Artificial intelligence proxy method and device based on large model

Abstract

The invention provides an artificial intelligence agency method and device based on a large model, and the method comprises the steps: carrying out the intention recognition of input information in user information in response to the user information sent by a user, and obtaining an intention recognition result; determining a target service scene type based on the intention recognition result, wherein the target service scene type at least comprises a first type of service scene or a second type of service scene; if the type of the target service scene is a first type of service scene, performing matching on the basis of the input information and a trigger keyword corresponding to a sub-service scene to determine an interface corresponding to the target service scene; if the target service scene type is a second type of service scene, performing intention recognition based on the input information to obtain theme information; determining an interface corresponding to the target business scene based on the theme information; and calling an interface of the target service scene to perform information processing, and returning response information. In the mode, the decision-making efficiency and the intelligent level of the artificial intelligence agent system in a complex scene are improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to an artificial intelligence agent method and device based on a large model. Background Art

[0002] With the rapid development of digitalization in various industries, artificial intelligence agent systems need to process massive amounts of information and make complex decisions in a variety of task scenarios. However, existing large models have obvious shortcomings in flexibly responding to diverse task requirements due to problems such as insufficient support for long-range dialogues, limited reasoning capabilities, incomplete coverage of professional knowledge domain corpora, and lack of timeliness of training data. Summary of the invention

[0003] In view of this, the purpose of the present invention is to provide an artificial intelligence agent method and device based on a large model, which can recognize the intention and dynamically match the information input by the user, so as to overcome the problems of the prior art in long-range conversations, insufficient reasoning ability and weak dynamic adaptability, thereby improving the decision-making efficiency and intelligence level of the artificial intelligence agent system in complex scenarios.

[0004] In a first aspect, an embodiment of the present invention provides an artificial intelligence agent method based on a large model, comprising: in response to user information sent by a user, performing intent recognition on input information in the user information to obtain an intent recognition result; the intent recognition result is used to characterize the intent type of the input information; based on the intent recognition result, a target business scenario type is determined, and the target business scenario type includes at least a first type of business scenario or a second type of business scenario, the first type of business scenario is used to characterize the call of a third-party tool API interface or an API interface of a device management platform connected to a physical device for communication to complete a predefined business, and the second type of business scenario is used to characterize a conversation task of knowledge question and answer or small talk; The first type of business scenario includes calling interfaces of multiple sub-business scenarios, and each sub-business scenario has a preset trigger keyword; the second type of business scenario includes at least a calling interface of a large model; if the target business scenario type is the first type of business scenario, the interface corresponding to the target business scenario is determined based on matching the input information and the corresponding trigger keywords of the sub-business scenario; if the target business scenario type is the second type of business scenario, the subject information is obtained based on the input information through intent recognition, and the subject information is used to characterize the professional field and keywords to which the input information belongs; the interface corresponding to the target business scenario is determined based on the subject information; the interface of the target business scenario is called to process the information and return the response information.

[0005] In a second aspect, an embodiment of the present invention provides an artificial intelligence agent system, including: an intent recognition module, which is used to respond to user information sent by a user, perform intent recognition on input information in the user information, and obtain an intent recognition result; the intent recognition result is used to characterize the intent type of the input information; a business scenario type determination module, which is used to determine the target business scenario type based on the intent recognition result, and the target business scenario type includes at least a first type of business scenario or a second type of business scenario, the first type of business scenario is used to characterize the call of a third-party tool API interface or an API interface of a device management platform that communicates with a physical device to complete a predefined business, and the second type of business scenario is used to characterize a conversation task of knowledge question and answer or small talk; wherein the first type The business scenario includes calling interfaces of multiple sub-business scenarios, and each sub-business scenario is preset with trigger keywords; the second-category business scenario includes at least a calling interface of a large model; a first processing module is used to, if the target business scenario type is a first-category business scenario, match the input information and the corresponding trigger keywords of the sub-business scenario to determine the interface corresponding to the target business scenario; a second processing module is used to, if the target business scenario type is a second-category business scenario, perform intent recognition based on the input information to obtain subject information, and the subject information is used to characterize the professional field and keywords to which the input information belongs; determine the interface corresponding to the corresponding target business scenario based on the subject information; a response module is used to call the interface of the target business scenario for information processing and return response information.

[0006] The embodiment of the present invention provides an artificial intelligence agent method and device based on a large model. In this technical solution, the user's dialogue task is quickly classified into the first and second business scenarios according to the intention recognition results, so as to achieve accurate adaptation to diverse user needs. When it is confirmed as the first business scenario, the keywords extracted from the user input information are matched to the corresponding sub-business scenario for processing, thereby improving the response efficiency. When it is confirmed as the second business scenario, the large model is used to perform secondary intent recognition and the multi-round dialogue mechanism dynamically updates the collaborative mechanism of intent parameters, thereby improving the accuracy of the field to which the dialogue task belongs and the positioning of the user's true intention. At the same time, in the process of supplementing the missing parameters, it is also determined whether the topic transfer occurs by judging the context consistency of the historical dialogue, and then switching and processing the adaptive dialogue task type based on the new topic, thereby improving the decision-making efficiency and intelligence level of the artificial intelligence agent system in complex scenarios.

[0007] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0008] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0010] Figure 1 A flow chart of an artificial intelligence agent method based on a large model provided by an embodiment of the present invention; Figure 2 A schematic diagram of an artificial intelligence agent device based on a large model provided by an embodiment of the present invention; Figure 3 A schematic diagram of another large model-based artificial intelligence agent device provided in an embodiment of the present invention.

[0011] Icons: 1-intention recognition module; 2-business scenario type determination module; 3-first processing module; 4-second processing module; 5-response module; 6-alarm confirmation module. DETAILED DESCRIPTION

[0012] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0013] To facilitate understanding of this embodiment, the embodiment of the present invention is described in detail below.

[0014] Embodiment 1: Figure 1 A flowchart of an artificial intelligence agent method based on a large model provided by an embodiment of the present invention. Figure 1 , an artificial intelligence agent method based on a large model, including: Step S101, in response to user information sent by the user, perform intent recognition on input information in the user information to obtain an intent recognition result; the intent recognition result is used to characterize the type of intent of the input information.

[0015] Here, the artificial intelligence agent system receives user information sent by the user in various forms, such as voice input, text input, image input or multimodal input, as the starting point of the interaction. If the user inputs voice information, the voice recognition module converts the voice into processable text information (recognizes the text). Based on the recognized text or directly input text, the semantic recognition method is used to extract the type of intention expressed by the user to obtain the intention recognition result. The intention recognition result is used to distinguish the categories of user needs, which are the first type of business scenario dialogue needs (such as setting an alarm, content on demand, robot route navigation or home appliance control, etc.) and the second type of business scenario dialogue needs (such as chatting or professional knowledge Q&A, etc.). For example: the user inputs "help me set an alarm for tomorrow", which matches the first type of business scenario of "setting an alarm"; the user inputs "nutritional analysis of bananas", which matches the second type of business scenario of "multimodal content analysis".

[0016] Step S102, determine the target business scenario type based on the intention recognition result, and the target business scenario type includes at least the first type of business scenario or the second type of business scenario. The first type of business scenario is used to characterize the call of a third-party tool API interface or an API interface of a device management platform that communicates with a physical device to complete a predefined business, and the second type of business scenario is used to characterize a knowledge question and answer or chat dialogue task; the first type of business scenario includes a calling interface of multiple sub-business scenarios, and each sub-business scenario is preset with a trigger keyword; the second type of business scenario includes at least a large model calling interface.

[0017] Here, the sub-business scenarios are task-oriented dialogue scenarios in different fields, and each sub-business scenario is associated with a trigger keyword for quickly locating the business sector. The first type of business scenario can be a smart home appliance control interface and a path navigation interface, etc.; the second type of business scenario is a large model call interface, etc. The predefined business includes at least but is not limited to device control, information query, task scheduling, path navigation, data processing or other operations that can be performed through the API interface. For example: if the intent recognition result is "set alarm", the corresponding trigger keywords may include "alarm", "time", "reminder", etc. It should be noted that the target business scenario type determined based on the intent recognition result here can be determined by keyword matching through the intent recognition result, and the user's dialogue demand type can be quickly classified to determine whether it is a first type of business scenario or a second type of business scenario, so as to quickly call the corresponding dialogue processing system to quickly and high-quality process the user's dialogue needs, and speed up the response to the user efficiency. In other embodiments, the input information can also be processed by the large model to determine its corresponding target business scenario type, and its dialogue demand type can be adaptively and accurately located to avoid locating the wrong business type through simple keyword matching. For example, if the user inputs information such as "I'm a little hungry and want to order food, but I don't know what to eat. I want a spicy and healthy menu. Do you have any good recommendations?" or "What's the weather like today? I want to go out and play. Do you have any good places to recommend?", if keyword matching is used, the above dialogue needs can belong to both the first type of business scenario (corresponding to ordering food or weather inquiries) and the second type of business scenario (Q&A about nutritional meal planning or travel recommendations), which can easily lead to wrong decisions. However, if large model intent recognition is used, it can be accurately located in the second type of business scenario.

[0018] Furthermore, the big model in this application includes at least the first big model and the second big model. The first big model is a general big model that performs well in general knowledge question and answer and small talk, such as GPT and Wenxin Yiyan, etc.; the second big model is an industry vertical big model that performs well in professional knowledge question and answer, such as the transportation construction industry vertical big model, the financial vertical big model, iFlytek Spark cognitive big model, Chitu big model, Wenxiu big model and Xingzhi X-intelligence, etc. After the secondary intention recognition, the corresponding business classification is carried out. If the dialogue task is to deal with general small talk or emotional care tasks, it emphasizes natural dialogue and semantic fluency, has strong generalization ability, and is more inclined to generate semantically coherent dialogues. If the deep understanding of specific domain knowledge is weak, the calling interface of the big model is determined to be the first big model, otherwise the calling interface of the big model is determined to be the second big model.

[0019] Step S103: If the target business scenario type is a first type of business scenario, an interface corresponding to the target business scenario is determined by matching the input information with a trigger keyword corresponding to the sub-business scenario.

[0020] Here, when the target business scenario type is judged to be the first type of business scenario, it means that the user hopes to complete a specific task or operation through the artificial intelligence agent system, such as: setting an alarm, controlling smart home devices, playing audio content (music, audio novels, etc.), querying the weather, page jump or navigation path, etc. Based on the keywords in the input information, the corresponding sub-business scenario is quickly located. For example, if the input information is "query Beijing's weather tomorrow", the API interface is called for the weather query sub-business; if the input information is "take me to the business center hall of XX", the API interface in the device management platform connected to the navigation robot communication is called based on the device unique identification code as the path navigation sub-business. If the path navigation is performed for a certain online virtual attraction or offline physical attraction, the preset attractions along the path navigation can be introduced with relevant content and information interaction can be generated with the user; if the input information is "turn on the lighting in the living room", the API interface is called for the smart home appliance control sub-business; if the input information is "play Jay Chou's song "Chrysanthemum Terrace", the API interface is called for the audio content playback sub-business; if the input information is "jump to "Draw"", the API interface is called for the display interface jump sub-business.

[0021] Step S104, if the target business scenario type is the second type of business scenario, then perform intent recognition based on the input information to obtain subject information, where the subject information is used to characterize the professional field and keywords to which the input information belongs; determine the interface corresponding to the target business scenario based on the subject information.

[0022] When the target business scenario type is judged to be the second type of business scenario, it means that the user wants to conduct professional knowledge Q&A or chat, etc. Among them, professional knowledge Q&A may be pension policy inquiries, health advice, etc.; chat may be emotional companionship or travel experience sharing, such as "How is the stock trend today, please give advice" is determined not to belong to the first type of business scenario based on the intention recognition result. In order to more accurately locate the professional field described by the user's dialogue needs, more professional knowledge is applied to process the dialogue task, the user's input information is intentionally recognized to obtain the theme information, and the noise information is filtered out to prepare for the subsequent high-quality prompt word reconstruction, avoiding making wrong business classification decisions based on the first intention recognition result after simple keyword matching.

[0023] Step S105, calling the interface of the target business scenario to process information and return response information.

[0024] Here, the target business interface is called to process the user request, for example: calling the alarm setting interface to set the alarm for the user-specified time; or calling the large model interface to process the user's question, etc., and the response information is generated according to the processing result returned by the interface, and returned to the user through the interaction method used by the user (such as voice, text, etc.).

[0025] The large-model-based artificial intelligence agent method disclosed in this embodiment ensures that the artificial intelligence agent system can efficiently respond to diverse user needs, especially in real-time dynamic tasks and complex scenarios, by means of business scenario matching, multi-round dialogue supplementation, and interface calls, providing a flexible and intelligent interactive experience.

[0026] In one embodiment, after calling the interface corresponding to the target business scenario, in order to further improve the accuracy of user intent recognition in the business scenario, it is necessary to check whether the input information or subject information is complete. If the parameters are incomplete, the user is supplemented with parameters through multiple rounds of dialogue until the input information or subject information corresponds to the intent parameters. Among them, the intent parameters are specific information related to the intent (such as time, location, task description, etc.). When the target business scenario type is a first-class business scenario or a second-class business scenario, step S105 includes the following steps S301-S304: Step S301: Based on the interface of the target business scenario, obtain all preset intent parameters corresponding to the target business scenario.

[0027] Here, according to the preset intent parameters of the target business scenario, check whether the intent parameters in the intent recognition result contain all the intent parameters required for the scenario. For example, in the "path navigation" scenario, the required parameters include "starting point" and "end point".

[0028] Step S302: Determine whether the input information or subject information contains all intent parameters corresponding to the target business scenario.

[0029] Specifically, if the user input information is "finalize a high-speed rail ticket to Beijing", the preset intention parameters in the target business scenario corresponding to travel ticket purchase include at least "departure place, destination, time". The required intention parameters are set to make the user's input information intention clear, so that the response information or processing results generated by the dialogue processing system are more in line with the user's expectations, thereby improving user experience and stickiness.

[0030] Step S303: If the input information or subject information does not contain all the intent parameters, it is determined that the input information or subject information is incomplete, and the input information or subject information is updated through multiple rounds of dialogue until the updated input information or subject information is complete.

[0031] Specifically, if the user inputs "help me set an alarm for 6 a.m. tomorrow", the intent recognition result is: set the alarm; it is determined as the first type of target business scenario and the interface for setting the alarm is called; in the "set alarm" scenario, the required parameters include "time" and "date", and based on the input information, it is determined that the current intent parameters are complete, and the alarm is set based on all intent parameters.

[0032] However, if the user enters the information "Help me set an alarm for tomorrow morning" in the above sub-business scenario, according to the above steps, it can be determined that the current intent parameters are incomplete and "time" is missing. After determining that the intent parameters are incomplete, analyze the missing parts of the parameters required for the target business scenario, and generate follow-up information for the missing parameters, asking: "OK, please tell me the specific time you want to set?". Supplement the user's intent parameters through multiple rounds of interaction. Then extract the supplementary information from the user's reply and update the intent parameters until all required parameters are complete. For example, if the user replies "tomorrow morning at 6 o'clock", update the intent parameters to "tomorrow", "6 o'clock in the morning" and "alarm". After multiple rounds of dialogue updates, check whether the intent parameters meet all the parameters required by the target business scenario.

[0033] Step S304: Process the updated complete input information or subject information as input information to obtain response information.

[0034] In one embodiment, the user information also includes identity information. In step S304, the step of updating the input information or subject information through multiple rounds of dialogue until the updated input information or subject information is complete includes the following steps S401-S403: Step S401, based on the identity information and the preset first identification information, determine whether there is corresponding historical dialogue information for the current dialogue; the first identification is the multi-round dialogue identification information.

[0035] Here, identity information is used to associate user historical conversation records, for example, through user accounts, device identifiers, or session identifiers, to ensure that the artificial intelligence agent system can correctly locate user historical conversation data. The first identifier is multi-round conversation identification information, which is used to distinguish whether the current conversation continues the previous topic. If it has the first identifier, it means that the conversation is part of a multi-round conversation and needs to be processed in combination with historical information; if it does not have the first identifier, it means that the current conversation is a new topic. Based on the first identifier, it is determined whether there is historical conversation information that can be associated with the current conversation. If it exists, enter the historical conversation information association processing process; if it does not exist, enter the follow-up information generation and parameter supplement process. Specifically, the user inputs: "Help me analyze the nutrition." Based on the user's identity information, it is determined that the current conversation does not have the first identifier, that is, there is no historical record; and it is determined that the intention is incomplete and the intention parameter supplement process needs to be entered.

[0036] Step S402, if there is no historical dialogue information in the current dialogue, based on the missing required parameters to be supplemented in the input information or subject information and the preset follow-up question template, generate follow-up question text information, and convert the follow-up question text information into follow-up question audio.

[0037] Here, analyze the required parameter list corresponding to the target business scenario, compare it with the intent parameters corresponding to the current input information or subject information, and find out the missing parameters to be supplemented. For example, in the above example: the required parameters in the nutritional analysis scenario include "analysis object" (such as food name), and the current intent parameters are missing "analysis object". Generate follow-up text information based on the preset follow-up question template of the target business scenario. For example: Follow-up question template: "Please tell me what food you want to analyze?". The generated follow-up text message: "OK, please tell me what food you want to perform nutritional analysis on, such as cake", is converted into voice format and sent to the user.

[0038] Step S403, add the first identification information to the current conversation, and send the follow-up audio and the first identification information to the user corresponding to the identity information, and update the input information or subject information according to the supplementary information received from the user's reply until the updated input information or subject information is complete.

[0039] Here, a first identifier is attached to the current conversation, marking it as part of a multi-round conversation. The identifier is bound to the user's identity information to ensure that subsequent conversations are contextually associated. The user replies with additional information in voice or text form, for example: the user replies: "cake." If the user replies in voice, it is converted into text through voice recognition, and then the supplementary parameters are extracted through the intent recognition module. Fill the supplementary information into the intent parameters. For example: the initial intent parameter is "nutritional analysis", which is updated to "nutritional analysis" and "cake" after supplementation. Check whether the intent parameters are complete. If complete, call the business interface; if there are still missing, repeat the questioning process until the parameters are complete.

[0040] In one embodiment, after step S401, the method further includes the following steps S501-S504: Step S501: If there is historical conversation information in the current conversation, obtain historical topic information corresponding to the historical conversation information; the historical topic information includes historical intention parameters and historical target business scenarios.

[0041] Here, the historical conversation information related to the current conversation is located through the first identifier. The historical conversation information may include the user's target business scenario in the previous conversation, the supplemented intent parameters, and other key data. Historical intent parameters: For example, in the historical conversation, the user enters "nutrition analysis" and supplements "cake" as the analysis object, then the historical intent parameters may be "nutrition analysis" and "cake". Historical target business scenario: records the target business scenario corresponding to the conversation, such as "nutrition analysis".

[0042] Step S502, determining whether the historical intention parameters and the parameters to be supplemented are contextually consistent.

[0043] Here, by analyzing the logical relationship between the historical intent parameters and the current intent parameters, we determine whether they are in the same or similar semantic range. For example, if the user previously discussed "nutritional analysis of cake" and currently mentions "change to bread", the two are consistent in context. If the user previously discussed "playing music" and currently mentions "analyzing the nutrition of bread", the context is inconsistent.

[0044] The judgment method may be a combination of keyword matching and business scenario matching. Keyword matching includes checking whether the historical intent parameters and the current intent keywords overlap or are semantically related. Business scenario matching includes whether the historical target business scenario is the same as the current business scenario or belongs to a similar category.

[0045] Step S503, if the historical intent parameters are contextually consistent with the parameters to be supplemented, determine whether the historical intent parameters include the parameters to be supplemented; if the historical intent parameters include the parameters to be supplemented, update the input information or subject information based on the historical intent parameters until the updated input information or subject information is complete; if the historical intent parameters do not include the parameters to be supplemented, and historical preference operation data is stored in the identity information, update the intent parameters based on the historical preference operation data until the updated input information or subject information is complete.

[0046] Here, if the historical intent parameters are contextually consistent with the parameters to be supplemented, that is, the historical session and the current session are on the same topic, check whether the historical intent parameters contain the parameters to be supplemented that are missing in the current scenario. If the historical intent parameters contain the parameters to be supplemented, directly fill the historical parameters into the intent parameters corresponding to the current input information or topic information. If the historical intent parameters do not include the parameters to be supplemented, and the historical preference operation data is stored in the identity information, you can refer to the historical preference data for supplementation.

[0047] Historical preference operation data refers to preference data accumulated by users over a long period of time. For example, music preference data may include genre, artist, era, theme, language, etc. In home appliance control, preference data may include air conditioner setting temperature, speaker volume, etc.

[0048] Specifically, the user inputs "play music", determines that the intent parameter in the current input information is incomplete, adds the first identification information to the current conversation, and sends the follow-up audio and the first identification information to the user.

[0049] The supplementary information received from the user's reply is "any music will do". The artificial intelligence agent system did not obtain the required parameters to be supplemented, but the music preference data "Jay Chou's songs" was stored in the identity information. The supplementary information was determined based on the music preference data to play Jay Chou's songs.

[0050] For another example, the user inputs "increase the air conditioning temperature", determines that the intent parameter in the current input information is incomplete, adds the first identification information to the current conversation, and sends the follow-up audio and the first identification information to the user. The supplementary information received from the user is "a higher temperature is fine", and the artificial intelligence agent system does not obtain the required parameters to be supplemented, but the home appliance control preference data "air conditioning temperature 24 degrees Celsius" is stored in the identity information. The supplementary information is determined based on the home appliance control preference data to adjust the air conditioning temperature to the user's preferred 24 degrees Celsius.

[0051] Step S504, if the historical intention parameter and the required parameter to be supplemented do not have contextual consistency, delete the first identification information corresponding to the historical conversation information; and return to the step of performing intent recognition on the input information in the user information to obtain the intent recognition result.

[0052] Here, if the context is inconsistent, it means that the user may have started a new topic. At this time, delete the first identifier that is not related to the current conversation to avoid misusing irrelevant historical information. After deleting the irrelevant first identifier, treat the current conversation as an independent single-round conversation and enter the judgment process of whether the intention parameter is complete. For example: the user previously discussed "nutritional analysis of cake" and currently inputs "play music", then delete the first identifier related to "nutritional analysis", and perform intent recognition based on "play music" to obtain the intention recognition result, and determine the target business scenario type based on the intention recognition result, and then execute the subsequent steps. For example, the user's first input information is A: I want to order a meal. The robot replies B: What do you want to eat and when will it be delivered. The user's second input information is A: I don't know what to eat either, can you match me with a nutritious and beauty package? At this time, the types of the two rounds of user input dialogues belong to the first type of business and the second type of business respectively. In the prior art, the robot cannot switch between different types of dialogues, and cannot recognize the information entered by the user for the second time, resulting in the forced termination of the dialogue demand. In this embodiment, a flexible and accurate multi-round dialogue mechanism is constructed by utilizing historical topic information, context consistency judgment and user preference data, which can efficiently supplement the user's missing intent parameters, and adaptively delete irrelevant historical information when switching contexts to avoid interfering with new topic processing. At the same time, when different types of dialogue tasks are carried out in an intersecting manner, the artificial intelligence agent system can also adaptively switch the dialogue processing system flexibly, and has a high accuracy rate in intent recognition. It does not only rely on the large model to process all dialogue request tasks, but can hand over some task-oriented dialogue requests to other dialogue systems for processing, reducing the load balancing pressure of the large model, and improving the response efficiency of the large model, thereby improving the user experience and interaction effect.

[0053] In one example, in order to make up for the deficiencies of the big model in terms of knowledge timeliness, processing of massive information, professional accuracy, and domain adaptability, the professional knowledge associated with the dialogue is reconstructed into the prompt words through prompt word engineering, and accurate and comprehensive background knowledge support is provided to the big model in a knowledge enhancement manner, so as to expand and supplement the knowledge capacity of the big model, so that the big model can not only understand the user's dialogue intention more deeply, but also generate more accurate and relevant information, and improve the decision-making ability and the accuracy of response feedback. Step S304 includes the following steps S601-S604: Step S601: perform similarity calculation based on the updated topic information and the corpus in the preset knowledge base, and determine whether there is related knowledge based on the similarity. The related knowledge is the text whose similarity corresponding to the corpus in the knowledge base is greater than a preset threshold.

[0054] Here, a knowledge base consisting of knowledge in various fields is formed based on professional knowledge from various industries. Similarity calculation is performed in the knowledge base based on the updated subject information. When the corresponding relevance of the corpus in the knowledge base is greater than the preset threshold, it means that the associated knowledge is related to the current topic. Using it as background knowledge support can allow the large model to better understand and process the user's input information.

[0055] Step S602: if there is associated knowledge, an updated first prompt word is generated based on the complete subject information, the associated knowledge and the preset first prompt word and the preliminary prompt word construction method.

[0056] Here, the related knowledge is informed to the big model by constructing prompt words to know the relevant background knowledge. Even the general model can improve its professional question-answering ability, avoiding the need for a large amount of high-quality annotated specific corpus data for fine-tuning the vertical big model. The difficulty, high cost and high professionalism of collecting high-quality corpus in specific fields have reduced the threshold for processing professional knowledge question-answering dialogue tasks for big models, and made the response text generated by the big model based on prompt words more professional and in line with user expectations. A prompt word template is preset for each subject information corresponding to the intent. The first prompt word template contains the structure and content framework of the generated prompt word. The template will provide guidance for the construction of the prompt word based on the scene-specific keywords and user input information. Based on the framework provided by the prompt word template, the first prompt word, the complete subject information and related knowledge, the updated first prompt word is reconstructed. For example, the preset template of the first prompt word can be: Please answer the question based on the role [role description], around the keyword [keyword], and the related knowledge [XXX]: [specific question]. Adapt the chat-type content and generate the first prompt word: Based on the role of finance major, please answer the question around the keywords: software product tax refund, related knowledge [XXX]: What is the process of software product tax refund? Step S603: if there is no associated knowledge, an updated second prompt word is generated based on the complete topic information, the preset second prompt and the preset prompt word construction method.

[0057] Here, if there is no related knowledge, it may not be a professional question and answer, but rather a casual conversation. In this case, the casual chat template is used to generate the basic prompt words. The second prompt word template is: Please answer the question based on the role [role description], around the keyword [keyword]: [specific question]. The casual chat content is adapted to generate the updated second prompt words: Please answer the question based on the role of emotional care expert, keyword: happy, plug-in call information, empty; question: I am really happy today, how about you? Step S604: Based on the first prompt word or the second prompt word as input information, the large model is used to process and generate and return response information.

[0058] Here, the prompt word is processed using a large model, and the generated response information is returned to the user. The response information can be voice or text data.

[0059] In one embodiment, in order to enrich the capabilities of the large model, expand the application scenarios of the large model, and improve the practicality and interactivity of the large model, step S602 includes the following steps S701-704: Step S701: According to the updated theme information, find out whether there is a preset plug-in keyword matching the theme information.

[0060] Here, the preset plug-ins can include two categories: information enhancement plug-ins and interaction enhancement plug-ins. Among them, information enhancement plug-ins can help users obtain real-time information, such as stock prices, latest news, etc. The corresponding plug-ins can be WebPilot (web information processing plug-in), Wikipedia, Wolfram (a combination of Wolfram Alpha's computational language and ChatGPT large language model), VoxScript (video content summary plug-in) and other plug-ins (the corresponding functional description is information query, obtain real-time information). Interaction enhancement plug-ins can read multimodal inputs such as knowledge base, PDF, pictures, voice, and generate multimodal outputs such as tables and mind maps. The corresponding plug-ins can be Banli Kanban, Zhixi Al Mind Map, Al PPT.cn, TreeMind Tree Map, Biyou ChatPPT and other plug-ins (the corresponding functional description is multimodal content generation, generating visual charts, project Kanban, mind maps and other contents). Search from the preset plug-in library to see if there are matching plug-in keywords, and the plug-in keywords are determined by the functional description of the plug-in. For example, the information enhancement plug-in contains keywords such as “real-time information” and “data query”, and the interaction enhancement plug-in contains keywords such as “multimodal input and output” and “mind map”.

[0061] Specifically, for the intent "analyze the nutritional ingredients of cake", matching the keyword "nutritional ingredient analysis" can find information enhancement plug-ins; for the intent "analyze the nutritional ingredients of cake and present it with a mind map", matching the keyword "mind map" can find interaction enhancement plug-ins such as "Zhixi AI Mind Map".

[0062] Step S702: if no preset plug-in keyword matching the subject information is found, an updated first prompt word is generated based on the complete subject information, associated knowledge, the preset first prompt word, and the preset prompt word construction method.

[0063] Here, when there is no preset plug-in, the updated first prompt word is generated based on the complete subject information, the preset first prompt word and the preset prompt word construction method. The preset template corresponding to the subject information is selected. The first prompt word may include a role description, a task goal and supplementary information. The role description is used to define the angle or expert identity of the answer generation. The task goal is used to clarify the subject or specific task to be analyzed. The supplementary information is used to list specific analysis directions or concerns. The associated knowledge can clarify the background knowledge of the current problem. Generate and output the complete first prompt word to ensure that the model can be effectively guided to complete the answer.

[0064] Specifically, input information: user intention: "Analyze the nutritional content of cake". Intent recognition result: analysis; subject information: cake, nutritional content; related knowledge [XXX]. The first prompt word template: 1. You are a [field expert] specializing in [task area], and your analysis should be based on [methods and principles]; 2. [Analysis objective 1]: Analyze [specific attributes 1] of [topic]; 3. [Analysis Objective 2]: [Task Requirements / Additional Attributes]; 4. [Analysis objective 3]: [Additional suggestions].

[0065] Based on the complete subject information, target business scenario type and target business scenario, determine the template filling part as follows: Field experts: Nutrition experts; Task area: Questions and answers about the effects of food on human health; Methods and principles: Based on scientific research, provide practical suggestions; Analysis goal 1: Analyze the nutrients contained in food and its harmful effects on the human body; Analysis Objective 2: Potential impact of continued consumption of this food on health; Analysis objective 3: Provide healthier food alternatives.

[0066] Generate the first prompt word: 1. You are a nutritionist who specializes in studying the effects of food on human health. Your analysis should be based on scientific research and provide practical dietary advice. 2. Relevant scientific literature includes "XXX". 3. Nutritional analysis: Analysis of the nutrients contained in cakes and their harmful effects on the human body. 4. Long-term effects: The potential effects of continued consumption of cakes on physical health. 5. Substitution suggestions: Provide healthier food alternatives. 6. The topic I want to talk to you about is: the nutritional content of cakes.

[0067] Step S703: If a preset plug-in keyword matching the theme information is found, the plug-in corresponding to the preset plug-in keyword is determined as the target plug-in, and the plug-in call information corresponding to the target plug-in is determined. An updated first prompt word is generated based on the complete theme information, plug-in call information, associated knowledge, preset first prompt words, and preset prompt word construction method.

[0068] Here, the plug-in function can be directly called through the plug-in call information, such as generating a mind map or a chart. The plug-in call information may include a plug-in identifier, a plug-in API address, and required plug-in call parameters. When a plug-in exists, specifically, the first prompt word generated is: 1. You are a nutritionist who specializes in the impact of food on human health. Your analysis should be based on scientific research and provide practical dietary advice. 2. Relevant scientific literature includes "XXX". 3. Nutritional analysis: Analyze the nutrients contained in the cake and its harmful effects on the human body. 4. Long-term effects: The potential impact of continued consumption of this food on your health. 5. Suggested substitutes: Provide healthier food alternatives. 6. The topic I want to talk to you about is: Analyze the nutritional content of cakes and present it in a mind map. 7. The calling parameter information of the relevant plug-in tool is: XXXXXX. Step S704: Based on the first prompt word or the second prompt word as input information, the large model is used to process and generate and return response information.

[0069] Here, the first prompt word or the second prompt word is input into the large model to generate a response that meets the user's needs. The returned information may be plain text, an image (such as a mind map), a data table, etc.

[0070] In this embodiment, by introducing the collaborative operation of plug-in calling, knowledge base loading and prompt word reconstruction, the functional scalability and professional response ability of the artificial intelligence agent system are significantly improved. When no plug-in is matched, the artificial intelligence agent system can use the general large language model to generate answers; when a plug-in is matched, the artificial intelligence agent system combines the knowledge base and the reconstructed prompt words to generate answers that better meet the user's expectations.

[0071] In one embodiment, step S602 includes the following steps S705-S707: Step S705: Search according to the updated theme information to see whether there is a preset plug-in keyword that matches the theme information.

[0072] Step S706: If no preset plug-in keyword matching the theme information is found, an updated second prompt word is generated based on the complete theme information, the preset second prompt word and the preset prompt word construction method.

[0073] Step S707: If a preset plug-in keyword matching the theme information is found, the plug-in corresponding to the preset plug-in keyword is determined as the target plug-in, and the plug-in call information corresponding to the target plug-in is determined; and an updated second prompt word is generated based on the complete theme information, the plug-in call information, the preset second prompt word and the preset prompt word construction method.

[0074] The construction method of the second prompt word here is just to remove the associated knowledge from the construction method of the first prompt word. The rest can be inferred from the construction method of the first prompt word above, and will not be repeated here.

[0075] In one embodiment, the method further includes steps S801-804: Step S801, if external information sent by an external device is received, determine the device type of the external device based on the device identifier in the external information; the device type includes risk warning devices and other devices; risk warning devices include life and health monitoring devices and intelligent security monitoring devices; the external information includes the source of information.

[0076] Here, the artificial intelligence agent system receives data information from external devices through the network, Bluetooth or other communication methods. External information includes device identifiers, information content and sending time. For example, a life and health monitoring device may send an "abnormal heart rate alert: heart rate 180 beats / minute". The artificial intelligence agent system parses the device identifier in the information to determine the device type of the external device. Risk warning devices can be life and health monitoring devices (smart bracelets, blood pressure monitors) and smart security monitoring devices (smoke detectors, smart door locks), etc. Other devices can be mobile devices (mobile phones), etc.

[0077] Specifically, if the device identifier is "HeartMonitor-001", the corresponding device type is life and health monitoring equipment. If the device identifier is "SmokeDetector-002", the corresponding device type is intelligent security monitoring equipment.

[0078] Step S802: If the device type is a life and health monitoring device, save the external information; based on the pre-set abnormal confirmation voice generation method, generate and return the abnormal confirmation voice.

[0079] Here, the information sent by the life and health monitoring equipment is saved, and the voice content is generated according to the preset template, for example: "Alert: Your heart rate is detected to be abnormal, currently 180 times / minute, do you need help?". The user is reminded of the abnormal situation at the first time, and the user is further confirmed through voice interaction whether he needs help. Specifically, after receiving the abnormal heart rate information, the artificial intelligence agent system generates a voice "The heart rate is detected to be too fast, do you need to notify the emergency contact?".

[0080] Step S803: If the device type is an intelligent security monitoring device, save the external information.

[0081] Here, information sent by devices such as smoke alarms and door lock alarms is saved to provide users with a basis for subsequent inquiries. Since such devices generally have built-in alarm functions (such as beeps), no voice confirmation is required. When a fire warning is received from a smoke alarm, the information is recorded: Fire alarm, thick smoke detected.

[0082] Step S804: if the device type is other devices, a prompt voice is generated and returned based on the information source and a preset prompt voice generation method.

[0083] Here, we extract the information source sent by external devices (such as SMS, Taobao push notifications), generate prompt voices based on the information source, and provide instant voice reminders to prevent users from missing important information. For example: mobile operator SMS: "You have a new SMS message."; Taobao notification: "You have a Taobao order update message.".

[0084] In one embodiment, if the device type is a life health monitoring device or other device, the step of generating and returning a prompt voice based on the information source and a preset prompt voice generation method includes: If an abnormal reply message from the user is received, it is identified whether the abnormal reply message includes a preset help keyword.

[0085] Here, the artificial intelligence agent system monitors the user's reply information to the abnormal confirmation voice, and extracts the user's voice or text information as abnormal reply information. The abnormal reply information may include a clear request for help, a negative response, or a vague expression. Specifically, the user replies: "Please help me contact a doctor!" or "No help is needed for the time being." The artificial intelligence agent system extracts the subject information through speech recognition or natural language processing. The preset help keywords may include: help, contact a doctor, emergency, help, and call 120, etc.

[0086] The user's abnormal reply information is matched with the preset help keywords, and the natural language processing technology (NLP) is used to process semantically ambiguous expressions. Specifically, the abnormal reply information is: "Help me contact the emergency contact." The matching result: the keywords "contact" and "emergency contact" are identified. If the abnormal reply information includes the preset help keywords, the preset alarm information is sent to the contact based on the preset contact information. Here, the preset alarm information is sent to the contact to pass the user's danger information to the relevant parties as soon as possible to ensure the user's safety. According to the mapping relationship between the user's identity information and the emergency contact, the user's emergency contact list or the contact information of the relevant danger department is queried. Based on the preset contact information, the preset alarm information is generated and sent. The alarm information may include a basic description of the abnormal event, device monitoring data, and user location (such as if the device supports GPS function). The sending method may be SMS, phone, and email.

[0087] If the abnormal reply does not include the preset help keywords, if the user explicitly replies "no help needed" or does not have help keywords, end the current conversation. Save the user's reply information and device data for subsequent queries. For example, the user replies: "I'm fine, thank you." The AI ​​agent system records the user information and ends the voice interaction.

[0088] This embodiment analyzes the abnormal information of external devices and generates abnormal confirmation or prompt voice, thereby achieving accurate monitoring and timely response to the user's life and health status or security anomalies.

[0089] The embodiment of the present invention provides an artificial intelligence agent method and device based on a large model, including: in response to user information sent by a user, performing intent recognition on input information in the user information to obtain an intent recognition result; the intent recognition result is used to characterize the intent type of the input information; according to the intent recognition result, searching for a trigger keyword matching the intent recognition result in a business scenario library; the business scenario library includes multiple business scenarios corresponding to each business scenario type, and each business scenario is provided with a trigger keyword; if a trigger keyword matching the intent recognition result is found, determining a target business scenario type according to the trigger keyword; the target business scenario type is a first type of business scenario or a second type of business scenario; the first type of business scenario is used to characterize a business scenario that completes a user's predefined or operation; the second type of business scenario is used to characterize a business scenario in a professional knowledge field or multimodal interaction; if a trigger keyword matching the intent recognition result is not found, determining the target business scenario type to be other types of business scenarios; other types of business scenarios are used to characterize business scenarios that do not belong to the first type of business scenarios and the second type of business scenarios; based on the target business scenario type and a pre-set correspondence between the business scenario and the business interface, determining a target business interface corresponding to the target business scenario type; calling the target business interface and returning a response message. In this way, the information input by the user can be used for intent recognition and dynamic scene matching, thereby improving the reasoning ability and dynamic adaptability of the artificial intelligence agent system, thereby improving the decision-making efficiency and intelligence level of the artificial intelligence agent system in complex scenarios.

[0090] Embodiment 2: Figure 2 A schematic diagram of an artificial intelligence agent device based on a large model provided in an embodiment of the present invention, referring to Figure 2 , an artificial intelligence agent device based on a large model, including: The intention recognition module 1 is used to respond to the user information sent by the user, perform intention recognition on the input information in the user information, and obtain an intention recognition result; the intention recognition result is used to characterize the type of the intention of the input information.

[0091] Business scenario type determination module 2 is used to determine the target business scenario type based on the intention recognition result. The target business scenario type includes at least the first type of business scenario or the second type of business scenario. The first type of business scenario is used to characterize the call of a third-party tool API interface or the API interface of a device management platform that communicates with a physical device to complete a predefined business. The second type of business scenario is used to characterize the dialogue tasks of knowledge questions and answers or small talk. The first type of business scenario includes calling interfaces of multiple sub-business scenarios, and each sub-business scenario is preset with trigger keywords. The second type of business scenario includes at least the calling interface of the large model.

[0092] The first processing module 3 is used to determine the interface corresponding to the target business scenario by matching the input information and the trigger keywords corresponding to the sub-business scenario if the target business scenario type is the first type of business scenario.

[0093] The second processing module 4 is used to obtain subject information by performing intent recognition based on the input information if the target business scenario type is a second type of business scenario. The subject information is used to characterize the professional field and keywords to which the input information belongs; and determine the interface corresponding to the target business scenario based on the subject information.

[0094] The response module 5 is used to call the interface of the target business scenario to process information and return response information.

[0095] In one embodiment, the response module 5 is further configured to: Based on the interface of the target business scenario, obtain all preset intent parameters corresponding to the target business scenario; Determine whether the input information or subject information contains all the intent parameters corresponding to the target business scenario; If the input information or subject information does not contain all intent parameters and it is determined that the input information or subject information is incomplete, the input information or subject information is updated through multiple rounds of dialogue until the updated input information or subject information is complete; Based on the updated complete input information or subject information as input information, response information is obtained.

[0096] In one embodiment, the user information also includes identity information, and the response module 5 is further used to: Based on the identity information and the preset first identification information, determining whether there is corresponding historical dialogue information for the current dialogue; the first identification is the multi-round dialogue identification information; If there is no historical conversation information in the current conversation, based on the missing required parameters to be supplemented in the input information or subject information and the preset follow-up question template, generate follow-up question text information, and convert the follow-up question text information into follow-up question audio; Add the first identification information to the current conversation, send the follow-up audio and the first identification information to the user corresponding to the identity information, and update the input information or subject information according to the supplementary information received from the user's reply until the updated input information or subject information is complete.

[0097] In one embodiment, the response module 5 is further configured to: If there is historical conversation information in the current conversation, obtain the historical topic information corresponding to the historical conversation information; the historical topic information includes historical intent parameters and historical target business scenarios; Determine whether the historical intention parameters and the parameters to be supplemented are contextually consistent; If the historical intent parameters are contextually consistent with the parameters to be supplemented, determine whether the historical intent parameters include the parameters to be supplemented; if the historical intent parameters include the parameters to be supplemented, update the input information or subject information based on the historical intent parameters until the updated input information or subject information is complete; if the historical intent parameters do not include the parameters to be supplemented, and the identity information stores historical preference operation data, update the input information or subject information based on the historical preference operation data until the updated input information or subject information is complete; If the historical intention parameters and the parameters to be supplemented do not have contextual consistency, delete the first identification information corresponding to the historical conversation information; and return to the step of performing intent recognition on the input information in the user information to obtain the intent recognition result.

[0098] In one embodiment, based on the updated complete subject information being processed as input information, the response module 5 is further configured to: Calculate similarity based on the updated topic information and the corpus in the preset knowledge base, and determine whether there is related knowledge based on the similarity. The related knowledge is the text whose similarity to the corpus in the knowledge base is greater than the preset threshold. If there is associated knowledge, an updated first prompt word is generated based on the complete subject information, the associated knowledge, the preset first prompt word and the preset prompt word construction method; If there is no associated knowledge, an updated second prompt word is generated based on the complete subject information, the preset second prompt word and the preset prompt word construction method; Based on the first prompt word or the second prompt word as input information, the large model is used to process and generate and return response information.

[0099] In one embodiment, before generating the updated first prompt word based on the complete subject information, the associated knowledge, the preset first prompt word and the preset prompt word construction method, the response module 5 is further used to: Search based on the updated theme information to see if there are preset plugin keywords that match the theme information; If no preset plug-in keyword matching the subject information is found, an updated first prompt word is generated based on the complete subject information, associated knowledge, the preset first prompt word and the preset prompt word construction method; If a preset plug-in keyword matching the subject information is found, the plug-in corresponding to the preset plug-in keyword is determined as the target plug-in, and the plug-in call information corresponding to the target plug-in is determined; and an updated first prompt word is constructed based on the complete subject information, related knowledge, plug-in call information, the preset first prompt word and the preset prompt word construction method.

[0100] In one embodiment, before generating the updated second prompt word based on the complete subject information, the preset second prompt word and the preset prompt word construction method, the response module 5 is further used to: Based on the updated theme information, check whether there are preset plug-in keywords that match the theme information.

[0101] If no preset plug-in keyword matching the theme information is found, an updated second prompt word is generated based on the complete theme information, the preset second prompt word and the preset prompt word construction method; If a preset plug-in keyword matching the theme information is found, the plug-in corresponding to the preset plug-in keyword is determined as the target plug-in, and the plug-in call information corresponding to the target plug-in is determined; and an updated second prompt word is generated based on the complete theme information, the plug-in call information, the preset second prompt word and the preset prompt word construction method.

[0102] In one embodiment, referring to Figure 3 The device of the artificial intelligence agent based on the large model also includes: an alarm confirmation module 6, the alarm confirmation module 6 is connected to the intention recognition module 1; the alarm confirmation module 6 is used to: If external information is received from an external device, the device type of the external device is determined based on the device identifier in the external information; the device types include risk warning devices and other devices; risk warning devices include life and health monitoring devices and intelligent security monitoring devices; the external information includes the source of the information.

[0103] If the device type is a life and health monitoring device, save the external information; based on the pre-set abnormal confirmation voice generation method, generate and return the abnormal confirmation voice.

[0104] If the device type is an intelligent security monitoring device, save external information.

[0105] If the device type is other device, a prompt voice is generated and returned based on the information source and the preset prompt voice generation method.

[0106] In one embodiment, the alarm confirmation module 6 is further used to: If an abnormal reply message from the user is received, it is identified whether the abnormal reply message includes a preset help keyword.

[0107] If the abnormal reply information includes preset help keywords, the preset alarm information is sent to the contact based on the preset contact information.

[0108] An embodiment of the present invention provides a device of an artificial intelligence agent based on a large model, which can perform intent recognition and dynamic scene matching on information input by a user, and has high decision-making efficiency and intelligence level in complex scenarios.

[0109] The computer program product provided in the embodiment of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the method described in the previous method embodiment. The specific implementation can be found in the method embodiment and will not be repeated here.

[0110] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system and device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0111] In addition, in the description of the embodiments of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0112] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0113] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0114] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. An artificial intelligence agent method based on a large model, characterized in that: include: In response to user information sent by the user, performing intent recognition on input information in the user information to obtain an intent recognition result; The intention recognition result is used to characterize the type of intention of the input information; Determine the target business scenario type based on the intention recognition result, the target business scenario type includes at least the first business scenario or the second business scenario, the first business scenario is used to characterize the call of a third-party tool API interface or an API interface of a device management platform that is connected to a physical device for communication to complete a predefined business, and the second business scenario is used to characterize a conversation task of knowledge question and answer or small talk; The first type of business scenario includes a calling interface of multiple sub-business scenarios, and each sub-business scenario is preset with a trigger keyword; the second type of business scenario includes at least a calling interface of a large model; If the target business scenario type is the first type of business scenario, matching the trigger keyword corresponding to the sub-business scenario based on the input information to determine the interface corresponding to the target business scenario; If the target business scenario type is the second type of business scenario, then performing intent recognition based on the input information to obtain subject information, wherein the subject information is used to characterize the professional field and keywords to which the input information belongs; and determining an interface corresponding to the target business scenario based on the subject information; Call the interface of the target business scenario to process information and return response information.

2. The large model-based artificial intelligence agent method according to claim 1, characterized in that: The calling of the interface of the target business scenario to process information and return response information includes: Based on the interface of the target business scenario, obtain all preset intent parameters corresponding to the target business scenario; Determining whether the input information or the subject information includes all the intention parameters corresponding to the target business scenario; If the input information or the subject information does not include all the intent parameters, it is determined that the input information or the subject information is incomplete, then the input information or the subject information is updated through multiple rounds of dialogue until the updated input information or the subject information is complete; The updated complete input information or the subject information is processed as input information to obtain response information.

3. The large model-based artificial intelligence agent method according to claim 2, characterized in that: The user information also includes identity information; The step of updating the input information or the subject information through multiple rounds of dialogue until the updated input information or the subject information is complete includes: Based on the identity information and preset first identification information, determining whether there is corresponding historical conversation information for the current conversation; the first identification is multi-round conversation identification information; If the historical conversation information does not exist in the current conversation, generating follow-up text information based on the missing required parameters to be supplemented in the input information or the subject information and a preset follow-up question template, and converting the follow-up text information into follow-up audio; Add the first identification information to the current conversation, send the follow-up audio and the first identification information to the user corresponding to the identity information, and update the input information or the subject information according to the supplementary information received from the user's reply until the updated input information or the subject information is complete.

4. The large model-based artificial intelligence agent method according to claim 3 is characterized in that: After the step of determining whether the current conversation has corresponding historical conversation information based on the identity information and the preset first identification information, the method further includes: If the current conversation has the historical conversation information, obtaining the historical topic information corresponding to the historical conversation information; the historical topic information includes historical intention parameters and historical target business scenarios; Determining whether the historical intention parameter and the required parameter to be supplemented have contextual consistency; If the historical intention parameter has contextual consistency with the required parameter to be supplemented, determine whether the historical intention parameter includes the required parameter to be supplemented; if the historical intention parameter includes the required parameter to be supplemented, update the input information or the subject information based on the historical intention parameter until the updated input information or the subject information is complete; if the historical intention parameter does not include the required parameter to be supplemented, and historical preference operation data is stored in the identity information, update the input information or the subject information based on the historical preference operation data until the updated input information or the subject information is complete; If the historical intention parameter and the required parameter to be supplemented do not have contextual consistency, delete the first identification information corresponding to the historical conversation information; and return to the step of performing intent recognition on the input information in the user information to obtain an intent recognition result.

5. The large model-based artificial intelligence agent method according to claim 2, characterized in that: The step of processing the updated complete subject information as input information to obtain response information includes: Calculate similarity based on the updated topic information and the corpus in the preset knowledge base, and determine whether there is associated knowledge based on the similarity, wherein the associated knowledge is text corresponding to the corpus in the knowledge base and the similarity is greater than a preset threshold; If the associated knowledge exists, generating an updated first prompt word based on the complete subject information, the associated knowledge, the preset first prompt word and the preset prompt word construction method; If the associated knowledge does not exist, generating an updated second prompt word based on the complete subject information, the preset second prompt word and the preset prompt word construction method; Based on the first prompt word or the second prompt word as input information, a large model is used to process and generate and return response information.

6. The large model-based artificial intelligence agent method according to claim 5, characterized in that: Before generating the updated first prompt word based on the complete subject information, the associated knowledge, the preset first prompt word and the preset prompt word construction method, the following further comprises: Searching according to the updated theme information to see whether there is a preset plug-in keyword matching the theme information; If the preset plug-in keyword matching the subject information is not found, then generating an updated first prompt word based on the complete subject information, the associated knowledge, the preset first prompt word and the preset prompt word construction method; If the preset plug-in keyword matching the subject information is found, the plug-in corresponding to the preset plug-in keyword is determined to be the target plug-in, and the plug-in calling information corresponding to the target plug-in is determined; and the updated first prompt word is constructed based on the complete subject information, the associated knowledge, the plug-in calling information, the preset first prompt word and the preset prompt word construction method.

7. The large model-based artificial intelligence agent method according to claim 5, characterized in that: The method further includes: generating an updated second prompt word based on the complete subject information, the preset second prompt word and the preset prompt word construction method; Searching according to the updated theme information to see whether there is a preset plug-in keyword matching the theme information; If the preset plug-in keyword matching the subject information is not found, generating an updated second prompt word based on the complete subject information, the preset second prompt word and the preset prompt word construction method; If the preset plug-in keyword matching the subject information is found, the plug-in corresponding to the preset plug-in keyword is determined to be the target plug-in, and the plug-in calling information corresponding to the target plug-in is determined; and the updated second prompt word is generated based on the complete subject information, the plug-in calling information, the preset second prompt word and the preset prompt word construction method.

8. The large model-based artificial intelligence agent method according to claim 1, characterized in that: The method further comprises: If external information sent by an external device is received, the device type of the external device is determined based on the device identifier in the external information; the device type includes risk warning devices and other devices; the risk warning devices include life and health monitoring devices and intelligent security monitoring devices; the external information includes the source of information; If the device type is the life and health monitoring device, save the external information; based on the preset abnormal confirmation voice generation method, generate and return the abnormal confirmation voice; If the device type is the intelligent security monitoring device, save the external information; If the device type is the other device, a prompt voice is generated and returned based on the information source and a preset prompt voice generation method.

9. The large model-based artificial intelligence agent method according to claim 8, characterized in that: After the step of generating and returning the abnormal confirmation voice, the method further includes: If an abnormal reply message from the user is received, identifying whether the abnormal reply message includes a preset help keyword; If the abnormal reply information includes the preset help keyword, the preset alarm information is sent to the contact person based on the preset contact information.

10. A device for an artificial intelligence agent based on a large model, characterized in that: include: An intention recognition module, configured to perform intention recognition on input information in the user information in response to the user information sent by the user, and obtain an intention recognition result; The intention recognition result is used to characterize the type of intention of the input information; A business scenario type determination module, used to determine a target business scenario type based on the intention recognition result, wherein the target business scenario type includes at least a first type of business scenario or a second type of business scenario, wherein the first type of business scenario is used to characterize the calling of a third-party tool API interface or an API interface of a device management platform that is connected to a physical device for communication to complete a predefined business, and the second type of business scenario is used to characterize a knowledge question and answer or small talk dialogue task; The first type of business scenario includes a calling interface of multiple sub-business scenarios, and each sub-business scenario is preset with a trigger keyword; the second type of business scenario includes at least a calling interface of a large model; A first processing module, configured to, if the target business scenario type is the first type of business scenario, determine an interface corresponding to the target business scenario by matching the trigger keyword based on the input information and the sub-business scenario; A second processing module is used for, if the target business scenario type is the second type of business scenario, performing intent recognition based on the input information to obtain subject information, wherein the subject information is used to characterize the professional field and keywords to which the input information belongs; and determining an interface corresponding to the target business scenario based on the subject information; The response module is used to call the interface of the target business scenario to process information and return response information.

Images