CN-122021246-A - Office large model system

CN122021246ACN 122021246 ACN122021246 ACN 122021246ACN-122021246-A

Abstract

The application discloses an office large model system, which comprises an infrastructure layer, a data resource layer, a public supporting layer, a basic application layer, a business application layer, a user interaction layer and a security and operation and maintenance management system, wherein the infrastructure layer is configured to provide bottom support for the office large model system, the data resource layer is configured to provide data for pre-training, fine tuning and business adaptation of the office large model system, the model capability layer comprises at least one of a single-mode office large model and a multi-mode office large model, the public supporting layer is configured to provide support for upper office applications, the basic application layer is configured to provide basic office applications, the business application layer is configured to provide various business scene applications, the user interaction layer is configured to provide a unified operation interface and entrance for various business application systems, and the security and operation and maintenance management system is configured to provide security and operation and maintenance guarantee for various layers of the office large model system. According to the scheme, the application scene of the office large model system can be widened as much as possible, and the support function of the office large model system is enriched, so that office requirements can be met as much as possible.

Inventors

LI ANG
GUO XIAOTIAN
XU QIUCHEN
Zhong Gun

Assignees

科大讯飞股份有限公司

Dates

Publication Date: 20260512
Application Date: 20251223

Claims (10)

1. An office large model system, comprising: An infrastructure layer configured to provide underlying support for the office large model system; A data resource layer configured to provide data for pre-training, fine tuning, business adaptation of the office large model system, the data comprising general data, domain specific data, and business data; The model capability layer comprises at least one of a single-mode office large model and a multi-mode office large model; a common support layer configured to provide support for upper office applications, including at least one of model management, knowledge base platforms; a base application layer configured to provide a base office application; the business application layer is configured to provide various business scene applications; The user interaction layer is configured to provide a unified operation interface and an entrance for various business application systems; And the security and operation and maintenance management system is configured to provide security and operation and maintenance guarantee for each level of the office large model system.
2. The office automation system of claim 1, wherein, in the case where the base office application comprises an office document, the model capability layer is configured to have processing capabilities of generating classes and understanding classes for the single-mode office automation model, and processing capabilities of image question-answering, image description, image-text retrieval, chart information processing, text-rendering, graphical-rendering, speech recognition, speech synthesis, and text-generated audio-visual, wherein the generating classes comprise at least one of text drafting, text rewriting, abstract extraction, text translation, text summarization, code generation, and formula generation, and the understanding classes comprise at least one of document structuring, text ordering, semantic understanding, logical reasoning, text question-answering, multi-round dialogue, fact consistency verification, keyword extraction, semantic extraction, tool call, and emotion analysis.
3. The office large model system according to claim 1, wherein the model management is configured to include at least one of model up-put and down-put, model switching, model update, access right control, model version management, personalized setting of large models, statistics, display or printing of use cases of office applications, restriction management of office large models according to different office applications, continuous iteration of models according to data and business demand changes, providing model API interface to application layer scheduling, and MCP service providing large model AI functions; And/or the knowledge base platform is configured to comprise at least one of supporting data of a batch import system, defining or expanding document bases of different service dimensions according to user requirements, supporting document retrieval and inquiry, document authority management, supporting real-time updating and synchronizing data, supporting identification and processing of different document formats, and supporting hierarchical storage management of data.
4. The office large model system of claim 1, wherein, in the case where the base office application comprises an office document, the base application layer comprises at least one of: in the case where the base office application includes intelligent authoring, the base application layer is configured to perform content generation in a natural language manner; In the case where the base office application includes intelligent typesetting, the base application layer is configured to typeset based on semantic recognition; In the case where the base office application includes intelligent collation, the base application layer is configured to identify errors in the document and to collate; In the case where the base office application includes automatically generating a presentation, the base application layer is configured to automatically generate a presentation; in the case where the base office application includes intelligent questions and answers, the base application layer is configured to understand questions in natural language and feed back results; in the case where the base office application includes intelligent reading, the base application layer is configured to assist the user in quickly understanding the document content; in the case where the base office application includes an intelligent search, the base application layer is configured to search according to a natural language or element pair; In the case that the basic office application comprises intelligent duplicate checking, the basic application layer is configured to search for documents with similar or similar contents by comparison; in the case where the base office application includes intelligent entry, the base application layer is configured to quickly build and enter data in a natural language manner; in the case where the base office application includes intelligent cleansing, the base application layer is configured to cleansing data in a natural language manner; in the case where the base office application includes intelligent computing, the base application layer is configured to issue data classification and statistical instructions in a natural language manner; In the case where the base office application includes intelligent analysis, the base application layer is configured to query, count, interpret, and visually analyze data in a natural language manner; in the case where the base office application includes smart form operations, the base application layer is configured to automatically generate code in natural language fashion and to quickly operate forms from the code to aid in data analysis.
5. The office large model system according to claim 1, wherein the construction flow of the office large model system comprises the following processes of planning and designing, model customized construction, model implementation construction and continuous improvement; The method comprises the following sub-processes of target definition, current situation assessment and scheme design, wherein the model customization construction comprises the following sub-processes of base model selection, data preparation, base model pre-training, industry fine tuning, model verification test and model reasoning optimization, the model implementation construction comprises the following sub-processes of knowledge base construction and association, calculation power construction and model deployment, and effect assessment, and the continuous improvement comprises the following sub-processes of service monitoring and user feedback.
6. The office large model system of claim 5, wherein the objective definition is configured to include at least one of defining core value and boundaries, defining key performance indicators, dismantling stage objectives and milestones; and/or the presence assessment is configured to include at least one of assessing resource suitability, establishing assessment baselines and methods, risk assessment; And/or the solution design is configured to include at least one of a delivery solution design, a demand problem tracking, a demand change.
7. The office large model system of claim 5, wherein the base model selection is configured to include at least one of a model parameter magnitude, a model modality, a model structure, wherein the model parameter magnitude is configured to be comprehensively considered in connection with office scene requirements and resource constraints, the model modality includes at least one of text unimodal, image unimodal, multi-modal, the model structure includes a decoder-type autoregressive model, a MoE architecture, a multi-modal fusion structure; And/or the data preparation is configured to comprise at least one of data acquisition, data analysis and conversion, data cleaning, data enhancement, data labeling, data management and data storage, wherein the data cleaning is configured to comprise at least one of content compliance identification and filtering, repeated data identification and deletion, noise content removal, document content quality grading and document content segmentation, and the data enhancement is configured to comprise at least one of the following; and/or the base model pre-training is configured to include at least one of a base technology support configured to include at least one of heterogeneous cluster virtualization, failure awareness, distributed file system and storage acceleration, a stability guarantee configured to include at least one of distributed communication network testing, tensor state monitoring in training, loss stability monitoring and alerting, asynchronous storage and checkpoint interrupt recovery, a training optimization configured to include at least one of data proportioning optimization, operator fusion and optimization acceleration, data and model hybrid parallelism policy, And/or the industry fine tuning is configured to include at least one of model fine tuning, office capability optimization, security alignment, wherein the model fine tuning is configured to include at least one of low rank adaptation training, multi-task fine tuning, partial order learning, reinforcement learning based on human feedback, the office capability optimization is configured to include at least one of data reasoning capability optimization, retrieval enhancement generation, instruction compliance, content security assurance, and the security alignment is configured to include at least one of model input/output compliance checking, data preference alignment, model value alignment; And/or the model verification test is configured to include at least one of a task effect test, a performance test, a security test; and/or the model inference optimization is configured to include at least one of employing at least one optimization technique of quantization, clipping, knowledge distillation to reduce inference resource consumption, employing at least parallel inference policies to increase computational efficiency, employing at least Pd-split architecture to reduce inference redundancy.
8. The office large model system of claim 5, wherein the knowledge base construction and association is configured to include at least one of knowledge base construction, knowledge base association, wherein the knowledge base construction is configured to include at least one of data access, knowledge acquisition, knowledge slicing, knowledge enhancement, knowledge indexing; The computing power dispatching platform construction and model deployment is configured to comprise at least one of heterogeneous computing clusters, storage and networks, wherein the heterogeneous computing clusters comprise at least one of hardware selection, cluster scale and resource management, and the storage comprises at least one of hierarchical storage strategies, distributed and redundant designs; and/or the effect evaluation is configured to include at least one of an application effect evaluation, a stress test, a security evaluation.
9. The office large model system of claim 5, wherein the continuous improvement is configured to include at least one of monitoring system service level agreement metrics, periodically polling, establishing user feedback channels to collect user ratings for generated content, periodically analyzing feedback data and re-evaluating model capabilities after feedback data as input data for model iterations and system optimization.
10. The office large model system of claim 5, wherein the planning and design is configured to technically direct and target the model customized construction and the model implementation construction with reference to the planning and design implementation; And/or the model customization construction is configured to provide an adaptive basis for the model implementation construction; and/or the model implementation construction is configured to verify whether the customized model meets the real requirements through actual use and effect evaluation, and reversely feed back the problems found in the implementation to promote the model secondary optimization; and/or the continuous improvement is configured to continuously monitor and operate and check the whole construction result, take improvement measures for the problems and feed back to at least one of the design and construction; And/or the construction mode of the office large model system comprises a full-self-building mode, a collaborative construction private deployment mode and a subscription mode.

Description

Office large model system Technical Field The application relates to the technical field of artificial intelligence, in particular to an office large model system. Background The intelligent terminal equipment in the office large model system energization office scene provides comprehensive and efficient office assistance for users, and improves the working efficiency of the users. Most of the existing office large model systems are based on a general large model, are difficult to adapt to application requirements of different scenes, and have single supporting functions, so that the office requirements are difficult to really meet. In view of this, how to widen the application scenario of the office large model system as much as possible and enrich the support functions of the office large model system to meet the office requirements as much as possible becomes a urgent problem to be solved. Disclosure of Invention The application mainly solves the technical problem of providing the office large model system, which can widen the application scene of the office large model system as much as possible and enrich the support function of the office large model system so as to meet the office requirement as much as possible. In order to solve the technical problems, the first aspect of the application provides an office large model system, which comprises an infrastructure layer, a data resource layer, a model capability layer, a public supporting layer, a basic application layer, a business application layer, a user interaction layer and a security and operation and maintenance management system, wherein the infrastructure layer is configured to provide bottom support for the office large model system, the data resource layer is configured to provide data for pre-training, fine tuning and business adaptation of the office large model system, the data comprises general data, domain specific data and business data, the model capability layer comprises at least one of a single-mode office large model and a multi-mode office large model, the public supporting layer is configured to provide support for upper office applications and comprises at least one of model management and a knowledge base platform, the basic application layer is configured to provide basic office applications, the business application layer is configured to provide various business scene applications, the user interaction layer is configured to provide uniform operation interfaces and inlets for various business application systems, and the security and operation and maintenance management system is configured to provide security and operation and maintenance guarantee for various levels of the office large model system. According to the scheme, on one hand, the data resource layer is configured to provide data for pre-training, fine tuning and business adaptation of the office large model system, the data comprises at least one of general data, domain specific data and business data, the model capability layer comprises at least one of a single-mode office large model and a multi-mode office large model, so that the data resource layer can provide different dimension data such as general data, domain specific data and business data for the model capability layer in all aspects of pre-training, fine tuning and business adaptation, the data resource layer is beneficial to improving the suitability of the office large model system to different business requirements and different data modes, on the other hand, the public support layer is configured to provide support for upper office applications, including at least one of model management and a knowledge base platform, the basic application layer is configured to provide basic office applications, and the business application layer is configured to provide various business scene applications, and as close as possible, and as a user interaction layer is configured to provide a uniform operation interface and an entrance to the business application system and a business application layer can provide a uniform operation interface and a uniform operation channel and a business system as possible under various business scenes, and the business system can also provide different security requirements and a uniform security channel as far as possible. Therefore, the application scene of the office large model system can be widened as much as possible, and the support function of the office large model system is enriched so as to meet the office requirement as much as possible. Drawings FIG. 1 is a schematic diagram of a framework of one embodiment of an office large model system of the present application; FIG. 2 is a simplified schematic diagram of one embodiment of the office large model system of FIG. 1; fig. 3 is a schematic diagram of the construction of an embodiment of the office large model system of the present application. Detailed Description The following describes embodiments of the present