Voice interaction method, voice interaction device and computer readable storage medium
By combining target large language model with sample thinking chain data and vehicle knowledge data, the problem of low training efficiency and high cost of large language model when updating the functions of automotive intelligent cockpit is solved, realizing efficient and low-cost API decision-making without retraining after API library update.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-31
- Publication Date
- 2026-04-14
AI Technical Summary
In existing technologies, large language models need to be retrained when updating the functions of automotive smart cockpits, resulting in low efficiency and high cost, and making it impossible to accurately decide which API to call.
Employing a target large language model and trained with sample thought chain data generated by a general large language model, it possesses deep reasoning capabilities, enabling it to understand the matching relationship between API description text and user intent. By combining vehicle knowledge data to generate target thought chain data, it achieves accurate judgment of APIs.
After the API library is updated, there is no need to retrain the target large language model, which improves efficiency, reduces costs, and ensures the accuracy of decision-making and response speed.
Smart Images
Figure CN121860077A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of vehicle technology, and in particular to a voice interaction method, a voice interaction device, and a computer-readable storage medium. Background Technology
[0002] The voice interaction process in a car's intelligent cockpit generally includes: recognizing voice requests as user command text through a voice recognition module; finding the most relevant APIs to the user command text in the application programming interface (API) library through a retrieval model; inputting the user command text and the found APIs into a dedicated large language model, which then decides which API to call to achieve the corresponding function and complete the voice interaction.
[0003] However, large language models in related technologies are generally trained by manually writing or collecting large amounts of paired data (e.g., tens of thousands of user command texts) and their corresponding API call results. Essentially, they are "memorizing" a huge and complex "user command text to API mapping table." The trained large language model only has a "memory" function, that is, associating user command texts with APIs and selecting the API associated with the user command text during the decision-making process.
[0004] However, the functions of automotive smart cockpits are frequently updated, and the API library is updated accordingly. Each time the API library is updated (added or modified), a large amount of user command text containing the new API must be manually annotated again, and training must be restarted. Otherwise, the large language model will not recognize the new API and therefore cannot accurately decide which API to call. However, each function update requires retraining the large language model, which is not only time-consuming and inefficient but also costly. Summary of the Invention
[0005] This application provides a voice interaction method, apparatus, and computer-readable storage medium.
[0006] This application provides a voice interaction method, the method comprising: Based on the user command text corresponding to the voice request forwarded by the vehicle, at least one alternative API related to the user command text is found in the pre-built API library. The target large language model is used to process the user instruction text, the at least one alternative API, and vehicle knowledge data to determine the target API from the at least one alternative API; Call the target API; The target large language model is trained using sample thought chain data generated by a general large language model. The general large language model has deep reasoning capabilities and can generate thought chain data. The target large language model can generate target thought chain data and natural language processing results carrying the associated information of the target API based on the user instruction text, the alternative APIs, and the vehicle knowledge data.
[0007] In this way, the target large language model, trained with sample thought chain data, acquires the deep reasoning capabilities of a general large language model. It possesses the ability to understand the semantic relationship between API description text and user intent (user command text), rather than simply memorizing patterns. After the API library is updated, the target large language model can reason by understanding the description text of the new API, accurately determining which API to call during the decision-making process. Therefore, when updating the functions of the automotive smart cockpit, it is not necessary to retrain the target large language model, resulting in high efficiency and low cost.
[0008] In some implementations, the step of finding at least one alternative API related to the user command text in a pre-built API library based on the user command text corresponding to the voice request forwarded by the vehicle includes: Speech recognition is used to generate the user instruction text from the voice request; The retrieval-augmented generation (RAG) technique is used to find at least one alternative API associated with the user command text.
[0009] Thus, RAG (Rapid API Search) can quickly narrow down the candidate API pool, avoiding the inefficiency caused by the target large language model processing a massive number of APIs in the API library. Deep inference using the target large language model then ensures the accuracy of API selection, guaranteeing response speed without sacrificing decision quality. RAG typically employs vector similarity retrieval. Vector similarity search matches the voice request with the API description text based on deep semantic similarity, rather than relying on superficial keyword overlap. Secondly, vector similarity search can complete retrieval from a massive API library in sub-second time, meeting the real-time requirements of in-vehicle interaction. Finally, when the API library is updated, only the description text of the new API needs to be converted into a vector and inserted into the existing index; there is no need to retrain the retrieval model, achieving "hot-swappable" functionality expansion with virtually zero maintenance costs.
[0010] In some implementations, the step of processing the user instruction text, the at least one alternative API, and vehicle knowledge data using a target large language model to determine the target API from the at least one alternative API includes: The user instruction text, the at least one alternative API, and the vehicle knowledge data are concatenated and filled into a preset prompt information template to obtain the target prompt information; Based on the target prompt information, generate the target thought chain data carrying the associated information of the target API.
[0011] In this way, by pre-setting prompt information templates, user instructions, alternative APIs and vehicle knowledge data with different structures are organized into a standardized format that the model can stably understand, eliminating input noise and significantly improving the stability and reliability of the target large language model inference.
[0012] In some implementations, the step of processing the user instruction text, the at least one alternative API, and vehicle knowledge data using a target large language model to determine the target API from the at least one alternative API further includes: Based on the target prompt information and the target thought chain data, the target API is filled with parameters to generate the natural language processing result carrying the associated information of the target API.
[0013] The target thinking chain data contains detailed parameter reasoning logic (e.g., why "a tiny bit" is mapped to relative adjustment rather than a specific value), which can directly guide the parameter filling of the target API, ensuring that the final natural language processing result is logically consistent with the previous reasoning, and avoiding the disconnect between decision-making and execution.
[0014] In some embodiments, the method includes: Based on the sample user instruction text corresponding to the sample voice request, determine at least one API related to the sample user instruction text; The general large language model is used to process the sample user instruction text, the at least one API, and the vehicle knowledge data to generate the sample thought chain data; The target large language model is fine-tuned based on the sample thought chain data, the sample user instruction text, the at least one API, and the vehicle knowledge data.
[0015] In this way, sample thought chain data is generated by the general large language model, and the target large language model is fine-tuned based on the sample thought chain data. This enables the target large model to acquire the deep reasoning ability, semantic understanding and environmental awareness capabilities of the general large language model, thereby improving the generalization ability of the target large language model and the accuracy of the output natural language processing results.
[0016] In some implementations, the step of processing the sample user instruction text, the at least one API, and the vehicle knowledge data using the general large language model to generate the sample thought chain data includes: The sample user instruction text, the at least one API, and the vehicle knowledge data are concatenated and filled into the sample preset prompt information template to obtain the sample target prompt information; The sample target prompt information is processed using the general large language model to generate the first basic thought chain data; If the length of the first basic thought chain data is less than or equal to a preset length threshold, the first basic thought chain data is determined to be the sample thought chain data.
[0017] In this way, by filtering the generated initial thought chain data through a preset length threshold, the long, verbose, or low-quality outputs containing irrelevant information that might be generated by the general large language model are effectively filtered out, ensuring that the training data is concise and has high information density. Thought chain data of appropriate length can be learned and imitated by the target large language model more efficiently, avoiding the additional computational overhead and attention distraction problems caused by excessively long sequences, accelerating the convergence speed of the target large language model, and improving the stability of the training process.
[0018] In some implementations, the sample user instruction text, the at least one API, and the vehicle knowledge data are processed using the general large language model to generate the sample thought chain data, and the process further includes: If the length of the first basic thought chain data is greater than the preset length threshold, the general large language model is used to perform secondary reasoning based on the sample target prompt information. If a second basic thought chain data with a length less than or equal to the preset length threshold is obtained within a preset number of inferences, the second basic thought chain data is determined as the sample thought chain data.
[0019] Thus, when the length of the first basic thought chain data exceeds a preset length threshold, a second reasoning process is used to obtain new first basic thought chain data. By combining the preset length threshold and the preset number of iterations, shorter second basic thought chain data is selected as sample thought chain data to improve the data quality of the sample thought chain data, thereby improving the computational efficiency and accuracy of the subsequent target large language model.
[0020] In some embodiments, the method further includes: If the second basic thought chain data is not obtained within the preset number of times, the configured third basic thought chain data is received. The third basic thinking chain data is identified as the sample thinking chain data.
[0021] This can improve the data quality of the sample thought chain data, thereby improving the computational efficiency and accuracy of the subsequent target large language model.
[0022] This application provides a voice interaction device. The device includes: The retrieval module is used to find multiple alternative APIs related to the user command text in a pre-built API library based on the user command text corresponding to the voice request forwarded by the vehicle. The processing module is used to train the sample thought chain data generated by the general large language model. The general large language model has deep reasoning ability and can generate thought chain data. The target large language model can generate target thought chain data and natural language processing results carrying the association information of the target API based on the user instruction text, the alternative API and the vehicle knowledge data. A calling module, which is used to call the target API.
[0023] This application provides a computer-readable storage medium storing a computer program that, when executed by one or more processors, implements the steps of the above-described method.
[0024] In the voice interaction device and computer-readable storage medium provided in this application, the target large language model, trained with sample thought chain data, acquires the deep reasoning ability of a general large language model. This means it possesses the ability to understand the semantic matching relationship between API description text and user intent (user command text), rather than simply memorizing patterns. After the API library is updated, the target large language model can reason by understanding the description text of the new API, accurately determining which API to call during the decision-making process. Thus, retraining the target large language model is unnecessary when updating the functions of the automotive intelligent cockpit, resulting in high efficiency and low cost.
[0025] Additional aspects and advantages of embodiments of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of embodiments of this application. Attached Figure Description
[0026] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, wherein: Figure 1 This is one of the flowcharts illustrating a voice interaction method according to certain embodiments of this application; Figure 2 This is a second flowchart illustrating a voice interaction method according to certain embodiments of this application; Figure 3 This is the third flowchart illustrating a voice interaction method according to certain embodiments of this application; Figure 4This is the fourth flowchart illustrating a voice interaction method according to certain embodiments of this application; Figure 5 This is the fifth flowchart illustrating a voice interaction method according to certain embodiments of this application; Figure 6 This is a flowchart of a voice interaction method according to certain embodiments of this application, number six. Figure 7 This is a schematic diagram of the process for obtaining sample thought chain data in some embodiments of this application; Figure 8 This is the seventh flowchart illustrating a voice interaction method according to certain embodiments of this application; Figure 9 This is a schematic diagram of the voice interaction process in some embodiments of this application. Detailed Implementation
[0027] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the embodiments of this application, and should not be construed as limiting the embodiments of this application.
[0028] The voice interaction process in a car's intelligent cockpit generally includes: recognizing voice requests as user command text through a voice recognition module; finding the most relevant APIs to the user command text in the API (application programming interface, API) library through a retrieval model; inputting the user command text and the found APIs into a dedicated large language model, which then decides which API to call to achieve the corresponding function and complete the voice interaction.
[0029] However, large language models in related technologies are generally trained by manually writing or collecting large amounts of paired data (e.g., tens of thousands of user command texts) and their corresponding API call results. Essentially, they are "memorizing" a huge and complex "user command text to API mapping table." The trained large language model only has a "memory" function, that is, associating user command texts with APIs and selecting the API associated with the user command text during the decision-making process.
[0030] However, the functions of automotive smart cockpits are frequently updated, and the API library is updated accordingly. For example, the "comprehensive seat adjustment" function (i.e., API) may be broken down into more specific APIs such as "seat cushion angle adjustment" and "backrest tilt adjustment," and even the parameter enumeration values of existing APIs may be adjusted. Each time the API library is updated (added or modified), a large amount of user command text containing the new API must be manually annotated again, and training must be restarted; otherwise, the large speech model will not recognize the new API and therefore cannot accurately decide which API to call. However, each function update requires retraining the large language model, which is not only time-consuming and inefficient but also costly. Therefore, each function update requires retraining both the retrieval model and the large language model, which is costly and slow.
[0031] Based on the above issues, please refer to Figure 1 This application provides a voice interaction method, the method including: 01: Based on the user command text corresponding to the voice request forwarded by the vehicle, find at least one alternative API related to the user command text in the pre-built API library; 02: The target large language model is used to process the user instruction text, the at least one alternative API, and vehicle knowledge data to determine the target API from the at least one alternative API; 03: Call the target API; The target large language model is trained using sample thought chain data generated by a general large language model. The general large language model has deep reasoning capabilities and can generate thought chain data. The target large language model can generate target thought chain data and natural language processing results carrying the associated information of the target API based on the user instruction text, the alternative APIs, and the vehicle knowledge data.
[0032] This application provides a voice interaction device. The voice interaction method of this application can be implemented by the voice interaction device of this application. The voice interaction device includes a retrieval module, a processing module, and a calling module. The retrieval module is used to find multiple candidate APIs related to the user command text in a pre-built API library based on the user command text corresponding to the voice request forwarded by the vehicle. The processing module is used to train a target large language model using sample thought chain data generated by a general large language model. The general large language model has deep reasoning capabilities and can generate thought chain data. The target large language model can generate target thought chain data carrying the association information of the target API and natural language processing results based on the user command text, the candidate APIs, and the vehicle knowledge data. The calling module is used to call the target API.
[0033] This application also provides a server, which includes a memory and a processor. The voice interaction method of this application can be implemented by the server of this application. The memory stores a computer program, and the processor executes steps 01-03 when executing the computer program.
[0034] This application also provides a computer-readable storage medium storing a computer program that, when executed by one or more processors, performs steps 01-03.
[0035] In the voice interaction method, voice interaction device, server, and computer-readable storage medium provided in the proposed implementation method, the target large language model is trained with sample thought chain data, acquiring the deep reasoning ability of a general large language model. This allows it to understand the semantic relationship between API description text and user intent (user command text), rather than simply memorizing patterns. After API library updates, the target large language model can reason by understanding the description text of the new API, accurately determining which API to call during the decision-making process. Thus, retraining the target large language model is unnecessary when updating the functions of the automotive smart cockpit, resulting in high efficiency and low cost.
[0036] Specifically, a voice request refers to the raw user voice audio signal captured by the microphone array within the vehicle's cockpit. It is the initial input to the entire interaction process. "Vehicle-forwarded voice request" emphasizes that this audio data is received by the cockpit domain controller or related modules and transmitted to this voice interaction system for processing. It has not yet undergone any semantic parsing; its content needs to be converted into understandable text by a subsequent automatic speech recognition module.
[0037] User command text refers to the natural language string that represents the user's intent in text form, converted from a "voice request" through automatic speech recognition technology. It is the core input object for semantic understanding, retrieval, and reasoning. For example, the voice request "open the car window" is converted into the text of the four characters "open the car window" after automatic speech recognition. It is a standard digital expression of the user's voice content.
[0038] An API refers to a specific vehicle function control unit that can be called by software. Each API corresponds to a vehicle function (such as opening a window or setting the air conditioning temperature) and includes a clear call name, parameter list, and specification.
[0039] The API library refers to a structured database that stores metadata for all available voice control APIs in a vehicle. Its core entries include not only the API name, but more importantly, the natural language description text for each API's function. The target large language model does not memorize API names, but rather understands their function by reading and comprehending the description text in the API library. For example, the description of the Seat_Ventilation_On API might be "Activate the ventilation function of the specified seat".
[0040] A general-purpose large language model refers to a class of foundational models with a large number of parameters, pre-trained on massive amounts of data, and optimized through complex alignment (such as instruction fine-tuning and reinforcement learning from human feedback). Its key characteristics lie in its powerful deep logical reasoning ability and ability to follow complex instructions. In the embodiments of this application, a general-purpose large language model specifically refers to a model capable of generating high-quality, detailed, and non-normalized thought chains (such as DeepSeek-R1). As a "teacher model," its core value is not direct deployment, but rather leveraging its superior reasoning capabilities to generate "teaching materials" (sample thought chain data) for training lightweight, vehicle-specific models.
[0041] Sample thought chain data refers to the detailed reasoning process text generated by the aforementioned "General Large Language Model" based on training samples (sample user command text, relevant API descriptions, and vehicle knowledge). "Sample" signifies that it is data used in the training phase. "Thought chain" emphasizes that its content is a step-by-step logical deduction. Its quality requirements are "detailed and non-standardized format": the content must fully analyze intent, compare options, and cite knowledge; in terms of form, it should avoid fixed templates to help the learning model grasp the logical core rather than superficial sentence structures. This is a key carrier in the knowledge distillation process.
[0042] The Target Large Language Model (DML) refers to a language model specifically designed for in-vehicle scenarios, trained using "sample thought chain data" through knowledge distillation techniques. It has a relatively small parameter set. The DML is deployed as an inference engine on the vehicle. It inherits the deep reasoning methodology of the teacher model but is smaller and faster. Its inputs are real-time user commands, retrieved API descriptions, and vehicle knowledge; its outputs are the target thought chain data and the final natural language processing results.
[0043] Vehicle knowledge data refers to structured or semi-structured data used to provide domain-specific constraints and contextual information during reasoning. Vehicle knowledge data serves as a bridge connecting general semantics with specific vehicle specifications. It includes: Static knowledge: vehicle configuration (such as whether there is a rear entertainment screen) and logical rules between functions (such as "disable video playback while driving").
[0044] Dynamic knowledge: Real-time vehicle status (vehicle speed, battery level), environmental conditions (outside temperature), and conversation history.
[0045] The target large language model will extract relevant knowledge and inject it into the reasoning process in text form to ensure that the decisions comply with vehicle safety and design specifications.
[0046] The target thought chain data refers to the text of the reasoning process for the current specific scenario generated by the "target large language model" deployed on the vehicle side when processing real-time requests. It is the intermediate product and interpretable output of the online reasoning of the target large language model. It will clearly explain why a certain one (i.e., the "target API") is selected from the alternative APIs and how the parameters are determined. Its content "carries the associated information of the target API". For example, it may write: "The user said 'It's too stuffy'. Combining the knowledge [the window is currently closed], it is preferred to choose 'open the window' rather than 'air conditioner external circulation' because...".
[0047] The natural language processing result refers to the highly structured machine-executable instructions finally output by the target large language model, usually in the format of <API name, {parameters: value}>. This is the final deliverable of the entire speech understanding pipeline. It converts the vague user natural language into precise and unambiguous software call instructions. For example, converting "Open the driver's window halfway" into {"API": "Window_Control", "Arguments": {"position": "driver", "action": "open", "percent": 50}}.
[0048] It should be noted that the target large language model is trained based on the sample thought chain data generated by the general large language model. Among them, the general large language model refers to a large language model with logical reasoning and environmental adaptation capabilities.
[0049] Using the sample thought chain data output by the general large language model as training data to train the target large language model can enable the target large language model to inherit the logical reasoning and environmental adaptation capabilities of the general large language model, so that the target large language model can understand the semantics of voice requests and APIs based on the user instruction text, alternative APIs, and vehicle knowledge data, also generate the corresponding target thought chain data, and accurately output the natural language processing result according to the target thought chain data.
[0050] It can be understood that during the process of vehicle cockpit function update or refined iteration, new sub-APIs that are split from the APIs in the original API library and only responsible for partial functions can be established, and the mapping relationship between the two can be established. For example, there is an API in the pre-built API library for turning on all the sounds inside the vehicle cockpit. Through function refinement iteration, an API for turning on the media sound inside the vehicle cockpit and an API for turning on the call sound inside the vehicle cockpit can be generated.
[0051] The target large language model, trained on sample thought chain data, acquires the deep reasoning capabilities of a general large language model. It possesses the ability to understand the semantic relationship between API description text and user intent (user command text), rather than simply memorizing patterns. After API library updates, the target large language model can reason by understanding the description text of the new API, accurately determining which API to call during the decision-making process. Thus, when updating the functions of an automotive smart cockpit, there is no need to retrain the target large language model, resulting in high efficiency and low cost.
[0052] In one example, after receiving the user command text forwarded by the vehicle, the two APIs most relevant to the user command text function can be selected from the pre-built vehicle API library and identified as candidate APIs to initially associate the voice request with the API. Then, the user instruction text, alternative APIs, and vehicle knowledge data are combined into a target Prompt according to a preset template to ensure that the model input format is consistent. Next, input the target large language model, analyze the core requirements of the request, verify the feasibility by combining vehicle knowledge, and judge the API compatibility one by one to clarify the parameter filling logic and output the target thinking chain data. Then, the target large language model can output structured NLU results based on the target thought chain data, ensuring that the API and parameters fully match the user's needs; Finally, the NLU results are sent to the corresponding in-vehicle functional modules so that the functional modules can call the API in the NLU results and pass in the parameters to perform specific operations; at the same time, the execution results are fed back to the user through the speech synthesis module to complete the voice interaction.
[0053] In summary, in this embodiment, a list of similar application interfaces can be determined based on the received voice request. Then, based on the pre-trained target large language model, combined with the voice request, the list of similar application interfaces, and vehicle knowledge data, the natural language processing result is determined, improving the accuracy of the natural language processing result. Voice interaction is then performed based on the natural language processing result. Compared to frequently iterated large language models, the pre-trained target large language model in this embodiment possesses the ability to understand application interfaces. When new application interfaces are generated due to updates to vehicle cockpit functions, the inference logic can be automatically adjusted to identify the newly added APIs without updating the target large language model. This reduces the training and maintenance costs of the target large language model and improves the user experience to some extent.
[0054] Please see Figure 2In some implementations, step 01 (finding multiple alternative APIs related to the user command text in a pre-built API library based on the user command text corresponding to the voice request forwarded by the vehicle) includes: 011: Speech recognition to generate user command text from voice requests; 012: RAG (retrieval-augmented generation) technology finds at least one alternative API associated with the user command text.
[0055] Thus, RAG (Rapid API Search) can quickly narrow down the candidate API pool, avoiding the inefficiency caused by the target large language model processing a massive number of APIs in the API library. Deep inference using the target large language model then ensures the accuracy of API selection, guaranteeing response speed without sacrificing decision quality. RAG typically employs vector similarity retrieval. Vector similarity search matches the voice request with the API description text based on deep semantic similarity, rather than relying on superficial keyword overlap. Secondly, vector similarity search can complete retrieval from a massive API library in sub-second time, meeting the real-time requirements of in-vehicle interaction. Finally, when the API library is updated, only the description text of the new API needs to be converted into a vector and inserted into the existing index; there is no need to retrain the retrieval model, achieving "hot-swappable" functionality expansion with virtually zero maintenance costs.
[0056] RAG (Reference-Action Generic) technology is a paradigm that combines information retrieval with text generation. Its core principle is "first retrieve relevant knowledge, then generate answers based on that knowledge." In this application's implementation, RAG specifically refers to its retrieval component. The system utilizes semantic vector retrieval to quickly (in sub-seconds) identify the few API descriptions most relevant to the user's commands from a massive API library, serving as "candidate knowledge" for subsequent deep reasoning, thereby greatly improving efficiency and ensuring real-time information delivery.
[0057] Vector similarity retrieval is a retrieval method based on semantic vectors for information matching. It converts text into vectors (a set of numbers) in a high-dimensional space and measures the semantic similarity between texts by calculating the geometric distance between vectors. Vector similarity retrieval is the specific technical means to implement RAG retrieval. It enables the system to understand the deep semantic connection between "I'm a little hot" and "seat ventilation," rather than simply matching identical words. Furthermore, its index updates (inserting vectors from new APIs) have extremely low cost and support "hot-plugging."
[0058] Specifically, the API library is a database of pre-built vehicle cockpit function APIs. The API library stores multiple APIs and the name, function description and parameter definition of each API according to unified rules.
[0059] In one example, user command text can be converted into vector format to eliminate ambiguity in natural language; API description vectors of all APIs can be extracted by calling a pre-built API library; semantic similarity between user command text vectors and all API description vectors can be calculated based on the RAG model; and multiple APIs with high similarity can be selected as candidate APIs.
[0060] For example, if the user command text is "turn off the car's sound", the semantic similarity between the user command text vector and all API description vectors can be calculated, and finally two APIs with high similarity in the library can be selected as candidate APIs: "Media_Voice_Open" (function description includes "turn on the sound or unmute") and "Headrest_Voice_Model_Close" (function description includes "turn off the headrest speakers").
[0061] It should be noted that in the embodiments of this application, the two alternative APIs are only for illustrative purposes and should not be construed as a limitation on the corresponding number. In other examples, the number of alternative APIs may also be 1, 4, etc., which is not limited here and can be set according to the actual situation.
[0062] Please see Figure 3 In some implementations, step 02 (processing user command text, at least one alternative API, and vehicle knowledge data using a target large language model to determine the target API from at least one alternative API) includes: 021: Concatenate the user command text, at least one alternative API, and vehicle knowledge data into the preset prompt information template to obtain the target prompt information; 022: Generate target mind chain data based on target prompts.
[0063] In this way, by pre-setting prompt information templates, user instructions, alternative APIs and vehicle knowledge data with different structures are organized into a standardized format that the model can stably understand, eliminating input noise and significantly improving the stability and reliability of the target large language model inference.
[0064] Specifically, the preset prompt template is a pre-defined text framework that can be used to standardize the organization logic of input information.
[0065] By calling the pre-stored preset prompt information template, user command text, alternative APIs and vehicle knowledge data can be filled in according to the preset prompt information template to form the input data of the target language big model with a unified structure, namely the target prompt information.
[0066] In one example, the target prompt information includes the functional definition of the target language big model, the output format, vehicle knowledge data, a list of available APIs (API library), and user instruction text.
[0067] Specifically, after the target prompt information is input into the target large language model, the target large language model can perform reasoning based on the target prompt information to generate target thought chain data containing complete reasoning logic.
[0068] During the reasoning process, the target large language model can combine vehicle knowledge data analysis and understand the semantics of user command text to analyze candidate APIs among the alternative APIs. By analyzing the degree of functional matching between user command text and candidate APIs, irrelevant APIs can be eliminated, generating target thought chain data with complete reasoning logic; Based on the target thought chain data, the target API that best matches the user command text can be determined from at least one alternative API.
[0069] It should be noted that if none of the APIs in the candidate APIs are compatible, the target large language model can directly output a preset prompt message, such as "None", and terminate the subsequent steps.
[0070] In some implementations, step 02 (processing user command text, at least one alternative API, and vehicle knowledge data using a target large language model to determine the target API from at least one alternative API) includes: 023: Based on the target prompt information and target thought chain data, perform parameter filling processing on the target API to output natural language processing results.
[0071] After obtaining the uniquely compatible target API, the target API can be parameter-filled based on the target thought chain data. This involves filling in the required parameters and generating a natural language processing result in a preset format, providing accurate data support for subsequent API calls. The target thought chain data contains detailed parameter reasoning logic (e.g., why "a tiny bit" is mapped to relative adjustment rather than a specific value), directly guiding the parameter filling of the target API. This ensures that the final natural language processing result is logically consistent with the initial reasoning, avoiding a disconnect between decision-making and execution.
[0072] In one example, after receiving the target prompt, the target large language model can first locate its own function according to the role definition in the template. For example, the target large language model can be located as an "in-vehicle language assistant," which needs to perform voice interaction based on the user's command text and the list of available APIs (API library); Then, reasoning is performed based on the target prompt information. First, it can be determined that "inside the car = all" based on vehicle knowledge data. Then, among the candidate APIs, the matching degree between each API and the user command text is analyzed to generate target thinking chain data. Then, based on the target mind chain data, from the self-selected APIs, the mismatched APIs are excluded, and the uniquely compatible target API is selected. For example, from the descriptions of [Headrest_Voice_Model_Close] and [Media_Voice_Open], the sound-related API [Media_Voice_Open] is selected. Finally, based on the retrieved API description and its parameter definitions, the corresponding slots are extracted from the user instructions, such as {"target_function":"sound","seat_position":"all"}, to generate natural language processing results.
[0073] In addition, to reduce the latency caused by generating thought chains for output, the output format in the target prompt information can be changed, and the output order of the target large language model can be adjusted. Instead of having the target large language model output the thought chain first and then the NLU, it can be changed to directly output the NLU and stop outputting the thought chain, so as to reduce the overall latency of the target large language model.
[0074] Thus, based on the target prompt information, target thought chain data is determined; based on the target thought chain data, the target API is selected from at least one candidate API; based on the target prompt information and the target thought chain data, parameter filling processing is performed on the target API to determine the natural language processing result. In this way, the target language model can first generate target thought chain data based on the target prompt information, and then, based on the target thought chain data, understand the semantics of the user command text and accurately match the corresponding target API to obtain accurate natural language processing results, providing precise data support for subsequent voice interaction execution.
[0075] Please see Figure 4 In some implementations, the voice interaction method further includes: 04: Based on the sample user instruction text corresponding to the sample voice request, determine at least one API related to the sample user instruction text; 05: Use a general large language model to process sample user command text, at least one API, and vehicle knowledge data to generate sample thought chain data; 06: Fine-tune the target large language model based on sample thought chain data, sample user instruction text, at least one API and vehicle knowledge data.
[0076] In this way, sample thought chain data is generated by the general large language model, and the target large language model is fine-tuned based on the sample thought chain data. This enables the target large model to acquire the deep reasoning ability, semantic understanding and environmental awareness capabilities of the general large language model, thereby improving the generalization ability of the target large language model and the accuracy of the output natural language processing results.
[0077] Fine-tuning refers to the process of further training the model parameters using a dataset specific to a particular domain or task, based on a pre-trained model, to adapt it to a new task. In the embodiments of this application, it specifically refers to supervised fine-tuning of the "target large language model" using "sample thought chain data". The training objective is to enable the target model to learn to generate reasoning processes and make correct decisions similar to the teacher model, thereby achieving the transfer of reasoning ability.
[0078] Specifically, sample voice requests can be acquired in advance, covering various functions and scenarios such as seat adjustment, audio control, and media playback in the vehicle cabin, to simulate user commands in the cabin scenario and provide a data foundation for the target large language model to learn the reasoning logic of different scenarios.
[0079] The same sample voice request can be converted into sample user instruction text through speech recognition, and then at least one API similar to the sample voice request can be retrieved from a pre-built API library through RAG retrieval.
[0080] The sample API can include multiple APIs and a functional description and parameter definition for each API, which are used to provide a candidate API pool for the sample thought chain data of the target large language model, so that the sample thought chain data can contain API and voice request adaptation analysis logic.
[0081] By inputting sample voice requests, sample APIs, and vehicle knowledge data into a general-purpose large language model, the model can parse the sample voice requests, determine the compatibility between the sample APIs and the sample voice requests, and determine the basis for API filling. This process yields sample thought chain data containing textual descriptions of voice request analysis, API compatibility judgments, and parameter filling basis, providing input data for training the target large language model.
[0082] The training of the target large language model is based on the pre-trained large language model. The pre-trained large language model has basic language understanding capabilities after pre-training. Based on the sample thought chain data of the general large language model, as well as sample voice requests, sample APIs and vehicle knowledge data, the pre-trained large language model can be fine-tuned to obtain the target large language model with API semantic understanding and environmental awareness capabilities.
[0083] Understandably, the sample thought chain data generated by the general large language model is quite detailed, which can accurately perceive the usage scenario of each API and provide corresponding reasons for the matching degree of each API, as well as for selection and exclusion.
[0084] Based on the sample thought chain data of the general large language model, after inputting sample voice requests, sample APIs, and vehicle knowledge data, the pre-trained large language model can learn to understand each API in the sample voice request from a semantic level and interpret the matching degree of each API to generate a COT similar to the general large language model. Then, when the description, name, and other parameters of each API in the sample API change, the model can select one API and exclude other APIs based on the COT interpretation model.
[0085] The target large language model obtained by training on sample thought chain data can inherit the ability of the general large language model to generate thought chain data, so as to understand the role of each API at a deeper level.
[0086] Compared to memorizing all APIs, since the training process of the target large language model and the actual voice interaction process are related to the preset API library, when new functions are added to the vehicle cockpit, the new APIs will be included in the preset API library. The thought chain samples of the new APIs can be generated through the existing process and fine-tuned without reconstructing the training framework. This ensures that the target large language model can adapt to cockpit function updates in a timely manner and meet users' voice interaction needs for new functions.
[0087] Thus, based on the sample voice request, the sample API related to the sample voice request is determined; based on the general large language model, the sample voice request, the sample API, and vehicle knowledge data, sample thought chain data is determined; based on the sample thought chain data, the sample voice request, the sample API, and the vehicle knowledge data, the pre-trained large language model is fine-tuned to obtain the target large language model. In other words, sample thought chain data can be automatically generated through the general large language model, and the pre-trained large language model can be fine-tuned based on the sample thought chain data to obtain a target large language model with API semantic understanding and environment awareness capabilities, thereby improving the generalization ability of the target large language model and the accuracy of the output natural language processing results.
[0088] Please see Figure 5 In some implementations, step 05 (processing sample user instruction text, at least one API, and vehicle knowledge data using a general large language model to generate sample thought chain data) includes: 051: Concatenate the sample user instruction text, at least one API, and vehicle knowledge data into the sample preset prompt information template to obtain the sample target prompt information; 052: A general large language model is used to process the sample target prompt information to generate the first basic thought chain data; 053: If the length of the first basic thought chain data is less than or equal to the preset length threshold, the first basic thought chain data is determined to be the sample thought chain data.
[0089] In this way, by filtering the generated initial thought chain data through a preset length threshold, the long, verbose, or low-quality outputs containing irrelevant information that might be generated by the general large language model are effectively filtered out, ensuring that the training data is concise and has high information density. Thought chain data of appropriate length can be learned and imitated by the target large language model more efficiently, avoiding the additional computational overhead and attention distraction problems caused by excessively long sequences, accelerating the convergence speed of the target large language model, and improving the stability of the training process.
[0090] Specifically, similar to the preset prompt information template, the sample voice request, sample API and vehicle knowledge data can be spliced together based on the sample preset prompt information template to form the input data of a general large language model with a unified structure, that is, the sample target prompt information.
[0091] In one example, after inputting the target prompt information into the general large language model, the model first parses the sample voice request based on the prompt information. For example, if the user command is "lower the right seat cushion in the third row a little bit", the key information is "right side of the third row" and "seat cushion". Next, the compatibility of the candidate API pool (sample API) with the sample voice requests is compared. For example, Seat_Value_Set is used to set specific values, but the user command "a tiny bit" does not require numbers, so it does not match and needs to be excluded; Seat_Gear_Set is used for relative adjustment and is compatible; Seat_Max_Min_Set is used for maximum and minimum value adjustment and is not compatible. Finally, the parameter filling logic is explained in detail. For example, device is "cushion", direction is "down" (because "lower" means downward), etc., generating basic thought chain data containing complete logic.
[0092] The general large language model understands sample voice requests and then interprets each API from a semantic level based on the candidate API pool and vehicle knowledge data. This explains why the model considers the API to be a match or not, the reasons for filling the slots of the final selected API, and generates the initial thought chain text, i.e. the first basic thought chain data.
[0093] Understandably, the first basic thought chain data may be too long, such as repetitive API analysis logic and irrelevant instruction interpretation, which will increase the storage of training data and the computational burden on the subsequent target large language model. It may even cause the target large language model to get stuck in learning irrelevant details, unable to accurately capture the core rules of API selection and parameter matching, resulting in a decline in training effect.
[0094] Based on this, sample thought chain data can be filtered by setting a preset length threshold. If the length of the first basic thought chain data is less than or equal to the preset length threshold, it can be considered that the length of the first basic thought chain data is appropriate and will not affect the computational efficiency of the target large language model. Therefore, the first basic thought chain data can be determined as sample thought chain data. By filtering shorter sample thought chain data, the data quality of the sample thought chain data can be improved, thereby improving the computational efficiency and accuracy of the subsequent target large language model.
[0095] Thus, based on the preset prompt information template, the sample voice request, sample API, and vehicle knowledge data are concatenated to determine the sample target prompt information. Based on the general large language model and the sample target prompt information, the first basic thought chain data is determined. If the length of the first basic thought chain data is less than or equal to a preset length threshold, it is determined as the sample thought chain data. In this way, the sample target prompt information, which concatenates the sample voice request, sample API, and vehicle knowledge data, is input into the general large language model to obtain first basic thought chain data containing complete logic. Combined with the preset length threshold, shorter first basic thought chain data is selected as sample thought chain data to improve the data quality of the sample thought chain data, thereby improving the computational efficiency and accuracy of the subsequent target large language model.
[0096] Please see Figure 6 In some implementations, the method further includes: 054: When the length of the first basic thinking chain data exceeds the preset length threshold, a general large language model is used to perform secondary reasoning based on the sample target prompt information; 055: If a second basic thought chain data with a length less than or equal to a preset length threshold is obtained within a preset number of inferences, the second basic thought chain data is determined as sample thought chain data.
[0097] In some implementations, the control module is further configured to perform secondary reasoning on the sample thought chain data based on the general large language model and sample target prompt information if the length of the first basic thought chain data is greater than a preset length threshold. The control module is also configured to determine the second basic thought chain data as the sample thought chain data if a second basic thought chain data with a length less than or equal to the preset length threshold is obtained within a preset number of reasoning iterations.
[0098] In some implementations, the processor is further configured to perform secondary inference on the sample thought chain data based on the general large language model and sample target prompt information, if the length of the first basic thought chain data is greater than a preset length threshold. The processor is also configured to determine the second basic thought chain data as the sample thought chain data if, within a preset number of inferences, a second basic thought chain data with a length less than or equal to the preset length threshold is obtained.
[0099] Specifically, the second basic thinking chain data refers to the thinking chain data obtained by secondary reasoning of the sample thinking chain data based on the general large language model and sample target prompt information when the length of the first basic thinking chain data exceeds a preset length threshold.
[0100] If the length of the first basic thought chain data exceeds a preset length threshold, it is considered that the first basic thought chain data is too long and may affect the computational efficiency of the target large language model, and therefore cannot be used as sample thought chain data.
[0101] By automatically invoking the general large language model, the sample target prompt information is used to perform secondary reasoning on the sample thinking chain data to obtain new basic thinking chain data, namely the second basic thinking chain data, and the number of repeated reasoning is recorded.
[0102] Understandably, the reasoning path and level of detail in the general-purpose large language model are random. In a single reasoning process, it may generate first basic thought chain data that exceeds the preset length threshold. Through multiple generation, the general-purpose large language model can try different reasoning simplification methods to generate second basic thought chain data.
[0103] For the second basic thought chain data, the length of the second basic thought chain data is determined by combining the length threshold and the number of repeated inferences, and whether the length of the second basic thought chain data is greater than the preset length threshold and whether the number of repeated inferences is within the preset number.
[0104] If a second basic thought chain data with a length less than or equal to a preset length threshold is obtained within a preset number of inferences, similar to the first basic thought chain data generated for the first time, the second basic thought chain data can be determined as sample thought chain data. By selecting shorter second thought chain data as sample thought chain data, the data quality of the sample thought chain data can be improved, thereby improving the computational efficiency and accuracy of the subsequent target large language model.
[0105] The following is Figure 7 For example, Figure 7 This is a schematic diagram illustrating the process of acquiring sample thought chain data, which explains the process of acquiring sample thought chain data in the embodiments of this application: First, by receiving the user's instruction (lower the right seat cushion in the third row slightly), the vehicle can obtain the sample voice request forwarded by the vehicle. Then, the user command is input into the RAG model for retrieval to obtain relevant APIs that are semantically similar to the user command, i.e., sample APIs; Then, the user commands and related APIs are concatenated using a prompt. That is, based on the sample preset prompt information template, the sample voice request, sample API and vehicle knowledge data are concatenated to determine the sample target prompt information. Then, the result of concatenating the prompts, i.e. the sample target prompt information, is input into the general large language model, and COT data, i.e. the first basic thinking chain data, is output. Then, by determining whether the COT length is greater than L (preset length threshold), COTs with a length greater than L are discarded, and new COT data is generated again through DeepSeek-R1 until the length of the COT data is less than or equal to L, so that COT data with a length less than or equal to L are identified as sample thought chain data. Finally, based on the sample thought chain data, the LLM (pre-trained large language model) was fine-tuned using the supervised fine-tuning (SFT) method to obtain an LLM model (target large language model) with API semantic understanding capabilities.
[0106] Thus, when the length of the first basic thought chain data exceeds a preset length threshold, secondary reasoning is performed on the sample thought chain data based on the general large language model and sample target prompts. If a second basic thought chain data with a length less than or equal to the preset length threshold is obtained within a preset number of reasoning iterations, this second basic thought chain data is identified as sample thought chain data. In this way, when the length of the first basic thought chain data exceeds the preset length threshold, secondary reasoning obtains the second basic thought data. By combining the preset length threshold and the preset number of iterations, shorter second basic thought chain data is selected as sample thought chain data, thereby improving the data quality of the sample thought chain data and ultimately improving the computational efficiency and accuracy of the subsequent target large language model.
[0107] Please see Figure 8 In some implementations, step 05 (processing sample user instruction text, at least one API, and vehicle knowledge data using a general large language model to generate sample thought chain data) further includes: 056: If the second basic thought chain data is not obtained within the preset number of attempts, receive the configured third basic thought chain data; 057: The third basic thinking chain data is determined as the sample thinking chain data.
[0108] In some implementations, the control module is further configured to receive configured third basic thought chain data if the second basic thought chain data is not obtained within a preset number of iterations. The control module is also configured to identify the third basic thought chain data as sample thought chain data.
[0109] In some implementations, the processor is further configured to receive configured third basic thought chain data if the second basic thought chain data is not obtained within a preset number of inference attempts. The processor is also configured to determine the third basic thought chain data as sample thought chain data.
[0110] Specifically, if the second basic thought chain data with a length greater than a preset length threshold is obtained through reasoning, and the number of secondary reasoning, i.e., repeated reasoning, is greater than a preset number, i.e., the second basic thought chain data is not obtained through reasoning within the preset number of times, then the configured third basic thought chain data can be received.
[0111] Understandably, the third basic thought chain data is thought chain data obtained through preset processing. For example, if the obtained thought chain data has a length greater than a preset length threshold, redundant expressions in the thought chain data can be manually reviewed and deleted to reduce the length of the thought chain data, thereby identifying the thought chain data as the third basic thought chain data.
[0112] Similar to generating the first basic thought chain data for the first time, the third basic thought chain data can be identified as the sample thought chain data, thereby improving the data quality of the sample thought chain data and thus improving the computational efficiency and accuracy of the subsequent target large language model.
[0113] Thus, if the second basic thought chain data is not obtained within a preset number of iterations, the configured third basic thought chain data is received; this third basic thought chain data is then designated as sample thought chain data. This process, where the second basic thought chain data is not obtained within the preset number of iterations, and the configured third basic thought chain data is received and designated as sample thought chain data, improves the data quality of the sample thought chain data, thereby enhancing the computational efficiency and accuracy of the subsequent target large language model.
[0114] Figure 9 This is a schematic diagram of the voice interaction process according to the embodiments of this application, hereinafter referred to as... Figure 9 Taking an example, the voice interaction process of the embodiments of this application will be explained: First, by receiving a user instruction (turn off the sound in the car), you can obtain the text of the user instruction forwarded by the vehicle. Then, the user command is input into the RAG model for retrieval to obtain relevant APIs that are semantically similar to the user command, i.e., candidate APIs; Then, the user command and related APIs are concatenated together, that is, based on the preset prompt information template, the user command text, alternative APIs and vehicle knowledge data are concatenated to determine the target prompt information; Finally, the concatenated prompt result (target hint information) is input into an LLM model (target large language model) capable of COT inference, such as the 7B model, which can output accurate...<API,Arguments> (Natural Language Processing Results).
[0115] This application also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it implements the steps of the voice interaction method described above.
[0116] It is understood that a computer program includes computer program code. Computer program code can be in the form of source code, object code, executable files, or some intermediate form. Computer-readable storage media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), and software distribution media, etc.
[0117] In this specification, the terms "specifically," "furthermore," "particularly," "understandably," etc., refer to specific features, structures, materials, or characteristics described in connection with embodiments or examples that are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0118] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of executable request code comprising one or more steps for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order according to the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.
[0119] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.
Claims
1. A voice interaction method, characterized in that, The method includes: Based on the user command text corresponding to the voice request forwarded by the vehicle, at least one alternative application interface related to the user command text is found in the pre-built application interface library. The target large language model is used to process the user instruction text, the at least one alternative application interface, and vehicle knowledge data to determine the target application interface from the at least one alternative application interface. Invoke the target application interface; The target large language model is trained using sample thought chain data generated by a general large language model. The general large language model has deep reasoning capabilities and can generate thought chain data. The target large language model can generate target thought chain data and natural language processing results carrying the associated information of the target application interface based on the user instruction text, the alternative application interface, and the vehicle knowledge data.
2. The method according to claim 1, characterized in that, The step of finding at least one candidate application interface related to the user command text in a pre-built application interface library based on the user command text corresponding to the voice request forwarded by the vehicle includes: Speech recognition is used to generate the user instruction text from the voice request; The at least one alternative application programming interface associated with the user instruction text is found using retrieval-enhanced production techniques.
3. The method according to claim 1, characterized in that, The step of processing the user instruction text, the at least one alternative application programming interface (API), and vehicle knowledge data using a target large language model to determine the target API from the at least one alternative API includes: The user instruction text, the at least one alternative application programming interface, and the vehicle knowledge data are concatenated and filled into a preset prompt information template to obtain the target prompt information; Based on the target prompt information, generate the target thought chain data carrying the associated information of the target application interface.
4. The method according to claim 3, characterized in that, The step of processing the user instruction text, the at least one alternative application programming interface (API), and vehicle knowledge data using a target large language model to determine the target API from the at least one alternative API also includes: Based on the target prompt information and the target thought chain data, the target application interface is filled with parameters to generate the natural language processing result carrying the associated information of the target application interface.
5. The method according to claim 1, characterized in that, The method includes: Based on the sample user instruction text corresponding to the sample voice request, determine at least one application programming interface related to the sample user instruction text; The general large language model is used to process the sample user instruction text, the at least one application programming interface, and the vehicle knowledge data to generate the sample thought chain data; The target large language model is fine-tuned based on the sample thought chain data, the sample user instruction text, the at least one application programming interface, and the vehicle knowledge data.
6. The method according to claim 5, characterized in that, The process of using the general large language model to process the sample user instruction text, the at least one application programming interface, and the vehicle knowledge data to generate the sample thought chain data includes: The sample user instruction text, the at least one application interface, and the vehicle knowledge data are concatenated and filled into the sample preset prompt information template to obtain the sample target prompt information; The sample target prompt information is processed using the general large language model to generate the first basic thought chain data; If the length of the first basic thought chain data is less than or equal to a preset length threshold, the first basic thought chain data is determined to be the sample thought chain data.
7. The method according to claim 6, characterized in that, The process of using the general large language model to process the sample user instruction text, the at least one application programming interface, and the vehicle knowledge data to generate the sample thought chain data further includes: If the length of the first basic thought chain data is greater than the preset length threshold, the general large language model is used to perform secondary reasoning based on the sample target prompt information. If a second basic thought chain data with a length less than or equal to the preset length threshold is obtained within a preset number of inferences, the second basic thought chain data is determined as the sample thought chain data.
8. The method according to claim 7, characterized in that, The process of using the general large language model to process the sample user instruction text, the at least one application programming interface, and the vehicle knowledge data to generate the sample thought chain data further includes: If the second basic thought chain data is not obtained within the preset number of times, the configured third basic thought chain data is received. The third basic thinking chain data is identified as the sample thinking chain data.
9. A voice interaction device, characterized in that, The device includes: The retrieval module is used to find at least one alternative application interface related to the user command text in a pre-built application interface library based on the user command text corresponding to the voice request forwarded by the vehicle. The processing module is configured to process the user instruction text, the at least one alternative application programming interface (API), and vehicle knowledge data using a target large language model to determine a target API from the at least one alternative API. The target large language model is trained using sample thought chain data generated by a general large language model, which possesses deep reasoning capabilities and can generate thought chain data. The target large language model can generate target thought chain data and natural language processing results carrying associated information about the target API based on the user instruction text, the alternative APIs, and the vehicle knowledge data. The calling module is used to call the target application interface.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by one or more processors, implements the method of any one of claims 1-8.