An end-side vectorization memory-based personalized AI execution system and method

Abstract

The application discloses an individualized AI execution system and method based on end-side vector quantization memory, and relates to the technical field of artificial intelligence and automatic control.The application constructs a user behavior vector library locally on a terminal, and vectorizes and stores offline user historical data, preference data and scene data;according to the current device state, user input or environmental trigger condition, vector similarity retrieval is completed on the end side to generate an individualized execution strategy matching user habits; according to the individualized execution strategy, automatic scheduling and execution of automatic tasks including but not limited to API calling, script running, process triggering, RPA operation, intelligent control instruction, automatic payment instruction, etc., are performed, and the execution result is updated to the local vector library to realize end-side closed-loop iteration.The application can complete local AI decision and automatic execution without relying on the cloud, has strong data privacy and fast response speed, is suitable for offline or privacy-first scenarios such as AI toys, smart home, IoT devices, vehicle-mounted terminals, automatic payment, local life consumption, etc., has a clear right protection scope, easy evidence for infringement, and has high practical and right protection values.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, edge-cloud collaboration, and automated execution technology. Specifically, it relates to an application system and method that implements vectorized memory storage and retrieval locally on a terminal device, and combines a cloud-based personalized planning engine and automated execution unit to complete privacy-preserving automated task execution. Background Technology

[0002] This application is a further improvement and extension of the technical solution of the prior application with application number 2026102583924. That prior application discloses a technical solution that performs vectorized memory storage and semantic retrieval locally on the terminal device, and uploads only a simplified context to the cloud, effectively protecting user privacy and reducing cloud costs.

[0003] However, the earlier application primarily addressed the issue of localized storage and retrieval of "memory." Existing AI execution systems, especially those combining AI agents and RPA technology, while capable of performing complex automated tasks, generally suffer from the following technical shortcomings: 1. High privacy risk: To complete complex automated tasks, the system needs to upload the complete user interaction history and business data to the cloud, which poses a risk of data leakage.

[0004] 2. High cost: Each task requires the transmission of a large amount of historical context, resulting in high bandwidth and cloud computing power consumption.

[0005] 3. Lack of personalization: Most cloud-based planning engines are general-purpose and cannot dynamically optimize task paths based on users' long-term and private personalized preferences, resulting in low execution efficiency and poor user experience.

[0006] 4. Poor hardware compatibility: Traditional agent architectures have high hardware performance requirements and are difficult to deploy on low-power embedded devices (such as AI toys and smart home terminals), which limits application scenarios. Summary of the Invention

[0007] (a) Technical problems to be solved This invention aims to address the technical problems faced by existing AI automated execution systems when deployed on edge devices, such as difficulties in privacy protection, high cloud costs, and low efficiency of personalized execution.

[0008] (II) Technical Solution A personalized AI execution system based on edge-side vectorized memory, characterized in that it includes: Terminal devices, deployed on the user side, include: • Data acquisition unit, used to acquire user voice, text or touch input data; • The local vectorization memory module is used to store the user's historical interaction data and its feature vectors locally on the terminal, and quickly retrieve the N most relevant historical records based on the current input data through semantic retrieval to generate a concise context. • A communication unit, used to upload the simplified context to the cloud and receive the execution results returned by the cloud; • A local cache unit is used to temporarily store the simplified context and task instructions when the network is interrupted, and automatically resume execution after the network is restored, ensuring that tasks are not lost in weak network or no network environment.

[0009] Cloud-based execution systems include: • A personalized task planning engine is used to receive the simplified context and dynamically generate a task execution path that matches the user's personalized needs based on the user's historical preference information therein; • An automated execution scheduling unit is used to complete automated operations based on the personalized execution path, using methods including but not limited to RPA, API interface, smart agent, timed scheduling, script execution, workflow engine, and rule engine. The operations include but are not limited to food delivery ordering, payment operations, device control, office process automation, and third-party service calls. • An execution feedback unit is used to return the execution status and final result to the terminal device.

[0010] A personalized AI execution method based on the above system includes the following steps: Step 1: The terminal device collects user input data and generates a concise context containing the current command and the most relevant historical records through the local vectorization memory module; Step 2: The terminal device uploads the simplified context to the cloud execution system; Step 3: The cloud-based personalized task planning engine plans a personalized task execution path based on the user's historical preferences in the simplified context; Step 4: The cloud-based automated execution scheduling unit completes the automated operation according to the execution path; Step 5: The execution result is returned to the terminal device, and the memory is updated.

[0011] (III) Beneficial Effects 1. Excellent in both privacy and cost: By completing memory storage and retrieval locally on the terminal, the cloud only receives a simplified context, which eliminates the risk of leakage of the user's full privacy data from the architectural level, while significantly reducing bandwidth and cloud computing power consumption.

[0012] 2. Significantly improved execution efficiency: The cloud-based planning engine can directly utilize users' historical preferences (such as operating habits and decision-making tendencies) to dynamically optimize the execution path. Compared with traditional general solutions, the execution path planning time is shortened by more than 30%, and the number of user interactions is reduced by 50%, which significantly improves the execution efficiency and user experience of automated tasks.

[0013] 3. Strong hardware adaptability: The core edge module of this invention is lightweight and can run efficiently on low-power embedded chips such as ESP32, and can be widely deployed in AI toys, smart home control, portable IoT devices and other scenarios.

[0014] 4. High robustness: Through local caching and resume mechanisms, tasks can still be guaranteed not to be lost in weak network or network outage environments, and will be automatically completed after the network is restored.

[0015] 5. Extremely high compatibility: Supports all automated execution methods such as RPA, API, Agent, scheduled operations, scripts, and workflows, covering the entire automation scenario. (See attached image for details) Figure 1 This is a side-end version system architecture diagram of the present invention. Figure 2 This is a flowchart of the side-end version method of the present invention. Detailed Implementation

[0016] Example 1: Personalized ordering based on AI toys The terminal is an AI toy with a built-in ESP32-S3 chip. The user inputs via voice: "Order me Kung Pao Chicken from that restaurant last time, but this time make it less spicy."

[0017] 1. Memory Retrieval: The toy's local vectorized memory module retrieves the history and finds that the "last restaurant" is "Chuanweiju", "Kung Pao Chicken" is recorded, and "less spicy" is a newly added preference. A simplified context is generated: "User instruction: Order Kung Pao Chicken from the last restaurant, current preference: less spicy; history: restaurant: Chuanweiju, flavor: mild".

[0018] 2. Cloud-based planning: The cloud-based personalized task planning engine receives a simplified context and automatically plans the task path based on the "mildly spicy" history and the new preference for "less spicy": First, log in to the food delivery platform and search for "Chuanweiju"; second, select "Kung Pao Chicken" and note "less spicy"; third, pay using the historical payment method.

[0019] 3. Automated execution: The cloud-based automated execution scheduling unit completes the ordering process according to the planned path through RPA, API interfaces, or agents.

[0020] 4. Feedback and Updates: The execution result is returned to the toy with a voice announcement, and the new command "less spice" is saved as a new preference in the local memory bank.

[0021] Example 2: Smart Home Automation Control in a Weak Network Environment The terminal is a low-power smart home control system. The user issues the command: "Away mode, turn off all lights and air conditioning."

[0022] 1. Local Cache: When the network is interrupted, the terminal generates a local simplified context (including commands and "away mode" history) and temporarily stores it in the local cache unit.

[0023] 2. Resumption of Download and Execution: After the network is restored, the system automatically uploads a simplified context. The cloud-based personalized engine plans an execution path that matches the user's habits based on their historical away-from-home patterns (e.g., turning off the lights first and then the air conditioner), and the automated execution scheduling unit completes the control.

[0024] 3. Result feedback: The execution results are synchronized to the terminal and the local memory is updated.

Claims

1. A personalized AI execution system based on edge-side vectorized memory, characterized in that, include: Terminal devices, deployed on the user side, include: The data acquisition unit is used to acquire user input data; The local vectorization memory module is used to store the user's historical interaction data and its feature vectors locally on the terminal, and generate a simplified context based on the current input data; A communication unit is used to upload the simplified context to the cloud; A local cache unit is used to temporarily store the simplified context and task instructions when the network is interrupted, and automatically resume the transmission after the network is restored; Cloud-based execution systems include: A personalized task planning engine is used to dynamically generate personalized task execution paths based on the user's historical preference information in the simplified context. An automated execution scheduling unit is used to complete automated operations based on the personalized execution path, using methods including but not limited to RPA, API interface, smart agent, timed scheduling, script execution, workflow engine, and rule engine. The operations include but are not limited to food delivery ordering, payment operations, device control, office process automation, and third-party service calls. The execution feedback unit is used to return the execution result to the terminal device.

2. The system according to claim 1, characterized in that, The local vectorization memory module includes: Lightweight vectorization model for converting user input data into feature vectors; A local vector database is used to store feature vectors and their original text from historical interaction data; The semantic retrieval unit is used to retrieve the N most relevant historical data based on the feature vector of the current input data to generate the simplified context.

3. The system according to claim 1, characterized in that, The personalized task planning engine generates personalized task execution paths by dynamically adjusting the order, parameters, or methods of task execution based on the user's historical operating habits, decision-making tendencies, or preferences.

4. The system according to claim 1, characterized in that, The terminal device is an embedded device, a smart toy, a smart home device, an IoT device, or a portable low-power terminal.

5. A personalized AI execution method based on edge-side vectorized memory, applied to the system described in any one of claims 1-4, characterized in that, Includes the following steps: Step 1: The terminal collects user input data and generates a simplified context through the local vectorization memory module; Step 2: The terminal uploads the simplified context to the cloud execution system; Step 3: The cloud-based personalized task planning engine generates personalized task execution paths based on the user's historical preferences in a simplified context; Step 4: The cloud-based automated execution scheduling unit completes the automated operation based on the personalized execution path; Step 5: The execution result is returned to the terminal device.

6. The method according to claim 5, characterized in that, The process of generating a simplified context in step 1 includes: vectorizing the user input data and retrieving the N historical data with the highest semantic similarity from the local vector database, and packaging them together with the current input.

7. The method according to claim 5, characterized in that, Step 3 generates a personalized task execution path, including: dynamically optimizing the task execution order, branch selection, or parameter configuration based on the user's historical preferences.

8. A terminal device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described in any one of claims 5-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method described in any one of claims 5-7.

Images