A method and system for building an intelligent agent with large models and intelligent recommendations

Abstract

The present invention relates to the field of intelligent recommendation technology, and more particularly to a method and system for constructing an intelligent agent for large models and intelligent recommendations. The construction method comprises: constructing a basic intelligent agent based on a large language model based on a prompting project; constructing a memory mechanism and a recommendation mechanism based on a Milvus vector database and the large language model, and constructing a feedback learning mechanism based on the recommendation mechanism; and embedding the memory mechanism, recommendation mechanism, and feedback learning mechanism into the basic intelligent agent to construct an intelligent agent, wherein the input of the intelligent agent is user information and the output is a recommendation list and recommendation explanations. The present invention utilizes a large language model to construct an intelligent recommendation intelligent agent capable of understanding and processing complex user needs, performing task planning, and utilizing tools such as recommendation systems. The intelligent agent is applied in the form of a virtual digital human to implement functions such as voice interaction with users.

Description

Technical Field

[0001] The present invention relates to the field of intelligent recommendation technology, and in particular to a method and system for constructing an intelligent agent for large models and intelligent recommendations. Background Art

[0002] Currently, intelligent recommendation systems are widely used in various fields, such as e-commerce, social media, music, and video platforms. These systems recommend relevant content to users based on their interests, behaviors, and preferences, thereby improving user experience and platform user stickiness. However, existing intelligent recommendation systems have the following major problems:

[0003] 1. Unable to understand complex user needs: Traditional recommendation systems usually make recommendations based on users' historical behavior and preferences, and have difficulty understanding and processing users' complex needs and personalized preferences.

[0004] 2. Lack of task planning and decision-making capabilities: Traditional recommendation systems lack task planning and decision-making capabilities and are unable to effectively decompose tasks and allocate resources according to user needs.

[0005] 3. Poor recommendation explainability: Traditional recommendation systems usually do not provide explanations for recommendations, making it difficult for users to understand the basis for the recommendation results, which reduces their trust in the recommendation system.

[0006] 4. Lack of continuous optimization mechanism: Traditional recommendation systems lack a continuous optimization mechanism, making it difficult to dynamically adjust recommendation strategies based on user feedback, resulting in recommendation results that are unable to meet users' ever-changing needs.

[0007] In order to solve the above problems, the present invention proposes a method for constructing an intelligent agent with large models and intelligent recommendation, as well as an application system of the constructed intelligent agent, aiming to build an intelligent recommendation agent that can understand and process complex user needs, perform task planning, and use tools such as recommendation systems, so as to achieve more accurate, personalized, and explainable intelligent recommendation services. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for constructing an intelligent agent with a large model and intelligent recommendation, using a large language model to build an intelligent recommendation agent that can understand and process complex user needs, perform task planning, and use tools such as recommendation systems.

[0009] To achieve the above objectives, the present invention provides, on the one hand, a method for constructing an intelligent agent for large models and intelligent recommendations, comprising:

[0010] Building a basic intelligent agent based on a large language model based on prompt engineering;

[0011] Building a memory mechanism and a recommendation mechanism based on the Milvus vector database and the large language model, and building a feedback learning mechanism based on the recommendation mechanism;

[0012] The memory mechanism, recommendation mechanism, and feedback learning mechanism are embedded in the basic intelligent agent to construct an intelligent agent, wherein the input of the intelligent agent is user information and the output is a recommendation list and a recommendation explanation.

[0013] Optionally, constructing the memory mechanism includes:

[0014] Building a long-term memory based on the Milvus vector database, and building a short-term memory based on the context of the large language model, wherein building the long-term memory includes:

[0015] Acquire user and item data and preprocess them, wherein the user and item data includes user data, item data, and user-item interaction data;

[0016] Perform feature extraction on the preprocessed user and item data, and vectorize the extracted features as well as the agent's behavior and results in each recommendation task to obtain vector data;

[0017] The vector data is stored in the Milvus vector database and indexed to complete the storage of long-term memory.

[0018] Optionally, establishing the recommendation mechanism includes:

[0019] Utilize the nearest neighbor search function of the Milvus vector database to filter the top n recommendations with the highest similarity to the user's vector data;

[0020] The recommendations are sorted and filtered to obtain a recommendation list, and recommendation explanations are generated based on the user information and the recommendation list.

[0021] Optionally, constructing the feedback learning mechanism includes:

[0022] Collecting user feedback on the recommendation list;

[0023] The feedback behavior is evaluated through the reward function, and the user satisfaction is obtained and stored as historical experience, and then added to the next recommendation task.

[0024] Optionally, the reward function assigns a value to the user feedback behavior, takes the average of several behavior assignments, and determines the user satisfaction corresponding to the average through a preset satisfaction interval.

[0025] To further achieve the above objectives, the present invention also provides an intelligent agent application system for large models and intelligent recommendations. The intelligent agent implementation constructed based on the above method includes:

[0026] A user request module is used to receive user requests and convert the format of the user requests;

[0027] The task planning module is used to decompose and plan tasks based on the formatted user requests and obtain task information;

[0028] An intelligent recommendation module, configured to input the task information into an intelligent agent and output a recommendation list and a recommendation explanation;

[0029] The memory storage module is used to record the agent's work log, including previous behaviors and external knowledge, as well as interaction records with users.

[0030] Optionally, the system receives requests and outputs recommendation results through voice interaction with the user.

[0031] Optionally, the system takes the form of a virtual digital human, which is designed and generated through 3Dmax by collecting user voice information and body information.

[0032] The beneficial effects of the present invention are:

[0033] The present invention provides a method and system for constructing an intelligent agent with a large model and intelligent recommendation, which can efficiently decompose and process the user's recommendation instructions; can realize the effective storage and rapid access of data by the intelligent agent through the combination of long-term memory and short-term memory; can enhance the relevance and personalization of recommendations through intelligent recommendation, and provide recommendation explanations that traditional recommendation systems cannot provide; an effective feedback learning mechanism can dynamically adjust the recommendation strategy based on the user's language and behavioral feedback, and realize continuous optimization of intelligent recommendation; the present invention integrates a large model and a recommendation system to construct an intelligent agent and its application system that can perform complex recommendation tasks, so as to achieve the purpose of user-agent completing recommendations in the form of dialogue and providing recommendation explanations. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0035] Figure 1 A flowchart of a method for constructing a large model and intelligent recommendation agent according to an embodiment of the present invention;

[0036] Figure 2 This is a flow chart of the feedback learning mechanism according to an embodiment of the present invention;

[0037] Figure 3 This is a voice interaction workflow diagram of an embodiment of the present invention;

[0038] Figure 4 This is a workflow diagram of an intelligent agent application for a large model and intelligent recommendation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0040] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] On the one hand, this embodiment provides a method for constructing an intelligent agent with a large model and intelligent recommendation, such as Figure 1 As shown, including:

[0042] Building a basic intelligent agent based on a large language model based on prompt engineering;

[0043] Build a memory mechanism and recommendation mechanism based on the Milvus vector database and large language model, and build a feedback learning mechanism based on the recommendation mechanism;

[0044] The memory mechanism, recommendation mechanism, and feedback learning mechanism are embedded in the basic intelligent agent to construct an intelligent agent, where the input of the intelligent agent is user information and the output is a recommendation list and recommendation explanation.

[0045] Furthermore, the basic intelligent agent based on the large language model is constructed based on the prompt engineering, including:

[0046] Prompt Chaining is used to build an agent based on a large language model. The core idea is to decompose a large problem into a series of smaller, more specific sub-problems. After determining the subtasks, the prompt words for the subtasks are provided to the language model. The resulting prompt words serve as part of the new prompt word, which enables the large model to complete very complex tasks.

[0047] The specific steps include: determining the overall task of the intelligent agent, namely understanding and responding to user needs, providing personalized services and continuously optimizing the user experience by integrating data and making intelligent recommendations; decomposing the overall task into the following: the intelligent agent waiting for user recommendation instructions, calling the recommendation system, the intelligent agent providing recommendation explanations, and the intelligent agent executing the feedback learning mechanism; designing initial prompt words for the three subtasks of the intelligent agent waiting for user recommendation instructions, the intelligent agent providing recommendation explanations, and the intelligent agent executing the feedback learning mechanism; replacing the prompt words of the large language model and running the above subtasks in turn, integrating the output of each subtask into the prompt words of the next task until all subtasks are completed.

[0048] Furthermore, building a memory mechanism includes:

[0049] Long-term memory is built based on the Milvus vector database, and short-term memory is built based on the context of the large language model. Building long-term memory includes:

[0050] Acquire user and item data and preprocess them, where the user and item data includes user data, item data, and user-item interaction data;

[0051] Perform feature extraction on the preprocessed user and item data, and vectorize the extracted features as well as the agent's behavior and results in each recommendation task to obtain vector data;

[0052] The vector data is stored in the Milvus vector database and indexed to complete the storage of long-term memory.

[0053] Specifically, data preprocessing is performed on user data, item data, and user-item interaction data, including noise removal and missing value processing. User feature extraction methods include user activity and matrix decomposition. The calculation formulas are as follows:

[0054] ;

[0055] in, Represents a user A collection of Indicates the An interactive collection of items, Indicates the number of items;

[0056] Matrix decomposition decomposes the user-item rating matrix R into two low-dimensional matrices and , is the user's potential factor, The latent factors of items are expressed as explicit ratings through user historical behaviors to mine the matching relationship between user interests and item attributes:

[0057] ;

[0058] in, represents the transposed matrix of the latent factor of the i-th item, is the user-item rating matrix.

[0059] The text data is segmented and encoded using BertTokenizer and BertModel from the transformers library to obtain the BERT model output vector. The steps for segmentation based on BertTokenizer include:

[0060] 1) Divide the original text content into atomic tokens, such as single Chinese characters:

[0061] ;

[0062] Among them, tokenize represents the word segmentation method provided by BertTokenizer; text represents the original text content; Indicates the tokens, .

[0063] 2) Map the token to its corresponding integer ID. The basic principle is as follows:

[0064] ;

[0065] Among them, dict represents the vocabulary list after BertTokenizer pre-training; [CLS] represents the beginning of sentence token; [SEP] represents the end of sentence / delimiter token.

[0066] Secondly, the steps of encoding based on BertModel include:

[0067] 1) Calculate the initial hidden state , the formula is as follows:

[0068] ;

[0069] in, represents the learnable weight matrix; represents a learnable bias; P represents positional encoding; S represents segment embedding, Indicates the integer ID after token mapping.

[0070] 2) Extracting semantic features from local to global through a multi-layer Transformer encoder :

[0071]

[0072] ;

[0073] in, represents the global layer number, express Semantic features of the layer, express The semantic features of the layer, LN represents the layer normalization layer; Att represents the self-attention mechanism, which captures the global dependencies of all positions in the sequence; FNN represents the feedforward network, which performs nonlinear transformation on the results of layer normalization and learns complex local feature interactions;

[0074] The discrete features are then vectorized through the nn.Embedding layer of the PyTorch framework, and the mapping relationship from discrete to continuous is:

[0075] ;

[0076] in, Represents a discrete category index, such as an item ID; is the weight matrix, represents the feature vector.

[0077] For each recommendation task, the agent’s historical behavior and experience summary are vectorized using the same BERT model mentioned above; the resulting vector data is stored in the Milvus database and indexed to optimize search efficiency, enabling storage and fast access of long-term memory data.

[0078] Furthermore, building a recommendation mechanism includes:

[0079] Use the nearest neighbor search function of the Milvus vector database to filter the top n recommendations with the highest similarity to the user's vector data;

[0080] Sort and filter recommendations, obtain a recommendation list, and generate recommendation explanations based on user information and the recommendation list.

[0081] Specifically, implementing recommendations based on vectorized embeddings includes:

[0082] 1) In the recall phase, the nearest neighbor search function of the Milvus vector database is used to efficiently filter out hundreds to thousands of items that are most similar to the user's embedding vector from a large number of items through the ANN algorithm; the optimization goal of this phase is to , from the item vector database Find the n most similar vectors, expressed as:

[0083] ;

[0084] To efficiently find the most proximate neighborhood from the database, the HNSW (Hierarchical Navigable Small World) indexing method of the ANN algorithm is adopted. It organizes data points into a multi-layer graph, where each layer contains a low-level subgraph and a high-level subgraph. Each node maintains a pointer jump pointing to other nodes to ensure that the graph structure is traversed within a constant number of steps. When inserting a new node, a greedy strategy is adopted and the jump pointer is updated. When searching, starting from the query point, traverse up / down along the jump pointer, gradually approaching the nearest neighbor.

[0085] 2) In the sorting stage, the feature embedding vectors of users and items are combined to use convolutional neural network (CNN) to predict the user's click probability or rating of the item; the top items with the highest click probability or rating are selected. Items are used as the recommendation list. The calculation mechanism at this stage is:

[0086] ;

[0087] in, Is the activation function; represents CNN feature extraction; Represents the outer product operation of user-item embedding, Indicates the user's click probability or rating of an item.

[0088] User information and the recommendation list are used as the context of the large language model input, and a recommendation explanation template is specified to guide the large language model to generate explanations. In this embodiment, the large language model can be a pre-trained BERT model, a PyTorch framework, other recommendation systems, and other tools.

[0089] Furthermore, building a feedback learning mechanism includes:

[0090] Collect user feedback on recommendation lists;

[0091] The feedback behavior is evaluated through the reward function, and the user satisfaction is obtained and stored as historical experience, which is then added to the next recommendation task.

[0092] The reward function assigns a value to the user's feedback behavior, takes the average of several behavior assignments, and determines the user satisfaction corresponding to the average through a preset satisfaction interval.

[0093] Specifically, such as Figure 2 As shown in Figure 2, the feedback learning mechanism includes:

[0094] Users, acting as the external environment for interaction, provide subjective feedback on the recommendation list. The reward function evaluates user behavioral feedback to determine user satisfaction. The agent generates a textual summary based on user language and satisfaction, which is stored in long-term memory as historical experience to guide the next recommendation task.

[0095] The reward function construction method includes:

[0096] 1) Set a rule-based reward function to assign a certain score to each user's behavior; the feedback behavior classification and assignment rules are shown in Table 1:

[0097] Table 1

[0098] Behavior Type Weight rules Example Assignment Show positive feedback Users actively express their love (such as collecting, rating ≥ 4 stars, purchasing) 5~10 Implicit positive feedback High user interaction with items (such as clicks, swipes, and dwell time) 0~5 Neutral feedback The user completes the browsing but does not interact further (e.g., only an impression but no click) 0.5 Explicit negative feedback The user clearly expresses dissatisfaction (e.g., rating ≤ 2 stars, reporting, canceling favorites) -5~-1 No feedback The user did not take any action on the recommended list 0

[0099] Therefore, users For items The feedback evaluation calculation formula can be shown as follows:

[0100] ;

[0101] in, Indicates the number of explicit positive feedbacks; Indicates the user's Get the first exposure time, unit: seconds, ; Indicates the user's The number of days from the last explicit negative feedback time to the current time, Represents a user For items Feedback evaluation assignment.

[0102] 2) Take the average of the total scores of multiple behaviors, determine which numerical range the average falls into, and obtain the user satisfaction level corresponding to the range. Therefore, the formula for calculating user satisfaction is as follows:

[0103] ;

[0104] in, Represents a user A collection of all items of historical interaction, Represents a user The satisfaction level and feedback level are shown in Table 2:

[0105] Table 2

[0106] Satisfaction range grade Business Implications Triggering Strategy [0.8,1.0] A-level Users are highly satisfied and the recommendation results are accurate Strengthen the recommendation of related items and increase exposure weight [0.6,0.8) Class B User satisfaction, but room for improvement Optimize recommendation diversity and introduce slight adjustments [0.4,0.6) C-level User satisfaction is average Trigger the re-ranking mechanism to reduce the weight of low-scoring items <0.4 D-Class The user is not satisfied and the recommendation results need to be adjusted urgently Trigger the re-ranking mechanism to reduce the weight of low-scoring items

[0107] On the other hand, this embodiment also provides an intelligent agent application system for large models and intelligent recommendations. The intelligent agent is implemented based on the above method. The specific workflow of the intelligent agent application is as follows: Figure 4 As shown, the application system includes the following modules:

[0108] User request module, used to receive user requests and convert the format of user requests;

[0109] The task planning module is used to decompose and plan tasks based on the formatted user requests and obtain task information;

[0110] The intelligent recommendation module is used to input task information into the agent and output a recommendation list and recommendation explanation;

[0111] The memory storage module is used to record the agent's work log, including previous behaviors and external knowledge, as well as interaction records with users.

[0112] Furthermore, the system receives requests and outputs recommendation results through voice interaction with users.

[0113] Specifically, the voice interaction function includes: receiving user voice, identifying keywords based on a small model, and triggering whisper processing to identify voice content; based on the LLaMA 3.1 model, providing feedback on the voice content and executing user instructions.

[0114] like Figure 3 As shown, this example uses the pyaudio library to record an audio file in WAV format. During the recording process, parameters such as the audio block size, audio format, channels, sampling rate, recording duration, and saved file name are set. For example, the audio block size is set to 1024, the audio format is 16-bit, mono, the sampling rate is 44.1kHz, the recording duration is 5 seconds, and the recorded audio file is saved as "output.wav".

[0115] Call the whisper model to transcribe the recorded audio file into text. This step is achieved by loading the whisper model and calling its transcribe method. The transcribed text can be used for subsequent natural language processing and command parsing.

[0116] Deploy the model locally through LM Studio and enable the API service. It can receive user voice commands and convert them into text for processing. Finally, it provides feedback to the user based on the model output and completes the user's command.

[0117] Furthermore, the system adopts the form of a virtual digital human, which is designed and generated through 3Dmax by collecting user voice information and body information.

[0118] Specifically, users are first required to enter key body parameters such as gender, age, height, weight, and BMI. These parameters can be entered through voice and text input, ensuring user convenience. The entered data is then stored in a user information database. The system extracts the user's body parameters from the database and uses them to accurately match a model in a 3D model library to select a model that matches the user's characteristics. This model is then transferred to the front-end page for display. Using three.js, the system loads an external 3D model file that matches the user's data and renders it into a high-quality scene. The model's size dynamically adjusts based on the browser window to ensure optimal display on different devices. Furthermore, the model's specific characteristics, such as body shape and proportions, are adjusted in real time based on the user's input data. To enhance the user experience, various interactive features have been added, such as model rotation, allowing users to observe the model from all angles. Furthermore, the useMemo hook of the React framework effectively avoids unnecessary re-rendering, significantly improving rendering performance and responsiveness.

[0119] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for constructing an intelligent agent for large models and intelligent recommendations, characterized in that: include: Building a basic intelligent agent based on a large language model based on prompt engineering; Building a memory mechanism and a recommendation mechanism based on the Milvus vector database and the large language model, and building a feedback learning mechanism based on the recommendation mechanism; Embed the memory mechanism, recommendation mechanism, and feedback learning mechanism into the basic intelligent agent to construct an intelligent agent, wherein the input of the intelligent agent is user information and the output is a recommendation list and a recommendation explanation; Constructing the memory mechanism includes: Building a long-term memory based on the Milvus vector database, and building a short-term memory based on the context of the large language model, wherein building the long-term memory includes: Acquire user and item data and preprocess them, wherein the user and item data includes user data, item data, and user-item interaction data; Perform feature extraction on the preprocessed user and item data, and vectorize the extracted features as well as the agent's behavior and results in each recommendation task to obtain vector data; Storing the vector data in the Milvus vector database and creating an index to complete long-term memory storage; Among them, data preprocessing is performed on user data, item data, and user-item interaction data, including noise removal and missing value processing; Feature extraction of preprocessed user and item data includes: User activity calculation: ； in, Represents a user A collection of Indicates the An interactive collection of items, Indicates the number of items; Matrix decomposition: Decompose the user-item rating matrix R into two low-dimensional matrices and , is the user's potential factor, The latent factors of items are expressed as explicit ratings through user historical behaviors to mine the matching relationship between user interests and item attributes: ； in, represents the transposed matrix of the latent factor of the i-th item, is the user-item rating matrix; Participle: Split the original text content into atomic tokens: ； Among them, tokenize represents the word segmentation method provided by BertTokenizer; text represents the original text content; Indicates the tokens, ; Map the token to its corresponding integer ID: ； Where dict represents the vocabulary list after BertTokenizer pre-training; [CLS] represents the beginning of sentence marker; [SEP] represents the end of sentence / delimiter marker; coding: Calculate the initial hidden state : ； in, represents the learnable weight matrix; represents the learnable bias term; P represents the position encoding; S represents the segment embedding, Indicates the integer ID after token mapping; Extracting semantic features from local to global through multi-layer Transformer encoder : ； in, represents the global layer number, express Semantic features of the layer, express The semantic features of the layer, LN represents the layer normalization layer; Att represents the self-attention mechanism, which captures the global dependencies of all positions in the sequence; FNN represents the feedforward network, which performs nonlinear transformation on the results of layer normalization and learns complex local feature interactions; The method for performing the vectorization processing is: ； in, Represents a discrete category index; is the weight matrix, represents the eigenvector; Building the recommendation mechanism includes: Utilize the nearest neighbor search function of the Milvus vector database to filter the top n recommendations with the highest similarity to the user's vector data; Sorting and filtering the recommendations to obtain a recommendation list, and generating recommendation explanations based on the user information and the recommendation list; Among them, using the neighbor search function of the Milvus vector database to filter the top n recommendations with the similarity to the user's vector data includes: ； Sorting and filtering the recommendations, obtaining a recommendation list, and generating a recommendation explanation based on the user information and the recommendation list include: ； in, Is the activation function; represents CNN feature extraction; Represents the outer product operation of user-item embedding, Indicates the user's click probability or rating of the item; Constructing the feedback learning mechanism includes: Collecting user feedback on the recommendation list; The feedback behavior is evaluated through a reward function to obtain user satisfaction, store it as historical experience, and incorporate it into the next recommendation task; The reward function assigns a value to the user's feedback behavior, takes the average of several behavior assignments, and determines the user satisfaction corresponding to the average value through a preset satisfaction interval; Among them, assign values ​​to user feedback behaviors: ； in, Indicates the number of explicit positive feedbacks; Indicates the user's Get the first exposure time, unit: seconds, ; Indicates the user's The number of days from the last explicit negative feedback time to the current time, Represents a user For items Feedback evaluation assignment; Calculate user satisfaction: ； in, Represents a user A collection of all items of historical interaction, Represents a user satisfaction.

2. An intelligent agent application system for large models and intelligent recommendations, based on the intelligent agent constructed by the method described in claim 1, characterized in that: include: A user request module is used to receive user requests and convert the format of the user requests; The task planning module is used to decompose and plan tasks based on the formatted user requests and obtain task information; An intelligent recommendation module, configured to input the task information into an intelligent agent and output a recommendation list and a recommendation explanation; The memory storage module is used to record the agent's work log, including previous behaviors and external knowledge, as well as interaction records with users.

3. The intelligent agent application system for large models and intelligent recommendations according to claim 2 is characterized in that: The system receives requests and outputs recommendation results through voice interaction with users.

4. The intelligent agent application system for large models and intelligent recommendations according to claim 2 is characterized in that: The system adopts the form of a virtual digital human, which is designed and generated through 3Dmax by collecting user voice information and body information.

Images