A method and system for collaborative interaction between an internal system based on privacy data protection and an external artificial intelligence model

Abstract

The application provides a method and system for collaborative interaction between an internal system and an external artificial intelligence model based on privacy data protection. The internal system obtains original natural language instructions and identifies private entity information, then converts the private entity information into logical reference identifiers. A semantic mapping table is constructed in the internal system to record the correspondence between logical reference identifiers and private entity information. Based on the interaction information between the current original natural language instruction and the internal system, the atomic capability set of the internal system is dynamically pruned and reorganized to obtain a dynamic capability combination. The logical reference identifier and the dynamic capability combination are sent to the external artificial intelligence model to generate a structured task execution plan. In a local virtual sandbox, the task execution plan returned by the external artificial intelligence model is simulated and executed using abstract data for formal verification. After receiving and parsing the verified task execution plan, an execution call is initiated to the backend private system layer.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more specifically, to a technology that integrates artificial intelligence, enterprise system integration and data security, particularly to a method and system for collaborative interaction between an internal system and an external artificial intelligence model based on privacy data protection. Background Technology

[0002] In recent years, general artificial intelligence (AI) technologies, represented by Large Language Models (LLM), have made groundbreaking progress. Their powerful natural language understanding, logical reasoning, and task planning capabilities have brought enormous potential for transforming enterprise software interaction methods and improving operational efficiency. Enterprises are eager to integrate these powerful AI capabilities into their internal proprietary information systems, such as Enterprise Resource Planning (ERP), Customer Relationship Management (CRM), and Transportation Management Systems (TMS), to enable complex data queries and business operations to be completed through natural language commands.

[0003] However, directly integrating external general-purpose AI models with internal enterprise systems poses a critical data security risk in practice, namely, the risk of sensitive data leakage: ① Users' natural language commands often contain a large amount of core sensitive enterprise data (such as customer names, contact information, contract amounts, etc.). If the original commands containing this data are sent directly to a third-party AI model, the data will fall outside the enterprise's control and may be recorded by the service provider, used for model retraining, or leaked. ② To fully execute user commands, it is necessary to call the internal system's API. However, if the internal system's API is directly exposed to external AI in order to introduce the "powerful functions of external AI," it is tantamount to making internal private data public, which is completely impractical.

[0004] Limitations of existing solutions: ① Simple data anonymization / masking: Replacing or masking sensitive words before sending severely compromises semantic integrity, preventing AI models from accurately understanding intent; furthermore, external AI may still access business logic, hindering genuine interaction with the backend system. ② Building a large private model: Extremely high cost and technical barriers, unaffordable for most enterprises. ③ External AI lacks understanding of internal system capabilities. Existing proxy-based privacy protection solutions, such as methods that only replace and restore keywords from user input (CN117725610A), while preventing sensitive word leakage to some extent, only focus on text-level privacy anonymization and fail to organically combine the planning capabilities of external AI with the execution capabilities of the internal system, thus failing to achieve truly intelligent business operations.

[0005] Therefore, there is an urgent need in this field for an innovative technical solution to resolve the fundamental contradiction between fully utilizing the powerful intelligence of external general AI and ensuring the absolute security of enterprise internal data, so as to achieve secure and efficient collaboration between external intelligence and internal systems. Summary of the Invention

[0006] To address the problems existing in the background technology, this invention proposes a method and system for collaborative interaction between an internal system and an external artificial intelligence model based on privacy data protection.

[0007] A method for collaborative interaction between an internal system and an external artificial intelligence model based on privacy data protection includes the following steps: S100. After obtaining the original natural language instructions and identifying the private entity information through the internal system, the private entity information is converted into a logical reference identifier without any actual meaning. S200. Construct a semantic mapping table in the internal system that records the correspondence between logical reference identifiers and private entity information; S300: Based on the interaction information between the current original natural language instructions and the internal system, dynamically trim and recombine the atomic capability set of the internal system to obtain the required dynamic capability combination. S400. Send the logical reference identifier and the dynamic capability combination to the API of an external artificial intelligence model, requesting it to generate a structured task execution plan composed of atomic capabilities and logical reference identifiers provided in the dynamic capability combination; S500: In a local virtual sandbox, abstract data is used to simulate the execution of the task execution plan returned by an external artificial intelligence model for formal verification. After receiving and parsing the verified task execution plan, S600 initiates an execution call to the backend private system layer.

[0008] Based on the above, in step S100, private entity information is identified by deterministic string matching through high-speed caching or direct querying of the internal system; or / and, a deterministic pattern matching method is used to identify private entity information by recognizing entities with fixed formats.

[0009] Based on the above, in step S200, the semantic mapping table also identifies the associated metadata for each logical reference, and the metadata includes at least the lifecycle TTL, access limit, and dependency relationship.

[0010] Based on the above, in step S300, the atomic capability set is a collection of all atomic capabilities of the internal system. An atomic capability is an API interface of the internal system that has been semantically encapsulated according to the system function definition, and includes at least the capability name, capability description, parameter definition, prerequisite dependencies, and subsequent effects. The dynamic capability combination is a subset of the atomic capability set that meets the required conditions, which is manifested as a structured specification document, including at least the unique identifier of each capability, natural language description, parameter definition, prerequisite dependencies, subsequent effects, and permission boundaries.

[0011] Based on the above, in step S400, the structured task execution plan is an arrangement of operations on a series of atomic capabilities available to the internal system.

[0012] Based on the above, in step S400, a system prompt word template is also constructed to constrain the output of the external artificial intelligence model; the logical reference identifier and the dynamic capability combination are sent to the API of the external artificial intelligence model through the system prompt word template, requesting it to generate a structured task execution plan composed of atomic capabilities and logical reference identifiers provided in the dynamic capability combination.

[0013] Based on the above, in step S500, formal verification includes at least type matching check, dependency integrity verification, permission boundary detection, and dead loop / deadlock detection.

[0014] Based on the above, in step S600, during the parsing process, after performing status verification on the encountered logical reference identifiers, the valid identifiers are restored to real business entities through the semantic mapping table, and the restored real parameters are passed to the internal API interface of the backend private system layer and executed.

[0015] Based on the above, for the execution flow of managing multi-step tasks, the execution result of the previous task step is used as the input of the next task step, and steps S100-S600 are repeated until all task steps are completed.

[0016] A privacy-preserving internal system and external artificial intelligence model collaborative interaction system includes a front-end interaction layer, a security proxy layer, and a back-end private system layer. After receiving raw natural language instructions through the front-end interaction layer, the system executes any of the above-mentioned privacy-preserving internal system and external artificial intelligence model collaborative interaction methods through the security proxy layer. After obtaining the task execution plan, the system initiates an execution call to the back-end private system layer.

[0017] This invention has outstanding substantive features and significant progress compared to the prior art, specifically: 1) Achieved absolute data security: Through the process of "semantic abstraction → dynamic capability building → task execution planning → sandbox verification → local execution," it ensures that the enterprise's core sensitive data never leaves the physical boundary of the enterprise's intranet. External AI only accesses meaningless logical reference identifiers and capability description documents (containing only capability definitions, not specific implementation logic and code), fundamentally eliminating the risk of data leakage.

[0018] 2) Achieved true intelligence and system integration: Seamlessly integrates the task planning capabilities of cutting-edge external general AI with the execution capabilities of internal enterprise systems. Users can express complex intentions using natural language, which are then logically decomposed and arranged by external AI, and finally executed by the local engine through secure API calls, providing a user experience far exceeding traditional interaction methods.

[0019] 3) Innovative security verification mechanism: Through sandbox pre-verification and logical reference identifier lifecycle management, formal verification and fine-grained access control of the execution process are realized, far exceeding the simple input / output filtering in existing technologies.

[0020] 4) Extremely low implementation cost and extremely high scalability: Enterprises do not need to build or fine-tune large models themselves. When new functions are added to the backend system, only a few lines of definition need to be added to the atomic capability set, without modifying any AI-related code, resulting in extremely low system iteration and maintenance costs.

[0021] 5) Universality and advancement of the solution: The architecture proposed in this invention is a general technical solution that does not depend on any specific external AI model or backend business system and can be widely used as a standard middleware. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the process of this invention.

[0023] Figure 2 This is a schematic diagram of the dynamic capability combination (JSON format illustration) generated by an embodiment of the present invention.

[0024] Figure 3 This is a schematic diagram illustrating an example of a system prompt word template according to an embodiment of the present invention.

[0025] Figure 4 This is a schematic diagram of a structured task execution plan according to an embodiment of the present invention.

[0026] Figure 5 This is a schematic block diagram illustrating the dynamic capability combination construction process according to an embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] like Figure 1 As shown, a method for collaborative interaction between an internal system and an external artificial intelligence model based on privacy data protection includes the following steps: S100, after obtaining the original natural language instructions and identifying private entity information through the internal system, the private entity information is converted into logical reference identifiers without actual meaning; S200, a semantic mapping table is constructed in the internal system to record the correspondence between logical reference identifiers and private entity information; S300, based on the interaction information between the current original natural language instructions and the internal system, the atomic capability set of the internal system is dynamically trimmed and recombined to obtain the required dynamic capability combination; S400, the logical reference identifiers and the dynamic capability combination are sent to the API of the external artificial intelligence model to request it to generate a structured task execution plan composed of atomic capabilities and logical reference identifiers provided in the dynamic capability combination; S500, in a local virtual sandbox, the task execution plan returned by the external artificial intelligence model is simulated using abstract data for formal verification; S600, after receiving and parsing the verified task execution plan, an execution call is initiated to the backend private system layer.

[0029] Specifically, the system in this embodiment includes at least: a front-end interaction layer, a security proxy layer, and a back-end private system layer. Its core security proxy layer includes the following key technical modules: 1) Semantic Abstraction and Logical Identifier Generation Module Working principle: Deployed within the enterprise intranet, it receives raw natural language commands through a front-end interaction layer. It scans and identifies private entity information contained within the commands in the local environment using at least one of the following methods: Method 1 (Dictionary Matching): Perform high-precision, deterministic string matching through caching or direct querying.

[0030] Method 2 (Rule Matching): Uses deterministic pattern matching methods such as regular expressions to identify entities with fixed formats.

[0031] Once a private entity is identified, a meaningless logical reference identifier (such as @TOKEN_ABC123@) is generated to replace the entity. Simultaneously, a semantic mapping table, existing only in local memory, records the correspondence between the "logical reference identifier" and the "real entity," and associates metadata (such as lifecycle TTL, access limit, dependencies, etc.) with each identifier. This process "purifies" the original instructions into generic instructions containing no sensitive data.

[0032] 2) Dynamic Capability Combination Construction Module How it works: Based on the current session context, user permissions, and session state, it dynamically trims and reassembles the set of atomic capabilities available in the backend private system, transforming them into a dynamic combination of capabilities that can be understood by the external AI model.

[0033] In this embodiment, atomic capabilities are semantically encapsulated API interfaces within the internal system, including at least a capability name (e.g., an English name that AI can understand), a capability description (e.g., telling AI: I am used to query customer orders), parameter definitions (e.g., telling AI: parameter A must be a code starting with @TOKEN_, and parameter B must be a date), prerequisite dependencies (e.g., telling AI: to use this API, you must first obtain the customer code), and post-processing effects (e.g., telling AI: after execution, an order list will be provided). All atomic capabilities of the backend private system constitute the atomic capability set. A dynamic capability composition is a structured documentation (e.g., JSON) including the capability name, capability description, parameter definitions, prerequisite dependencies, and post-processing effects for each capability. This dynamic capability composition not only describes the capabilities themselves but also expresses the data flow and constraints between capabilities. The external AI model performs path planning on this dynamic capability composition, rather than performing simple semantic matching on the text.

[0034] 3) Intent-Planning Analysis Module How it works: The "purified instructions" and "dynamic capability combination" are sent together to the API of an external general AI model, requesting it to generate a structured task execution plan consisting of atomic capabilities and logical reference identifiers provided in the capability combination.

[0035] In this embodiment, the external model acts as a "task planner" or "compiler," outputting an arrangement of a series of atomic capabilities (intermediate representation layer), rather than the final data.

[0036] 4) Sandbox pre-validation module Working principle: In a local virtual sandbox, abstract data is used to simulate the execution of the task execution plan returned by the external model for formal verification, including at least: type matching check, dependency integrity verification, permission boundary detection, and deadlock / infinite loop detection.

[0037] In this embodiment, the actual local execution process only begins after the simulated execution passes. This represents a qualitative shift from "passive filtering" to "active verification."

[0038] 5) Local execution engine Working principle: Receives and parses the verified task execution plan. During the parsing process, when encountering a logical reference identifier, it performs the following: Status verification: Check the metadata of the identifier (such as whether it has expired, whether it has been consumed, and whether the dependency relationship is satisfied); Parameter reverse parsing: Restore valid identifiers to real business entities through semantic mapping tables; Local API call: Pass the restored real parameters to the internal API interface of the backend private system layer and execute the call; State management and workflow: manage the execution flow of multi-step tasks, using the execution result of the previous step as the input of the next step, until all steps are safely completed.

[0039] The following example illustrates this: Suppose a company has deployed a Transportation Management System (TMS) with its atomic capability set defined and loaded, including capabilities such as query_orders_by_customer (query customer orders) and find_max_value_in_list (find the maximum amount).

[0040] Step 1: Receive user instructions The front-end interaction layer receives the user's input command: "Please check all the orders from 'XXX Technology' company last month and see which order has the highest amount." Step 2: Semantic Abstraction and Logical Identifier Generation The semantic abstraction module of the security proxy layer accesses the TMS customer database and identifies "XXX Technology" as a customer entity. It generates a logical reference identifier @TOKEN_CUST_001@, records {@TOKEN_CUST_001@: "XXX Technology"} in the memory-mapped table, and sets the metadata for this identifier: {ttl: 3600, max_uses: 5, depends_on: [ ]} (time-to-live: 3600 seconds, maximum usage: 5 times, dependencies: none). The cleansed instruction becomes: "Help me check all the orders from @TOKEN_CUST_001@ company last month and see which order has the highest amount." Step 3: Dynamic Capability Combination Construction like Figure 5 As shown, the core of this step lies in selecting necessary atomic capabilities that meet the constraints from the full set of atomic capabilities within the internal system to form a dynamic capability combination. The system does not statically send all atomic capabilities, but rather dynamically prunes them based on the current context, filtering out the usable capability set using existing dynamic pruning algorithms. The generated dynamic capability combination (illustrated in JSON format) is shown below. Figure 2 As shown.

[0041] Through this process, the dynamic combination of capabilities sent to the external AI model is the "minimum necessary capability set". This not only significantly reduces token consumption, but more importantly, it prevents the model from planning operation paths it does not have the authority to execute or erroneous instructions with missing dependencies from the source, thus achieving the security goal of "planning is compliance".

[0042] Step 4: Request an external model planning task The key to this step is to design specific prompt engineering templates that force the external model to output a strictly structured intermediate representation (IR), rather than natural language.

[0043] 1) Prompt build strategy: The system combines the "purified user instructions" with the "dynamic capability combination generated in step 3" and injects it into a preset system prompt, including system function constraints and conditional constraints. An example of a prompt template is shown below. Figure 3 As shown.

[0044] 2) Parsing and Verification Upon receiving the string returned by the external model, the local system performs the following operations: Cleaning: Remove any existing Markdown code block markers.

[0045] Analysis: Attempting to parse the string into a JSON object.

[0046] Schema validation: Verify whether the JSON structure contains a steps array and whether each step contains required fields such as action and parameters.

[0047] Whitelist matching: Check if the operation in each step exists in the dynamic capability combination provided in step 3.

[0048] If any of the above steps fails (such as parsing errors, missing fields, or unauthorized actions), a retry mechanism will be immediately triggered (up to N times) or the process will terminate directly with an error message to prevent unauthorized plans from entering the sandbox verification process.

[0049] 3) Example of planning results The intent-planning parsing module successfully received the structured task execution plan (IR), such as Figure 4 As shown.

[0050] Technical effect description: Strong constraints: By using explicit whitelists of operations and variable format constraints (such as @ and $ prefixes), the possibility of generating illusionary instructions in large models is eliminated.

[0051] Explicit data flow: By using the output_var (output variable of this step) and $var_XXX (output variable of the previous step) referencing mechanism, the data dependencies between steps are explicitly defined, enabling subsequent sandbox validation to accurately track the data flow.

[0052] Machine readability: Plain JSON output ensures that the local system can parse and execute it directly without natural language processing (NLP), improving the system's determinism and response speed.

[0053] Step 5: Sandbox Pre-validation This is the key step that distinguishes this invention from existing technologies. The system constructs a virtual execution environment (VEE) in memory without accessing the real database, and performs formal verification of the IR plan.

[0054] The process specifically includes the following four stages: Phase 1: Environment Initialization 1) Constructing Virtual Memory: The system creates an independent virtual memory object in memory to simulate the declaration and storage of variables; it is initialized to an empty state and does not contain any real business data.

[0055] 2) Load the Mock Executor: Inject a mock executor that exists only in memory. This executor does not contain any real database drivers or network request libraries; it only has type inference and logical calculation capabilities.

[0056] Phase Two: Context Injection Inject the necessary validation dependencies into the sandbox, but not the actual data: 1) Inject logical identifier metadata: Only inject the metadata of the logical identifier into the sandbox, including: lifecycle (TTL), maximum number of uses, and permission whitelist.

[0057] 2) Injection capability definition: The definition of the combination of capabilities to load the backend private system, which only includes the input / output type definition of the interface, and does not include the specific implementation code.

[0058] Phase Three: Step-by-Step Deduction and Verification The sandbox performs the following triple check for each step in the IR plan, following the step_id order: 1) Status and permission verification Existence check: Iterate through the parameters of the current step and check if the referenced logical identifier exists.

[0059] Lifecycle check: Compare the current system time with the TTL (Time To Live) in the metadata. If it has expired, immediately interrupt the process and display an expiration message.

[0060] Permission boundary check: Checks whether the operation type (Action) of the current step is within the allowed operation whitelist of this logical identifier. Example: If this identifier only allows "Query", but the plan attempts to "Delete", the verification will fail.

[0061] 2) Type consistency check Parameter type matching: Infer the type of the input parameter (identifier, variable reference, or literal) and compare it with the expected type defined in the current Action.

[0062] Variable dependency check: If the parameter references a variable from the previous step (such as $var_orders), check if the variable name already exists in virtual memory. If it does not exist (i.e., the dependency is undefined), throw a dependency error message.

[0063] 3) Logical reachability and deadlock detection State trajectory recording: Maintain a list of execution trajectories, recording the characteristic signatures of the executed steps (such as operation name + parameters).

[0064] Loop detection: Before executing each step, check if the current state signature already exists in the trajectory. Purpose: To prevent infinite recursion or dead loops in the plan generated by large models (e.g., step A calls step B, and step B calls back to step A).

[0065] Simulation state deduction: If all the above checks pass, a "simulation output object" that conforms to the defined type is generated and stored in virtual memory for verification in the next step.

[0066] Phase Four: Verification Result Determination 1) End-to-end pass: When all steps pass the above triple checks successfully and no anomalies are triggered.

[0067] Action: The sandbox returns VerificationResult = PASS and marks the IR plan as "verified".

[0068] Transfer: Transfer control to step 6 (local execution engine).

[0069] 2) Verification interruption: An anomaly is detected in any stage (1), 2), or 3 of stage 3).

[0070] Action: The sandbox immediately stops simulation and does not perform any real operations.

[0071] Feedback: Capture specific exception types (such as TokenExpiredException or TypeError) and return general security error messages to the user (such as "Instruction planning is invalid, please re-enter"), thereby achieving zero-risk protection for the underlying system.

[0072] Technical effect description: Through the specific sandbox pre-verification process described above, this invention achieves "verification before execution".

[0073] Defense Illusion: Even if an external AI generates an IR containing flawed logic or malicious instructions, it will be intercepted in Phase 3 and unable to reach the internal network.

[0074] Zero data residue: Since the sandbox only uses simulated objects and metadata, the entire verification process is completed in memory and leaves no trace after it ends, ensuring absolute physical isolation.

[0075] Step 6: Execute the call locally The local execution engine receives the plan and parses the first step: Verify the metadata of @TOKEN_CUST_001@: not expired, not exceeded the usage limit, and dependencies satisfied.

[0076] The query mapping table restores the actual customer name to "XXX Technology".

[0077] Invoke the internal API tms.api.query_orders(customer_name="XX Technology", time_range="last_month") to obtain the order list and store it in the variable var_orders.

[0078] Execute the second step: Extract data from var_orders, call the local function find_max_value_in_list, compare the amount field, and obtain the order with the highest amount.

[0079] The result is stored in var_max_order.

[0080] After the execution is completed, according to the metadata settings, update the usage count of the logical reference identifier (e.g., reduce it from 5 to 4), or decide whether to release it according to the TTL.

[0081] Step 7: Return the result The final result (order information with the highest amount) is returned to the front-end interaction layer and displayed to the user.

[0082] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claims involved.

Claims

1. A method for collaborative interaction between an internal system and an external artificial intelligence model based on privacy data protection, characterized in that, Including the following steps: S100. After obtaining the original natural language instructions and identifying the private entity information through the internal system, the private entity information is converted into a logical reference identifier without any actual meaning. S200. Construct a semantic mapping table in the internal system that records the correspondence between logical reference identifiers and private entity information; S300: Based on the interaction information between the current original natural language instructions and the internal system, dynamically trim and recombine the atomic capability set of the internal system to obtain the required dynamic capability combination. S400. Send the logical reference identifier and the dynamic capability combination to the API of an external artificial intelligence model, requesting it to generate a structured task execution plan composed of atomic capabilities and logical reference identifiers provided in the dynamic capability combination; S500: In a local virtual sandbox, abstract data is used to simulate the execution of the task execution plan returned by an external artificial intelligence model for formal verification. After receiving and parsing the verified task execution plan, S600 initiates an execution call to the backend private system layer.

2. The method for collaborative interaction between an internal system and an external artificial intelligence model based on privacy data protection according to claim 1, characterized in that, In step S100, private entity information is identified by deterministic string matching through high-speed caching or direct querying of the internal system; or / and, by using deterministic pattern matching to identify entities with fixed formats.

3. The method for collaborative interaction between an internal system and an external artificial intelligence model based on privacy data protection according to claim 1, characterized in that, In step S200, the semantic mapping table also identifies associated metadata for each logical reference, and the metadata includes at least the lifecycle TTL, access limit, and dependencies.

4. The method for collaborative interaction between an internal system and an external artificial intelligence model based on privacy data protection as described in claim 1, characterized in that: In step S300, the atomic capability set is a collection of all atomic capabilities of the internal system. An atomic capability is an API interface of the internal system that has been semantically encapsulated according to the system function definition, and includes at least a capability name, capability description, parameter definition, prerequisite dependencies, and subsequent effects. The dynamic capability combination is a subset of the atomic capability set that meets the required conditions, which is represented as a structured specification document, including at least a unique identifier for each capability, a natural language description, parameter definition, prerequisite dependencies, subsequent effects, and permission boundaries.

5. The method for collaborative interaction between an internal system and an external artificial intelligence model based on privacy data protection according to claim 1, characterized in that: In step S400, the structured task execution plan is an orchestration of operations on a series of atomic capabilities available to the internal system.

6. The method for collaborative interaction between an internal system and an external artificial intelligence model based on privacy data protection according to claim 1, characterized in that: In step S400, a system prompt word template is also constructed to constrain the output of the external artificial intelligence model; the logical reference identifier and the dynamic capability combination are sent to the API of the external artificial intelligence model through the system prompt word template, requesting it to generate a structured task execution plan composed of atomic capabilities and logical reference identifiers provided in the dynamic capability combination.

7. The method for collaborative interaction between an internal system and an external artificial intelligence model based on privacy data protection according to claim 1, characterized in that: In step S500, formal verification includes at least type matching check, dependency integrity verification, permission boundary detection, and dead loop / deadlock detection.

8. The method for collaborative interaction between an internal system and an external artificial intelligence model based on privacy data protection according to claim 1, characterized in that: In step S600, during the parsing process, after performing status verification on the encountered logical reference identifiers, the valid identifiers are restored to real business entities through the semantic mapping table, and the restored real parameters are passed to the internal API interface of the backend private system layer and executed.

9. The method for collaborative interaction between an internal system and an external artificial intelligence model based on privacy data protection according to claim 1, characterized in that: For managing the execution flow of multi-step tasks, the execution result of the previous task step is used as the input of the next task step, and steps S100-S600 are repeated until all task steps are completed.

10. A collaborative interaction system between an internal system and an external artificial intelligence model based on privacy data protection, characterized in that: It includes a front-end interaction layer, a security proxy layer, and a back-end private system layer. After receiving the original natural language instructions through the front-end interaction layer, the security proxy layer executes the method of collaborative interaction between the internal system and the external artificial intelligence model based on privacy data protection as described in any one of claims 1-9. After obtaining the task execution plan, it initiates an execution call to the back-end private system layer.

Images