A recommendation agent system, a recommendation method, and a recommendation device

Abstract

The present application discloses a recommendation agent system, a recommendation method, and a recommendation device, relating to the field of agent technology. A perception module acquires first dialogue data and / or multi-modal environment data, a memory module acquires historical human-computer interaction data related to a user intention and a target object to be recommended, a planning module retrieves a reference workflow that matches the user intention, and generates an initial agent workflow according to the data acquired by the perception module, the historical human-computer interaction data, the target object to be recommended, and the reference workflow. An execution module is called to execute the initial agent workflow, and it is determined whether to update the initial agent workflow based on second dialogue data and / or return data generated during the execution process. If so, the initial agent workflow is updated, and the updated agent workflow is executed starting from the self-update position to complete the recommendation. The present application can perform dynamic planning on the recommendation task by combining multiple types of data, improving the planning accuracy and the recommendation accuracy rate.

Description

Technical Field

[0001] This application relates to the field of intelligent agent technology, and in particular, to a recommendation intelligent agent system, a recommendation method, and a recommendation device. Background Art

[0002] Intelligent recommendation systems are an important part of digital marketing operations. They can screen the most suitable products from a vast amount of information according to users' needs and interests, improving user satisfaction and loyalty.

[0003] The process of traditional intelligent recommendation systems for making recommendations is as follows: obtaining historical interaction data between users and front-end applications, and using a recommendation system model based on the historical interaction data to obtain recommendation results. Since historical interaction data can only reflect users' past product preferences, the recommendation results obtained based on historical recommendation interaction data may not be consistent with users' current product preferences, resulting in inaccurate recommendation results. Summary of the Invention

[0004] In view of the above problems, this application provides a recommendation intelligent agent system, a recommendation method, and a recommendation device to achieve the purpose of improving the accuracy of recommendation results. The specific solutions are as follows:

[0005] In a first aspect of this application, a recommendation intelligent agent system is provided, including: a perception module, a memory module, a planning module, and an execution module;

[0006] The perception module is configured to obtain first dialogue data and / or multi-modal environment data collected during human-computer interaction, where the first dialogue data includes at least one round of dialogue data;

[0007] The memory module is configured to obtain historical human-computer interaction data related to user intentions and target objects to be recommended, where the user intentions reflect the recommended objects that users tend to;

[0008] The planning module is configured to retrieve a reference workflow that matches the user intentions, and generate an initial intelligent agent workflow according to the data obtained by the perception module, the historical human-computer interaction data, the target objects to be recommended, and the reference workflow. The reference workflow refers to an orchestration workflow used for successful recommendation of objects in the scenario related to the user intentions, or an artificial orchestration workflow that matches the user intentions and verifies the recommendation effect. The initial intelligent agent workflow uses the reference workflow as a thought chain for inference of the recommendation large model and is a workflow orchestrated by the recommendation large model;

[0009] The planning module is further configured to call the execution module to execute the initial agent workflow, and determine whether to update the initial agent workflow based on the second dialogue data and / or return data generated during the execution process. If so, the initial agent workflow is updated based on the first dialogue data, the second dialogue data, and the reference workflow, and the execution module is called to execute the updated agent workflow starting from the update position to complete the recommendation. The second dialogue data is at least one round of dialogue data based on a step-by-step conversational guidance strategy with the user.

[0010] In a possible implementation, when generating the initial agent workflow according to the data obtained by the perception module, the historical human-computer interaction data, the target object to be recommended, and the reference workflow, the planning module is specifically configured to:

[0011] Fill the data obtained by the perception module, the historical human-computer interaction data, the target object to be recommended, and the reference workflow into the first prompt template of the recommendation large model to obtain the first prompt instruction prompt;

[0012] Input the first prompt instruction prompt into the recommendation large model to obtain the initial agent workflow output by the recommendation large model;

[0013] Among them, the training data used by the recommendation large model in the training stage includes: training multi-modal environment data and / or training dialogue data, training historical human-computer interaction data, training target objects to be recommended, training reference workflows, and corresponding training initial agent workflows.

[0014] In a possible implementation, the training data used by the recommendation large model in the training stage further includes: pre-configured custom workflows and workflow generation rules, where the workflow generation rules are used to instruct the recommendation large model to generate the required workflow following the execution logic and execution order in the custom workflow.

[0015] In a possible implementation, before the memory module obtains the historical human-computer interaction data related to the user intention and the target object to be recommended, the perception module is further configured to:

[0016] Determine the user intention according to the data obtained by the perception module.

[0017] In a possible implementation, when determining the user intention according to the data obtained by the perception module, the perception module is specifically configured to:

[0018] Fill the data obtained by the perception module into the second prompt template of the recommendation large model to obtain a second prompt instruction prompt;

[0019] Input the second prompt instruction prompt into the recommendation large model to obtain the user intention output by the recommendation large model;

[0020] Among them, the training data used by the recommendation large model in the training stage includes: training multi-modal environment data and / or training dialogue data, and corresponding training user intentions.

[0021] In a possible implementation, when the memory module obtains the historical human-computer interaction data related to the user intention and the target object to be recommended, it specifically is used for:

[0022] Obtain the historical human-computer interaction data from the first memory bank according to the user intention;

[0023] Obtain the target object to be recommended from the second memory bank according to the user intention.

[0024] In a possible implementation, the multi-modal environment data includes one or more of image data, video data, and audio data.

[0025] In a possible implementation, the historical human-computer interaction data is a content summary of the historical dialogue data between the user and the recommendation assistant.

[0026] The second aspect of this application provides a recommendation method, which is applied to a recommendation intelligent agent system, including:

[0027] Obtain the first dialogue data and / or multi-modal environment data collected during the human-computer interaction process, where the first dialogue data includes at least one round of dialogue data;

[0028] Obtain the historical human-computer interaction data related to the user intention and the target object to be recommended, where the user intention reflects the recommended object that the user favors;

[0029] Retrieve the reference workflow that matches the user intention, and generate an initial intelligent agent workflow according to the first dialogue data and / or the multi-modal environment data, the historical human-computer interaction data, the target object to be recommended, and the reference workflow. The reference workflow refers to the orchestration workflow used when successfully recommending objects in the scenario related to the user intention, or the manual orchestration workflow that matches the user intention and verifies the recommendation effect. The initial intelligent agent workflow is a workflow planned and arranged by the recommendation large model with the reference workflow as the thinking chain used for the inference of the recommendation large model;

[0030] Execute the initial agent workflow, and determine whether to update the initial agent workflow based on the second dialogue data and / or return data generated during the execution process. If so, update the initial agent workflow based on the first dialogue data, the second dialogue data, and the reference workflow, and start executing the updated agent workflow from the update position to complete the recommendation. The second dialogue data is at least one round of dialogue data based on a step-by-step dialogue guidance strategy with the user.

[0031] The third aspect of this application provides a recommendation device, which is applied to a recommendation agent system and includes:

[0032] A data acquisition unit, configured to acquire first dialogue data and / or multimodal environment data collected during human-computer interaction, where the first dialogue data includes at least one round of dialogue data;

[0033] A memory extraction unit, configured to acquire historical human-computer interaction data related to the user's intention and the target object to be recommended, where the user's intention reflects the recommended object preferred by the user;

[0034] A workflow orchestration unit, configured to retrieve a reference workflow that matches the user's intention, and generate an initial agent workflow according to the first dialogue data and / or the multimodal environment data, the historical human-computer interaction data, the target object to be recommended, and the reference workflow. The reference workflow refers to an orchestration workflow used for successful recommendation of objects in the scenario related to the user's intention, or an artificial orchestration workflow that matches the user's intention and verifies the recommendation effect. The initial agent workflow is a workflow orchestrated by the recommendation large model by using the reference workflow as the thinking chain for the recommendation large model inference;

[0035] A dynamic planning unit, configured to execute the initial agent workflow, and determine whether to update the initial agent workflow based on the second dialogue data and / or return data generated during the execution process. If so, update the initial agent workflow based on the first dialogue data, the second dialogue data, and the reference workflow, and start executing the updated agent workflow from the update position to complete the recommendation. The second dialogue data is at least one round of dialogue data based on a step-by-step dialogue guidance strategy with the user.

[0036] With the above technical solution, for the recommended intelligent agent system provided by this application, the perception module acquires the first conversation data and / or multi-modal environment data collected during the human-computer interaction process, determines the user intention based on the acquired data, the memory module acquires the historical human-computer interaction data related to the user intention and the target object to be recommended, the planning module retrieves the reference workflow that matches the user intention, generates an initial intelligent agent workflow based on the data acquired by the perception module, the historical human-computer interaction data, the target object to be recommended, and the reference workflow, calls the execution module to execute the initial intelligent agent workflow, and determines whether to update the initial intelligent agent workflow based on the second conversation data and / or the returned data generated during the execution process. If so, updates the initial intelligent agent workflow based on the first conversation data, the second conversation data, and the reference workflow, and calls the execution module to execute the updated intelligent agent workflow starting from the updated position to complete the recommendation.

[0037] Since the first conversation data and the multi-modal environment data are the data generated when the user needs a recommendation, the first conversation data and the multi-modal environment data can better reflect the user's current preferences, while the historical human-computer interaction data can reflect the user's past interests. Therefore, when making a recommendation, the recommended intelligent agent system provided by this application simultaneously refers to the first conversation data, the multi-modal environment data, and the historical human-computer interaction data, which can more accurately determine the recommended object preferred by the user and improve the recommendation accuracy.

[0038] The target object to be recommended provides a recommendation direction for the recommendation process. Planning the initial intelligent agent workflow based on the target object to be recommended and then making a product recommendation based on the initial intelligent agent workflow can guide the user to establish an interest in the target object to be recommended during the recommendation process, improve the probability of successful recommendation of the target object to be recommended. At the same time, the existence of the target object to be recommended makes active recommendation possible.

[0039] Furthermore, the historical human-computer interaction data in this application is the long-term memory data of the user. Combining the historical human-computer interaction data with the target object to be recommended enables this application to plan a long-term recommendation strategy and improve the probability of successful recommendation.

[0040] Furthermore, this application can draw on the successful recommendation experience of objects in scenarios related to user intent, and use the reference workflow used when successfully recommending the object as the thought chain for the recommendation large model to reason. This enables the recommendation large model to draw on the successful recommendation thinking mode, improving the accuracy of the automatic planning of the recommendation large model. Based on the more accurate initial agent workflow arranged by the recommendation large model for recommendation, the recommendation accuracy is improved. Even if the choreography workflow used when successfully recommending an object in the scenario related to user intent cannot be retrieved, this application can still use the manual choreography workflow that matches the user intent and has verified the recommendation effect as the reference workflow, enabling the initial agent workflow generated based on the reference workflow to make more effective recommendations and improving the recommendation success rate.

[0041] In addition, regardless of whether the current recommendation task is complex, this application can retrieve the corresponding reference workflow. After learning the recommendation thinking mode of this reference workflow, it generates the initial agent workflow, enabling the initial agent workflow to provide a better recommendation strategy for the current recommendation task, improving the recommendation accuracy and success rate. Compared with the existing recommendation solutions that can only solve simple recommendation tasks, this application can provide an accurate initial agent workflow to solve complex recommendation tasks, with a wider application range.

[0042] Furthermore, during the process of making recommendations based on the initially planned agent workflow provided by this application, the recommendation agent system can adopt a step-by-step conversational guidance strategy to conduct one or more rounds of conversations with the user to obtain second conversation data. The addition of the second conversation data and the return data generated during the execution of the initial agent workflow enables this application to better understand the user intent, and then dynamically update the initial agent workflow based on the understanding of the user intent, making the updated agent workflow more in line with the user intent and improving the recommendation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In combination with the accompanying drawings and with reference to the following specific embodiments, the above and other features, advantages, and aspects of the various embodiments of the present disclosure will become more apparent. Throughout the accompanying drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the original components and elements are not necessarily drawn to scale.

[0044] Figure 1 It is a schematic diagram of a system architecture provided by this application;

[0045] Figure 2 It is a schematic diagram of an optional hardware structure of the terminal 100 provided by this application;

[0046] Figure 3 It is a schematic diagram of the structure of a server 200 provided by this application;

[0047] Figure 4 A structural schematic diagram of a recommendation agent system provided for this application;

[0048] Figure 5 A schematic diagram showing the interaction of each module during dynamic programming based on a recommendation large model provided for this application;

[0049] Figure 6 A flowchart schematic diagram of a recommendation method provided for this application;

[0050] Figure 7 A structural schematic diagram of a recommendation device provided for this application;

[0051] Figure 8 A structural schematic diagram of an electronic device provided for this application. Detailed implementation manners

[0052] The embodiments of this application will be described below with reference to the accompanying drawings in the embodiments of this application. The terms used in the embodiments part of this application are only used to explain the specific embodiments of this application, rather than aiming to limit this application.

[0053] The embodiments of this application will be described below with reference to the accompanying drawings. Those of ordinary skill in the art will know that with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.

[0054] The terms "first", "second", etc. in the specification, claims and above-mentioned drawings of this application are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that such terms can be interchanged under appropriate circumstances, which is only a way of distinguishing when describing objects with the same attributes in the embodiments of this application. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion, so that a process, method, system, product or device including a series of units does not have to be limited to those units, but may include other units not clearly listed or inherent to these process, method, product or device.

[0055] See Figure 1 , Figure 1 shows a schematic diagram of a system architecture. The system may include a terminal 100 and a server 200. Among them, the server 200 may include one or more servers ( Figure 1 illustrated by taking one server as an example), and the server 200 may provide the method provided in the embodiments of this application for one or more terminals.

[0056] Among them, an application corresponding to a third-party system can be installed on the terminal 100. The above application and web page can provide an interface. The terminal 100 can receive relevant parameters input by the user on the interface and send the above parameters to the server 200. The server 200 can obtain a processing result based on the received parameters and return the processing result to the terminal 100.

[0057] It should be understood that in some alternative implementations, the terminal 100 can also complete the action of obtaining the processing result based on the received parameters by itself, without the need for the server to cooperate. The embodiments of the present application do not limit this.

[0058] Next, the product form of the terminal 100 will be described. Figure 1 in the terminal 100;

[0059] The terminal 100 in the embodiments of the present application can be a mobile phone, a tablet computer, a wearable device, a vehicle-mounted device, an augmented reality (AR) / virtual reality (VR) device, a laptop computer, an ultra-mobile personal computer (UMPC), a netbook, a personal digital assistant (PDA), etc. The embodiments of the present application do not impose any restrictions on this.

[0060] Figure 2 FIG. shows an optional schematic hardware structure of the terminal 100.

[0061] Referring to Figure 2 as shown, the terminal 100 may include a radio frequency unit 110, a memory 120, an input unit 130, a display unit 140, a camera 150 (optional), an audio circuit 160 (optional), a speaker 161 (optional), a microphone 162 (optional), a headphone jack 163 (optional), a processor 170, an external interface 180, a power supply 190, and other components. Those skilled in the art can understand that Figure 2 merely examples of the terminal or multifunctional device, and do not constitute a limitation on the terminal or multifunctional device, and may include more or fewer components than shown in the figure, or combine certain components, or different components.

[0062] The input unit 130 can be used to receive input numerical or character information and generate key signal inputs related to the user settings and function control of the portable multifunctional device. Specifically, the input unit 130 can include a touch screen 131 (optional) and / or other input devices 132. The touch screen 131 can collect touch operations of the user thereon or nearby (such as operations of the user using any suitable object such as a finger, a joint, a stylus, etc. on or near the touch screen), and drive corresponding connection devices according to a preset program. The touch screen can detect the touch action of the user on the touch screen, convert the touch action into a touch signal and send it to the processor 170, and can receive and execute commands sent by the processor 170; the touch signal at least includes contact coordinate information. The touch screen 131 can provide an input interface and an output interface between the terminal 100 and the user. In addition, various types such as resistive, capacitive, infrared, and surface acoustic wave can be used to implement the touch screen. In addition to the touch screen 131, the input unit 130 can also include other input devices. Specifically, the other input devices 132 can include, but are not limited to, one or more of a physical keyboard, function keys (such as volume control keys, switch keys, etc.), a trackball, a mouse, a joystick, etc.

[0063] Among them, the input device 132 can receive input data, etc.

[0064] The display unit 140 can be used to display information input by the user or information provided to the user, various menus of the terminal 100, an interactive interface, file display, and / or the playback of any multimedia file.

[0065] The memory 120 can be used to store instructions and data. The memory 120 mainly includes a storage instruction area and a storage data area. The storage data area can store various data, such as multimedia files, texts, etc.; the storage instruction area can store software units such as an operating system, applications, instructions required for at least one function, or their subsets and extended sets. It can also include a non-volatile random access memory; it provides the processor 170 with management of hardware, software, and data resources in the computing processing device, supports control software and applications. It is also used for the storage of multimedia files and the storage of running programs and applications.

[0066] The processor 170 is the control center of the terminal 100, connecting various parts of the entire terminal 100 through various interfaces and lines. By running or executing instructions stored in the memory 120 and invoking data stored in the memory 120, it executes various functions of the terminal 100 and processes data, thereby exercising overall control over the terminal device. Optionally, the processor 170 may include one or more processing units; preferably, the processor 170 may integrate an application processor and a modem processor. Among them, the application processor mainly processes the operating system, user interface, application programs, etc., and the modem processor mainly processes wireless communications. It can be understood that the above-mentioned modem processor may not be integrated into the processor 170 either. In some embodiments, the processor and the memory may be implemented on a single chip, and in some embodiments, they may also be separately implemented on independent chips. The processor 170 can also be used to generate corresponding operation control signals, send them to corresponding components of the computing and processing device, read and process data in the software, especially read and process the data and programs in the memory 120, so that each functional module therein executes corresponding functions, thereby controlling the corresponding components to act according to the requirements of the instructions.

[0067] Among them, the memory 120 can be used to store software codes related to the recommendation method, and the processor 170 can execute the steps of the recommendation method or schedule other units (such as the above-mentioned input unit 130 and display unit 140) to implement corresponding functions.

[0068] The radio frequency unit 110 (optional) can be used to receive and send information or signals during a call. For example, after receiving the downlink information of the base station, it is sent to the processor 170 for processing; in addition, the uplink data is sent to the base station. Generally, the RF circuit includes but is not limited to antennas, at least one amplifier, transceiver, coupler, low noise amplifier (LNA), duplexer, etc. In addition, the radio frequency unit 110 can also communicate with network devices and other devices through wireless communication. This wireless communication can use any communication standard or protocol, including but not limited to Global System of Mobile Communication (GSM), General Packet Radio Service (GPRS), Code Division Multiple Access (CDMA), Wideband Code Division Multiple Access (WCDMA), Long Term Evolution (LTE), email, Short Messaging Service (SMS), etc.

[0069] Among them, in the embodiments of the present application, the radio frequency unit 110 can send data to the server 200 and receive the processing result sent by the server 200.

[0070] It should be understood that the radio frequency unit 110 is optional and can be replaced by other communication interfaces, such as a network interface.

[0071] The terminal 100 further includes a power supply 190 (such as a battery) for powering each component. Preferably, the power supply can be logically connected to the processor 170 through a power management system, so as to implement functions such as management of charging, discharging, and power consumption management through the power management system.

[0072] The terminal 100 further includes an external interface 180, which can be a standard Micro USB interface or a multi-pin connector, and can be used to connect the terminal 100 to other devices for communication, or can be used to connect a charger to charge the terminal 100.

[0073] Although not shown, the terminal 100 may further include a flash, a Wireless Fidelity (WiFi) module, a Bluetooth module, sensors with different functions, etc., which will not be elaborated here. Some or all of the methods described below can be applied to the terminal 100 as Figure 2 shown.

[0074] Next, the product form of the server 200 will be described Figure 1 in;

[0075] Figure 3 A schematic structural diagram of a server 200 is provided, as Figure 3 shown. The server 200 includes a bus 201, a processor 202, a communication interface 203, and a memory 204. The processor 202, the memory 204, and the communication interface 203 communicate with each other through the bus 201.

[0076] The bus 201 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of representation, Figure 3 only a thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0077] The processor 202 can be any one or more of processors such as a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), a Micro Processor (MP), or a Digital Signal Processor (DSP).

[0078] The memory 204 can include volatile memory, such as Random Access Memory (RAM). The memory 204 can also include non-volatile memory, such as Read-Only Memory (ROM), flash memory, a Hard Drive Drive (HDD), or a Solid State Drive (SSD).

[0079] Among them, the memory 204 can be used to store software code related to the recommendation method, and the processor 202 can execute the steps of the recommendation method of the chip or schedule other units to implement corresponding functions.

[0080] It should be understood that the above terminal 100 and server 200 can be centralized or distributed devices, and the processors in the above terminal 100 and server 200 (such as processor 170 and processor 202) can be hardware circuits (such as an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), a general-purpose processor, a DSP, a microprocessor, or a microcontroller, etc.), or a combination of these hardware circuits. For example, the processor can be a hardware system with the function of executing instructions, such as a CPU, a DSP, etc., or a hardware system without the function of executing instructions, such as an ASIC, an FPGA, etc., or a combination of the above hardware system without the function of executing instructions and the hardware system with the function of executing instructions.

[0081] See Figure 4 , which is a schematic structural diagram of a recommendation agent system provided by this application. As Figure 4 shown, the recommendation agent system includes a sensing module 51, a memory module 53, a planning module 52, and an execution module 54.

[0082] Among them, the perception module 51 is used to obtain the first dialogue data and / or multi-modal environment data collected during the human-computer interaction process. The first dialogue data includes at least one round of dialogue data;

[0083] The memory module 53 is used to obtain the historical human-computer interaction data related to the user's intention and the target object to be recommended. The user's intention reflects the recommended object that the user prefers;

[0084] The planning module 52 is used to retrieve the reference workflow that matches the user's intention, and generate an initial agent workflow based on the data obtained by the perception module, the historical human-computer interaction data, the target object to be recommended, and the reference workflow. The reference workflow refers to the orchestration workflow used for successful recommendation of objects in the scenario related to the user's intention, or the manual orchestration workflow that matches the user's intention and verifies the recommendation effect. The initial agent workflow is the thinking chain used by the recommendation large model for reasoning, and is the workflow orchestrated by the recommendation large model;

[0085] The planning module 52 is also used to call the execution module to execute the initial agent workflow, and determine whether to update the initial agent workflow based on the second dialogue data and / or the return data generated during the execution process. If so, update the initial agent workflow based on the first dialogue data, the second dialogue data, and the reference workflow, and call the execution module to execute the updated agent workflow starting from the self-update position to complete the recommendation. The second dialogue data is at least one round of dialogue data based on the step-by-step dialogue guidance strategy with the user.

[0086] The above process will be introduced in detail below.

[0087] A recommendation assistant is set in the recommendation agent system provided by this application. The user can have one or more rounds of conversations with the recommendation assistant, and the dialogue data of the first or multiple rounds of conversations is used as the first dialogue data of this application.

[0088] The perception module 51 in this application can also obtain multi-modal environment data, which is the environment data related to the user's current recommendation task. Optionally, the multi-modal environment data includes one or more of image data, video data, and audio data.

[0089] In the actual application scenario, the multi-modal environment data can be the environment data input by the user or other environment data.

[0090] The environmental data input by the user described above is, for example: data such as images and videos that the user views or clicks on the page browsed during the current recommendation task. The data that the user views or clicks on is the environmental data input by the user. Since the environmental data that the user views and clicks on is often the data that the user is more concerned about and interested in, recommending based on these data helps to improve the recommendation accuracy.

[0091] Other environmental data is, for example: during the process of the user's conversation with the recommendation assistant, data such as images, videos, and audios related to the conversation are automatically generated based on the conversation content.

[0092] Of course, the multi-modal environmental data can also be other, and the present application does not make specific limitations.

[0093] The above-mentioned memory module 53 can obtain relevant historical human-computer interaction data and the target object to be recommended based on the user's intention.

[0094] Optionally, the historical human-computer interaction data can be the historical conversation data between the user and the recommendation assistant, or the content summary of the historical conversation data between the user and the recommendation assistant. Here, the content summary refers to the useful information in the historical conversation data, such as user preferences, user historical events (such as where to travel, which movies have been watched, which awards have been won, etc.).

[0095] The above-mentioned target object to be recommended refers to the target object that the recommendation intelligent system intends to recommend to the user. In the actual application scenario, the recommendation intelligent system provided by the present application may recommend the target object to be recommended to the user, or may not recommend the target object to be recommended to the user. For example, in some recommendation scenarios, the recommendation intelligent system discovers that the user is interested in the target object to be recommended during the interaction with the user through the recommendation assistant, then it may recommend the target object to be recommended to the user; in other recommendation scenarios, the recommendation intelligent system discovers that the user is not interested in the target object to be recommended during the interaction with the user through the recommendation assistant, then it gives up recommending the target object to be recommended to the user.

[0096] Optionally, the above-mentioned target object to be recommended can be pre-set in the recommendation intelligent system.

[0097] Optionally, the above-mentioned target object to be recommended can be the target content to be recommended (such as text, pictures, audios, videos, books, movies, games, online packages, etc.) or the target product to be recommended (such as products in commercial fields such as goods and services). Of course, the target object to be recommended can also be other, and the present application does not make specific limitations.

[0098] Optionally, the process of the memory module 53 obtaining historical human-computer interaction data related to the user's intention and the target object to be recommended may specifically include: obtaining historical human-computer interaction data from the first memory bank according to the user's intention; obtaining the target object to be recommended from the second memory bank according to the user's intention.

[0099] That is, in this embodiment, a first memory bank and a second memory bank can be established in advance. The first memory bank is used to store interaction data when the user interacts with the recommendation assistant, and the second memory bank is used to store pre-configured target objects to be recommended. Then, after determining the user's intention, the embodiment of the present application can obtain historical human-computer interaction data related to the user's intention from the first memory bank and obtain target objects to be recommended related to the user's intention from the second memory bank.

[0100] The above planning module 52 can retrieve a reference workflow that matches the user's intention. Here, the reference workflow refers to an orchestration workflow used for successful recommendation of objects in a scenario related to the user's intention, or a manually orchestrated workflow that matches the user's intention and has verified the recommendation effect. Among them, "verifying the recommendation effect" means performing a recommendation simulation on the manually orchestrated workflow and achieving a successful recommendation during the recommendation simulation. The scenario related to the user's intention refers to a scenario whose similarity to the user's intention is greater than a preset similarity threshold.

[0101] Optionally, the objects in the scenario related to the user's intention may be recommendation content in the scenario related to the user's intention (such as text, pictures, audio, video, books, movies, games, online packages, etc.), or recommended products in the scenario related to the user's intention (such as products in commercial fields such as goods and services). Of course, the objects in the scenario related to the user's intention may also be others, and the present application does not make specific limitations.

[0102] Optionally, the reference workflow includes the orchestration results of tool usage, logical judgment, and model invocation. Of course, the reference workflow may also include other orchestration results, and the present application does not make specific limitations.

[0103] Optionally, the process of the planning module 52 "retrieving a reference workflow that matches the user's intention" may include: generating a query vector based on the user's intention and querying from a pre-generated vector database to obtain a reference workflow that matches the user's intention.

[0104] Specifically, the recommendation agent system will pre-store a series of reference workflows in the vector database. These workflows are encoded by the vectorization model and transformed into vectorized representations to support the efficient retrieval and matching of the recommendation agent system. In this recommendation task, the planning module 52 analyzes the user's intention, transforms it into a query vector, matches the relevant workflows in the vector database, and finds the workflow that best matches the requirements of this recommendation task in the vector database as the reference workflow that matches the retrieved user intention. This retrieval process combines semantic similarity and task requirements to ensure that the retrieved reference workflow meets the user's intention.

[0105] Furthermore, the planning module 52 can also call the recommendation large model to generate an initial agent workflow based on the data obtained by the perception module, historical human-computer interaction data, the target object to be recommended, and the reference workflow.

[0106] Here, the initial agent workflow uses the reference workflow as the thought chain for the recommendation large model to reason, and is the workflow arranged by the recommendation large model. That is, the recommendation large model can learn the successful recommendation thinking mode of the reference workflow and apply the learned thinking mode to the current recommendation task, enabling the recommendation large model to break down the current recommendation task into multiple subtasks and arrange the multiple subtasks according to the thinking mode provided by the reference workflow to obtain the arranged initial agent workflow.

[0107] Since the recommendation large model can draw on the successful recommendation thinking mode of the reference workflow to generate the initial agent workflow, it improves the accuracy of the automatic arrangement of the recommendation large model. Furthermore, based on the initial agent workflow for recommendation, the recommendation accuracy can be improved.

[0108] In an optional embodiment, in order to further improve the recommendation accuracy, the planning module 52 can also perform further dynamic planning on the initial agent workflow.

[0109] Specifically, during the process of the planning module 52 calling the execution module 54 to execute the initial agent workflow, it can also have one or more rounds of conversations with the user through the recommendation assistant. Optionally, it can adopt the step-by-step conversational guidance strategy generated by the recommendation large model to have one or more rounds of conversations with the user to obtain the second conversation data. Here, the step-by-step conversational guidance strategy can guide the user to provide more valuable information during the conversation, so as to more accurately understand the user's intention based on the valuable information and then more accurately recommend relevant objects to the user.

[0110] Meanwhile, data may also be returned during the execution of the initial agent workflow. For example, some data may be retrieved from the recommendation database during the execution of the initial agent workflow, and some data may also be retrieved from the Internet engine. Then, the data retrieved from the recommendation database and the data retrieved from the Internet engine can both be used as the returned data. Of course, the returned data can also be other things, which are not limited in this application.

[0111] After the planning module 52 obtains the second dialogue data and the returned data, it can determine whether to update the initial agent workflow based on the second dialogue data and / or the returned data. If so, it updates the initial agent workflow based on the first dialogue data, the second dialogue data, and the reference workflow, and then calls the execution module 54 to execute the updated agent workflow starting from the update position. During the execution of the updated agent workflow, it can also have another round or multiple rounds of conversations with the user through the recommendation assistant to obtain the second dialogue data. In addition, data may also be returned during the execution of the initial agent workflow. Then, it can determine whether to update the agent workflow after the first update based on the second dialogue data and the returned data. If so, it performs a second update on the agent workflow after the first update, and so on until the recommendation is completed.

[0112] In this embodiment, the process of updating the initial agent workflow based on the first dialogue data, the second dialogue data, and the reference workflow can be completed within the recommendation large model. Then, the first dialogue data and the second dialogue data can be input as the dialogue context into the fourth prompt template of the recommendation large model. At the same time, the reference workflow and the initial agent workflow can also be input into the fourth prompt template to obtain the fourth prompt instruction prompt. The fourth prompt instruction prompt is input into the recommendation large model to obtain the updated agent workflow output by the recommendation large model.

[0113] It should be noted that the above "fourth prompt template" is only used to distinguish from the following "first prompt template", "second prompt template", and "third prompt template", and is not used to limit the order; similarly, the "fourth" in the "fourth prompt instruction prompt" is only used to distinguish from the following "first four prompt instruction prompt", "second prompt instruction prompt", and "third prompt instruction prompt", and is not used to limit the order.

[0114] To make those skilled in the art better understand the above dynamic update process, an example of an initial agent workflow that sequentially executes subtasks 11, 12, 13, 14, and 15 is used for explanation.

[0115] First, execute subtask 11. After executing subtask 11, if it is determined that the initial agent workflow is not updated, then execute subtask 12. After executing subtask 12, if it is determined that the initial agent workflow is updated, then update subtask 13 to subtask 23, update subtask 14 to subtask 24, and update subtask 15 to subtask 25. Then execute subtask 23. After executing subtask 23, if it is determined that an update is needed, then update subtask 24 to subtask 34 and update subtask 25 to subtask 35. Next, execute subtask 34 and determine again whether to update. If not, then execute subtask 35 to obtain the recommended object and recommend it to the user.

[0116] That is, the planning module 52 can think and adjust the plan during the execution of the initial agent workflow, making the updated agent workflow more accurate.

[0117] In some scenarios, if it is determined based on the second dialogue data and / or return data generated during the execution process that the initial agent workflow is not updated, then this application can complete the recommendation based on the initial agent workflow.

[0118] Optionally, in some scenarios, as more data is generated during the recommendation process, the user's intention may change. To avoid incorrect recommendations and a poor recommendation experience for the user at this time, in this embodiment, after the second dialogue data and / or return data are generated during the execution process, the second dialogue data and / or return data can be combined to analyze the user's intention again, and the historical human-computer interaction data, the target object to be recommended, and the reference workflow are re-queried according to the newly obtained user intention, and the initial agent workflow is regenerated and executed until the recommendation is completed.

[0119] It should be noted that the above "completing the recommendation" can be successfully recommending the target object to be recommended to the user, or abandoning the target object to be recommended and recommending other interesting objects to the user, or the user actively ending the conversation and ending the recommendation process (possibly without successfully recommending relevant objects to the user), and so on.

[0120] The recommendation intelligent agent system provided by this application, the perception module obtains the first conversation data and / or multi-modal environment data collected during the human-computer interaction process, determines the user intention according to the obtained data, the memory module obtains the historical human-computer interaction data related to the user intention and the target object to be recommended, the planning module retrieves the reference workflow that matches the user intention, generates an initial intelligent agent workflow according to the data obtained by the perception module, the historical human-computer interaction data, the target object to be recommended and the reference workflow, calls the execution module to execute the initial intelligent agent workflow, and determines whether to update the initial intelligent agent workflow based on the second conversation data and / or the returned data generated during the execution process. If so, update the initial intelligent agent workflow based on the first conversation data, the second conversation data and the reference workflow, and call the execution module to execute the updated intelligent agent workflow starting from the update position to complete the recommendation.

[0121] Since the first conversation data and the multi-modal environment data are the data generated when the user needs a recommendation, the first conversation data and the multi-modal environment data can better reflect the user's current preferences, while the historical human-computer interaction data can reflect the user's past interests. Therefore, the recommendation intelligent agent system provided by this application refers to the first conversation data, the multi-modal environment data and the historical human-computer interaction data at the same time when making a recommendation, and can more accurately determine the recommended object that the user prefers, improving the recommendation accuracy.

[0122] The target object to be recommended provides a recommendation direction for the recommendation process. Planning the initial intelligent agent workflow based on the target object to be recommended and then making a product recommendation based on the initial intelligent agent workflow can guide the user to establish an interest in the target object to be recommended during the recommendation process, increasing the probability of successful recommendation of the target object to be recommended. At the same time, the existence of the target object to be recommended makes active recommendation possible.

[0123] Furthermore, the historical human-computer interaction data in this application is the long-term memory data of the user. Combining the historical human-computer interaction data with the target object to be recommended enables this application to plan a long-term recommendation strategy, increasing the probability of successful recommendation.

[0124] Furthermore, the present application can draw on the successful recommendation experience of objects in scenarios related to user intent, and use the reference workflow used when successfully recommending the object as the thought chain for the recommendation large model to reason, enabling the recommendation large model to draw on the thought pattern of successful recommendation, improving the accuracy of the automatic planning of the recommendation large model, and making recommendations based on the initial agent workflow with higher accuracy compiled by the recommendation large model, thereby improving the recommendation accuracy. Even if the choreography workflow used when successfully recommending an object in the scenario related to user intent cannot be retrieved, the present application can still use the manual choreography workflow that matches the user intent and has verified the recommendation effect as the reference workflow, enabling the initial agent workflow generated based on the reference workflow to make more effective recommendations, thereby improving the recommendation success rate.

[0125] In addition, regardless of whether the current recommendation task is complex, the present application can retrieve the corresponding reference workflow, learn the recommendation thought pattern of the reference workflow, and then generate the initial agent workflow, enabling the initial agent workflow to provide a better recommendation strategy for the current recommendation task, thereby improving the recommendation accuracy and success rate. Compared with the existing recommendation solutions that can only solve simple recommendation tasks, the present application can provide an accurate initial agent workflow to solve complex recommendation tasks, with a wider application scope.

[0126] Furthermore, during the process of making recommendations based on the initial agent workflow based on automatic planning, the recommendation agent system provided by the present application can adopt a step-by-step conversational guidance strategy to have one or more rounds of conversations with the user to obtain second conversation data. The addition of the second conversation data and the return data generated during the execution of the initial agent workflow enables the present application to better understand the user intent, and then dynamically update the initial agent workflow based on the understanding of the user intent, making the updated agent workflow more matching the user intent and improving the recommendation accuracy.

[0127] The following introduces the process of the planning module 52 generating the initial agent workflow based on the data obtained by the perception module, the historical human-computer interaction data, the target object to be recommended, and the reference workflow.

[0128] In this embodiment, the recommendation large model can be pre-trained, and the parameters of the recommendation large model can be fine-tuned through the pre-training process, enabling the recommendation large model to plan and reason out a more accurate initial agent workflow. Here, the training data used by the recommendation large model during the training stage includes: training multi-modal environment data and / or training conversation data, training historical human-computer interaction data, training target objects to be recommended, training reference workflows, and the corresponding training initial agent workflows.

[0129] Optionally, the pre-trained recommended large model in this embodiment can be Tongyi Qianwen 7B (Qwen-7B) large model or Llama-7B large model.

[0130] After training, this embodiment can fill the data obtained by the perception module, historical human-computer interaction data, the target object to be recommended, and the reference workflow into the first prompt template of the recommended large model to obtain the first prompt instruction prompt, and then input the first prompt instruction prompt into the recommended large model to obtain the initial agent workflow output by the recommended large model.

[0131] Optionally, the process of "filling the data obtained by the perception module, historical human-computer interaction data, the target object to be recommended, and the reference workflow into the first prompt template of the recommended large model" can include: obtaining the first prompt template, where the first prompt template includes a first basic data input slot, an interaction data input slot, an object input slot, and a workflow input slot, and the first prompt template is used to instruct the recommended large model to understand the user's intention based on the data in the included slots and give the initial agent workflow; filling the data obtained by the perception module into the first basic data input slot, filling the historical human-computer interaction data into the interaction data input slot, filling the target object to be recommended into the object input slot, and filling the reference workflow into the workflow input slot to obtain the first prompt instruction prompt.

[0132] This embodiment fine-tunes the recommended large model to enable the large model to improve the task planning ability in the recommendation scenario; fills the reference workflow into the preset first prompt template to provide the background information required for the recommended large model to reason and generate a new workflow. Finally, the recommended large model reasons based on the first prompt template and the reference workflow to generate a new and customized initial agent workflow, realizing the intelligent planning and optimization of tasks, enabling the recommended large model to still plan a reasonable and accurate initial agent workflow when facing complex recommendation tasks, and thus improving the recommendation accuracy.

[0133] In another possible implementation, to avoid the recommended large model generating an unreasonable initial agent workflow, this application can pre-configure a custom workflow and a workflow generation rule, and before using the recommended large model to generate the initial agent workflow, first train the recommended large model based on the pre-configured custom workflow and workflow generation rule, so that the recommended large model can follow the execution logic and execution order in the custom workflow to generate the required workflow under the instruction of the workflow generation rule.

[0134] Optionally, the custom workflow includes the following subtasks: Start: initial state; SQL: internal database retrieval; Search: retrieval for the internal vector library or external application programming interface (API); Retrieve: recommend that the large model perform a recovery action based on the Search retrieval results; Recall: recall operation; Rerank: reordering; Analyze: data summary and analysis; Finish: task end.

[0135] In the above custom workflow, the retrieval and search processes of each subtask specify which tools need to be used (including recommendation tools, retrieval tools, and how to read the external application programming interface API), what logical judgments need to be made, which large models need to be called, etc. Therefore, making recommendations according to the execution order and logic of these subtasks can retrieve the required data faster during the recommendation process and improve the recommendation efficiency.

[0136] Optionally, specifically in this embodiment, the recommended large model "generates the required workflow following the execution logic and order in the custom workflow" includes generating a workflow that includes the above initial state, task end, and at least one subtask between the initial state and the task end, and the workflow follows the execution logic and order of the included subtasks.

[0137] In the embodiment of the present application, the recommended large model is trained based on the workflow generation rules and the custom workflow, so that the recommended large model can follow the execution logic and order in the custom workflow when planning and executing the recommendation task, avoiding generating unreasonable initial agent workflows, such as generating initial agent workflows that skip intermediate retrieval and search steps, and improving the performance and accuracy of the recommended large model in complex tasks.

[0138] In a possible implementation, before the memory module 53 obtains the historical human-computer interaction data related to the user intention and the target object to be recommended, the recommended agent system provided by the present application can also determine the user intention through the perception module 51.

[0139] Specifically, the perception module 51 can also determine the user intention according to the data obtained by itself (i.e., the first dialogue data and / or multi-modal environment data).

[0140] Optionally, the perception module 51 may call the recommendation large model to determine the user intention, that is, the perception module 51 may fill the data obtained by itself into the second prompt template of the recommendation large model to obtain the second prompt instruction prompt, and input the second prompt instruction prompt into the recommendation large model to obtain the user intention output by the recommendation large model. Here, the training data used by the recommendation large model in the training stage includes: training multi-modal environment data and / or training dialogue data, and the corresponding training user intention.

[0141] Optionally, the above second prompt template includes a second basic data input slot, and the second prompt template is used to instruct the recommendation large model to reason based on the data in the second basic data input slot and give the user intention. Then, the present application can fill the data obtained by the perception module 51 into the second basic data input slot to obtain the above second prompt instruction prompt.

[0142] In the embodiment of the present application, by virtue of the powerful analysis and reasoning ability of the recommendation large model, the user intention is inferred and analyzed by the recommendation large model, which improves the accuracy of the user intention and further improves the recommendation accuracy.

[0143] Continuing the previous introduction, the historical human-computer interaction data obtained by the memory module 53 can be the content summary of the historical dialogue data between the user and the recommendation assistant. In the embodiment of the present application, there are various methods for extracting the content summary from the historical dialogue data, and the following are provided but not limited to.

[0144] First, the keyword extraction method, such as using the keyword extraction algorithm in natural language processing technology to extract the most important words or phrases from the historical dialogue data, and these words or phrases can represent the main content of the historical dialogue data.

[0145] Second, use the automatic summarization technology, such as the statistics-based summarization method, the graph-based summarization method or the deep learning-based summarization method, to automatically generate the content summary of the historical dialogue data.

[0146] Third, use the method of artificial intelligence or machine learning model, such as training a machine learning model or a deep learning model, and output the content summary of the historical dialogue data based on the model.

[0147] Fourth (preferred solution), use the recommendation large model to obtain the content summary of the historical dialogue data, and this process may include: filling the historical dialogue data into the third prompt template of the recommendation large model to obtain the third prompt instruction prompt, inputting the third prompt instruction prompt into the recommendation large model to obtain the content summary output by the recommendation large model; here, the training data used by the recommendation large model in the training stage includes: training dialogue data and the corresponding training content summary.

[0148] Optionally, the above third prompt template includes a dialogue data input slot, and the third prompt template is used to instruct the recommendation large model to give a specific content summary in combination with the data in the dialogue data input slot. Then, the present application can fill the historical dialogue data into the dialogue data input slot to obtain the above third prompt instruction prompt.

[0149] It should be noted that, in addition to the above method, other methods can also be used to obtain the content summary of the historical dialogue data, and the present application does not make specific limitations.

[0150] See Figure 5 , which is an optional schematic diagram of the interaction of each module during dynamic programming provided by the embodiment of the present application.

[0151] Such as Figure 5 , the perception module 51 includes a user interaction module 511 and a multimodal perception module 512. The user interaction module 511 is used to obtain the first dialogue data, and the multimodal perception module 512 (optionally, the multimodal perception module 512 can be a multimodal interaction large model) is used to obtain multimodal environment data.

[0152] The above first dialogue data and multimodal environment data can be sent to the dynamic programming executor 521 in the planning module 52. The dynamic programming executor 521 can call the recommendation large model 522 to determine the user intention according to the first dialogue data and multimodal environment data. Then, the dynamic programming executor 521 obtains the historical human-computer interaction data related to the user intention from the first memory bank 531 in the memory module 53, obtains the target object to be recommended related to the user intention from the second memory bank 532, and obtains the reference workflow matching the user intention from the workflow library 524. After that, the dynamic programming executor 521 calls the recommendation large model 522 again according to the first dialogue data, multimodal environment data, historical human-computer interaction data, target object to be recommended, and reference workflow, and combines with the orchestration object 541 to generate an initial intelligent agent workflow.

[0153] After the initial intelligent agent workflow is generated, the initial intelligent agent workflow can be executed through the execution module 54. Based on the second dialogue data and / or return data generated during the execution process, it is determined whether the initial intelligent agent workflow is to be updated. If so, the initial intelligent agent workflow is updated according to the first dialogue data, the second dialogue data and the reference workflow. If it is determined that the initial intelligent agent workflow is not to be updated, the initial intelligent agent workflow continues to be executed. Until the initial intelligent agent workflow or the updated intelligent agent workflow is executed, the recommended related objects are fed back to the user or the user ends the conversation. When the workflow evaluation module 523 determines through evaluation that the related objects are successfully recommended based on the final updated intelligent agent workflow or the initial intelligent agent workflow, the final updated intelligent agent workflow or the initial intelligent agent workflow is written to the workflow library 524.

[0154] It should be noted that there may be a situation where the initial intelligent agent workflow generated by the recommendation model 522 is the same as the reference workflow, or there may be a situation where the final updated intelligent agent workflow is the same as the reference workflow. Since the reference workflow has been stored in the workflow library, even if the workflow evaluation module 523 determines through evaluation that the relevant objects are successfully recommended based on the final updated intelligent agent workflow or the initial intelligent agent workflow, the final updated intelligent agent workflow or the initial intelligent agent workflow will not be written to the workflow library 524.

[0155] Optionally, the above-mentioned choreography objects include but are not limited to the following: tools used in the recommendation process, models called, and judgment logic used.

[0156] Optionally, the tools used in the recommendation process include but are not limited to: the first recall tool Recall0, the second recall tool Recalln, the reranking tool Rerank, database retrieval tools, API call tools, knowledge base retrieval tools, vector library retrieval tools, tools for writing and executing Python scripts, tools for filtering data using Structured Query Language (SQL), and tools for analyzing API results.

[0157] In summary, this application flexibly arranges the orchestration objects based on the automatic planning capability of the recommendation agent system to generate the initial agent workflow, and dynamically plans and adjusts the initial agent workflow based on the continuous recommendation capability of the recommendation agent system, so that the updated agent workflow is more accurate, thereby improving the user satisfaction and recommendation success rate of conversational recommendations.

[0158] The present application also provides a recommendation method, which can be applied to the recommendation agent system described above. In order to make those skilled in the art more understandable, the recommendation method of the embodiment of the present application is described in detail below with reference to the accompanying drawings.

[0159] Referring to Figure 6 , Figure 6 which is a schematic flowchart of a recommendation method provided by an embodiment of the present application. The method may include:

[0160] Step S601, obtaining first dialogue data and / or multi-modal environment data collected during the human-computer interaction process.

[0161] Among them, the first dialogue data includes at least one round of dialogue data.

[0162] Step S602, obtaining historical human-computer interaction data related to the user's intention and the target object to be recommended.

[0163] Among them, the user's intention reflects the recommended object that the user prefers.

[0164] Step S603, retrieving a reference workflow that matches the user's intention, and generating an initial agent workflow based on the first dialogue data and / or multi-modal environment data, historical human-computer interaction data, the target object to be recommended, and the reference workflow.

[0165] Among them, the reference workflow refers to an orchestration workflow used for successful recommendation of objects in a scenario related to the user's intention, or an artificial orchestration workflow that matches the user's intention and verifies the recommendation effect. The initial agent workflow is a thought chain that uses the reference workflow for inference of the recommendation large model, and is a workflow planned and arranged by the recommendation large model.

[0166] For example, in the scenario of package recommendation, the user's intention is "the current 5G package is too expensive and doesn't seem cost-effective. Comparing with other products on the market, I hope to get a recommendation for a package with high cost performance". Then the generated initial agent workflow can be: query the relevant database to obtain the user's package name; call the search engine to query the prices of the same products in the market; according to the user's usage behavior preferences in the user profile, find the appropriate packages sorted by the top n (i.e., the top n most suitable packages) through the recommendation engine; compare the cost performance of the competitor packages retrieved by the search engine in turn and select the optimal package; generate the reply content.

[0167] Step S604, execute the initial agent workflow, and determine whether to update the initial agent workflow based on the second dialogue data and / or return data generated during the execution process. If so, update the initial agent workflow based on the first dialogue data, the second dialogue data, and the reference workflow, and start executing the updated agent workflow from the update position to complete the recommendation.

[0168] Among them, the second dialogue data is at least one round of dialogue data based on the step-by-step dialogue guidance strategy and the user.

[0169] In the recommendation method provided by this application, since the first conversation data and the multi-modal environment data are generated when the user needs a recommendation, the first conversation data and the multi-modal environment data can better reflect the user's current preferences, while the historical human-computer interaction data can reflect the user's past interests. Therefore, when making a recommendation, this application simultaneously refers to the first conversation data, the multi-modal environment data, and the historical human-computer interaction data, which can more accurately determine the recommended object that the user prefers and improve the recommendation accuracy.

[0170] The proposed recommended target object provides a recommendation direction for the recommendation process. Planning the initial agent workflow based on the proposed recommended target object and then making product recommendations based on the initial agent workflow can guide the user to build an interest in the proposed recommended target object during the recommendation process, increasing the probability of successful recommendation of the proposed recommended target object. At the same time, the existence of the proposed recommended target object makes active recommendation possible.

[0171] Furthermore, the historical human-computer interaction data in this application is the user's long-term memory data. Combining the historical human-computer interaction data with the proposed recommended target object enables this application to plan a long-term recommendation strategy and increases the probability of successful recommendation.

[0172] Moreover, this application can draw on the successful recommendation experience of objects in scenarios related to the user's intention. The reference workflow used when successfully recommending this object is used as the thought chain for the recommendation large model to reason. This enables the recommendation large model to draw on the successful recommendation thinking mode, improving the accuracy of the automatic planning of the recommendation large model. Making recommendations based on the initial agent workflow with higher accuracy compiled by the recommendation large model improves the recommendation accuracy. Even if the choreography workflow used when successfully recommending an object in a scenario related to the user's intention cannot be retrieved, this application can still use the manually choreographed workflow that matches the user's intention and has verified the recommendation effect as the reference workflow, enabling the initial agent workflow generated based on the reference workflow to make more effective recommendations and increasing the recommendation success rate.

[0173] In addition, regardless of whether the current recommendation task is complex, this application can retrieve the corresponding reference workflow. After learning the recommendation thinking mode of this reference workflow, it generates the initial agent workflow, enabling the initial agent workflow to provide a better recommendation strategy for the current recommendation task and improving the recommendation accuracy and success rate. Compared with the existing recommendation solutions that can only solve simple recommendation tasks, this application can provide an accurate initial agent workflow to solve complex recommendation tasks and has a wider application range.

[0174] Furthermore, during the process of making recommendations based on the initial agent workflow with automatic planning, the present application can adopt a step-by-step conversational guidance strategy to conduct one or more rounds of conversations with the user to obtain second conversation data. The addition of the second conversation data and the return data generated during the execution of the initial agent workflow enables the present application to better understand the user's intention. Furthermore, based on the understanding of the user's intention, the initial agent workflow is dynamically updated, making the updated agent workflow more in line with the user's intention and improving the accuracy of the recommendation.

[0175] The above introduced a recommendation method provided by an embodiment of the present application. Next, a device for executing the above recommendation method will be introduced.

[0176] Please refer to Figure 7 , Figure 7 , which is a schematic structural diagram of a recommendation device provided by an embodiment of the present application. This recommendation device can be applied to the above-mentioned recommendation agent system. As Figure 7 shown, this recommendation device includes:

[0177] A data acquisition unit 801, configured to acquire first conversation data and / or multi-modal environment data collected during human-computer interaction, where the first conversation data includes at least one round of conversation data;

[0178] A memory extraction unit 802, configured to acquire historical human-computer interaction data related to the user's intention and the target object to be recommended, where the user's intention reflects the recommended object that the user prefers;

[0179] A workflow orchestration unit 803, configured to retrieve a reference workflow that matches the user's intention, and generate an initial agent workflow based on the first conversation data and / or multi-modal environment data, historical human-computer interaction data, the target object to be recommended, and the reference workflow. Here, the reference workflow refers to an orchestration workflow used for successful recommendation of objects in a scenario related to the user's intention, or an artificial orchestration workflow that matches the user's intention and verifies the recommendation effect. The initial agent workflow uses the reference workflow as the thinking chain for the recommendation large model to reason, and is a workflow orchestrated by the recommendation large model;

[0180] A dynamic planning unit 804, configured to execute the initial agent workflow, determine whether to update the initial agent workflow based on the second conversation data and / or return data generated during the execution process. If so, update the initial agent workflow based on the first conversation data, the second conversation data, and the reference workflow, and execute the updated agent workflow starting from the update position to complete the recommendation. Here, the second conversation data is at least one round of conversation data based on the step-by-step conversational guidance strategy and the user.

[0181] The recommendation device provided in this application corresponds to the recommendation method and recommendation agent system provided above. For details, please refer to the above introduction and will not be elaborated here.

[0182] An electronic device is also provided in an embodiment of this application. Refer to Figure 8 As shown, it shows a schematic structural diagram of an electronic device suitable for implementing the electronic device in the embodiment of this application. The electronic device in the embodiment of this application may include, but is not limited to, fixed terminals such as mobile phones, laptop computers, PDAs (Personal Digital Assistants), PADs (Tablet Computers), desktop computers, and the like. Figure 8 The electronic device shown is only an example and should not impose any limitation on the functions and usage scope of the embodiment of this application.

[0183] As Figure 8 shown, the electronic device may include a processing device (such as a central processing unit, a graphics processing unit, etc.) 901, which may perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 902 or the program loaded from the storage device 908 into the random access memory (RAM) 903. When the electronic device is powered on, various programs and data required for the operation of the electronic device are also stored in the RAM 903. The processing device 901, the ROM 902, and the RAM 903 are connected to each other through a bus 904. The input / output (I / O) interface 905 is also connected to the bus 904.

[0184] Generally, the following devices may be connected to the I / O interface 905: an input device 906 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 907 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 908 including, for example, a memory card, a hard disk, etc.; and a communication device 909. The communication device 909 may allow the electronic device to communicate with other devices wirelessly or wiredly to exchange data. Although Figure 8 the electronic device with various devices is shown, it should be understood that it is not required to implement or have all the shown devices. More or fewer devices may be implemented or had alternatively.

[0185] An embodiment of this application also provides a computer program product including computer-readable instructions. When the computer-readable instructions run on the electronic device, the electronic device is enabled to implement any of the recommendation methods provided in the embodiment of this application.

[0186] An embodiment of this application also provides a computer-readable storage medium. The storage medium carries one or more computer programs. When the one or more computer programs are executed by the electronic device, the electronic device can be enabled to implement any of the recommendation methods provided in the embodiment of this application.

[0187] In addition, it should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in the attached drawings of the device embodiments provided in this application, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines.

[0188] Through the description of the above embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware. Of course, it can also be implemented by dedicated hardware including application-specific integrated circuits, dedicated CPUs, dedicated memories, dedicated components, etc. Generally, functions completed by computer programs can be easily implemented by corresponding hardware, and the specific hardware structures used to implement the same function can also be diverse, such as analog circuits, digital circuits, or dedicated circuits. However, for this application, in more cases, software program implementation is a better implementation method. Based on such an understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disc of a computer, and includes several instructions to enable a computer device (which can be a personal computer, training device, or network device, etc.) to execute the methods described in various embodiments of this application.

[0189] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product.

[0190] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center by wire (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wirelessly (such as infrared, wireless, microwave, etc.). The computer-readable storage medium may be any available medium that can be stored by a computer or a data storage device such as a training device or a data center that includes one or more integrated available media. The available medium may be a magnetic medium (such as a floppy disk, hard disk, magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

Claims

1. A recommended intelligent agent system, characterized in that, Including: A perception module, a memory module, a planning module, and an execution module; The perception module is used to obtain first dialogue data and / or multi-modal environment data collected during human-computer interaction, where the first dialogue data includes at least one round of dialogue data; The memory module is used to obtain historical human-computer interaction data related to the user's intention and the target object to be recommended, where the user's intention reflects the recommended object preferred by the user; The planning module is used to retrieve a reference workflow that matches the user's intention, and call a recommendation large model to generate an initial agent workflow based on the data obtained by the perception module, the historical human-computer interaction data, the target object to be recommended, and the reference workflow. The reference workflow refers to an orchestration workflow used for successful recommendation of objects in the scenario related to the user's intention, or an artificial orchestration workflow that matches the user's intention and verifies the recommendation effect. The initial agent workflow uses the reference workflow as the thinking chain for the recommendation large model to reason, and is the workflow orchestrated by the recommendation large model. Among them, the reference workflow contains the orchestration results of tool use, logical judgment, and model call. The recommendation large model can break down the current recommendation task into multiple subtasks, and orchestrate the multiple subtasks according to the thinking mode provided by the reference workflow to obtain the orchestrated initial agent workflow; The planning module is further used to call the execution module to execute the initial agent workflow, and determine whether to update the initial agent workflow based on the second dialogue data and / or return data generated during the execution process. If so, update the initial agent workflow based on the first dialogue data, the second dialogue data, and the reference workflow, and call the execution module to execute the updated agent workflow starting from the update position to complete the recommendation. The second dialogue data is at least one round of dialogue data based on a step-by-step dialogue guidance strategy with the user; among them, it is judged whether to update the workflow after the subtask is completed; When the planning module generates an initial agent workflow according to the data obtained by the perception module, the historical human-computer interaction data, the target object to be recommended, and the reference workflow, it specifically is used for: Filling the data obtained by the perception module, the historical human-computer interaction data, the target object to be recommended, and the reference workflow into the first prompt template of the recommendation large model to obtain a first prompt instruction prompt; Inputting the first prompt instruction prompt into the recommendation large model to obtain the initial agent workflow output by the recommendation large model; Among them, the training data used by the recommended large model in the training phase includes: training multi-modal environment data and / or training dialogue data, training historical human-computer interaction data, training target objects to be recommended, training reference workflows, and corresponding initial intelligent agent workflows. It also includes: pre-configured custom workflows and workflow generation rules, where the workflow generation rules are used to instruct the recommended large model to generate the required workflows following the execution logic and execution order in the custom workflows. The custom workflows include several sub-tasks: initial state; internal database retrieval; retrieval of internal vector libraries or external application programming interfaces; based on the retrieval results of internal vector libraries or external application programming interfaces, the recommended large model performs recovery actions; recall operations; re-ranking; data summary and analysis; task end.

2. The recommended intelligent agent system according to claim 1, wherein Before the memory module obtains the historical human-computer interaction data related to the user's intention and the target object to be recommended, the perception module is further used for: Determine the user's intention based on the data obtained by the perception module.

3. The recommended intelligent agent system according to claim 2, wherein When the perception module determines the user's intention based on the data obtained by the perception module, it specifically is used for: Fill the data obtained by the perception module into the second prompt template of the recommended large model to obtain a second prompt instruction prompt; Input the second prompt instruction prompt into the recommended large model to obtain the user's intention output by the recommended large model; Among them, the training data used by the recommended large model in the training phase includes: training multi-modal environment data and / or training dialogue data, and corresponding training user intentions.

4. The recommended intelligent agent system according to claim 1, wherein When the memory module obtains the historical human-computer interaction data related to the user's intention and the target object to be recommended, it specifically is used for: Obtain the historical human-computer interaction data from the first memory bank according to the user's intention; Obtain the target object to be recommended from the second memory bank according to the user's intention.

5. The recommended intelligent agent system according to claim 1, characterized in that, The multi-modal environment data includes one or more of image data, video data, and audio data.

6. The recommended intelligent agent system according to claim 1, characterized in that The historical human-computer interaction data is a content summary of the historical dialogue data between the user and the recommendation assistant.

7. A recommendation method, characterized in that, Applied to the recommendation intelligent agent system, it includes: Obtain the first dialogue data and / or multi-modal environment data collected during the human-computer interaction process, where the first dialogue data includes at least one round of dialogue data; Obtain the historical human-computer interaction data related to the user's intention and the target object to be recommended, where the user's intention reflects the recommended object preferred by the user; Retrieve the reference workflow that matches the user intent, and call the recommendation large model to generate an initial agent workflow based on the first dialogue data and / or the multi-modal environment data, the historical human-computer interaction data, the target object to be recommended, and the reference workflow. The reference workflow refers to an orchestration workflow used for successful recommendation of objects in the scenario related to the user intent, or a manually orchestrated workflow that matches the user intent and verifies the recommendation effect. The initial agent workflow uses the reference workflow as the thinking chain for the recommendation large model to reason, and is a workflow orchestrated by the recommendation large model. Among them, the reference workflow contains the orchestration results of tool usage, logical judgment, and model call. The recommendation large model can break down the current recommendation task into multiple subtasks, and orchestrate the multiple subtasks according to the thinking mode provided by the reference workflow to obtain the orchestrated initial agent workflow. Execute the initial agent workflow, and determine whether to update the initial agent workflow based on the second dialogue data and / or the return data generated during the execution process. If so, update the initial agent workflow based on the first dialogue data, the second dialogue data, and the reference workflow, and start executing the updated agent workflow from the update position to complete the recommendation. The second dialogue data is at least one round of dialogue data based on a step-by-step dialogue guidance strategy with the user. Among them, it is determined whether to update the workflow after executing the subtask. Generate an initial agent workflow based on the first dialogue data and / or the multi-modal environment data, the historical human-computer interaction data, the target object to be recommended, and the reference workflow, specifically including: Fill the first dialogue data and / or the multi-modal environment data, the historical human-computer interaction data, the target object to be recommended, and the reference workflow into the first prompt template of the recommendation large model to obtain the first prompt instruction prompt. Input the first prompt instruction prompt into the recommendation large model to obtain the initial agent workflow output by the recommendation large model. Among them, the training data used by the recommendation large model in the training stage includes: training multi-modal environment data and / or training dialogue data, training historical human-computer interaction data, training target objects to be recommended, training reference workflows, and corresponding training initial agent workflows. It also includes: pre-configured custom workflows and workflow generation rules. The workflow generation rules are used to instruct the recommendation large model to generate the required workflow following the execution logic and execution order in the custom workflow. The custom workflow includes several subtasks: initial state; internal database retrieval; retrieval for the internal vector library or external application programming interface; according to the retrieval results of the internal vector library or external application programming interface retrieval, the recommendation large model performs a recovery action; recall operation; re-ranking; data summary analysis; task end.

Images