An agent cross-domain knowledge hybrid retrieval method and system for privacy protection

Abstract

This invention provides a privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval method and system, relating to the fields of artificial intelligence and data security technology. The method includes: responding to a query request initiated by a cloud-based intelligent agent, parsing the privacy tags embedded in the request workflow definition; and concurrently executing the following two operations through an asynchronous task scheduler: sending the public subquery to a cloud-based knowledge retrieval service, and routing the private subquery to a privacy computing agent client deployed in the target local environment; the privacy computing agent client executing the private subquery in a local private knowledge base; and at a cloud aggregation point, aggregating the decrypted ciphertext knowledge fragment set and the public knowledge fragment set into unified knowledge data, and returning it to the intelligent agent. This invention achieves efficient, secure, and integrated retrieval and utilization of distributed heterogeneous knowledge bases by a cloud-based intelligent agent while ensuring that the original privacy data does not leave the local security boundary.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and data security technology, specifically to a method and system for cross-domain knowledge hybrid retrieval of intelligent agents with privacy protection. Background Technology

[0002] With the popularization of Large Language Model (LLM) technology and the in-depth development of agent applications, integrating external knowledge bases with agent systems has become a key path to improve their accuracy and reliability in vertical domain applications. This integration aims to provide agents with real-time and accurate external knowledge (i.e., knowledge context data) by using retrieval-enhanced generation technology, thereby reducing model "illusion" and improving their performance in professional tasks (such as customer service, data analysis, and decision support).

[0003] In existing technical solutions, the knowledge retrieval architecture of intelligent agents mainly presents two mainstream forms: The first type is a centralized cloud-native architecture, which stores all knowledge data (including public knowledge and sensitive corporate data) in a unified cloud vector database. Intelligent agents retrieve and call knowledge through cloud APIs. However, its core drawback is that data security and privacy cannot be guaranteed. For the core business departments of finance, healthcare, government, and large group enterprises, their operational data, customer information, technical documents, etc. are all highly sensitive assets. Uploading this data to the cloud in plaintext will face serious problems such as loss of data sovereignty, potential leakage risks, and difficulty in meeting compliance audit requirements, which seriously limits the applicability of this architecture in scenarios with high security requirements.

[0004] The second type is a localized closed architecture. To address data security challenges, some solutions choose to deploy the agent and its entire associated knowledge base entirely in the user's local or private environment to achieve physical data isolation. While this architecture achieves absolute data security and controllability, it leads to serious "data silos" and "capability silos" problems. On the one hand, the agent cannot access and utilize public, non-sensitive shared knowledge (such as product manuals, industry regulations, and general technical encyclopedias) existing in the cloud or other departments of the group, resulting in an incomplete knowledge system and limited decision-making capabilities. On the other hand, this architecture fragments the development and operation processes: the development, debugging, version management, and monitoring of the agent cannot be efficiently coordinated between cross-regional teams, significantly increasing the complexity and cost of operation and maintenance, and sacrificing the system's scalability and agility.

[0005] Most existing technologies focus on the security of parameter exchange during the model training phase, while neglecting the technical challenges faced by agents in real-time and collaboratively retrieving distributed and heterogeneous knowledge sources during the inference phase. This often leads to developers having to piece together multiple systems when building complex agents that require both public knowledge assistance and the processing of sensitive data, resulting in process fragmentation, performance loss, and security risks.

[0006] Therefore, existing technologies cannot achieve real-time, secure, and efficient retrieval and fusion of knowledge scattered across multiple security domains (cloud and local) by cloud-based intelligent agents while ensuring that user data does not leave the local environment. Summary of the Invention

[0007] To address the aforementioned issues, this invention proposes a privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval method and system. While ensuring that the original privacy data does not leave the local security boundary, it enables cloud-based intelligent agents to efficiently, securely, and uniformly retrieve and utilize distributed heterogeneous knowledge bases. This effectively resolves the technical contradiction in existing architectures that makes it difficult to balance data privacy protection, system scalability, and the completeness of knowledge utilization.

[0008] According to some embodiments, the present invention adopts the following technical solution: A privacy-preserving method for intelligent agents to perform cross-domain knowledge retrieval, comprising: In response to a query request initiated by a cloud-based intelligent agent, the privacy tag embedded in the workflow definition of the request is parsed; based on the privacy tag, the original query request is split into a public subquery pointing to a public knowledge base in the cloud and a private subquery pointing to a specified local private knowledge base. The following two operations are executed concurrently through an asynchronous task scheduler: the public subquery is sent to the cloud knowledge retrieval service, and the private subquery is routed to the privacy computing agent client deployed in the target local environment; The privacy computing agent client executes a private subquery in a local private knowledge base and encrypts the resulting set of knowledge fragments before returning it to the cloud. At the cloud aggregation point, the system receives a set of plaintext public knowledge fragments from the cloud knowledge retrieval service and a set of ciphertext knowledge fragments from the privacy computing agent client. It then aggregates the decrypted ciphertext knowledge fragments and the public knowledge fragments into unified knowledge data, which is used as knowledge context data and returned to the intelligent agent.

[0009] According to some embodiments, the present invention adopts the following technical solution: A privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval system includes: The tag parsing module is configured to: respond to a query request initiated by a cloud-based intelligent agent, parse the privacy tags embedded in the workflow definition of the request; and, based on the privacy tags, split the original query request into a public subquery pointing to a cloud-based public knowledge base and a private subquery pointing to a specified local private knowledge base. The task scheduling module is configured to concurrently execute the following two operations through an asynchronous task scheduler: send the public subquery to the cloud knowledge retrieval service, and route the private subquery to the privacy computing agent client deployed in the target local environment; The local query module is configured such that the privacy computing agent client executes a private subquery in the local private knowledge base and returns the queried knowledge fragment set to the cloud in encryption. The knowledge aggregation module is configured to receive a set of plaintext public knowledge fragments from a cloud knowledge retrieval service and a set of ciphertext knowledge fragments from a privacy computing agent client at a cloud aggregation point. It then aggregates the decrypted ciphertext knowledge fragments and the public knowledge fragments into unified knowledge data, which is returned to the intelligent agent as knowledge context data.

[0010] According to some embodiments, the present invention adopts the following technical solution: A computer program product includes a computer program that, when executed by a processor, implements the aforementioned privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval method.

[0011] According to some embodiments, the present invention adopts the following technical solution: A non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval method.

[0012] According to some embodiments, the present invention adopts the following technical solution: An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval method.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention stores sensitive data in a local knowledge base and employs a privacy-label-based query parsing and asynchronous parallel scheduling mechanism, enabling concurrent execution of retrieval processes in the cloud and local knowledge bases. This avoids the cumulative latency caused by serial retrieval and significantly improves the system performance and real-time response of cross-domain knowledge retrieval.

[0014] This invention enables developers to orchestrate complex intelligent agent workflows involving both public and private knowledge in a visual manner on a unified cloud platform, realizing a collaborative paradigm of "development in the cloud, data on-premises." This greatly reduces the development threshold and debugging costs in hybrid environments. At the same time, the centralized log management and monitoring capabilities in the cloud significantly simplify the operation and maintenance complexity of distributed systems, providing an agile and low-cost new model for intelligent agent development and operation and maintenance.

[0015] The hybrid retrieval architecture naturally supports horizontal scaling. Enterprise groups can centrally manage public knowledge in the cloud to empower the entire system, while each subsidiary or department can independently maintain its local private knowledge base. New local nodes can be accessed through a standardized secure proxy gateway. Without modifying the core intelligent agent logic, the knowledge map can be elastically expanded, perfectly adapting to the tree-like management structure of large organizations and enhancing the system's scalability and architectural flexibility. Attached Figure Description

[0016] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0017] Figure 1 This is a framework diagram of Example 1.

[0018] Figure 2 This is a flowchart of the intelligent agent development process in Example 1.

[0019] Figure 3 This is a flowchart of the intelligent agent debugging and operation process in Example 1. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0023] Example 1 One embodiment of the present invention provides a privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval method, comprising: S1. In response to the query request initiated by the cloud-based intelligent agent, parse the privacy tag embedded in the workflow definition of the request; based on the privacy tag, split the original query request into a public subquery pointing to the cloud-based public knowledge base and a private subquery pointing to the specified local private knowledge base; S2. Execute the following two operations concurrently via an asynchronous task scheduler: send the public subquery to the cloud knowledge retrieval service, and route the private subquery to the privacy computing agent client deployed in the target local environment; S3. The privacy computing agent client executes a private subquery in a local private knowledge base and encrypts the resulting set of knowledge fragments before returning it to the cloud. S4. At the cloud aggregation point, receive the set of plaintext public knowledge fragments from the cloud knowledge retrieval service and the set of ciphertext knowledge fragments from the privacy computing agent client. Aggregate the decrypted set of ciphertext knowledge fragments and the set of public knowledge fragments into unified knowledge data, which is then returned to the intelligent agent as knowledge context data.

[0024] As one embodiment, the present invention provides a privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval method. Under the premise of ensuring that the original privacy data does not leave the local security boundary, it enables cloud-based intelligent agents to efficiently, securely, and uniformly retrieve and utilize distributed heterogeneous knowledge bases. It effectively solves the technical contradiction that existing architectures struggle to balance data privacy protection, system scalability, and the completeness of knowledge utilization. The specific implementation process is described below.

[0025] This method uses, for example Figure 1 The framework includes two deployment locations: the cloud and local. Deployed in the cloud are the intelligent agent development platform, hybrid search engine, cloud knowledge retrieval service, result aggregation engine, and intelligent agent output module. Deployed locally is the privacy computing agent client module, whose specific functions are as follows: The intelligent agent development platform, located in the cloud, provides a visual intelligent agent orchestration interface, an online environment for designing, developing, debugging, and deploying intelligent agents, and unified cloud-based management and maintenance of intelligent agents.

[0026] The hybrid search engine, located in the cloud, receives query requests from the cloud, splits the query requests, and calls the cloud and local knowledge base systems in parallel.

[0027] Cloud-based knowledge retrieval services are located in the cloud and are used to store publicly available knowledge data and provide data query and retrieval services.

[0028] The privacy computing agent client is located on the local client and stores users' sensitive localized knowledge data, providing services such as data query and retrieval, data anonymization, and local knowledge data management.

[0029] The result aggregation engine, located in the cloud, is responsible for aggregating the knowledge fragment sets returned by the cloud knowledge base and the local knowledge base to form unified and ordered knowledge context data, which is then sent to the large model of the intelligent agent to perform reasoning.

[0030] The intelligent agent output module, located in the cloud, calls upon a large model and outputs the analysis results of that model.

[0031] Based on the above framework, a specific implementation example is provided, including an agent development phase and an agent debugging and operation phase. The agent development phase mainly involves establishing a local private knowledge base and designing the agent, while the agent debugging and operation phase mainly involves reasoning based on query requests.

[0032] I. Agent development phase, such as Figure 2 As shown, it includes: Step 1: Users create a private local knowledge base on their local client, upload private and sensitive data to the local knowledge base, and the local knowledge base will anonymize the data to ensure the security of local knowledge data. 1.1 Client Installation and Initialization: Users download and install the privacy computing agent client software package in the target local environment (such as an enterprise intranet server, a physical host of a specific department, or a secure container). The installation process includes generating a unique client identity (ClientID) and a digital certificate.

[0033] Upon startup, the client will complete two-way authentication registration with the security proxy gateway of the cloud-based intelligent agent development platform through the pre-set cloud registration address on its first run. After successful registration, the local client will be incorporated into the platform's unified management view and become a usable "private knowledge domain node".

[0034] 1.2 Knowledge Base Creation and Configuration: Users can create a new private knowledge base through the management interface provided by the local client, and name it (such as "B Sales Department Customer Data"). They can also set access control policies (such as allowed roles or IP ranges).

[0035] In the configuration, users can select or customize privacy anonymization rules, for example: Field-level anonymization: Specify that the "Customer Name" field should be generalized (e.g., retain the last name) before being returned to the cloud.

[0036] Differential privacy: Configure Laplace noise parameters for numeric fields (such as "transaction amount") to ensure that no single original data record can be inferred from the returned results.

[0037] Fragment-level filtering: Set rules to filter out text fragments containing highly sensitive keywords.

[0038] 1.3 Data Upload and Vectorization Processing: Users can upload private data files (such as Excel spreadsheets, PDF contracts, and exported files from internal databases) through the client interface to ensure that the data does not leave the local machine.

[0039] The client's built-in embedding model or a localized model service called through a secure channel parses, chunks, and converts the uploaded document into a vector.

[0040] These vectors, along with the original text fragments (or the text after initial cleaning), are stored in a local vector database (such as a locally deployed ChromaDB, Milvus Lite, or FAISS index). At this point, a private, searchable local knowledge base is established.

[0041] Step 2: Developers design and develop intelligent agents in the cloud, using visual orchestration to mark the agent's workflow, and marking the types of knowledge referenced in the workflow, namely the cloud knowledge base and the local knowledge base created in Step 1, forming privacy tags: 2.1 Login and Project Creation: Developers log in to the unified cloud-based intelligent agent development platform through a browser, create a new intelligent agent project (e.g., "Sales Data Analysis Assistant"), and enter the visual workflow orchestration canvas.

[0042] 2.2 Workflow orchestration and knowledge base node drag-and-drop: On the canvas, developers drag and drop the required logic nodes from the component library, such as "user question reception", "condition judgment", "large model call", "response generation", etc., and define the execution flow by connecting them.

[0043] When the process requires knowledge retrieval, the developer drags a "Knowledge Base Retrieval" node from the component library to the corresponding position on the canvas, selects the node, and makes key configurations in the right-hand properties panel, including: (1) Knowledge source selection: Select a knowledge source from the drop-down menu, including: Cloud-based public knowledge bases: such as "company product white paper library" and "industry regulations library".

[0044] Registered local private knowledge bases: The list dynamically displays all client knowledge bases that have completed registration (step 1), such as "B Sales Department Customer Data".

[0045] (2) Privacy label marking: Based on the selected knowledge source, the node is automatically labeled with the corresponding privacy label (such as public data or private data [Client ID: B Sales Department Customer Data]). This label will be used as the basis for subsequent routing.

[0046] (3) Query construction: Developers can configure how upstream variables (such as the user's original question) are transformed into query statements or query vectors for the knowledge base.

[0047] 2.3 Constructing a complex hybrid retrieval workflow: Developers orchestrate complex processes involving multiple knowledge base retrieval nodes, for example: The first node retrieves general product information from the cloud-based public knowledge base. The second node dynamically routes users to the local private knowledge base corresponding to the user's department based on their identity, and retrieves their internal data. The results from the two nodes are subsequently aggregated into a "result aggregation" node.

[0048] The entire workflow definition (including node logic, connection relationships, and privacy tags for each node) is saved as an executable orchestration script or configuration template.

[0049] Step 3: Complete other configuration items for the agent, such as the large model referenced, to complete the agent's development. 3.1 Large Model and Inference Parameter Configuration: In the overall configuration area of ​​the agent, the developer selects the large language model (LLM) that the agent will use. The options may include various models hosted in the cloud (such as GPT-4 and Wenxin Yiyan) or proprietary models deployed in a private environment through a security gateway. The developer also configures the inference parameters of the model, such as temperature and maximum output length.

[0050] Ensure that in the process configuration, the output of the final "Result Aggregation" node is correctly connected to the "Knowledge Context Data (Context)" input of the "Large Model Call" node.

[0051] 3.2 Testing and Debugging: Using the platform's built-in debugger, developers can input test questions and run the agent.

[0052] The debugger clearly displays the execution flow: how requests are split, which cloud and local client they are sent to, how the returned plaintext and ciphertext fragments are aggregated, and the complete context of the final input LLM. All logs are centrally viewed in the cloud, making it easy to troubleshoot cross-origin issues.

[0053] 3.3 Release and Deployment: Once development is complete, the developer clicks "Publish," and the agent's metadata (workflow definition, configuration, privacy tag mapping) is saved in the cloud. The agent is then given a unique API endpoint or dialogue interface.

[0054] Administrators can bind the access permissions of this agent to a specific local client (private knowledge base). This means that when a user invokes this agent, the part of their request involving private retrieval will only be routed to the corresponding local knowledge base client that they have the right to access.

[0055] At this point, an intelligent agent capable of "hybridly retrieving public knowledge from the cloud and specified local private knowledge" has been developed and is ready to run.

[0056] II. The debugging and operation phase of the intelligent agent, such as Figure 3 As shown, it includes: Step 1: Query Initiation and Resolution: 1.1 Request Reception and Workflow Activation: End users (such as sales managers) can initiate a query request to the published agent (such as "Sales Data Analysis Assistant") via API endpoints or dialog interfaces, such as "Analyze the sales data of product A last month for me".

[0057] The cloud-based intelligent agent platform receives the request, which is associated with a predefined workflow. It then activates the workflow instance to begin execution. The activated workflow has embedded privacy tags (including public and private data) through a visual orchestration interface.

[0058] 1.2 Privacy Tag Identification and Query Breakdown: When the workflow reaches the pre-arranged "Knowledge Base Retrieval" node, the hybrid search engine loads the configuration of the current node and parses out the embedded "Privacy Tags," including the following: A single label for public data: If the current node is labeled with only one public data label, the entire query is treated as a public subquery and directly routed to the cloud knowledge retrieval service.

[0059] A single label for private data: If the current node is labeled with only one private data label, the entire query is treated as a private subquery. The engine will bind and verify the query with the current user's identity / session token to confirm whether they have the right to access the target local knowledge base (i.e., the client identified by the Client ID). After confirmation, the subquery and authorization credentials will be routed together through the security proxy gateway to the corresponding privacy computing proxy client.

[0060] Hybrid Tags: If the workflow is designed for parallel retrieval, the hybrid retrieval engine will intelligently split the original query into multiple subqueries based on tags; for example, the original query may be split into a common subquery (such as "get product A specifications") and a private subquery (such as "get sales data of product A from sales department B last month").

[0061] Step 2, Parallel Asynchronous Scheduling: Its built-in asynchronous task scheduler performs two operations concurrently: (i) sending public subqueries directly to the cloud-based knowledge retrieval service; and (ii) routing private subqueries through a cloud-based, secure proxy gateway with two-way TLS authentication to a privacy computing proxy client deployed in the local network environment of Sales Department B. This concurrent design aims to minimize additional latency caused by cross-domain network communication.

[0062] 2.1 Cloud-based knowledge retrieval service: The cloud platform directly retrieves the cloud-based knowledge vector library, extracts open public knowledge fragments from the cloud, reorders them, and sends them to the next processing node, specifically: 1. Vectorization and Retrieval After receiving a common subquery, the cloud-based knowledge retrieval service first uses an embedding model that matches the cloud-based knowledge base to convert the query text into a query vector. This query vector is then used to perform an approximate nearest neighbor search in the cloud-based vector knowledge base to retrieve the Top-K most relevant knowledge fragments and their metadata (source documents, scores, etc.).

[0063] 2. Fragment Reordering To improve accuracy, the K initially retrieved fragments are reordered, based on the relevance score between the current query and each fragment. These processed plaintext public knowledge fragment sets are sent to the cache or message queue of the "result aggregation engine" to await aggregation.

[0064] 2.2 Local Privacy Computation and Retrieval In the local environment, after receiving a private subquery, the privacy-preserving computation agent client performs a search in its local private vector knowledge base to obtain a set of initial knowledge fragments (such as relevant sales record fragments). Next, the client performs two key privacy-preserving operations: First, it performs privacy-de-identification processing on sensitive fields in the fragments, such as direct personal identifiers (e.g., name, employee ID) (e.g., replacing them with role tags or adding Laplace noise); then, it encrypts the de-identified knowledge fragment set using a pre-negotiated session key with the cloud, generating an ciphertext knowledge fragment set; finally, it returns it to the cloud via a secure channel.

[0065] Step 3: Cloud Aggregation and Knowledge Integration At the knowledge aggregation point in the cloud, the result aggregation engine simultaneously receives a set of plaintext public knowledge fragments from the cloud-based knowledge retrieval service and a set of ciphertext knowledge fragments from the local privacy computing agent client. The engine first decrypts the received ciphertext to recover the anonymized private knowledge fragments; then, it executes a reordering algorithm; finally, it integrates all knowledge fragments (including public and private) into an ordered and unified knowledge context data.

[0066] Step 4: Agent Reasoning and Response Generation: Ultimately, the agent inputs this knowledge context data, which integrates public knowledge in the cloud and private knowledge locally, along with the user's original query, into the backend Large Language Model (LLM) for reasoning. Based on comprehensive contextual information, the LLM generates an accurate and reliable response (e.g., "Based on product A specifications and sales data from region B, last month's sales revenue was XXX, with the main growth coming from... Recommendation..."), and returns it to the end user through the agent's output module. This achieves effective retrieval and utilization of cross-domain knowledge while strictly protecting the original privacy data from leaving the local machine.

[0067] This embodiment provides a privacy-preserving method for cross-domain knowledge hybrid retrieval of intelligent agents, used in an intelligent agent development platform with cloud-developed and local storage isolated features, achieving the following objectives: Private knowledge never leaves the user's terminal, ensuring data security and preventing data leakage; Intelligent agents can dynamically combine and call data from cloud-based public knowledge bases and local private knowledge bases; Developers can develop intelligent agent applications that support both cloud and local knowledge bases through a unified and user-friendly visual interface design.

[0068] The solution proposed in this embodiment can balance data security and rapid response of intelligent agents. At the same time, the unified management, scheduling and deployment of intelligent agents in the cloud ensures low operation and maintenance costs for intelligent agents, which can effectively solve the technical problems in the current field.

[0069] Example 2 One embodiment of the present invention provides a privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval system, comprising: The tag parsing module is configured to: respond to a query request initiated by a cloud-based intelligent agent, parse the privacy tags embedded in the workflow definition of the request; and, based on the privacy tags, split the original query request into a public subquery pointing to a cloud-based public knowledge base and a private subquery pointing to a specified local private knowledge base. The task scheduling module is configured to concurrently execute the following two operations through an asynchronous task scheduler: send the public subquery to the cloud knowledge retrieval service, and route the private subquery to the privacy computing agent client deployed in the target local environment; The local query module is configured such that the privacy computing agent client executes a private subquery in the local private knowledge base and returns the queried knowledge fragment set to the cloud in encryption. The knowledge aggregation module is configured to receive a set of plaintext public knowledge fragments from a cloud knowledge retrieval service and a set of ciphertext knowledge fragments from a privacy computing agent client at a cloud aggregation point. It then aggregates the decrypted ciphertext knowledge fragments and the public knowledge fragments into unified knowledge data, which is returned to the intelligent agent as knowledge context data.

[0070] Example 3 One embodiment of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval method.

[0071] Example 4 In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided for storing computer instructions. When the computer instructions are executed by a processor, they implement the privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval method.

[0072] Example 5 One embodiment of the present invention provides an electronic device, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval method.

[0073] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0074] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0075] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A privacy-oriented agent cross-domain knowledge hybrid retrieval method, characterized in that, include: In response to a query request initiated by a cloud-based intelligent agent, the privacy tag embedded in the workflow definition of the request is parsed. Based on the privacy tag, the original query request is split into a public subquery pointing to a public knowledge base in the cloud and a private subquery pointing to a specified local private knowledge base. The following two operations are executed concurrently through an asynchronous task scheduler: the public subquery is sent to the cloud knowledge retrieval service, and the private subquery is routed to the privacy computing agent client deployed in the target local environment; The privacy computing agent client executes a private subquery in a local private knowledge base and encrypts the resulting set of knowledge fragments before returning it to the cloud. At the cloud aggregation point, the system receives a set of plaintext public knowledge fragments from the cloud knowledge retrieval service and a set of ciphertext knowledge fragments from the privacy computing agent client. It then aggregates the decrypted ciphertext knowledge fragments and the public knowledge fragments into unified knowledge data, which is used as knowledge context data and returned to the intelligent agent.

2. The privacy-oriented agent cross-domain knowledge hybrid retrieval method of claim 1, wherein, The privacy tags are knowledge data types dynamically marked by developers in the visual orchestration of the intelligent agent workflow. They include public data and private data. Public data points to a public knowledge base in the cloud, while private data points to a local private knowledge base.

3. The privacy-oriented agent cross-domain knowledge hybrid retrieval method of claim 2, wherein, It also includes setting the large model used by the agent in the visual orchestration of the agent workflow, with options including but not limited to GPT-4 and Wenxin Yiyan.

4. The privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval method as described in claim 1, characterized in that, The private subquery is routed to the target local environment through a secure proxy gateway deployed in the cloud; a secure channel is established between the secure proxy gateway and the privacy computing proxy client, and all communication is conducted through this channel.

5. The privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval method as described in claim 1, characterized in that, The specific operations of the privacy computing agent client are as follows: Execute a private subquery locally to obtain an initial set of knowledge fragments; The initial knowledge fragment set is subjected to privacy desensitization processing to generate a desensitized knowledge fragment set; Using the session key negotiated with the cloud, the de-identified knowledge fragment set is encrypted to generate the ciphertext knowledge fragment set and returned.

6. The privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval method as described in claim 1, characterized in that, It also includes the intelligent agent inputting the knowledge context data and the user's original query into a large model for reasoning to generate a final response.

7. A privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval system, characterized in that, include: The tag parsing module is configured to: parse the privacy tags embedded in the workflow definition of the request in response to a query request initiated by the cloud-based intelligent agent; Based on the privacy tag, the original query request is split into a public subquery pointing to a public knowledge base in the cloud and a private subquery pointing to a specified local private knowledge base. The task scheduling module is configured to concurrently execute the following two operations through an asynchronous task scheduler: send the public subquery to the cloud knowledge retrieval service, and route the private subquery to the privacy computing agent client deployed in the target local environment; The local query module is configured such that the privacy computing agent client executes a private subquery in the local private knowledge base and returns the queried knowledge fragment set to the cloud in encryption. The knowledge aggregation module is configured to receive a set of plaintext public knowledge fragments from a cloud knowledge retrieval service and a set of ciphertext knowledge fragments from a privacy computing agent client at a cloud aggregation point. It then aggregates the decrypted ciphertext knowledge fragments and the public knowledge fragments into unified knowledge data, which is returned to the intelligent agent as knowledge context data.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval method according to any one of claims 1-6.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement a privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval method as described in any one of claims 1-6.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform a privacy-preserving intelligent agent cross-domain knowledge hybrid retrieval method as described in any one of claims 1-6.

Images