Multi-model calling method, device and equipment based on intention judgment

Abstract

The application discloses a multi-model calling method and device based on intention judgment and equipment, wherein the method comprises the following steps: constructing a user portrait based on original data of each business system, constructing an initial model portrait for each initial model, and constructing a mapping relationship between the user portrait label and the initial model portrait; obtaining request data input by a user, analyzing the request data by using a classification model to determine request characteristics, and determining a user intention according to the request data and the user portrait label; screening a target initial model for processing the request data according to the user intention, the initial model portrait and the mapping relationship; searching for a context of the user intention by using a RAG system; and inputting the user intention and the context of the user intention into the target initial model to obtain reply data. The application can accurately determine the intention of a user when the user interacts with a large model, and accurately call a corresponding initial model according to the intention, so that the generated result is more accurate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a multi-model calling method and device based on intention judgment and equipment. BACKGROUND

[0002] In specific application scenarios, a large model LLM with corresponding solution capabilities is often used by fine-tuning. However, due to the need for constant switching when reading user usage, it is very inconvenient. The existing solutions are almost all around a single large model (LLM), whether it is multi-intention recognition, domain adaptation, dialogue context reasoning or keyword matching, all of which are implemented under a single model framework. There is a lack of multi-model dynamic cooperation mechanism, and different models with different expertise cannot be flexibly called to improve the overall answer quality and efficiency according to user intentions. And the dynamic management of user portraits or model portraits in the existing solutions is limited, mainly through static keywords, labels, and domain fine-tuning to improve the accuracy of intention recognition. There is no continuous reinforcement and iterative update process for user portraits and model portraits, that is, user and model portrait labels cannot be updated according to user usage habits and feedback data, and cannot meet the personalized and variable needs of users.

[0003] In summary, the current single model cannot meet the needs of high-quality answers according to the variable needs of users. SUMMARY

[0004] The present disclosure provides a multi-model calling method and device based on intention judgment, equipment and storage medium, to at least solve the above technical request data existing in the prior art.

[0005] According to a first aspect of the present application, a multi-model calling method based on intention judgment is provided, the method comprising:

[0006] Based on the original data of each business system, a user portrait is constructed, and an initial model portrait is constructed for each initial model, and a mapping relationship between each user portrait label and each initial model portrait is determined; wherein the user portrait label represents a business identity label, and the initial model portrait represents a business label;

[0007] Obtain the request data input by the user, analyze the request data using a classification model to determine the request characteristics, and determine the user intention according to the request data and the user portrait label;

[0008] According to the user intention, the initial model portrait and the mapping relationship, a target initial model for processing the request data is selected;

[0009] The RAG system is used to retrieve the context of the user intention;

[0010] input the user intention and the context of the user intention into the target initial model to obtain reply data.

[0011] In an implementation, the original data of each business system is used to construct a user portrait, including:

[0012] Collecting original data of each business system;

[0013] At least one of data cleaning processing, data integration processing, data specification processing, and data transformation processing is performed on the original data to obtain processed original data;

[0014] The user portrait is constructed based on the processed original data.

[0015] In an implementation, the original data includes business problem data; and the initial model portrait is constructed for each initial model, including:

[0016] The answer data of the initial model is obtained based on the business problem data;

[0017] The portrait label of the initial model is determined according to the answer data of the initial model corresponding to each business system;

[0018] The initial model portrait is constructed based on the portrait label.

[0019] In an implementation, after obtaining the reply data, the method further includes:

[0020] Obtaining user feedback on the reply data;

[0021] Extracting data features in the user feedback;

[0022] Determining a mapping relationship between each user portrait label and each initial model portrait based on the data features.

[0023] In an implementation, the method further includes:

[0024] Enhancing the user portrait label and the initial model portrait, specifically including:

[0025] Extracting intention features from each round of request data of the user and the large model;

[0026] Updating the user portrait based on the user feedback and the intention features;

[0027] Adjusting label weights of a corresponding business field of the initial model portrait based on the user feedback to update the initial model portrait.

[0028] In an implementable manner, the analyzing the request data by using the classification model to determine the request feature, and determining the user intention according to the request data and the user portrait label comprises:

[0029] The analyzing the request data by using the classification model determines the problem feature of the request data; the problem feature comprises a business field and a task type label;

[0030] The user portrait and the problem feature are synthesized to obtain the user intention.

[0031] In an implementable manner, the screening the target initial model used for processing the request data according to the user intention, the initial model portrait and the mapping relationship comprises:

[0032] The initial model portrait is screened for the first time according to the user intention, to obtain the remaining model after screening;

[0033] The similarity between the user intention and the model portrait of the remaining model is calculated, the remaining model is sorted according to the similarity, and a preset number of recommended models in the front of the sorting are obtained;

[0034] The preset number of recommended models are scored and sorted according to the mapping relationship, and the target initial model is determined according to the sorting structure.

[0035] In an implementable manner, the method further comprises:

[0036] Analyzing the user feedback of each round of user and large model interaction, the selection of the target initial model and the reply data, to enhance the user portrait label and the initial model portrait.

[0037] According to a second aspect of the present application, a multi-model calling device based on intention judgment is provided, the device comprises:

[0038] The portrait construction module is used for constructing a user portrait based on the original data of each business system, constructing an initial model portrait for each initial model, and determining the mapping relationship between each user portrait label and each initial model portrait; wherein the user portrait label represents a business identity label, and the initial model portrait represents a business skill label;

[0039] The intention determination module is used for obtaining the request data input by the user, analyzing the request data by using the classification model to determine the request feature, and determining the user intention according to the request data and the user portrait label;

[0040] The screening model module is used for screening the target initial model used for processing the request data according to the user intention, the initial model portrait and the mapping relationship;

[0041] a data generation module configured to search for a context of the user intention by using a RAG system;

[0042] a reply output module configured to input the user intention and the context of the user intention into the target initial model to obtain reply data.

[0043] According to a third aspect of the present application, an electronic device is provided, comprising:

[0044] at least one processor; and

[0045] a memory in communication with the at least one processor; wherein

[0046] the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described in the present application.

[0047] According to a fourth aspect of the present application, a non-transitory computer readable storage medium storing computer instructions is provided, the computer instructions being used to enable the computer to perform the method described in the present application.

[0048] According to a fifth aspect of the present application, a computer program product is provided, comprising computer programs or instructions, which, when executed by a processor, implement the method described in the present application.

[0049] By using the technical solution of the present application, the user intention is analyzed, the most suitable model is dynamically called according to the strengths and field capabilities of different initial models, the performance bottleneck of a single model in a variety of demand scenarios is broken through, and the accuracy and flexibility of the answer are improved.

[0050] It should be understood that the data described in this part are not intended to identify key or important features of the embodiments of the present application, nor are they used to limit the scope of the present application. Other features of the present application will become apparent through the following description. BRIEF DESCRIPTION OF DRAWINGS

[0051] The above and other objects, features and advantages of the exemplary embodiments of the present application will be more apparent from the following detailed description taken in conjunction with the accompanying drawings, in which:

[0052] In the drawings, the same or corresponding reference numerals indicate the same or corresponding parts.

[0053] Figure 1 An implementation flowchart of a multi-model calling method based on intention judgment in an embodiment of the present application is shown;

[0054] Figure 2A structural block diagram of a multi-model calling device based on intention judgment in the embodiment of the application is shown.

[0055] Figure 3 A component structure schematic diagram of an electronic device in the embodiment of the application is shown. DETAILED DESCRIPTION

[0056] To make the objectives, characteristics and advantages of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0057] To make the objectives, technical solutions and advantages of the present application more obvious, the present application will be described in further detail below with reference to the drawings. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0058] Although the existing solutions focus on improving the accuracy and domain adaptability of intention recognition, they mostly focus on offline training, fine-tuning and template design, and do not have a clear online closed-loop feedback mechanism. Even if multi-round interaction is mentioned, it is mostly done in a fixed step, and there is no systematic positive and negative feedback driven dynamic self-correction mechanism. Without automatic processing of user negative feedback (punishment mechanism), it is impossible to convert the user's negative evaluation into dynamic updates of the user portrait and model portrait, resulting in difficulty in adapting to user demand evolution in long-term use.

[0059] The existing technology still mainly focuses on multi-intent classification and answer generation through the same model for multi-intent, and lacks sufficient support for complex business scenarios involving different tasks and requiring different specialized models. Without introducing a multi-model orchestrator (scheduler) or workflow mechanism, there is a lack of flexible scheduling and multi-module collaborative processing capability when facing diversified and complex user needs.

[0060] In addition, although some existing solutions mention extracting information from context and knowledge base, the main methods are still limited to simple preprocessing, templates or data cleaning, which are not systematic RAG frameworks. Retrieval augmented generation (RAG) is not fully utilized to improve the accuracy and robustness of answers.

[0061] This application provides an implementation method for a multi-model invocation method based on intent judgment, which can realize a unified interactive page. When the user interacts, the system judges the user's intent and invokes different models according to the intent to achieve the best user experience.

[0062] The following description, in conjunction with the accompanying drawings, introduces a multi-model invocation method, apparatus, and device based on intent determination provided in this application.

[0063] like Figure 1 As shown, this application provides a multi-model invocation method based on intent determination, the method comprising:

[0064] S101, construct user profiles based on the original data of each business system, and construct initial model profiles for each initial model, and determine the mapping relationship between each user profile tag and each initial model profile; wherein, the user profile tag represents a business identity tag, and the initial model profile represents a business expertise tag.

[0065] The business systems provided in this application can include those from various industries. Taking the financial industry as an example, financial industry business systems include: transaction systems, risk management systems, settlement systems, customer management systems, etc. These systems collect raw data from the financial industry through corresponding business data collection platforms and can save the raw data to a database. Similarly, legal industry business systems include case management systems, trial document production systems, case database management systems, electronic case file systems, and human resource management systems. Likewise, these systems collect raw data from the legal industry through corresponding data collection platforms.

[0066] Users are assigned business identity tags based on raw data. These tags can include departments, functions, and common business scenarios. Business identity tags can be obtained from user registration information, industry backend databases, or the company's internal permission system.

[0067] When constructing initial model profiles for each initial model, specifically, a small amount of business problem data is used to test the response quality of each initial model. Response quality includes the accuracy, professionalism, style appropriateness, and response speed of the model's responses. Initial models for specific industries can be obtained through business problems from various industries. Due to the industry-specific nature of these business problems, initial model profiles are derived. In other words, the initial model profiles possess business tags indicating their expertise in those industries.

[0068] It can be understood that by corresponding the business identity tag of the user with the initial model good at the business tag, the initial model good at processing the corresponding business of the user can be determined, or the initial model associated with the user request data.

[0069] S102, obtaining the request data input by the user, analyzing the request data by using the classification model to determine the request feature, and determining the user intention according to the request data and the user portrait tag;

[0070] It can be understood that in addition to the various initial models good at various industries in the present application, a classification model capable of determining the user intention is also included. The classification model can be a large language model (LLM).

[0071] Specifically, after obtaining the request data input by the user, it can be understood that the type of the request data, the industry to which it belongs, etc. cannot be determined at this time, so the classification model needs to be used to analyze the request data first, so as to determine the request feature of the request data, wherein the request feature includes the corresponding business field and task type tag. Then the corresponding business field and task type tag are matched with the user portrait tag to determine the specific user intention.

[0072] S103, according to the user intention, the initial model portrait, and the mapping relationship, screening a target initial model for processing the request data;

[0073] After determining the user intention, the business field, the task type, the initial model portrait, and the mapping relationship are determined, the target initial model good at processing the business field and the task corresponding to the user intention can be found, so as to determine the target initial model good at processing the business field and the task, so as to process the request data of the user by using the target initial model, and more accurate reply data can be obtained.

[0074] S104, retrieving the context of the user intention by using the RAG system;

[0075] It should be noted that the RAG system (retrieval augmented generation system) is a technology combining a large language model (LLM) and an external knowledge source, which aims to enhance the ability of the LLM to generate coherent and context-related responses. The RAG system provided in the present application can be a Faiss vector retrieval system. The Faiss vector retrieval system can retrieve the context related to the request data in the database to obtain the retrieval result.

[0076] In the present application, before inputting the request data into the target initial model, the context of the request data is retrieved by using the RAG system first to generate the context of the user intention.

[0077] S105, input the user intention and the context of the user intention into the target initial model to obtain reply data.

[0078] After obtaining the related context and request data, the retrieved context and request data are then input into the target initial model to obtain reply data.

[0079] The method for calling multiple models based on intention judgment provided in the application first constructs a user portrait, an initial model portrait and a mapping relationship between the user portrait and the initial model portrait of each business system, then analyzes user request data through a classification model to determine a user intention, and then filters out a target initial model for processing the request data according to the user intention, the initial model portrait and the mapping relationship. Before inputting the target initial model, the user intention is searched by using a RAG system to obtain a context of the user intention, and then the request data, the user intention and the context of the user intention are input into the target initial model to obtain reply data which is displayed to the user.

[0080] The method for calling multiple models based on intention judgment provided in the application dynamically calls the most suitable model according to the strengths and field capabilities of different models by analyzing the user intention, breaks through the bottleneck of performance limitation of a single model in a diversified demand scenario, and improves the accuracy and flexibility of the answer.

[0081] In some embodiments, the user portrait is constructed based on the raw data of each business system, including:

[0082] Raw data of each business system is collected;

[0083] At least one of data cleaning processing, data integration processing, data specification processing and data transformation processing is performed on the raw data to obtain processed raw data;

[0084] The user portrait is constructed based on the processed raw data.

[0085] In the application, after collecting the raw data of each business system, at least one of cleaning, integration, specification and data transformation processing is performed on the raw data, so as to remove redundant data in the raw data and unify the data, thereby facilitating the construction of the user portrait.

[0086] In some embodiments, the raw data includes business problem data; and the initial model portrait is constructed for each initial model, including:

[0087] Reply data of the initial model is obtained based on the business problem data;

[0088] The portrait label of the initial model is determined according to the reply data of the initial model of each business system;

[0089] construct an initial model portrait based on the portrait label.

[0090] In the present application, the original data includes business problem data, and the initial model uses a small set of business problem data to evaluate the answer quality of each model during initialization, such as accuracy, professionalism, style appropriateness, response speed, etc. The performance of each initial model in different business fields and task types is recorded as a model portrait label, such as "good at financial compliance" and "good at technical document summary", so as to obtain an initial model portrait. It can be understood that these labels are stored in a vectorized manner, which can facilitate subsequent matching with user intent features.

[0091] In some embodiments, after obtaining the reply data, the following steps are further included:

[0092] Obtaining user feedback on the reply data;

[0093] Extracting data features in the user feedback;

[0094] Determining the mapping relationship between each user portrait label and each initial model portrait based on the data features.

[0095] Specifically, after the target initial model gives the reply data, the user can give feedback on the reply, which can be positive feedback (such as good reviews) or negative feedback (such as bad reviews), and then input the feedback text into the large language model for semantic understanding. For example, it is negative feedback, then the large language model extracts the core problem points of negative feedback, such as inaccurate answer, redundant content, inconsistent style, domain error, incomplete data, etc. Then map these problem points to structured information, such as: {problem type: poor accuracy, domain: financial compliance, model: model B}. Then get the analysis result, and use the analysis result to adjust the user portrait or the initial model portrait. In addition, the mapping relationship between each user portrait label and each initial model portrait can also be determined based on the data features.

[0096] For example, if the feedback indicates that the answer is too long and does not meet the user's expectations, the weight of the "prefer concise answer" label in the user portrait is increased, and the adaptation score of the model responsible for the answer in this preference dimension is reduced. If the feedback indicates that model B's answer in the financial compliance field is not satisfactory, the ability score of model B in this field is downgraded, and the "historical bad feedback in this field" label is marked to reduce the priority of model B in this field in subsequent recall. After repeated negative feedback, the corresponding label in the initial model portrait will be cumulatively reduced until the model is basically not recommended or needs to be significantly improved in the relevant field before its weight is restored; the user portrait reflects the aversion to such answer characteristics, thereby continuously optimizing the next round of recommendation strategy.

[0097] For the mapping relationship, for example, if the user portrait shows that the user likes concise and refined answers and is satisfied with such answers to model A in the past, then model A is scored higher in the ranking and is more likely to be selected as the target initial model.

[0098] In some embodiments, the method for calling multiple models based on intent judgment provided by the present application further comprises:

[0099] Enhancing the user portrait label and the initial model portrait, specifically including:

[0100] Extracting intent features from each round of request data of the user and the large model;

[0101] Updating the user portrait based on user feedback and intent features;

[0102] Adjusting the label weight of the corresponding business field of the initial model portrait based on user feedback to update the initial model portrait.

[0103] Specifically, with multiple rounds of user interaction, the system will extract the "final intent features" after each user question. The final intent features include domain classification, such as financial analysis, legal consultation, etc.; task type, such as simple question and answer, complex analysis, creative generation; style requirements, such as concise, detailed, and tabular information.

[0104] Then, according to the user feedback as described above, the user feedback includes implicit feedback such as usage frequency, click behavior and explicit feedback such as user rating, the user portrait is continuously refined, specifically as follows:

[0105] If the user frequently asks questions related to a specific field, the weight of the label of this field in the user portrait is increased;

[0106] If the user prefers concise answers multiple times (such as giving low ratings to lengthy answers), the "preference for short answers" label in the user portrait is strengthened;

[0107] If the user is satisfied with the output of a specific model, the user portrait records such model features, indicating that the user tends to use models with such answer style and domain capabilities.

[0108] The present application can also dynamically supplement and update the initial model portrait according to user feedback, specifically as follows:

[0109] If a certain initial model performs beyond expectations in subsequent use, for example, continuously receives good ratings in "risk assessment" tasks, the label weight of this model in this field is increased;

[0110] If an initial model encounters multiple negative feedback in a certain field, the corresponding field capability score of the model is reduced, and a "low credibility" or "cautionary call" label is marked in the model portrait.

[0111] The present application reduces the weight or fine-tunes the relevant model or portrait label by performing semantic understanding and structural analysis on the user's negative feedback, so that the system actively avoids bad solutions in the next similar problem and continuously improves the interaction quality.

[0112] In some embodiments, the use of a classification model to analyze the request data to determine the request features, and to determine the user's intention based on the request data and the user's portrait label, includes:

[0113] Using a classification model to analyze the request data to determine the problem features of the request data; the problem features include business domain and task type labels;

[0114] The user portrait and the problem features are synthesized to obtain the user's intention.

[0115] Specifically, the intention recognition of the user's new input in the present application can be a classification model that maps the user's request to the corresponding business domain and task type label, and then synthesizes the intention features based on the user's portrait dynamic vector and the current problem features to obtain a clear and structured final intention representation, such as "user: financial analyst + preference for concise answers, question: financial compliance report summary".

[0116] In some embodiments, the filtering of the target initial model for processing the request data based on the user's intention, the initial model portrait, and the mapping relationship includes:

[0117] The initial model portrait is first filtered based on the user's intention to obtain the remaining models after filtering;

[0118] The similarity between the user's intention and the model portrait of the remaining models is calculated, and the remaining models are sorted based on the similarity to obtain a preset number of recommended models in the front of the sorting;

[0119] The preset number of recommended models are scored and sorted based on the mapping relationship, and the target initial model is determined based on the sorting structure.

[0120] Specifically, the present application introduces real-time recommendation algorithm technology to recall and sort the models. Specifically, it includes initial screening and secondary screening, where the primary screening is coarse recall, and the secondary screening can be achieved through similarity matching, specifically as follows,

[0121] Primary screening: According to the user intention feature, quickly filter out obviously unsuitable models in the model image library, for example, by filtering out models with too large domain gap, no relevant ability label, etc.

[0122] Secondary screening: Use Embedding similarity calculation on the remaining initial models, compare the user intention vector with the model ability vector, and select the Top-N models closest to the current intention.

[0123] Then use real-time recommendation algorithm optimization for fine sorting: Introduce real-time recommendation strategy in Top-N model candidates: for example, use RL-based or RNN / Transformer for Recommender-based methods to consider the user's current session context, historical use preferences, and latest feedback, and perform real-time scoring and ranking on the initial models.

[0124] Among them, if the user portrait shows that the user likes short and concise answers and is satisfied with the past answers of model A, then model A will score higher in the ranking and be more likely to be selected as the target initial model.

[0125] The present application uses real-time recommendation to continuously optimize and iteratively select the initial model for online evaluation and ranking, quickly adapts to the latest user demand preferences and business context at each user query, and realizes instant and agile model selection decision-making. Thus, instant and flexible decision optimization is realized in the user interaction process, making the system quickly adapt to user preference changes and improving the accuracy and satisfaction of answers.

[0126] In some embodiments, it also includes:

[0127] Analyzing user feedback, target initial model selection and reply data of each round of user and large model interaction to enhance the user portrait label and the initial model image.

[0128] It can be understood that the multi-model calling method based on intention judgment provided by the present application can continuously optimize and iteratively select the user portrait and the initial model image. The user feedback, model selection and final answer performance of each round of interaction are recorded and analyzed for subsequent continuous optimization. For example, the user portrait and model image are summarized and cleaned regularly to ensure the timeliness and accuracy of the labels. Or it can also accumulate over time, and through meta-analysis, such as aggregating multiple rounds of feedback data to let the system automatically discover potential user demand shifts (such as users shifting from financial compliance to market analysis) or model quality trends (such as a model improving in a certain field after version update), and then automatically adjust the recommendation strategy and priority.

[0129] The user preference and model capability label are updated according to the user question and feedback after each interaction of the application, the user portrait and model portrait are continuously evolved with use, and the user demand and model specialty are more accurately matched in subsequent questions. The adaptive matching between the user demand and the model capability is realized, and the personalized and variable demand is met.

[0130] As shown above, the multi-model calling method based on intention judgment provided by the application is no longer limited to a single vertical field, and can be easily expanded to various business scenarios through the introduction of multi-model collaboration and dynamic portrait updating, and remains stable and efficient in a complex and variable business demand environment.

[0131] As shown in Figure 2 The application further provides a multi-model calling device based on intention judgment, and the device comprises:

[0132] The portrait construction module 201 is configured to construct a user portrait based on original data of each business system, construct an initial model portrait for each initial model, and determine a mapping relationship between each user portrait label and each initial model portrait, wherein the user portrait label represents a business identity label, and the initial model portrait represents a business skill label;

[0133] The intention determination module 202 is configured to obtain request data input by a user, analyze the request data to determine a request feature by using a classification model, and determine a user intention according to the request data and the user portrait label;

[0134] The screening model module 203 is configured to screen a target initial model for processing the request data according to the user intention, the initial model portrait, and the mapping relationship;

[0135] The data generation module 204 is configured to search for a context of the user intention by using a RAG system;

[0136] The reply output module 205 is configured to input the user intention and the context of the user intention into the target initial model to obtain reply data.

[0137] The application provides a multi-model calling device based on intention judgment. The portrait construction module 201 constructs user portraits based on original data of each business system, constructs initial model portraits for each initial model, and determines a mapping relationship between each user portrait label and each initial model portrait. The user portrait label represents a business identity label, and the initial model portrait represents a business label that is good at. The intention determination module 202 obtains request data input by a user, determines request characteristics by using a classification model to analyze the request data, and determines a user intention according to the request data and the user portrait label. The screening model module 203 screens a target initial model for processing the request data according to the user intention, the initial model portrait, and the mapping relationship. The data generation module 204 generates a context of the user intention by using a RAG system for retrieval. The reply output module 205 inputs the user intention and the context of the user intention into the target initial model to obtain reply data.

[0138] According to the embodiments of the application, the application further provides an electronic device and a readable storage medium.

[0139] The electronic device includes at least one processor, and a memory connected with the at least one processor in communication. The memory stores instructions executable by the at least one processor. The instructions are executed by the at least one processor to enable the at least one processor to perform the multi-model calling method based on intention judgment. The computer instructions are used to enable the computer to perform the multi-model calling method based on intention judgment.

[0140] The application further provides a computer program product, including computer programs / instructions, which are executed by a processor to implement the multi-model calling method based on intention judgment.

[0141] Figure 3 A schematic block diagram of an example electronic device 800 that can be used to implement embodiments of the application is shown. The electronic device is intended to represent various forms of digital computers, such as laptops, desktops, tablets, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular telephones, smartphones, wearable devices, and other similar computing devices. The components shown here, their connections and relationships, and their functions, are meant to be examples only, and are not meant to limit implementations of the application described and / or claimed in this document.

[0142] As Figure 3As shown, the device 800 includes a computing unit 801 that can perform various appropriate actions and processes in accordance with a computer program stored in a read-only memory (ROM) 802 or a computer program loaded from a storage unit 808 into a random access memory (RAM) 803. In the RAM 803, various programs and data required for the operation of the device 800 can also be stored. The computing unit 801, the ROM 802, and the RAM 803 are connected to each other through a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0143] A plurality of components in the device 800 are connected to the I / O interface 805, including: an input unit 806, such as a keyboard, a mouse, etc.; an output unit 807, such as various types of displays, speakers, etc.; a storage unit 808, such as a magnetic disk, an optical disk, etc.; and a communication unit 809, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 809 allows the device 800 to exchange information / data with other devices through a computer network, such as the Internet, and / or various telecommunication networks.

[0144] The computing unit 801 can be various general and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 801 performs various methods and processes described above, such as the intent-judgment-based multi-model invocation method. For example, in some embodiments, the intent-judgment-based multi-model invocation method can be implemented as a computer software program that is tangibly embodied in a machine-readable medium, such as the storage unit 808. In some embodiments, part or all of the computer program can be loaded and / or installed on the device 800 via the ROM 802 and / or the communication unit 809. When the computer program is loaded into the RAM 803 and executed by the computing unit 801, one or more steps of the intent-judgment-based multi-model invocation method described above can be performed. Alternatively, in other embodiments, the computing unit 801 can be configured to perform the intent-judgment-based multi-model invocation method by any other appropriate means, such as by means of firmware.

[0145] The various embodiments of the systems and techniques described above can be implemented in digital electronic circuitry, integrated circuitry, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a system on a chip (SOC), a load programmable logic device (CPLD), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0146] Program code for carrying out methods of the present application can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the program code, when executed by the processor or controller, produces a means for implementing the functions / acts specified in the flowcharts and / or block diagrams. The program code can be executed entirely on a machine, partially on a machine, partially on a machine as a stand-alone software package, or entirely on a remote machine or server.

[0147] In the context of the present application, a machine-readable medium can be a tangible medium that contains or stores a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include but is not limited to an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of the machine-readable storage medium will include one or more lines of electrical connections, portable computer disks, hard disk drives, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or Flash memory), optical fibers, portable compact disc read-only memories (CD-ROMs), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0148] To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0149] The systems and techniques described here can be implemented in a computing system that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0150] The computer system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The server can be a cloud server, a server of a distributed system, or a server combined with a blockchain.

[0151] The above description is only a specific implementation of the present application. The protection scope of the present application is not limited to this, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A multi-model calling method based on intention judgment, characterized in that, The method comprises: constructing a user portrait based on original data of each business system, and constructing an initial model portrait for each initial model, and determining a mapping relationship between each user portrait label and each initial model portrait; wherein the user portrait label represents a business identity label, and the initial model portrait represents a business proficiency label; obtaining request data input by a user, analyzing the request data by using a classification model to determine a request feature, and determining a user intent according to the request data and the user portrait label; screening a target initial model for processing the request data according to the user intent, the initial model portrait, and the mapping relationship; retrieving a context of the user intent by using an RAG system; inputting the user intent and the context of the user intent into the target initial model to obtain reply data.

2. The method of claim 1, wherein, The method comprises: collecting original data of each business system; performing at least one of data cleaning processing, data integration processing, data reduction processing, and data transformation processing on the original data to obtain processed original data; constructing a user portrait based on the processed original data.

3. The method of claim 1, wherein, The original data comprises business problem data; and the method comprises: obtaining answer data of an initial model based on the business problem data; determining an image label of the initial model according to the answer data of the initial model of each business system; constructing an initial model portrait based on the image label.

4. The method of claim 1, wherein, After obtaining the reply data, the method further comprises: obtaining user feedback on the reply data; extracting a data feature in the user feedback; determining a mapping relationship between each user portrait label and each initial model portrait based on the data feature.

5. The method of claim 4, wherein, The method further comprises: enhancing the user portrait label and the initial model portrait, specifically comprising: extracting an intent feature from each round of request data of a user and a large model; updating the user portrait based on user feedback and the intent feature; updating the initial model portrait based on user feedback to adjust a label weight of a corresponding business field of the initial model portrait.

6. The method of claim 1, wherein, The method comprises: analyzing the request data by using a classification model to determine a request feature, and determining a user intent according to the request data and the user portrait label; analyzing the request data by using a classification model to determine a problem feature of the request data; the problem feature comprises a business field and a task type label; 7. The method of claim 1, wherein, synthesizing an intent feature from the user portrait and the problem feature to obtain the user intent. The method comprises: performing a first screening on the initial model portrait according to the user intent to obtain a remaining model after screening; calculating a similarity between the user intent and a model portrait of the remaining model, sorting the remaining model according to the similarity to obtain a preset number of recommended models in a front position; 8. The method of claim 4, wherein, scoring and sorting the preset number of recommended models according to the mapping relationship, and determining a target initial model according to a sorting structure. The method further comprises: analyzing user feedback, selection of a target initial model, and reply data of each round of user interaction with a large model to enhance the user portrait label and the initial model portrait.

9. A multi-model invocation device based on intent determination, characterized in that, The device comprises: a portrait construction module configured to construct a user portrait based on raw data of each business system, construct an initial model portrait for each initial model, and determine a mapping relationship between each user portrait label and each initial model portrait; wherein the user portrait label represents a business identity label, and the initial model portrait represents a business proficiency label; an intent determination module configured to obtain request data input by a user, analyze the request data to determine request features using a classification model, and determine a user intent based on the request data and the user portrait label; a screening model module configured to screen a target initial model for processing the request data based on the user intent, the initial model portrait, and the mapping relationship; a data generation module configured to search and generate a context of the user intent using a RAG system; a reply output module configured to input the user intent and the context of the user intent into the target initial model to obtain reply data.

10. An electronic device, comprising: at least one processor; and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 8.

Images