Large-language-model twin system

Abstract

A large-language-model (LLM) twin system (100) can include: a user interface (200) for receiving a query from a user; a digital twin (150) for simulating and verifying the parameters of a plurality of devices in a factory; and a plurality of LLM agents including a manager agent (110) for controlling other LLM agents, communicating with the LLM agents, receiving a query, interpreting the query by using one or more of the LLM agents to update the query, transmitting the updated query to one or more of the other LLM agents to receive an answer to the query, receiving the answer from the one or more of the other LLM agents, and transmitting the answer to the user via the user interface (200).

Claims

1. A large language model twin system comprising: a user interface configured to receive queries from a user; a digital twin configured to simulate each of several pieces of equipment in a factory; and several large language model (LLM) agents, wherein the several LLM agents include a manager agent configured to communicate with other LLM agents among the several LLM agents and the digital twin; and the manager agent is configured to, in response to a user entering the query into the user interface, interpret the query using one or more of the other LLM agents to update the query; send the updated query to one or more of the other LLM agents to receive a response to the query; receive the response from one or more of the other LLM agents; and send the response to the user via the user interface.

2. In response that the query relates to a function of one of the plurality of devices of the factory, the manager agent is configured to request the digital twin to simulate the response from one or more LLM agents in order to validate the response, and in response that the digital twin has validated the response by simulating the response, the digital twin is configured to transmit the response to the manager agent, the large-scale language model twin system according to claim 1.

3. The large-scale language model twin system according to claim 2, wherein the query relates to an error in a first device among the plurality of devices, the response is a change to the parameters of the first device, the digital twin is configured to simulate the change to the parameters of the first device, determine whether the change to the parameters corrects the error, simulate the updated parameters until the error is resolved in response to the error not being corrected, and transmit the updated parameters to the user interface via the manager agent.

4. The large-scale language model twin system according to claim 1, wherein the manager agent is configured to determine which of the plurality of LLM agents can answer the query, the query relates to one of the plurality of devices, and the manager agent is further configured to update the parameters of the device based on the determined answer.

5. The large-scale language model twin system according to claim 1, wherein the manager agent is configured to provide a prompt to the user via the user interface when the query is too ambiguous, and the prompt includes a request asking the user to rephrase the query.

6. The large-scale language model twin system according to claim 1, wherein the manager agent is configured to calculate an initial answer, and then transmit the initial answer to one or more of the LLM agents and the digital twin for verification, the digital twin is configured to simulate the initial answer and other possible answers, determine any adjustments to the devices among the devices required to obtain the updated answer, and transmit the answer to the manager agent.

7. The large-scale language model twin system according to claim 6, wherein the one or more other LLM agents are trained on internal data and user data, the internal data includes at least one of operational data, maintenance data and system specifications, and the user data includes at least one of troubleshooting data and project files.

8. The Large-Scale Language Model Twin System according to claim 1, wherein the plurality of LLM agents further include expert agents, machine agents, and line agents, the expert agents being configured to be trained with data relating to changes in settings from specifications in order to operate the plurality of devices, the machine agents being configured to store and analyze data from each of the plurality of devices in order to optimize the operation of the plurality of devices, and the line agents being configured to manage the entire production line of the factory and optimize productivity in real time.

9. The large-scale language model twin system according to claim 1, wherein the manager agent is configured to determine whether to use an open LLM or a closed LLM to answer the query made by the user by interpreting the query, determining whether the query relates to confidential information, using the open LLM if the query does not relate to confidential information, and using the closed LLM if the query relates to confidential information, wherein the closed LLM is a closed internal network and the open LLM is an open network connected to the outside of the internal network.

10. The large-scale language model twin system according to claim 1, further comprising a RAG (Retrievable Augmented Generation) configured to provide data from an external knowledge base to the LLM twin system, wherein the RAG indexes text in the external knowledge base by converting it into a numerical representation that the LLM twin system can understand, retrieves the most relevant documents from the external knowledge base based on the determined similarity to the query, combines the retrieved documents with the query to form a new augmented query, and is configured to supply the augmented query to the plurality of LLM agents.

11. A user interface configured to receive queries from a user; a digital twin configured to simulate each of several pieces of equipment in a factory; and several Large Language Model (LLM) agents, wherein the several LLM agents include a manager agent configured to communicate with other LLM agents among the several LLM agents and the digital twin; and motion agents configured to control the several pieces of equipment in the factory, including the (one or more) processing paths and control parameters, via Supervisory Control and Data Acquisition (SCADA) software, wherein the user enters the query into the user interface, and in response that the query is a request for a manufacturing plan for an item to be produced, the manager agent searches for and determines the item to be produced, automatically generates a product design for the product based on the determined item, including forming a process image, transfers the product design to the motion agent to generate processing paths, and receives the processing paths from the motion agent. A large-scale language model twin system configured to produce the product in the factory using the generated product design and processing paths.

12. The large-scale language model twin system according to claim 11, wherein the manager agent is further configured to interpret the query using one or more of the other LLM agents to update the query.

13. The large-scale language model twin system according to claim 12, wherein the manager agent is further configured to provide a prompt to the user via the user interface when the query is too ambiguous, the prompt including a request asking the user to rephrase the query.

14. The Large-Scale Language Model Twin System according to claim 11, wherein the plurality of LLM agents further include expert agents, machine agents, and line agents, the expert agents being configured to be trained with data relating to changes in settings from specifications in order to operate the plurality of devices, the machine agents being configured to store and analyze data from each of the plurality of devices in order to optimize the operation of the plurality of devices, and the line agents being configured to manage the entire production line of the factory and optimize productivity in real time.

15. The large-scale language model twin system according to claim 11, wherein the manager agent is configured to determine whether to use an open LLM or a closed LLM to answer the query made by the user by interpreting the query, determining whether the query relates to confidential information, using the open LLM if the query does not relate to confidential information, and using the closed LLM if the query relates to confidential information, wherein the closed LLM is a closed internal network and the open LLM is an open network connected to the outside of the internal network.

16. A user interface configured to receive prompts from a user; a digital twin configured to simulate each of several pieces of equipment in a factory; and several Large Language Model (LLM) agents, wherein the LLM agents include a manager agent configured to communicate with other LLM agents among the LLM agents and the digital twin; an expert agent; a machine agent; and a line agent, wherein the expert agent is configured to be trained with data on changes in settings from specifications to operate the several pieces of equipment; the machine agent is configured to store and analyze data from each of the several pieces of equipment to optimize the operation of the several pieces of equipment; the line agent is configured to manage the entire production line of the factory and optimize productivity in real time; and in response to a user entering a query into the user interface, the manager agent interprets the query using one or more of the other LLM agents to update the query; and sends the updated query to one or more of the other LLM agents to receive an answer to the query. A large-scale language model twin system configured to, in response to the query relating to a function of one of the plurality of devices of the factory, request the digital twin to simulate the answer from one or more other LLM agents to verify the answer, receive the verified answer from the digital twin, and transmit the verified answer to the user via the user interface.

17. The large-scale language model twin system according to claim 16, wherein the manager agent is configured to calculate an initial answer and then transmit the initial answer to the one or more other LLM agents and the digital twin for verification, the digital twin is configured to simulate the initial answer and other possible answers, determine any adjustments to the devices among the plurality of devices required to obtain the updated answer, and transmit the answer to the manager agent.

18. The large-scale language model twin system according to claim 17, wherein the manager agent is configured to provide a prompt to the user via the user interface when the query is too ambiguous, and the prompt includes a request asking the user to rephrase the query.

19. The large-scale language model twin system according to claim 16, wherein the manager agent is configured to determine whether to use an open LLM or a closed LLM to answer the query made by the user by interpreting the query, determining whether the query relates to confidential information, using the open LLM if the query does not relate to confidential information, and using the closed LLM if the query relates to confidential information, wherein the closed LLM is a closed internal network and the open LLM is an open network connected to the outside of the internal network.

20. The large-scale language model twin system according to claim 16, further comprising a RAG (Search Enhancement Generation) configured to provide data from an external knowledge base to the LLM twin system, wherein the RAG indexes text in the external knowledge base by converting it into a numerical representation that the LLM twin system can understand, retrieves the most relevant documents from the external knowledge base based on the determined similarity to the query, and combines the retrieved documents with the query to form a new enhanced prompt, which is supplied to the plurality of LLM agents.

Images