System for designing large language model application workflows

Abstract

A system for designing large language model application workflows, including a process analysis module, a component mapping module and a concatenation module. The process analysis module uses a plurality of predefined system prompts to detect, analysis and disassemble a question message input by a user, to decompose the question message into multiple component contents. After the component mapping module understands the function or goal of each component content, it maps the component content to predefined related function modules based on the function or goal. The concatenation module pairs and concatenates the mapped function modules to generate a structured workflow. Through the system of the present invention, the user only needs to input the question content, and the system can break it down into intentions or prompt words to produce a set of LLM application workflow for the user.

Description

[Technical Field]

[0001] This invention relates to an AI model application field, and in particular to a system for designing a workflow for a large language model. [Previous Technology]

[0002] The large language model (LLM) application development process involves users dividing the target LLM application into necessary components, and then finding suitable programs for each component to build the complete application service. Currently, the common practice is manual design, where users manually break down the desired LLM application from beginning to end, and develop and write corresponding programs for each component, such as input processing, data cleaning, and model inference, to complete the LLM application service development. This process requires users to have a very clear understanding of the LLM application's workflow and a considerable level of background knowledge. However, not all users are professionals, so they often spend a lot of time designing and assembling the target application's workflow modules.

[0003] While some tools currently offer modular workflow building tools for LLM applications, they still require users to have certain background knowledge and a clear understanding of the workflow before and after designing the application system. In addition, it is difficult for users with less background knowledge to get started. Therefore, if there were a tool that could automatically build a prototype of the workflow module desired by the user based on the input prompts, and only need to make some additions according to other requirements, it would not only greatly reduce the threshold for users to build LLM applications, but also reduce the basic construction time of the workflow.

[0004] In view of the above, the present invention addresses the deficiencies of the prior art and future needs by proposing a system for designing a large-scale language model workflow. The specific architecture and implementation method are detailed below: [Summary of the Invention]

[0005] The main objective of this invention is to provide a system for designing a large-scale language model workflow. It utilizes a semantic optimization engine and a large-scale language model to analyze and decompose the user's input question content (or prompt words) and its construction intent, so as to break down complex question content into more specific component content.

[0006] Another objective of this invention is to provide a system for designing a large-scale language model workflow, which is supplemented by related modularization functions to map the content of the split components to predefined functional modules, complete the pairing of intent and functional modules, and perform collaborative connection of input and output formats between the preceding and following functional modules to construct a structured workflow.

[0007] Another objective of the present invention is to provide a system for designing a large-scale language model workflow. Users only need to input the question content, and the system will decompose it into intents or prompt words to generate a set of LLM application workflows that users need, thus solving the problem of the large amount of time spent by users in setting up background knowledge and system construction for setting up LLM applications.

[0008] To achieve the above objectives, the present invention provides a system for designing a large-scale language model workflow, comprising: a workflow analysis module, which uses predefined multiple system prompts to detect, analyze, and decompose a user-input question message to decompose the question message into multiple component contents; a component mapping module, connected to the workflow analysis module, which, after understanding the function or goal of each component content, maps the component content to relevant functional modules based on the function or goal, wherein the functional modules are predefined; and a concatenation module, connected to the component mapping module, which pairs and concatenates the mapped functional modules to generate a structured workflow.

[0009] According to an embodiment of the present invention, the component content includes multiple intentions and / or multiple prompt words.

[0010] According to an embodiment of the present invention, the process analysis module includes a large language model for the purpose of detecting question messages and decomposing the question messages into component content with sequential relationships.

[0011] According to an embodiment of the present invention, the process analysis module includes a semantic optimization engine for analyzing the semantic structure of component content and automatically adjusting the semantics of component content according to different domains or contexts to improve the accuracy of the mapping function module.

[0012] According to an embodiment of the present invention, the component content is mapped to the functional module in a one-to-one or one-to-many manner.

[0013] According to an embodiment of the present invention, the component mapping module is designed based on a user requirement or a number of successful workflows in the past, and adjusts a weight mapped to the functional module.

[0014] According to an embodiment of the present invention, the serial connection module first compares the input and output formats of each functional module and then connects them in series according to the restrictions of the input and output formats.

[0015] According to an embodiment of the present invention, when the input and output formats of the functional modules are incompatible, the serial module automatically generates a solution or outputs an error message.

[0016] According to an embodiment of the present invention, a test module is further included, which is connected to the serial module. After the serial module generates a structured workflow, the test module automatically runs the structured workflow to perform the test, and automatically adjusts or optimizes the structured workflow based on a test result.

Implementation Method

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0019] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0020] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0021] It should also be further understood that the term "and / or" as used in the specification of this invention and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0022] This invention provides a system for designing a large-scale language model workflow, housed in a network-connected electronic device. The electronic device includes a storage medium, an input / output device, and a processor. The storage medium stores multiple modules of the system of this invention. The processor is coupled to the storage medium and the input / output device, and accesses and executes the modules of the system of this invention. The storage medium can be memory, flash memory, hard disk, etc., the input / output device can be a mouse, keyboard, handwriting tablet, etc., and the processor can be a central processing unit (CPU), digital signal processor (DSP), microprocessor (MPU), graphics processing unit (GPU), or microcontroller (MCU). When a user sends a question message, the system of this invention can decompose the question message into specific intents or prompt words to more accurately match them to the corresponding functional modules. These functional modules are then connected one by one to form a structured workflow. Users do not need to manually select multiple prompt projects or toolkits. Even users outside the engineering field can easily build large-scale language models that meet the workflow requirements.

[0023] Please refer to Figure 1, which is a block diagram of the system 10 for designing the workflow of a large-scale language model according to the present invention. The system 10 for designing the workflow of a large-scale language model includes a workflow analysis module 12, a component mapping module 14, and a connection module 16. The workflow analysis module 12 is connected to the component mapping module 14, and the component mapping module 14 is connected to the connection module 16. When a user inputs a question message 30, the workflow analysis module 12 uses predefined multiple system prompts 22 to detect, analyze, and decompose the question message 30, and determines which system prompts are contained in the question message, so as to decompose the question message 30 into multiple component contents. These component contents include multiple intents and / or multiple prompts. The intents are what the user wants to do, such as data collection or content comparison, while the prompts are words that the system 10 for designing the workflow of a large-scale language model can recognize, such as file upload, file database, and result output. These intents and prompts also correspond to system prompts 22. The component mapping module 14 can understand the function or goal of each component content, and map the component content to the relevant functional modules based on the understood function or goal. These functional modules are also predefined. Finally, the concatenation module 16 pairs and concatenates the mapped functional modules to generate a structured workflow that matches the user's query message.

[0024] Please refer to Figure 2, a block diagram of another embodiment of the system for designing a large language model workflow. The system 10 for designing a large language model workflow includes a workflow analysis module 12, a component mapping module 14, and a concatenation module 16. The workflow analysis module 12 includes a large language model 122 and a semantic optimization engine 124. When a user inputs a question message 30, the large language model 122 in the workflow analysis module 12 uses predefined plural system prompts 22 to detect, analyze, and decompose the question message 30 to detect the purpose of the question message and decompose the question message into component content with a sequential relationship. If there is component content that the large language model 122 cannot decompose, the semantic optimization engine 124 analyzes the semantic structure of the component content and automatically adjusts the semantics of the component content according to different domains or contexts. In this way, it helps to improve the accuracy of the subsequent mapping function module 24. The component mapping module 14 is used to analyze the construction intent, that is, to understand the specific functions or goals involved in each component content. Then, based on the understood functions or goals, the component mapping module 14 matches the component content with each functional module 24. If the matching results are consistent, the component content is mapped to the matching functional module 24. Each component content corresponds to one or more functional modules 24 in the system, ensuring that the system can select the most appropriate function to process according to the user's needs. After mapping to each functional module 24, the concatenation module 16 matches the input and output formats of these functional modules 24 and concatenates them according to the constraints of the input and output formats. For example, if the output format of functional module A is the same as the input format of functional module B, it means that functional modules A and B can be paired, and the order of functional module A is before that of functional module B. And so on, so that all functional modules 24 can be concatenated to generate a structured workflow that matches the user's query message. In this way, if there are more needs in the future, we can simply add or remove from this structured workflow, saving a lot of time and effort in infrastructure construction.

[0025] If it is found that the input / output format of any function module 24 is incompatible with the input / output format of other function modules 24, the serial module 16 will automatically generate a solution or output an error message, requiring the user to re-enter the query message.

[0026] The system 10 for designing the workflow of a large-scale language model also includes a test module 18, which is connected to the serial module 16. After the serial module 16 generates a structured workflow, the test module 18 automatically runs the structured workflow to perform tests and generates a test result. The test module 18 then automatically adjusts or optimizes the structured workflow based on the test result to improve the stability and efficiency of the system operation.

[0027] In one embodiment of the present invention, the large language model 122 used by the process analysis module 12 is an instruction-type language model, such as LLaMA3-8B-Instruct.

[0028] Figure 3 is a schematic diagram of one embodiment of the process analysis module in this invention. If the user inputs a question message 30 as "I want to build an LLM file retrieval tool", the process analysis module 12 receives the question message 30 and then uses predefined system prompts 22 to analyze the intent of the question message 30, such as "Your task is to help me detect and analyze the intent of the input sentence and divide it into different modules with a sequential order". Then, the process analysis module 12 can decompose the question message 30 into multiple component contents 32, including six component contents 32 such as file upload, data processing, file database, user input, content comparison and result output. The process analysis module 12 then provides these six component contents 32 to the component mapping module.

[0029] Next, the component mapping module can find the functional modules that match the content of these six components from the predefined functional modules, and map the component content to the corresponding functional modules. It should be noted that when the process analysis module 12 is split, it will not necessarily be split into very concise parts or into parts with the exact same name as the functional modules. Instead, it will be a meaningful sentence. For example, the action of "file upload" may be broken down by the process analysis module 12 into "I want to upload a file". After the component mapping module understands the component content 32 of "I want to upload a file", it will then map it to the functional module of "file upload".

[0030] In summary, the system for designing a large-scale language model workflow provided by the present invention has the following advantages: 1. It uses a semantic optimization engine and a large-scale language model to analyze and decompose the user's input question content, so as to break down the complex question content into more specific intents or prompt words; 2. It maps the decomposed component content to predefined functional modules, completes the pairing of intents / prompt words with functional modules, and performs collaborative connection of input and output formats between functional modules before and after, so as to construct a complete workflow. Users do not need to have background knowledge of LLM.

[0031] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of the present invention. Therefore, all equivalent variations or modifications made in accordance with the features and spirit described in the claims of the present invention should be included within the scope of the patent application of the present invention. [Simplified Explanation of the Diagram]

[0017] Figure 1 is a block diagram of the system for designing the workflow of a large-scale language model according to the present invention. Figure 2 is a block diagram of another embodiment of the system for designing the workflow of a large-scale language model according to the present invention. Figure 3 is a schematic diagram of one embodiment of the workflow analysis module in the present invention.

Claims

1. A system for designing workflows for large-scale language models, comprising: A workflow analysis module detects, analyzes, and breaks down a user-input query message using predefined multiple system prompts, decomposing the query message into multiple component contents; a component mapping module connects to the workflow analysis module, which understands the function or goal of each component content and maps the component content to a related multiple functional modules based on the function or goal, wherein the functional modules are predefined; and a concatenation module connects to the component mapping module, which pairs and concatenates the mapped functional modules to generate a structured workflow.

2. A system for designing a large language model workflow as described in claim 1, wherein the components include plural intentions and / or plural cue words.

3. A system for designing a large language model workflow as described in claim 1, wherein the workflow analysis module includes a large language model for detecting the purpose of the query message and decomposing the query message into components with a sequential relationship.

4. A system for designing a large language model workflow as described in claim 1, wherein the workflow analysis module includes a semantic optimization engine for analyzing the semantic structure of the content of the components and automatically adjusting the semantics of the content of the components according to different domains or contexts to improve the accuracy of mapping the functional modules.

5. A system for designing a large language model workflow as described in Request 1, wherein the content of such components is mapped to such functional modules in a one-to-one or one-to-many manner.

6. A system for designing a large language model workflow as described in claim 1, wherein the component mapping module adjusts a weight mapped to the functional modules based on a user requirement or a previously successful multiple workflow design.

7. A system for designing a large language model workflow as described in claim 1, wherein the concatenation module is first paired with the input / output formats of each of the functional modules and then concatenated according to the constraints of the input / output formats.

8. A system for designing a large language model workflow as described in claim 7, wherein when the input and output formats of the functional modules are incompatible, the concatenated module automatically generates a solution or outputs an error message.

9. The system for designing a large language model workflow as described in claim 1 further includes a test module connected to the serial module. After the serial module generates the structured workflow, the test module automatically runs the structured workflow for testing and automatically adjusts or optimizes the structured workflow based on a test result.

Images