the view model through data binding and command mode.

3. According to claim 1, a work assistance system developed based on an AI large model is characterized in that: The MVVM architecture is composed of the following parts: Model: refers to the domain model that represents the real state content, or refers to the data access layer that represents the content. The model is independent of the specific platform or framework, defines the business logic and data operations, and is usually implemented through interfaces or abstract classes; View: refers to the structure, layout, and appearance that the user sees on the screen. The view is related to a specific platform or framework and is responsible for rendering the user interface and responding to user input events. A view is usually a XAML file that references the properties and commands defined in the view model through data binding; View model: refers to the abstraction of a view that exposes public properties and commands. The view model is a value converter that is responsible for exposing data objects from the model for easy management and presentation of objects. The view model is also responsible for processing user input events and calling business logic and data operations in the model. The view model usually implements the INotifyPropertyChanged interface to notify the view to update when properties change.

4. According to claim 1, a work assistance system developed based on an AI large model is characterized in that: In the asynchronous processing and responsive programming modules, asynchronous processing refers to executing some time-consuming or resource-dependent operations, such as network requests, file reading and writing, and database operations, without blocking the current thread; responsive programming is a programming paradigm based on data flow and change propagation, which processes asynchronous events and data in a declarative manner without worrying about specific implementation details and synchronization issues.

5. According to claim 1, a work assistance system developed based on an AI large model is characterized in that: The RxJava is a Java library based on the Reactive Extensions specification, which provides a series of classes and methods for composing and consuming asynchronous operations and events using responsive data streams.

6. The work assistance system developed based on the AI ​​big model according to claim 1 is characterized in that: The core concepts of RxJava include Observable, Observer, Operator, Scheduler, Flowable, Single, Maybe and Completable; Observable represents an observable data stream that can emit zero or more data, errors or completion signals to observers; Observer represents an observer that can subscribe to an Observable and process the data, errors or completion signals emitted by it; Operator represents an operator that transforms, filters and combines the data emitted by Observable and returns a new Observable; Scheduler represents a scheduler that specifies an Observable. The thread in which servable and Observer run, thereby realizing asynchronous processing and concurrency control; the Flowable represents an Observable that supports the back pressure mechanism, which controls the speed of the data flow according to the processing power of the observer to avoid memory overflow or data loss; the Single represents an Observable that only emits a single data or error signal, which is suitable for scenarios with only one result, such as network requests; the Maybe represents an Observable that emits a single data, error or completion signal, which is suitable for scenarios where there may be no results, such as database queries; the Completable represents an Observable that only emits an error or completion signal, which is suitable for scenarios that do not care about the result but only care about whether the task is completed, such as file deletion.

7. The work assistance system developed based on the AI ​​big model according to claim 1 is characterized in that: The Dagger Hilt provides a set of standard Android Dagger components, which are automatically generated according to the class path without manual creation and maintenance; provides a set of annotations to easily bind Dagger modules to corresponding components without manually writing component interfaces and injection methods; provides a way to use another annotation in the Android framework class to automatically inherit a base class to achieve dependency injection without manually calling Dagger components; provides some special types of components and scopes to better manage the life cycle and dependencies of objects at the application level, activity level, fragment level, etc.; provides some testing tools to use Hilt-generated components in tests and provide methods to add or replace test bindings.

8. The work assistance system developed based on the AI ​​big model according to claim 1 is characterized in that: The Room is a Jetpack architecture component in Android that provides an abstraction layer for data persistence and operations on SQLite databases.

9. The work assistance system developed based on the AI ​​big model according to claim 1 is characterized in that: The testing and continuous integration / continuous deployment modules improve the quality and reliability of the code while shortening the development and release cycles. Testing refers to various types of inspections on the code to ensure that it meets the expected functionality and performance. Continuous integration refers to frequently merging the developer's code into the main branch and performing automated building and testing. Continuous deployment refers to automatically deploying the tested code to the production environment.

Images