An interaction method and system for calling a small model by a large model based on an HTTP protocol
Patent Information
- Application Number
- CN202610980305.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-02
- Publication Date
- 2026-10-09
AI Technical Summary
此类错误在高维场景下呈指数级增长，不仅可能触发参数取值范围异常，更会导致下游小模型的推理输出结果出现错误
[0038]（1）通过标准化服务部署与结构化数据交互，大幅缩短了大模型与小模型间的通信试错周期，提升了模型间数据传输速率，使调用效率显著提升800%。
Smart Images
Figure CN122884560A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of data processing, HTTP protocol, and artificial intelligence technology, and in particular, it is an interactive method and system for a large model to call a small model based on the HTTP protocol. Background Technology
[0002] Currently, large models have demonstrated outstanding performance in general scenarios such as semantic understanding and text summarization. However, when applying large models to solve specific tasks in specialized fields, such as image classification, object detection, object tracking, semantic segmentation, and instance segmentation, their results are often unsatisfactory. Although traditional specialized models (small models) have developed mature solutions in these vertical fields, how to effectively integrate the decision-making capabilities of large models with the specialized characteristics of small models has become a focus of academic attention.
[0003] The researchers proposed a "large model + small model" collaborative framework, whose technical implementation path is as follows: the large model acts as the intelligent hub, while specialized small models are encapsulated as standardized API services. The large model dynamically generates API call parameters to drive the small models to execute tasks by parsing user requirements and combining interface description documents such as OpenAPI Schema. However, in practical applications, this architecture has been found to have the following problems:
[0004] 1. Inaccurate Parameter Mapping: For high-dimensional parameter interfaces with dimensions often reaching tens or even hundreds, large models struggle to accurately map requested parameter values during a single inference process. A typical problem manifests as parameter misalignment, where the value of parameter B is incorrectly assigned to parameter A. Such errors increase exponentially in high-dimensional scenarios, potentially triggering abnormal parameter value ranges and causing errors in the inference output of downstream smaller models.
[0005] 2. Lack of Data Standardization: Parameters generated by large models often do not conform to the constraints of the application programming interface (API) of small models. For example, consider feature parameters containing enumerated values: when the API requires categorical feature labels ∈ {X,Y,Z}, the user-input "y" is easily passed directly by the large model instead of being converted to a compliant value. Such non-standardized data not only causes call anomalies but also undermines the input distribution assumptions of the small model, leading to prediction bias.
[0006] 3. Reduced Call Efficiency: The combined effect of the first two types of problems significantly reduces the success rate of API calls for small models. Systems typically require multiple rounds of trial and error (an average of 3-5 retries) to complete a valid call, and in some scenarios, even infinite loops occur. Publicly available experimental data shows that for API calls with more than 20 parameter dimensions, the first-call success rate is less than 42%, severely restricting the real-time performance of large model systems.
[0007] The aforementioned problems essentially stem from the limitations of large models in processing structured data and understanding domain knowledge constraints. Future research should focus on constructing a parameter validation layer, developing domain-adaptive constraint decoding algorithms, and designing more efficient error recovery mechanisms to achieve deep synergy between the intelligent central processing unit of large models and the execution capabilities of specialized smaller models. Summary of the Invention
[0008] The purpose of this invention is to address the problems and deficiencies of the existing technology by proposing an interactive method and system for a large model to call a small model based on the HTTP protocol. According to the inference service function of the small model and the call description information, a fixed-format body request is first encapsulated based on the POST method of the HTTP protocol, and then the small model is accessed to return the inference result. This facilitates subsequent processing by the large model and can greatly reduce the difficulty of large models calling traditional small model inference services to perform data interaction and solve specific tasks in professional fields.
[0009] The technical solution to achieve the purpose of this invention is: an interaction method for a large model to call a small model based on the HTTP protocol, the method comprising:
[0010] Multiple small models are encapsulated as service components for standardized deployment, and access interfaces are provided to listen for network traffic requests, receive data requests, perform inference, and return inference results.
[0011] A small model information database is constructed and maintained, in which the interface call description information of each small model is registered and updated in real time.
[0012] The large model acquires the task request information input by the user, understands and decomposes the task request information according to preset prompts, and plans the calling steps of the small model.
[0013] If it is determined that a small model needs to be called according to the planned calling steps, then the small model information database is queried to obtain the interface call description information of the corresponding small model, and a network request body containing user data is encapsulated according to the input information of the small model call, a data transmission request is initiated to call the corresponding small model service component for inference, and the returned inference result is received.
[0014] If it is determined that the small model does not need to be called, the large model will directly respond to the task request information and return the result.
[0015] Furthermore, the process of encapsulating multiple small models into service-oriented components for standardized deployment includes:
[0016] The small model is encapsulated into a service component using the RESTful API interface specification;
[0017] The service components are deployed on one or more servers using containerization or localization methods, and the IP address, port number, request body format, and return body format of the server where the small model is located are exposed to the outside world.
[0018] Furthermore, the interface call description information recorded in the small model information database includes: functional description information, IP address, port number, request path, request parameters, and return body structure for each small model;
[0019] Furthermore, the small model information database also records dynamic runtime data for performance optimization, including: the time when the large model calls the small model, the IP address of the large model, the number of calls, the duration, and the call status.
[0020] Furthermore, the task request information input by the user includes one or more multimodal files such as documents, images, videos, and audio.
[0021] The encapsulation of the network request body including user data includes: encapsulating the multimodal file into the HTTP request body.
[0022] Furthermore, the data interaction between the large model and the small model service components is based on HTTP POST requests, and the request and return information are both in JSON string format.
[0023] Furthermore, when initiating a data transmission request to call the corresponding small model service component, the maximum number of dimensions of the request parameter interface for each small model shall not exceed 5 dimensions, and the network request body shall only contain the URI request path of the corresponding small model, the multimodal file uploaded by the user, or the result of the inference output of the previous small model.
[0024] Furthermore, before encapsulating the network request body, the method further includes:
[0025] A parameter dictionary is pre-generated based on the application interface description information of the small model;
[0026] The retrieval algorithm is used to search the parameter dictionary to determine the most relevant alternative values to the user input and to eliminate interference items. The large model then determines the corresponding standard parameter values based on the alternative alternative values.
[0027] Furthermore, the method also includes an anomaly recovery step:
[0028] When a large model fails to call a small model, an exception recovery mechanism is triggered, automatically restarting the request and initiating a retry.
[0029] Furthermore, the network request body also includes an access key for secure access to the encrypted deployment of the small model.
[0030] On the other hand, an interactive system for a large model to call a small model based on the HTTP protocol is provided, including a processor configured to perform the following steps:
[0031] Multiple small models are encapsulated as service components for standardized deployment, and access interfaces are provided to listen for network traffic requests, receive data requests, perform inference, and return inference results.
[0032] A small model information database is constructed and maintained, in which the interface call description information of each small model is registered and updated in real time.
[0033] The large model acquires the task request information input by the user, understands and decomposes the task request information according to preset prompts, and plans the calling steps of the small model.
[0034] If it is determined that a small model needs to be called according to the planned calling steps, then the small model information database is queried to obtain the interface call description information of the corresponding small model, and a network request body containing user data is encapsulated according to the input information of the small model call, a data transmission request is initiated to call the corresponding small model service component for inference, and the returned inference result is received.
[0035] If it is determined that the small model does not need to be called, the large model will directly respond to the task request information and return the result.
[0036] On the other hand, a small model inference deployment device is provided, on which a small model inference program is deployed. When the small model inference program is executed by a processor, it implements the steps executed by the small model side or the service side in the interaction method of calling the small model based on the large model.
[0037] Compared with the prior art, the significant advantages of this invention are:
[0038] (1) By standardizing service deployment and structured data interaction, the communication trial and error cycle between large and small models is greatly shortened, the data transmission rate between models is improved, and the calling efficiency is significantly improved by 800%.
[0039] (2) By controlling the dimension of the interface request parameters to within 5 dimensions and introducing mechanisms such as parameter dictionary and retrieval algorithm, the problem of parameter misalignment and non-standard data anomaly when the large model processes high-dimensional interfaces is effectively solved, and the overall call accuracy is improved by 90%.
[0040] (3) Unify the deployment standard of small models, encapsulate them into standardized service components, provide clear access interfaces to the outside world, greatly reduce the difficulty of large models calling traditional dedicated small models to perform specific tasks in vertical fields, thereby saving system development time and cost investment.
[0041] (4) The system can automatically trigger the abnormal recovery mechanism when the data interaction fails. By re-initiating the request and executing automatic retry, the call dead loop that may be caused by multiple rounds of trial and error is avoided, which effectively improves the real-time performance and stability of the system.
[0042] (5) The small model information database can not only maintain and update the call path in real time, but also provide convenience for analyzing performance bottlenecks by recording dynamic data such as call time, number of calls and call status in detail. At the same time, the database supports carrying access keys in the request body, thereby ensuring the access security of encrypted small models and core data.
[0043] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description
[0044] Figure 1 This is a schematic diagram of the interaction method and system for a large model to call a small model based on the HTTP protocol in one embodiment.
[0045] Figure 2 This is a schematic diagram of the basic process of an interaction method and system for a large model to call a small model based on the HTTP protocol in one embodiment. Detailed Implementation
[0046] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0047] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.
[0048] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.
[0049] In one embodiment, combined Figure 1 and Figure 2 This paper provides an interactive method for a large model to call a small model based on the HTTP protocol. The method includes:
[0050] Multiple small models are encapsulated as service components for standardized deployment, and access interfaces are provided to listen for network traffic requests, receive data requests, perform inference, and return inference results.
[0051] A small model information database is constructed and maintained, in which the interface call description information of each small model is registered and updated in real time.
[0052] The large model acquires the task request information input by the user, understands and decomposes the task request information according to preset prompts, and plans the calling steps of the small model.
[0053] If it is determined that a small model needs to be called according to the planned calling steps, then the small model information database is queried to obtain the interface call description information of the corresponding small model, and a network request body containing user data is encapsulated according to the input information of the small model call, a data transmission request is initiated to call the corresponding small model service component for inference, and the returned inference result is received.
[0054] If it is determined that the small model does not need to be called, the large model will directly respond to the task request information and return the result.
[0055] Furthermore, in one embodiment, the step of encapsulating multiple small models into service-oriented components for standardized deployment includes:
[0056] The small model is encapsulated into a service component using the RESTful API interface specification;
[0057] The service components are deployed on one or more servers using containerization or localization methods, and the IP address, port number, request body format, and return body format of the server where the small model is located are exposed to the outside world.
[0058] Furthermore, in one embodiment, the interface call description information recorded in the small model information database includes: functional description information of each small model, IP address, port number, request path, request parameters and return body structure, and the data in the database can be added, deleted, modified and queried to update the small model information in real time;
[0059] Furthermore, the small model information database also records dynamic runtime data for performance optimization, including: the time when the large model calls the small model, the IP address of the large model, the number of calls, the duration, and the call status, which facilitates the analysis of performance bottlenecks and the optimization of model performance.
[0060] Furthermore, in one embodiment, the task request information input by the user includes one or more multimodal files selected from documents, images, videos, and audio.
[0061] The encapsulation of the network request body including user data includes: encapsulating the multimodal file into the HTTP request body.
[0062] Furthermore, in one embodiment, the data interaction between the large model and the small model service components is based on the HTTP protocol POST request, and the format of the request and return information is JSON string format.
[0063] Furthermore, in one embodiment, when initiating a data transmission request to call the corresponding small model service component, the number of dimensions of the request parameter interface for each small model shall not exceed 5 dimensions, and the network request body shall only contain the URI request path of the corresponding small model, the multimodal file uploaded by the user, or the result of the inference output of the previous small model, etc.
[0064] Furthermore, in one embodiment, the method further includes, before encapsulating the network request body:
[0065] A parameter dictionary is pre-generated based on the application interface description information of the small model;
[0066] The retrieval algorithm is used to search the parameter dictionary to determine the most relevant alternative values to the user input and to eliminate interference items. The large model then determines the corresponding standard parameter values based on the alternative alternative values.
[0067] Furthermore, in one embodiment, the method further includes an anomaly recovery step:
[0068] When a large model fails to call a small model, an exception recovery mechanism is triggered, automatically restarting the request and initiating a retry.
[0069] Furthermore, in one embodiment, the network request body also includes an access key for secure access to the encrypted small model, ensuring the security of data and the model.
[0070] The following is a detailed explanation:
[0071] According to the small model inference service function and call description information, this application embodiment first encapsulates a fixed-format body request based on the HTTP protocol POST method, then accesses the small model and returns the inference result, which facilitates subsequent processing of the large model and can greatly reduce the difficulty for the large model to call the traditional small model inference service to perform data interaction and solve specific tasks in professional fields.
[0072] In existing technologies, the parameter mapping inaccuracies that exist when a large model calls a small model essentially stem from the limitations of the large model in processing unstructured data and understanding domain knowledge constraints. In this application, the small model is deployed in a standardized manner, and the parameters called by the small model use structured data, with the number not exceeding five, thus reducing the difficulty for the large model to solve specific tasks in a professional domain.
[0073] Traditional calling methods suffer from numerous optional values for fields and a lack of data standardization, making it difficult for large models to determine the corresponding optional values based on user input. This application uses a retrieval algorithm to identify the most relevant alternative optional values to the user input, removing many distracting terms. This facilitates the subsequent determination of the correct correspondence by the large model, thereby determining the standard parameter values. In this embodiment, a parameter dictionary is first generated based on the API documentation, and then a retrieval algorithm is used for searching, greatly reducing the difficulty for the large model to subsequently determine the values of each request parameter.
[0074] When a user uploads an underwater sonar image or audio file to the dialog interface, along with the message "Please analyze what type of target this sonar image or audio file represents," the large model analyzes the keywords in the dialogue, determines that a target classification agent should be invoked, and encapsulates the image or audio file into the request body (e.g., {"img":"underwater_object.jpg","audio":"null","time":"2025-10-01 12:00:00"}). The target classification agent then sends the request via Ethernet using the HTTP protocol to the target classification inference service at the specified IP and port. Upon receiving the image or audio file, the target classification inference service initiates its inference process and outputs the inference result (e.g., ...).
[0075] {
[0076] "result": [
[0077] {"type":"Ohio-class","confidence":"0.95"},
[0078] {"type":"Los Angeles level","confidence":"0.65"},
[0079] {"type":"Seawolf-class","confidence":"0.35"},
[0080] {"type":"Virginia level","confidence":"0.25"},
[0081] {"type":"whale","confidence":"0.15"}
[0082] {...}
[0083] ],
[0084] "infer_time":"300ms"
[0085] }
[0086] The results are returned to the target classification agent, which then returns the results to the large model. The large model organizes and synthesizes the results and finally presents them on the dialogue interface.
[0087] In one embodiment, an interactive system for a large model to call a small model based on the HTTP protocol is provided, including a processor configured to perform the following steps:
[0088] Multiple small models are encapsulated as service components for standardized deployment, and access interfaces are provided to listen for network traffic requests, receive data requests, perform inference, and return inference results.
[0089] A small model information database is constructed and maintained, in which the interface call description information of each small model is registered and updated in real time.
[0090] The large model acquires the task request information input by the user, understands and decomposes the task request information according to preset prompts, and plans the calling steps of the small model.
[0091] If it is determined that a small model needs to be called according to the planned calling steps, then the small model information database is queried to obtain the interface call description information of the corresponding small model, and a network request body containing user data is encapsulated according to the input information of the small model call, a data transmission request is initiated to call the corresponding small model service component for inference, and the returned inference result is received.
[0092] If it is determined that the small model does not need to be called, the large model will directly respond to the task request information and return the result.
[0093] For specific limitations on the interaction system of a large model calling a small model based on the HTTP protocol, please refer to the limitations on the interaction method of a large model calling a small model based on the HTTP protocol above, which will not be repeated here.
[0094] In one embodiment, an interactive device for a large model to call a small model based on the HTTP protocol is provided, the device comprising:
[0095] The first module is used to encapsulate multiple small models into service components for standardized deployment, provide access interfaces to listen for network traffic requests, receive data requests, perform inference, and return inference results.
[0096] The second module is used to: build and maintain a small model information database, wherein the interface call description information of each small model is registered and updated in real time in the small model information database;
[0097] The third module is used to: acquire task request information input by the user in the large model, understand and decompose the task request information according to preset prompt words, and plan the calling steps of the small model;
[0098] If it is determined that a small model needs to be called according to the planned calling steps, then the small model information database is queried to obtain the interface call description information of the corresponding small model, and a network request body containing user data is encapsulated according to the input information of the small model call, a data transmission request is initiated to call the corresponding small model service component for inference, and the returned inference result is received.
[0099] If it is determined that the small model does not need to be called, the large model will directly respond to the task request information and return the result.
[0100] Specific limitations regarding the interactive device for calling a smaller model based on the HTTP protocol can be found in the limitations of the interactive method for calling a smaller model based on the HTTP protocol mentioned above, and will not be repeated here. Each module in the aforementioned interactive device for calling a smaller model based on the HTTP protocol can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each module.
[0101] In one embodiment, a small model inference deployment device is provided, on which a small model inference program is deployed. When the small model inference program is executed by a processor, it implements the steps executed by the small model side or the service side in the interaction method of calling the small model based on the large model.
[0102] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements:
[0103] Multiple small models are encapsulated as service components for standardized deployment, and access interfaces are provided to listen for network traffic requests, receive data requests, perform inference, and return inference results.
[0104] A small model information database is constructed and maintained, in which the interface call description information of each small model is registered and updated in real time.
[0105] The large model acquires the task request information input by the user, understands and decomposes the task request information according to preset prompts, and plans the calling steps of the small model.
[0106] If it is determined that a small model needs to be called according to the planned calling steps, then the small model information database is queried to obtain the interface call description information of the corresponding small model, and a network request body containing user data is encapsulated according to the input information of the small model call, a data transmission request is initiated to call the corresponding small model service component for inference, and the returned inference result is received.
[0107] If it is determined that the small model does not need to be called, the large model will directly respond to the task request information and return the result.
[0108] For specific limitations on each step, please refer to the limitations on the interaction methods for large models calling small models based on the HTTP protocol mentioned above, which will not be repeated here.
[0109] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being implemented when executed by a processor:
[0110] Multiple small models are encapsulated as service components for standardized deployment, and access interfaces are provided to listen for network traffic requests, receive data requests, perform inference, and return inference results.
[0111] A small model information database is constructed and maintained, in which the interface call description information of each small model is registered and updated in real time.
[0112] The large model acquires the task request information input by the user, understands and decomposes the task request information according to preset prompts, and plans the calling steps of the small model.
[0113] If it is determined that a small model needs to be called according to the planned calling steps, then the small model information database is queried to obtain the interface call description information of the corresponding small model, and a network request body containing user data is encapsulated according to the input information of the small model call, a data transmission request is initiated to call the corresponding small model service component for inference, and the returned inference result is received.
[0114] If it is determined that the small model does not need to be called, the large model will directly respond to the task request information and return the result.
[0115] For specific limitations on each step, please refer to the limitations on the interaction methods for large models calling small models based on the HTTP protocol mentioned above, which will not be repeated here.
[0116] In summary, the method of this invention utilizes a large model to break down the task according to the task request and plan the steps for calling the small model. If a small model needs to be called, an HTTP request body containing multimodal files is encapsulated based on the input information of the small model, and a POST request is initiated to call the small model inference service, returning the inference result. If a small model does not need to be called, the method directly answers and returns the result. This method significantly reduces the difficulty of interactively solving specific tasks in a professional domain by having a large model call traditional small model inference services.
[0117] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention without departing from its spirit and scope should be included within the protection scope of the present invention.
Claims
1. A method for interaction between a large model and a small model based on the HTTP protocol, characterized in that, The method includes: Multiple small models are encapsulated as service components for standardized deployment, and access interfaces are provided to listen for network traffic requests, receive data requests, perform inference, and return inference results. A small model information database is constructed and maintained, in which the interface call description information of each small model is registered and updated in real time. The large model acquires the task request information input by the user, understands and decomposes the task request information according to preset prompts, and plans the calling steps of the small model. If it is determined that a small model needs to be called according to the planned calling steps, then the small model information database is queried to obtain the interface call description information of the corresponding small model, and a network request body containing user data is encapsulated according to the input information of the small model call, a data transmission request is initiated to call the corresponding small model service component for inference, and the returned inference result is received. If it is determined that the small model does not need to be called, the large model will directly respond to the task request information and return the result.
2. The interaction method for a large model calling a small model based on the HTTP protocol according to claim 1, characterized in that, The process of encapsulating multiple small models into service-oriented components for standardized deployment includes: The small model is encapsulated into a service component using the RESTful API interface specification; The service components are deployed on one or more servers using containerization or localization methods, and the IP address, port number, request body format, and return body format of the server where the small model is located are exposed to the outside world.
3. The interaction method for a large model calling a small model based on the HTTP protocol according to claim 1, characterized in that, The interface call description information recorded in the small model information database includes: functional description information, IP address, port number, request path, request parameters, and return body structure of each small model; Furthermore, the small model information database also records dynamic runtime data for performance optimization, including: the time when the large model calls the small model, the IP address of the large model, the number of calls, the duration, and the call status.
4. The interaction method for a large model calling a small model based on the HTTP protocol according to claim 1, characterized in that, The task request information entered by the user includes one or more multimodal files such as documents, images, videos, and audio. The encapsulation of the network request body including user data includes: encapsulating the multimodal file into the HTTP request body.
5. The interaction method for a large model calling a small model based on the HTTP protocol according to claim 1, characterized in that, The data interaction between the large model and the small model service components is based on HTTP POST requests, and both the request and return information are in JSON string format.
6. The interaction method for a large model calling a small model based on the HTTP protocol according to claim 1, characterized in that, When initiating a data transmission request to call the corresponding small model service component, the maximum number of dimensions of the request parameter interface for each small model shall not exceed 5 dimensions, and the network request body shall only contain the URI request path of the corresponding small model, the multimodal file uploaded by the user, or the result of the inference output of the previous small model.
7. The interaction method for a large model calling a small model based on the HTTP protocol according to claim 1, characterized in that, Before encapsulating the network request body, the method further includes: A parameter dictionary is pre-generated based on the application interface description information of the small model; The retrieval algorithm is used to search the parameter dictionary to determine the most relevant alternative values to the user input and to eliminate interference items. The large model then determines the corresponding standard parameter values based on the alternative alternative values. The method also includes an anomaly recovery step: When a large model fails to call a small model, an exception recovery mechanism is triggered, automatically restarting the request and initiating a retry.
8. The interaction method for a large model calling a small model based on the HTTP protocol according to claim 1, characterized in that, The network request body also contains an access key for secure access to the encrypted deployment of the small model.
9. An interactive system for a large model to call a small model based on the HTTP protocol, characterized in that, Includes a processor, which is configured to perform the following steps: Multiple small models are encapsulated as service components for standardized deployment, and access interfaces are provided to listen for network traffic requests, receive data requests, perform inference, and return inference results. A small model information database is constructed and maintained, in which the interface call description information of each small model is registered and updated in real time. The large model acquires the task request information input by the user, understands and decomposes the task request information according to preset prompts, and plans the calling steps of the small model. If it is determined that a small model needs to be called according to the planned calling steps, then the small model information database is queried to obtain the interface call description information of the corresponding small model, and a network request body containing user data is encapsulated according to the input information of the small model call, a data transmission request is initiated to call the corresponding small model service component for inference, and the returned inference result is received. If it is determined that the small model does not need to be called, the large model will directly respond to the task request information and return the result.
10. A small-model inference deployment device, characterized in that, The device is equipped with a small model inference program, which, when executed by the processor, implements the steps performed by the small model side or the service side in the interaction method based on the large model calling the small model as described in any one of claims 1 to 8.