Cloud tourism personalized route recommendation method based on multi-agent collaboration and RAG retrieval

Abstract

The invention provides a cloud tourism personalized route recommendation method based on agent collaboration and RAG retrieval. The method comprises the steps that a user collects guide input query data of an agent according to user preferences to obtain query information; based on the query information, obtaining retrieval information through an intelligent RAG retrieval model and a large language LLM model in the intelligent RAG retrieval model; obtaining a plurality of high-quality personalized routes based on the retrieval information; and the user selects one high-quality personalized route as a cloud tourism personalized route. According to the method, human team decision making can be simulated through multi-agent division cooperation, the recommendation process is accelerated, the intelligent RAG retrieval model can be combined with the knowledge graph to answer complex problems such as cloud tourism route planning, and a mature cloud tourism personalized route scheme is provided for a user in a short time.

Description

Technical Field

[0001] The present invention belongs to the technical field of cloud tourism route recommendation, and specifically relates to a cloud tourism personalized route recommendation method based on multi-agent collaboration and RAG retrieval. Background Art

[0002] Traditional tourism, typically arranged through travel agencies, offers significant advantages in terms of convenience, cost control, and safety. However, drawbacks such as rigid itineraries, fluctuating service quality, and a lack of personalization are driving more tourists toward emerging options like independent travel, customized tours, and smart tourism. The integration of traditional tourism with emerging technologies may address some of its shortcomings and maintain its market value.

[0003] Furthermore, the accelerated pace of life, fragmented time, and busy office workers make it difficult for people to find leisure time for travel. Furthermore, traditional tourism is often prohibitive due to high costs, itinerary expenses, and limited travel time and space. Physical discomfort during travel can further detract from the experience. To enhance travel experiences, a new model, cloud tourism, has emerged. Cloud tourism transcends the constraints of time and space, offering a whole new travel experience through low cost, high convenience, and cutting-edge technology.

[0004] Patent publication number CN117273860A discloses a method for processing conversational recommendations and feedback based on big data. The method identifies a user's cloud tourism destinations and constructs a cloud tourism model based on big data technology. Feature analysis is performed on the acquired cloud tourism record data to generate the user's travel feature data. Route planning is performed based on the feature data to generate a cloud tourism route. Cloud tourism is presented to the user based on the cloud tourism route, and user interaction session information is received. Feedback analysis is performed, and based on the feedback analysis results, the cloud tourism model and cloud tourism route are tracked and optimized. While this patent document enables autonomous route planning, route tracking and optimization are relatively weak, and the flexibility in adjusting cloud tourism routes is limited, which in turn creates inconvenience for the user experience.

[0005] To address the issue of flexible cloud tourism routes, previous route recommendation methods employed a crude division of labor and superficially collected preferences. These methods often relied on a single entity to handle the entire process, making it difficult to master every aspect and uncover niche preferences. Furthermore, cloud tourism information is highly variable, and updates are slow to respond to temporary changes in user needs. Furthermore, factors such as the adjustment of tourist attraction opening hours during peak seasons make it difficult to update original plans in a timely manner.

[0006] Furthermore, cloud tourism's scalability is limited, making it difficult to integrate emerging resources as new tourism trends emerge. Low processing efficiency leads to an overwhelmed response to a large number of requests and delayed recommendations. Recommendation quality is poor, with shallow personalization and difficulty uncovering hidden preferences. Cloud tourism's information extraction is crude, making it difficult to understand semantic content. Cloud tourism is isolated from external knowledge sources, making it difficult to connect multiple aspects of information about attractions. It lacks real-time and dynamic adaptability, making it difficult to incorporate and integrate emerging tourism products due to its rigid architecture, failing to meet users' desire for novel experiences. Furthermore, due to insufficient personalization, search strategies cannot be deeply customized based on multi-dimensional factors. Users may receive inappropriate attraction and route recommendations. Cloud tourism is also poor at tapping into users' potential needs, making inefficient use of relevant data and difficult to customize unique routes. Summary of the Invention

[0007] The purpose of the present invention is to solve the problems of poor flexibility and limited scalability of cloud tourism route recommendation in the existing technology, and to provide a cloud tourism personalized route recommendation method based on multi-agent collaboration and RAG retrieval. The user preference collection agent more accurately matches the user's multiple preferences. The intelligent RAG retrieval model mines personalized routes from a large amount of cloud tourism route information. The route planning agent recommends personalized routes based on the similar initial needs of different users. The evaluation and optimization agent tracks the changes in the tourism market in real time and adjusts the recommended routes in time. It not only realizes the division of labor and collaboration of multiple agents to simulate human team decision-making and accelerate the recommendation process, but also enables the intelligent RAG retrieval model to combine with the knowledge graph to answer complex problems such as cloud tourism route planning, and provide users with mature cloud tourism route solutions in a short time.

[0008] To achieve the above objectives, the technical solutions provided by the present invention are:

[0009] A personalized route recommendation method for cloud tourism based on agent collaboration and RAG retrieval, including:

[0010] Step 1: The user preference collection agent guides the user to input query data to obtain query information;

[0011] Step 2: Based on the query information, the intelligent RAG retrieval model and the large language LLM model therein are used to obtain retrieval information; specifically, the query information is input into the intelligent RAG retrieval model to determine whether it needs intelligent reconstruction to obtain a judgment result, and new query information is obtained based on the judgment result; the new query information is input into the intelligent RAG retrieval model, and the large language LLM model therein, based on internal data or associated external data, answers the new query information to obtain a retrieval result; the intelligent RAG retrieval model generates and outputs the retrieval information based on the retrieval result;

[0012] Step 3: Acquire multiple high-quality personalized routes based on the search information; specifically, plan multiple cloud travel routes based on the search information using a route planning agent, evaluate the cloud travel routes using an evaluation and optimization agent, and then select multiple high-quality personalized routes from the multiple cloud travel routes based on the evaluation results, and then recommend the high-quality personalized routes to the user;

[0013] Step 4: The user selects one of the high-quality personalized routes as the cloud tourism personalized route; if the user changes the attractions on the cloud tourism personalized route during the cloud tourism, return to step 2 and readjust to a new personalized tourism route.

[0014] As a further limitation of the present invention, the user preference collection agent in step 1 may be used in the following specific applications:

[0015] The user preference collection agent guides the user to input query information through an interactive interface to collect the travel preference information input by the user; wherein the travel preference information includes the type of attractions that the user likes and the estimated time of play; the types of attractions include historical sites, natural scenery, and modern urban landscapes;

[0016] The user preference collection agent guides the user to input query information through an interactive interface and collects the user's travel preference information; the travel preference information will be input into the intelligent RAG retrieval model as query information; wherein, the travel preference information includes but is not limited to the favorite attraction type and estimated tour time.

[0017] As a further limitation of the present invention, the specific application of the intelligent RAG retrieval model in step 2 includes:

[0018] Inputting the query information into the intelligent RAG retrieval model to determine whether intelligent reconstruction is required to obtain a determination result, wherein the intelligent reconstruction includes:

[0019] Structuring the travel preference information input by the user through the interactive interface into the query information, and transmitting the query information to the intelligent RAG retrieval model;

[0020] After receiving the query information, the intelligent RAG retrieval model uses internal data or searches external data to retrieve the retrieval information.

[0021] As a further limitation of the present invention, the specific application of the intelligent RAG retrieval model in step 2 also includes:

[0022] The large language LLM model in the intelligent RAG retrieval model answers the new query information and retrieves data based on its internal data or external data associated with the internal data to obtain retrieval results; specifically:

[0023] In the process of searching various resources through internal and external data, the intelligent RAG retrieval model not only searches existing scenic spot document data, but also searches multi-source information related to the scenic spot document data; the multi-source information includes tourism websites, geographic information systems and official data of the tourism bureau.

[0024] As a further limitation of the present invention, the structure of the intelligent RAG retrieval model in step 2 includes:

[0025] Adding a resource retrieval agent to the large language LLM model for reference by the intelligent RAG retrieval model. The resource retrieval agent actively thinks and makes decisions during the information retrieval and information generation process to generate accurate query predictions.

[0026] The large language LLM model in the intelligent RAG retrieval model selects qualified candidate resources from a large number of travel route resources as the retrieval information.

[0027] As a further limitation of the present invention, the training process of the intelligent RAG retrieval model in step 2 includes:

[0028] The intelligent RAG retrieval model is fine-tuned using a Transformer-based bidirectional encoder representation model to adapt it to semantic understanding in the tourism field. Specifically:

[0029] The training data required for fine-tuning the intelligent RAG retrieval model process covers a large number of tourism-related questions and corresponding resource description text pairs, allowing the intelligent RAG retrieval model to learn how to accurately extract information from the knowledge graph of external data based on the questions.

[0030] As a further limitation of the present invention, the route planning agent in step 3 may be used in the following specific applications:

[0031] The route planning agent is applied to plan the cloud travel route according to the search information, specifically:

[0032] (1) The route planning agent plans multiple initial cloud tourism routes based on the search information;

[0033] (2) The route planning agent uses a genetic algorithm to iteratively optimize the multiple initial cloud tourism routes by selection, crossover or mutation to plan a cloud tourism route; wherein the iterative optimization is specifically as follows:

[0034] Step a: Calculate the fitness of each of the initial cloud tourism routes; specifically, the fitness is calculated by scoring each of the initial cloud tourism routes based on the flow of visitors to tourist attractions, the time cost required for tourism, and the degree of user interest compliance as criteria for measuring the quality of cloud tourism routes;

[0035] Step b: performing multiplication or evolution processing on the initial cloud tourism routes one by one by selection, crossover or mutation to obtain a plurality of new cloud tourism routes; specifically, the multiplication or evolution process is to perform crossover or mutation processing on the initial cloud tourism routes based on fitness;

[0036] Step c: Calculate the fitness of each of the new cloud travel routes, and select and retain travel routes with high fitness for further reproduction and evolution;

[0037] Step d: Repeat steps a to c to iterate and optimize until the termination condition is met to obtain the cloud travel route.

[0038] As a further limitation of the present invention, the evaluation and optimization agent in step 3 may be used in the following specific applications:

[0039] (1) The evaluation and optimization agent performs route evaluation on a plurality of the cloud tourism routes; specifically, the evaluation and optimization agent predicts the satisfaction of the cloud tourism routes using a user satisfaction prediction model, scores the cloud tourism routes using historical user data or expert data, optimizes and adjusts the cloud tourism routes based on the satisfaction prediction results and the scoring results, and re-evaluates the cloud tourism routes based on the optimization and adjustment results;

[0040] (2) the evaluation and optimization agent selects a plurality of high-quality personalized routes from the plurality of cloud tourism routes according to the evaluation results;

[0041] (3) The evaluation and optimization agent recommends multiple high-quality personalized routes to the user.

[0042] As a further limitation of the present invention, the personalized cloud tourism route recommendation in step 4 includes:

[0043] Step 41: The interactive feedback agent displays the multiple high-quality personalized routes output in step 3;

[0044] Step 42: The user selects one of the high-quality personalized routes as the personalized cloud tourism route. If a certain scenic spot in the personalized cloud tourism route needs to be replaced, the process returns to step 2 and restarts to optimize the personalized cloud tourism route based on the scenic spot to be replaced.

[0045] As a further limitation of the present invention, the personalized cloud tourism route recommendation in step 4 further includes:

[0046] Step 43: After the user completes the cloud tourism experience based on the personalized cloud tourism route, the evaluation and feedback agent obtains the user's experience information of participating in this cloud tourism. Based on the experience information, the evaluation and feedback agent provides corresponding optimization information to the service provider of the corresponding tourism model or tourism route on the personalized cloud tourism route.

[0047] The advantages of the present invention are:

[0048] This invention uses a user preference collection agent to more accurately match diverse user preferences. An intelligent RAG retrieval model extracts personalized routes from a large amount of cloud tourism route information. A route planning agent recommends differentiated personalized routes based on the similar initial needs of different users. An evaluation and optimization agent tracks tourism market changes in real time and promptly adjusts recommended routes, allowing users to access the latest scenic spot information and respond to emergencies such as weather and equipment anomalies. This invention not only enables multi-agent collaboration to simulate human team decision-making and accelerate the recommendation process, but also enables the intelligent RAG retrieval model to combine knowledge graphs to answer complex questions such as cloud tourism route planning, providing users with comprehensive personalized cloud tourism route solutions in a short period of time.

[0049] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0051] Figure 1 : A flow chart of a cloud tourism personalized route recommendation method based on agent collaboration and RAG retrieval provided by the present invention;

[0052] Figure 2 : The multi-agent collaborative architecture diagram provided by the present invention;

[0053] Figure 3 : The route planning agent provided by the present invention adopts the genetic algorithm to carry out the reproduction and evolution process diagram;

[0054] Figure 4 : Schematic diagram of the intelligent RAG search provided by the present invention;

[0055] Figure 5 : A structural diagram of the Transformer-based bidirectional encoder representation model (BERT) fine-tuning model in the example provided by the present invention;

[0056] Figure 6 :The present invention provides Figure 5 Block diagram of the Transformer Encoder module of the Bidirectional Encoder Representation Model (BERT) fine-tuning model structure. DETAILED DESCRIPTION

[0057] The following describes in detail embodiments of the present invention. The embodiments are exemplary and intended to explain the present invention, but are not to be construed as limiting the present invention.

[0058] See also Figure 1 The embodiment of the present invention provides a cloud tourism personalized route recommendation method based on agent collaboration and RAG retrieval, including:

[0059] Step 1: The user preference collection agent guides the user to input query data to obtain query information.

[0060] In some preferred embodiments, the user preference collection agent in step 1 of the present invention includes the following: the user preference collection agent guides the user to enter query information through an interactive interface to collect the user's travel preference information; the travel preference information includes preferred attraction types and estimated visit duration; attraction types include historical sites, natural scenery, and modern urban landscapes; the user preference collection agent guides the user to enter query information through an interactive interface to collect the user's travel preference information; the travel preference information is input as query information into the intelligent RAG retrieval model; the travel preference information includes but is not limited to preferred attraction types and estimated visit duration. For example, if a user specifies a preferred attraction type, such as "historical sites," the model expands this into a series of related keywords, such as "museums," "ancient buildings," and "historical sites." The model then leverages the semantic understanding capabilities of the large language model (LLM) to analyze the user's specific requirements for "historical sites," such as whether they require historical sites from a specific period or whether they have specific cultural background requirements. It also combines internal and external data to retrieve scenic spot document data, tourism website information, geographic information system data and official data related to "historical sites", and generates search information including historical sites and attractions based on the search results.

[0061] If the user specifies an estimated tour duration, such as "one hour," the model plans a reasonable tour route based on the "one-hour" time limit, taking into account the opening hours of attractions and recommended visit times. Leveraging the optimization capabilities of the large language LLM model, the recommended route is ensured to be feasible and maximize the user's travel experience. It also retrieves route and attraction information relevant to the "one-hour" tour duration, including recommended visit times and schedules. Based on the search results, the model generates search information suitable for a one-hour tour.

[0062] The intelligent RAG retrieval model also generates and outputs retrieval information based on the above processing, including: tourist attractions and routes related to user preference information, and may also include relevant travel suggestions, precautions, etc., to help users better plan their cloud travel itineraries.

[0063] Step 2: Based on the query information, the search information is obtained through the intelligent RAG search model and the large language LLM model within it; specifically, the query information is input into the intelligent RAG search model to determine whether it needs to be intelligently reconstructed to obtain a judgment result, and new query information is obtained based on the judgment result; the new query information is input into the intelligent RAG search model, and the new query information is answered through the large language LLM model within it based on internal data or associated external data to obtain a search result; the intelligent RAG search model generates and outputs search information based on the search results.

[0064] It should be noted that the traditional RAG retrieval model and the large language LLM model are separate models. The traditional RAG retrieval model is responsible for search enhancement, while the large language LLM model is responsible for answer generation. The intelligent RAG retrieval model described in the embodiment of the present invention integrates the large language LLM model into the retrieval process, applying it to both the retrieval and answer generation processes of the model.

[0065] The above judgment criteria of the embodiment of the present invention include: (1) The intelligent RAG retrieval model first evaluates the clarity of the query information. a) If the query information contains vague or non-specific words, or the structure of the query information is unclear, the intelligent RAG retrieval model will determine that the query information needs to be intelligently reconstructed; for example, if the query information entered by the user is "I want to go to an interesting place", the model will consider the query information too vague and need to further clarify the user's intention; the model further evaluates the completeness of the query information. b) If the query information lacks key information, such as the specific type of attraction, travel time or budget, the model will determine that the query information needs to be intelligently reconstructed. For example, if the query information entered by the user is "I want to go to a historical site", but does not mention the specific travel time or location, the model will consider the query information incomplete and need to be supplemented with more information. The model will also evaluate the semantic consistency of the query information. c) If the query information contains contradictory or inconsistent information, the model will determine that the query information needs to be intelligently reconstructed. For example, if the query information entered by the user is "I want to go to a historical site, but preferably a modern urban landscape", the model will consider that the query information has semantic contradictions and needs to further clarify the user's intention. (2) The model evaluates the retrievability of the query information. If a keyword or phrase in a query is difficult to find matching resources in existing data sources, the model will determine that the query requires intelligent reconstruction. For example, if the user enters the query "I want to visit a hidden attraction that only locals know about," the model will determine that the query is less searchable and requires further clarification of the user's intent.

[0066] In some preferred embodiments, the specific application of the intelligent RAG retrieval model in step 2 of the embodiment of the present invention includes:

[0067] (1) Inputting the query information into the intelligent RAG retrieval model to determine whether intelligent reconstruction is required to obtain a judgment result, wherein the intelligent reconstruction includes: structuring the travel preference information input by the user through the interactive interface into query information, and transmitting the query information to the intelligent RAG retrieval model; after the intelligent RAG retrieval model receives the query information, it uses internal data or searches external data to retrieve the retrieval information.

[0068] (2) The large language LLM model in the intelligent RAG retrieval model answers new query information and retrieves data based on its internal data or external data associated with the internal data to obtain retrieval results. Specifically, the intelligent RAG retrieval model searches various resources through internal and external data. It not only searches existing scenic spot document data, but also searches multi-source information related to scenic spot document data. The multi-source information includes tourism websites, geographic information systems and official data of tourism bureaus.

[0069] For example, see Figure 2 In actual application, the user first uses the interactive interface provided by the user preference collection agent to collect the user's input travel preference information, including but not limited to the favorite types of attractions (historical sites, natural scenery, modern urban landscapes, etc.), estimated travel time period, etc.; then the above travel preference information is structured and passed to the cloud tourism resource retrieval agent. The cloud tourism resource retrieval agent uses the retrieval technology in the trained intelligent RAG retrieval model based on the travel preference information to connect with external tourism databases, knowledge graphs and other network information, generate precise query statements based on user needs, screen out qualified candidate resources from a large number of travel route resources, and then package the details of the candidate resources and send them to the route planning agent.

[0070] Please continue reading Figure 2 , the route planning agent integrates all candidate resources and uses Figure 3 The genetic algorithm shown in this paper treats all tourist routes as "individual organisms," with each route as a small organism. It first treats all possible routes as a large "biological population," using criteria such as visitor volume, time cost, and user interest compatibility as metrics for route quality (i.e., fitness). Each route is scored, allowing the routes to "reproduce" and "evolve." Specifically, there are two methods for "reproduction" and "evolution": one is "crossover" (for example, given two good routes, one with attractions ABC and the other with attractions DEF, the route planning agent swaps some attractions between the two routes, potentially generating a new route, such as attractions AEC); the other is "mutation" (for example, if a small organism suddenly mutates, one route has attractions ABC, and the route planning agent changes the order to BAC). New possibilities are generated through the genetic algorithm, and these newly generated routes are scored by the evaluation and optimization agent using pre-set criteria. The good routes are retained for further "reproduction" and "evolution," while the bad ones are eliminated, thus generating multiple routes. It's important to note that the route planning agent needs to consider practical factors such as attraction opening hours during planning to ensure the feasibility of the travel route. Furthermore, the evaluation and optimization agent evaluates cloud travel routes based on pre-set evaluation metrics to obtain preliminary routes. From these preliminary routes, it selects two or three high-quality routes. If a user requests a change of route because they've already visited a particular attraction, the process is immediately restarted for targeted optimization. The interactive feedback agent provides a visual representation of the preliminary route and collects user feedback. If the user is dissatisfied, the process is re-triggered to make targeted improvements to the preliminary route.

[0071] In some preferred embodiments, further, the intelligent RAG retrieval model in step 2 of the embodiment of the present invention has a structure including: adding a resource retrieval agent on the basis of the big language LLM model for the intelligent RAG retrieval model to reference, the resource retrieval agent actively thinks and makes decisions during the information retrieval and information generation process, and generates accurate query estimates; the big language LLM model in the intelligent RAG retrieval model selects qualified candidate resources from a large number of travel route resources as retrieval information.

[0072] The above-mentioned precise query prediction of the embodiment of the present invention includes: (1) The natural language processing module conducts an in-depth analysis of the user's query information and extracts the key entities, concepts and intentions therein. For example, if the user query is "I want to go to a place with historical sites", the module will identify "historical sites" as a key entity, and "go" indicates that the user has the intention to travel. (2) Based on the extracted key entities and concepts, the intelligent RAG queries internal and external knowledge bases and retrieves relevant information from multiple data sources. This includes obtaining evaluations and recommendations of scenic spots from travel websites, obtaining the geographical coordinates and surrounding facilities of scenic spots from geographic information systems, obtaining the opening hours of scenic spots from official data of tourism bureaus, etc., and obtaining related entities, attributes and relationships. For example, searching for scenic spots related to "historical sites", their geographical locations, historical backgrounds and other information in the knowledge base. (3) Combining the information obtained in the above steps, the resource retrieval agent generates precise query predictions. These query predictions include not only scenic spots directly related to the user's query, but also related routes, precautions, etc., to provide comprehensive travel planning support. (4) Obtaining qualified candidate resources. The Big Language LLM model performs semantic understanding and logical analysis on the user's query information to clarify the user's needs and preferences. For example, if the user's query is "I want to go to a place with historical sites", the model will understand that the user wants to find tourist attractions and related routes that contain historical sites. Based on the understanding of the query information, the intelligent RAG formulates retrieval strategies for different data sources, including keyword extraction, query expansion, and filtering conditions. For example, the keyword "historical sites" is extracted and the query scope is expanded to related types of attractions, such as "museums" and "ancient buildings". The retrieved results are evaluated and sorted to provide users with the best retrieval information based on relevance and importance. For example, the retrieved historical sites are sorted based on factors such as the rating of the attractions and user reviews. (5) Combining the information obtained in the above steps, the Big Language LLM model generates a list of candidate resources that meet the conditions. These candidate resources include tourist attractions, routes, and related services related to the user's query. For example, the generated candidate resource list may include historical sites, as well as tourist routes and related service information that include these attractions.

[0073] For example, when the resource search agent of the cloud tourism embodiment of the present invention performs query analysis, since the user input forms are various and the questions asked by each person are different, please refer to Figure 4 The intelligent RAG retrieval model in this instance will determine whether the personalized route recommendation question queried by the user needs to be intelligently reconstructed. If not, the user question will be directly answered using the large language LLM model; if necessary, the original user query question will no longer be directly copied, but will undergo careful analysis and reconstruction to transform the vague or complex question into a more precise and searchable structural form; the question after intelligent reconstruction will then be sent to the large language LLM model for answer.

[0074] For further information, please refer to Figure 4 The intelligent RAG retrieval model of this embodiment has multi-source data retrieval capabilities, namely flexible data retrieval capabilities, including real-time data (dynamically adjusted based on the current context) and internal documents (precisely matching the organization's internal knowledge with external data sources, and acquiring the latest information from the internet in real time). The Large Language Model (LLM) analyzes the received question and determines whether it can answer the user's question based on currently known internal documents. If so, it directly provides an answer based on the content of the internal document. If no information can be found that can answer the question, additional data sources (including user context data and external network data) are needed to more comprehensively answer the user's question. (For example, a user may ask, "How does the weather affect tourist attractions?") Traditional RAG retrieval models may not be able to answer this question based on existing internal documentation. When encountering unanswerable questions, traditional RAG retrieval models often simply tell the user "no solution." After obtaining the answer, the intelligent RAG retrieval model in this example will judge the answer information to determine whether it is correct or relevant to the question. If it is correct or highly relevant to the question, the answer is returned to the user. If it is incorrect or less relevant, the question is rewritten and analyzed and queried again. The previous step is repeated, actively seeking supplementary information sources and attempting to regenerate a better answer until the answer is correct for the user's question. It should be noted that the intelligent RAG retrieval model in this example is not satisfied with simply providing a single answer. Instead, it continuously optimizes through multiple rounds of iteration, generates multiple candidate answers, evaluates the accuracy and relevance of each answer, and re-queries or adjusts the generation strategy when necessary.

[0075] The Big Language Model (LLM) analyzes incoming questions to determine whether they can be answered based on currently available internal documentation. If relevant information can be found, the answer is provided directly based on the internal documentation. If no relevant information can be found, additional data sources (such as external data or documents, including user context and external network data) are needed to provide a more comprehensive answer. Furthermore, the intelligent RAG retrieval model in this example identifies the need, intelligently retrieves and integrates the required weather information from external network data, performs comprehensive analysis, and generates insightful answers.

[0076] In some preferred embodiments, further, the training process of the intelligent RAG retrieval model in step 2 of the embodiment of the present invention includes:

[0077] A Transformer-based bidirectional encoder representation model is used to fine-tune the intelligent RAG retrieval model, making it adaptable to semantic understanding in the tourism field. Specifically, the training data required for fine-tuning the intelligent RAG retrieval model covers a large number of tourism-related questions and corresponding resource description text pairs, allowing the intelligent RAG retrieval model to learn how to accurately extract information from the knowledge graph of external data based on the questions.

[0078] For example, see Figure 5 and Figure 6 In this embodiment of the present invention, the intelligent RAG retrieval model is trained using a Transformer-based bidirectional encoder representation model (BERT) to fine-tune the model to adapt it to the semantics of the tourism domain. Specifically, the training data includes a large number of tourism-related question and corresponding resource description text pairs, allowing the intelligent RAG retrieval model to learn how to accurately extract information from the knowledge graph based on the questions. Furthermore, to ensure that the intelligent RAG retrieval model remains up-to-date, a rigorous and efficient monitoring system is established using web crawler technology. Regular internet inspections are conducted to accurately capture every dynamic change in the tourism resource landscape. This allows for the monitoring of new special tourism route development projects launched by travel agencies, newly discovered niche scenic spots, and changes in open routes at tourist attractions due to policy adjustments or seasonal changes. Once these changes are detected, the external system initiates an update mechanism to optimize and update the knowledge graph and retrieval index in real time, ensuring that every recommendation provided to users is highly timely, allowing users to obtain the latest travel information and enjoy a better travel planning experience.

[0079] In practical applications, a large number of travel-related questions and corresponding resource description text pairs can be collected from various channels, such as travel forums, Q&A websites, and travel guide websites. The collected data is cleaned to remove duplicate, irrelevant, and low-quality data. The text is then preprocessed, including token segmentation, stop word removal, and stemming, to improve model training efficiency and accuracy. The training data is then annotated to clarify the key information in the resource description text corresponding to each question. For example, key information such as attraction type, travel time, and budget in the question is annotated, along with the corresponding information in the resource description text, such as attraction, route, and service.

[0080] The Bidirectional Encoder Representation (BERT) model has powerful semantic understanding capabilities and can provide good initial parameters for subsequent fine-tuning. Therefore, this embodiment of the present invention selects a pre-trained Transformer-based Bidirectional Encoder Representation (BERT) model as the basis for the intelligent RAG retrieval model. Model parameters, such as the hidden layer size, number of attention heads, and dropout rate, are set based on the characteristics and needs of the tourism sector. Questions and corresponding resource description text pairs in the training data are converted into the model input representations. Specifically, the BERT model's tokenizer is used to encode the question and resource description text, generating the corresponding input ID sequence, attention mask, and segment embedding. These are then input into the BERT model for forward propagation. The model generates corresponding semantic representations based on the input question and resource description text, and uses the cross-entropy loss function to measure the difference between the model's predicted resource description text and the actual annotated resource description text. Backpropagation is performed based on the calculated loss value to update the model parameters. The AdamW optimization algorithm is also used, with appropriate learning rates and weight decay rates to avoid overfitting and improve the model's generalization ability.

[0081] It should be noted that during training, loss and accuracy are recorded every 100 training steps. Model performance and convergence are monitored using metrics such as loss, precision, and recall. When the loss stops decreasing significantly within a certain number of training steps, the learning rate is halved to promote further model optimization. Training is stopped early if the model's performance on the validation set does not improve over several consecutive training cycles to prevent overfitting. During model training, not only traditional text data is used, but also innovative multi-source data, such as user reviews from travel websites and map data from geographic information systems, is incorporated. By integrating these multi-source data, the model can more comprehensively understand semantic information in the tourism domain, improving the accuracy and richness of retrieval results. During model fine-tuning, a multi-task learning strategy is employed to simultaneously optimize multiple related tasks, such as question classification and resource description generation. This strategy enables the model to share knowledge across different tasks, improving its overall performance and generalization. Knowledge distillation techniques are also incorporated to transfer knowledge from large pre-trained models to smaller models. Through the above training method, the model can reduce the consumption of computing resources while maintaining high performance, making it more suitable for practical applications.

[0082] Step 3: Obtain multiple high-quality personalized routes based on the search information. Specifically, a route planning agent plans several cloud travel routes based on the search information. The cloud travel routes are first evaluated by the evaluation and optimization agent. Based on the evaluation results, multiple high-quality personalized routes are screened from the multiple cloud travel routes, and then the high-quality personalized routes are recommended to the user.

[0083] In some preferred implementations, the route planning agent in step 3 of the embodiment of the present invention may be used in the following ways:

[0084] The route planning agent is used to plan the cloud travel route based on the retrieved information. Specifically:

[0085] (1) The route planning agent plans multiple initial cloud tourism routes based on the retrieved information;

[0086] (2) The route planning agent uses a genetic algorithm to iteratively optimize multiple initial cloud tourism routes by selection, crossover or mutation to plan the cloud tourism route. The iterative optimization is as follows:

[0087] Step a: Calculate the fitness of each initial cloud tourism route. Specifically, the fitness is calculated by scoring each initial cloud tourism route based on the visitor flow at tourist attractions, the time cost of travel, and the degree of user interest compliance.

[0088] Step b: performing reproduction or evolution processing on the initial cloud tourism routes one by one by selection, crossover or mutation to obtain a plurality of new cloud tourism routes; specifically, the reproduction or evolution process is to perform crossover or mutation processing on the initial cloud tourism routes based on fitness;

[0089] Step c: Calculate the fitness of each new cloud travel route, and select the travel routes with high fitness to continue reproduction and evolution;

[0090] Step d: Repeat steps a to c to iterate and optimize until the termination condition is met to obtain the cloud travel route.

[0091] For example, this embodiment transmits multiple preliminary travel routes generated by the route planning agent to the evaluation and optimization agent. Based on the user satisfaction prediction model (i.e., by analyzing user evaluation data of a large number of similar travel routes in the past, discovering the common patterns that affect user satisfaction, and then giving corresponding satisfaction prediction scores to the newly generated preliminary travel routes, this score can serve as an important reference for judging the quality of the route), the route planning agent in this example also invites some users with similar travel experiences or an expert system to score the preliminary travel routes. The user scoring results and the expert system scoring results are combined to jointly optimize and adjust the preliminary travel routes, and finally 2-3 high-quality travel routes are screened and fed back to the interactive feedback agent.

[0092] Furthermore, in some preferred embodiments, the evaluation and optimization agent in step 3 may be used to:

[0093] (1) The evaluation and optimization agent evaluates several cloud tourism routes. Specifically, the evaluation and optimization agent predicts the satisfaction of the cloud tourism routes through the user satisfaction prediction model, scores the cloud tourism routes based on historical user data or expert data, optimizes and adjusts the cloud tourism routes based on the satisfaction prediction results and the scoring results, and evaluates the cloud tourism routes again based on the optimization and adjustment results.

[0094] (2) The evaluation and optimization agent selects multiple high-quality personalized routes from several cloud tourism routes based on the evaluation results;

[0095] (3) The evaluation and optimization agent recommends multiple high-quality personalized routes to users.

[0096] Specifically, the evaluation and optimization agent described above in the embodiment of the present invention first performs a preliminary evaluation of several cloud tourism routes. This includes predicting the satisfaction of cloud tourism routes through a user satisfaction prediction model, and scoring cloud tourism routes through historical user data or expert data. The user satisfaction prediction model is based on a large amount of historical user data, analyzing past users' evaluations of similar tourism routes, and predicting the current user's satisfaction with the recommended route. The prediction results are presented in the form of scores, with higher scores indicating higher predicted satisfaction. By utilizing historical user rating data on similar routes and introducing expert professional evaluations of routes, a comprehensive score is generated for each cloud tourism route. The scoring criteria include factors such as the quality of attractions, route rationality, and time arrangement. Based on the satisfaction prediction results and scoring results, the evaluation and optimization agent performs a comprehensive analysis of each cloud tourism route.

[0097] Based on the evaluation results, the evaluation and optimization agent develops optimization strategies for each cloud tour route. These strategies include adjusting the order of attractions, adding or removing certain attractions, and other optimizations. Based on these optimization strategies, the cloud tour route is optimized. These adjustments involve replanning routes, adjusting schedules, and updating attraction information. The evaluation and optimization agent then re-evaluates the optimized cloud tour route. This involves re-using the user satisfaction prediction model to predict satisfaction and re-rating it based on historical user data or expert data. The evaluation results are compared with the pre-optimization results to analyze the effectiveness of the optimization. If the optimized route shows a significant improvement in both satisfaction prediction and rating, the optimization is considered successful. If the improvement is not significant or some aspects decline, the evaluation and optimization agent will re-develop the optimization strategy and make further adjustments. Through the close connection and iterative process between route evaluation and route optimization, the evaluation and optimization agent can continuously adjust and optimize the cloud tour route, ensuring that the final recommended route not only meets the user's preferences but also optimizes time scheduling, attraction selection, and other aspects, thereby providing the best cloud tour experience.

[0098] Step 4: The user selects a high-quality personalized route as the cloud tourism personalized route; if the user changes the attractions on the cloud tourism personalized route during the cloud tourism, return to step 2 and readjust to the new tourism personalized route.

[0099] In some preferred embodiments, the personalized cloud tourism route recommendation in step 4 of the embodiment of the present invention includes:

[0100] Step 41: The interactive feedback agent displays multiple high-quality personalized routes output in step 3;

[0101] Step 42: The user selects one of the high-quality personalized routes as the personalized cloud travel route. If a scenic spot in the personalized cloud travel route needs to be changed, the process returns to step 2 and restarts to optimize the personalized cloud travel route based on the desired scenic spot.

[0102] Step 43: After the user completes the cloud tourism experience based on the personalized cloud tourism route, the evaluation and feedback agent obtains the user's experience information of participating in this cloud tourism. Based on the experience information, the evaluation and feedback agent provides corresponding optimization information to the service provider of the corresponding tourism model or tourism route on the personalized cloud tourism route.

[0103] For example, the interactive feedback agent displays the final two to three high-quality travel routes as recommended routes in a user interface format using a combination of images and text. These recommended routes include a map-marked itinerary, photo previews of scenic spots, and a detailed itinerary schedule. If a user requests a change of itinerary because they have already visited a particular scenic spot, at least part of the process in this example's personalized route recommendation method for cloud tourism based on agent collaboration and RAG retrieval is immediately restarted to optimize the recommended route.

[0104] More specifically, after a user completes their cloud travel experience according to a recommended route, the feedback agent proactively establishes a communication channel with the user, initiating a question-and-answer exchange (for example: a) the feedback agent asks the user about their intuitive feelings during the just-concluded cloud travel experience, guiding them to share the wonderful moments or shortcomings; b) inquires whether the planning of the entire recommended cloud travel route is smooth and reasonable, and whether the transitions between attractions on the cloud travel route are abrupt or poorly connected; c) inquires from a technical perspective whether the image quality and sound effects during the cloud travel process are reasonable).

[0105] Furthermore, the feedback agent comprehensively aggregates and categorizes user reviews and submits them as reports to the corresponding service providers (e.g., merchants) along the travel route. Based on this genuine feedback, the service providers can quickly identify the problem areas and formulate targeted corrective measures to continuously improve the quality of services and products along the cloud travel journey, providing subsequent users with a more high-quality and seamless cloud travel experience.

[0106] As can be seen, the embodiments of the present invention use a user preference collection agent to more accurately match users' diverse preferences. The intelligent RAG retrieval model extracts personalized routes from a large amount of cloud tourism route information. The route planning agent recommends personalized routes based on the similar initial needs of different users. The evaluation and optimization agent tracks tourism market changes in real time and promptly adjusts recommended routes, allowing users to grasp the latest scenic spot information and respond to emergencies such as abnormal weather and equipment. In addition, the embodiments of the present invention not only implement multi-agent division of labor and collaboration to simulate human team decision-making and accelerate the recommendation process, but also enable the intelligent RAG retrieval model to combine with knowledge graphs to answer complex questions such as cloud tourism route planning, providing users with mature cloud tourism route solutions in a short period of time.

[0107] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present invention, and these modifications or replacements should all be included in the scope of protection of the present invention.

Claims

1. A personalized route recommendation method for cloud tourism based on agent collaboration and RAG retrieval, characterized by: include: Step 1: The user preference collection agent guides the user to input query data to obtain query information; Step 2: Based on the query information, the intelligent RAG retrieval model and the large language LLM model therein are used to obtain retrieval information; specifically, the query information is input into the intelligent RAG retrieval model to determine whether it needs intelligent reconstruction to obtain a judgment result, and new query information is obtained based on the judgment result; the new query information is input into the intelligent RAG retrieval model, and the large language LLM model therein, based on internal data or associated external data, answers the new query information to obtain a retrieval result; the intelligent RAG retrieval model generates and outputs the retrieval information based on the retrieval result; Step 3: Acquire multiple high-quality personalized routes based on the search information; Specifically, a route planning agent plans several cloud travel routes based on the search information, and an evaluation and optimization agent first evaluates the cloud travel routes. Then, based on the evaluation results, a plurality of high-quality personalized routes are screened from the plurality of cloud travel routes, and the high-quality personalized routes are then recommended to the user. Step 4: The user selects one of the high-quality personalized routes as the cloud tourism personalized route; If the user changes the scenic spots on the personalized cloud travel route during the cloud travel, the process returns to step 2 and readjusts to a new personalized travel route.

2. A cloud tourism personalized route recommendation method based on agent collaboration and RAG retrieval according to claim 1, characterized in that: The user preference collection agent in step 1 may be used in the following specific applications: The user preference collection agent guides the user to input query information through an interactive interface to collect the travel preference information input by the user; wherein the travel preference information includes the type of attractions that the user likes and the estimated time of play; the types of attractions include historical sites, natural scenery, and modern urban landscapes; The user preference collection agent guides the user to input query information through an interactive interface and collects the user's travel preference information; the travel preference information will be input into the intelligent RAG retrieval model as query information; wherein, the travel preference information includes but is not limited to the favorite attraction type and estimated tour time.

3. The method for recommending personalized routes for cloud tourism based on agent collaboration and RAG retrieval according to claim 1, characterized in that: The specific application of the intelligent RAG retrieval model in step 2 includes: Inputting the query information into the intelligent RAG retrieval model to determine whether intelligent reconstruction is required to obtain a determination result, wherein the intelligent reconstruction includes: Structuring the travel preference information input by the user through the interactive interface into the query information, and transmitting the query information to the intelligent RAG retrieval model; After receiving the query information, the intelligent RAG retrieval model uses internal data or searches external data to retrieve the retrieval information.

4. The method for recommending personalized routes for cloud tourism based on agent collaboration and RAG retrieval according to claim 3 is characterized in that: The specific application of the intelligent RAG retrieval model in step 2 also includes: The large language LLM model in the intelligent RAG retrieval model answers the new query information and retrieves data based on its internal data or external data associated with the internal data to obtain retrieval results; specifically: In the process of searching various resources through internal and external data, the intelligent RAG retrieval model not only searches existing scenic spot document data, but also searches multi-source information related to the scenic spot document data; the multi-source information includes tourism websites, geographic information systems and official data of the tourism bureau.

5. The method for recommending personalized routes for cloud tourism based on agent collaboration and RAG retrieval according to claim 1, characterized in that: The structure of the intelligent RAG retrieval model in step 2 includes: Adding a resource retrieval agent to the large language LLM model for reference by the intelligent RAG retrieval model. The resource retrieval agent actively thinks and makes decisions during the information retrieval and information generation process to generate accurate query predictions. The large language LLM model in the intelligent RAG retrieval model selects qualified candidate resources from a large number of travel route resources as the retrieval information.

6. A cloud tourism personalized route recommendation method based on agent collaboration and RAG retrieval according to any one of claims 1, 3, 4, and 5, characterized in that: The training process of the intelligent RAG retrieval model in step 2 includes: The intelligent RAG retrieval model is fine-tuned using a Transformer-based bidirectional encoder representation model to adapt it to semantic understanding in the tourism field. Specifically: The training data required for fine-tuning the intelligent RAG retrieval model process covers a large number of tourism-related questions and corresponding resource description text pairs, allowing the intelligent RAG retrieval model to learn how to accurately extract information from the knowledge graph of external data based on the questions.

7. The method for recommending personalized routes for cloud tourism based on agent collaboration and RAG retrieval according to claim 1, characterized in that: The route planning agent in step 3 has the following specific applications: The route planning agent is applied to plan the cloud travel route according to the search information, specifically: (1) The route planning agent plans multiple initial cloud tourism routes based on the search information; (2) The route planning agent uses a genetic algorithm to iteratively optimize the multiple initial cloud tourism routes by selection, crossover or mutation to plan a cloud tourism route; wherein the iterative optimization is specifically as follows: Step a: Calculate the fitness of each of the initial cloud tourism routes; specifically, the fitness is calculated by scoring each of the initial cloud tourism routes based on the flow of visitors to tourist attractions, the time cost required for tourism, and the degree of user interest compliance as criteria for measuring the quality of cloud tourism routes; Step b: performing multiplication or evolution processing on the initial cloud tourism routes one by one by selection, crossover or mutation to obtain a plurality of new cloud tourism routes; specifically, the multiplication or evolution process is to perform crossover or mutation processing on the initial cloud tourism routes based on fitness; Step c: Calculate the fitness of each of the new cloud travel routes, and select and retain travel routes with high fitness for further reproduction and evolution; Step d: Repeat steps a to c to iterate and optimize until the termination condition is met to obtain the cloud travel route.

8. The method for recommending personalized routes for cloud tourism based on agent collaboration and RAG retrieval according to claim 1, characterized in that: The evaluation and optimization agent in step 3 has the following specific applications: (1) The evaluation and optimization agent performs route evaluation on a plurality of the cloud tourism routes; specifically, the evaluation and optimization agent predicts the satisfaction of the cloud tourism routes using a user satisfaction prediction model, scores the cloud tourism routes using historical user data or expert data, optimizes and adjusts the cloud tourism routes based on the satisfaction prediction results and the scoring results, and re-evaluates the cloud tourism routes based on the optimization and adjustment results; (2) the evaluation and optimization agent selects a plurality of high-quality personalized routes from the plurality of cloud tourism routes according to the evaluation results; (3) The evaluation and optimization agent recommends multiple high-quality personalized routes to the user.

9. The method for recommending personalized routes for cloud tourism based on agent collaboration and RAG retrieval according to claim 1, characterized in that: The personalized cloud tourism route recommendation in step 4 includes: Step 41: The interactive feedback agent displays the multiple high-quality personalized routes output in step 3; Step 42: The user selects one of the high-quality personalized routes as the personalized cloud tourism route. If a certain scenic spot in the personalized cloud tourism route needs to be replaced, the process returns to step 2 and restarts to optimize the personalized cloud tourism route based on the scenic spot to be replaced.

10. A cloud tourism personalized route recommendation method based on agent collaboration and RAG retrieval according to claim 9, characterized in that: The personalized route recommendation for cloud tourism in step 4 further includes: Step 43: After the user completes the cloud tourism experience based on the personalized cloud tourism route, the evaluation and feedback agent obtains the user's experience information of participating in this cloud tourism. Based on the experience information, the evaluation and feedback agent provides corresponding optimization information to the service provider of the corresponding tourism model or tourism route on the personalized cloud tourism route.

Images