Commercial real estate question-answering content generation system based on retrieval enhancement technology

Abstract

The present invention belongs to the field of artificial intelligence technology, and specifically relates to a commercial real estate question-and-answer content generation system based on retrieval enhancement technology. The system includes an original data processing module, a text content segmentation module, a knowledge base construction module, a query rewriting module, a question planning and decomposition module, an intention routing module, a hybrid retrieval module, and a content generation module. The present invention has the characteristics of being more efficient and more accurate, and also greatly improving the user experience.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and in particular relates to a commercial real estate question-and-answer content generation system based on retrieval enhancement technology. Background Art

[0002] Commercial real estate companies have accumulated a wealth of knowledge and data assets in their daily operations. These data cover all aspects, from project development, plaza positioning, store leasing, business layout to operation management and customer service. These valuable data assets are the core elements to promote the digital transformation of enterprises. Their efficient management, in-depth analysis and innovative application will directly determine the success of the transformation. Therefore, how to effectively manage, deeply analyze and make full use of these valuable data assets constitutes the core challenges and opportunities for commercial real estate companies to achieve digital transformation. Intelligent solutions, such as advanced models, algorithms, technology platforms and innovative product designs, are important means to achieve this goal. They can help integrate scattered data resources, stimulate the potential value of data, and provide strong knowledge support, intelligent retrieval, question answering and decision assistance for the daily operations of commercial real estate companies. In this context, the intelligent question and answer assistant (RAG Intelligent Assistant) in the field of retail and commercial real estate came into being. It serves as a bridge connecting the internal data assets of enterprises with end users, allowing users to conduct natural language interactive queries through voice or text input.

[0003] RAG (Retrieval-Augmented Generation) is a technical framework that combines information retrieval (Retrieval) and generative models (Generation). It uses artificial intelligence (AI), machine learning (ML) and other cutting-edge technologies to quickly retrieve relevant information from huge knowledge bases and data sets, and present it to users in an intuitive and easy-to-understand form. Compared with conventional pre-trained natural language processing (NLP) models, RAG has shown excellent results in many aspects of natural language processing. It not only improves the accuracy and pertinence of users' information acquisition, but also improves retrieval efficiency, providing enterprise management with data-driven insights to help make smarter business decisions. Although retrieval-augmented generation technology has greatly improved the quality of model-generated content, existing retrieval-augmented generation technology still has many problems, including the following shortcomings:

[0004] 1. Existing solutions are not suitable for commercial real estate

[0005] Since the effectiveness of the RAG model depends largely on the quality of the data source on which its retrieval part relies. RAG models trained in different industries are generally not universal. The characteristics of the commercial real estate field determine that professional knowledge and data in related fields must be available to meet its diverse knowledge retrieval, data query and analysis suggestion needs. Although existing related research and patents have considered using domain data to fine-tune the basic large model to construct personalized, domain-specific large models, they are still insufficient to actually solve the relatively complex needs of commercial real estate in investment promotion, operation and other scenarios, such as intelligent analysis suggestions for operating conditions, shopping center indicator query, acquisition and execution of internal computing tools, etc.; in addition to relying on the semantic understanding and task decomposition capabilities of the large model, the above capabilities also require seeking solutions from professional knowledge and historical experience. In addition, the data in the commercial real estate field consists of a variety of structures, including plain text, tables, pictures, etc. Various business plans, investment brochures, etc. are also presented in a regular chapter structure. Document segmentation based on chapter or directory structure can better build a knowledge base in the commercial real estate field, thereby improving the efficiency and accuracy of retrieval. Existing technologies often lack research on this. For example, the document with patent publication number CN118820394A mentions text adaptive segmentation, but it focuses more on learning how to effectively segment text into smaller units.

[0006] 2. Insufficient ability to handle complex tasks

[0007] The document with patent publication number CN118377844A mentions the use of natural language processing technology for semantic analysis and conversion, including rewriting queries and splitting them into more manageable sub-queries, but it still relies on the capabilities of natural language processing technology or the large language model itself, and does not fully consider the role of historical business rules; in addition, the documents with patent publication numbers CN118377844A and CN118820394A both mention the use of intelligent routing or dynamic routing to classify and locate questions. The former also relies on the capabilities of intelligent agents or large models themselves, while the latter mainly performs simple classification of questions, identifies users' expectations for the "accuracy", "rigor" and "creativity" of the answer content, and calls different RAG modules according to actual conditions.

[0008] From this, we can see that although the existing trained RAG model can decompose complex questions into simple sub-questions through the model's own understanding ability and question classification ability in question answering, and then retrieve relevant document slices for each sub-question in the knowledge base, thereby enhancing the performance of the dialogue system or text generation system in natural language processing tasks. However, most of the existing RAG models only handle a single question. For task queries involving multiple steps or requiring logical planning, it may be difficult to effectively decompose the task and arrange the execution order, especially when complex business rule logic is involved, which affects the accuracy and efficiency of the system response.

[0009] 3. Insufficient diversity and adaptability of generated content and presentation formats

[0010] Existing RAG systems lack the diversity of generated content and presentation forms. Although the document with patent publication number CN118520867A mentions combining structured, semi-structured and unstructured search results to output the final annual report analysis answers, the presentation and output formats are still mainly text reports, and lack personalized answers for users. The RAG retrieval system in the commercial real estate field needs to be able to intelligently select the most appropriate output format and support multiple forms of content replies to meet the business needs of users with different job backgrounds in different scenarios.

[0011] Therefore, it is very important to design a commercial real estate question and answer content generation system based on retrieval enhancement technology that is more efficient, more accurate, and has a significantly improved user experience. Summary of the invention

[0012] The present invention aims to overcome the problems in the prior art, such as the existing retrieval enhancement technology, the incompatibility of the existing solutions with the business field, the insufficient ability to handle complex tasks, and the insufficient diversity and adaptability of the generated content and presentation forms. It provides a commercial real estate question and answer content generation system based on retrieval enhancement technology that is more efficient, more accurate, and has a greatly improved user experience.

[0013] In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical solutions:

[0014] The commercial real estate question-answering content generation system based on retrieval enhancement technology includes:

[0015] The raw data processing module is used to parse and process the collected multi-source heterogeneous data in the field of commercial real estate into text content;

[0016] The text content slicing module is used to slice the parsed text content;

[0017] The knowledge base construction module is used to classify the sliced ​​documents according to the corresponding document content, then vectorize the document slices and store them in different knowledge bases; at the same time, briefly describe the knowledge content of each different knowledge base;

[0018] The query rewriting module is used to perform reference disambiguation on the current user input query and rewrite it into instructions with clear and explicit semantics;

[0019] The question planning and decomposition module is used to guide the large model to determine whether the rewritten user query is a complex task instruction and decompose it by setting a reasonable prompt word template;

[0020] The intention routing module is used to identify the intention of the disassembled user task instructions, obtain the intention recognition results, and search from the knowledge base that meets the user's intention;

[0021] The hybrid retrieval module is used to obtain the comprehensive similarity score of documents by building a business keyword library and designing scoring rules based on the combination of semantic vector similarity and text matching method based on statistical word frequency information; at the same time, several document slices with the highest similarity scores are selected as relevant documents for user queries;

[0022] The content generation module is used to retrieve relevant documents based on the query input by the user, and provide users with customized question responses and analysis suggestions while taking into account the user's background information.

[0023] Preferably, in the raw data processing module, the multi-source heterogeneous data includes documents, PPTs, pictures, and audio and video.

[0024] Preferably, the text content cutting module specifically includes the following process:

[0025] By detecting the chapter structure of unstructured documents, the initial slicing is first performed according to chapters, and then the documents within the chapters are subjected to regular slicing operations. Subsequently, the chapter information of the slice is added to the header of the sliced ​​document in the form of a title, thereby re-establishing a direct connection between the document content and the corresponding chapter information.

[0026] Preferably, in the query rewriting module, if the user's input query contains business-related keywords, the user's input query will be supplemented with business keyword annotations in combination with the business keyword annotation library.

[0027] Preferably, the problem planning and decomposition module specifically includes the following process:

[0028] If the current rewritten user query is a simple task instruction, no modification is made; if the current rewritten user query is a complex task instruction, several business entity nodes related to the instruction and the operation relationships between the corresponding entities are matched from the pre-built business logic knowledge graph and input into the big model;

[0029] By utilizing the task decomposition and planning capabilities of the large model, the original instructions are ultimately decomposed into several subtasks;

[0030] Based on the business rule map obtained from the summary of previous business, the execution logic of each subtask is comprehensively planned; if it is judged according to the business rule map that there is a mutual dependency between the subtasks, the subtasks are executed serially, otherwise the subtasks are executed in parallel.

[0031] Preferably, the intention routing module specifically includes the following process:

[0032] We use manually annotated intent recognition training samples as seed data, construct synthetic data as training sets, and use the distributed training framework Deepspeed to fine-tune the model, thus obtaining a commercial real estate-specific intent recognition model for intent recognition of disassembled user task instructions.

[0033] The intention of user task instructions is recognized through a proprietary intention recognition model in the commercial real estate field, and the user task instructions are routed to the specific content generation sub-module downstream of the system, and retrieved from the knowledge base that meets the user's intention; at the same time, based on different intention recognition results, specific format templates are selected from the content template library for use by subsequent content generation modules.

[0034] Preferably, the hybrid search module specifically includes the following process:

[0035] S1, train a specific named entity recognition model through business corpus and build a business keyword library; use the keyword importance scoring model to assign a corresponding weight score to each keyword; for the keywords that co-occur in the user query statement and the document slices in the knowledge base, calculate the keyword correlation score between the query statement and each document slice in the knowledge base; the business keyword retrieval score is equal to the sum of the scores of each business keyword contained in the user query input text, which is specifically expressed as the following formula:

[0036] ;

[0037] in, Indicates the importance score of keywords. Represent the user's input query text and a slice document in the knowledge base respectively; Indicates the business keyword library, Represents the words in the user query contained in the business keyword library; represents a keyword importance scoring function extracted based on business rules. The stronger the business attribute of the keyword satisfies the function, the higher the output score is;

[0038] S2, calculates the similarity score between the user query and the knowledge base document slice based on the semantic text vector embedding and the text matching method based on statistical word frequency information; the similarity score calculation method based on the text embedding vector is as follows:

[0039] ;

[0040] in, Represents user query With Slice Document The similarity score of represents a word embedding model used to convert the text description of sliced ​​documents and user queries into semantic vectors; represents the two-norm;

[0041] S3, learn a set of weight coefficients based on the constructed business training data, and add the document slice similarity scores obtained by each retrieval method according to the weights to obtain the final similarity score , and select several document slices with the highest scores as relevant documents for the user's query. The specific calculation method is as follows:

[0042] ;

[0043] in, Query for users With Slice Document Similarity score represented by sparse vector of texts by counting word frequencies; Indicates the maximum similarity score based on statistical word frequency; Represents the minimum similarity score based on statistical word frequency; , , They are the similarity score based on text embedding, the similarity score based on statistical word frequency, and the weight coefficient based on keyword importance score; Indicates the maximum value of the keyword importance score; Indicates the minimum value of the keyword importance score.

[0044] Preferably, the content generation module includes a knowledge question and answer submodule, an intelligent retrieval submodule, a case matching submodule and a general question and answer submodule;

[0045] The knowledge question and answer submodule is used to provide detailed answers to questions about business knowledge, rules and regulations, and document regulations within the business scenario;

[0046] The intelligent search submodule is used to display information about commercial plazas and merchant brands in the real estate field, and to conduct comparative analysis from multiple dimensions based on specific content;

[0047] The case matching submodule is used to find corresponding excellent solution cases for the pain points that occur in the operation of the commercial plaza;

[0048] The general question and answer submodule is used to answer questions from users about general scenarios that are not related to the business.

[0049] Preferably, the system further comprises:

[0050] Information storage module, used to process and store the user's historical conversation information;

[0051] The slice filtering module is used to perform relevance filtering on the retrieved document slices.

[0052] Compared with the prior art, the present invention has the following beneficial effects: (1) With the help of the deep understanding and generalization ability of the large language model for natural language, combined with the rich information of multiple rounds of historical conversations of users, the present invention designs a set of effective prompt word templates to accurately perform semantic disambiguation on the current user query statement, and rewrite it into a clearer and more general query statement with more semantics, thereby significantly improving the accuracy and recall rate of downstream knowledge retrieval; further, in the face of user queries with complex task requirements, on the basis of fully considering historical information, the business entity nodes related to the instructions and the operational relationships between the corresponding entities are matched from the pre-constructed business logic knowledge graph, and this series of information is input into the large model. By using the task decomposition and planning capabilities of the large model, the original instructions are finally decomposed into a series of simple and logically clear subtasks; according to the business logic relationship of the subtasks, the subsequent processes of each task are executed serially or in parallel to ensure the efficient and smooth operation of the system;

[0053] (2) The present invention proposes a document slicing method based on chapter structure. The method first detects the directory structure information of the unstructured document and divides it into chapters during the preliminary slicing. Then, the document content within the chapter is subjected to conventional slicing operations. Next, the information of the chapter to which each document slice belongs is embedded in the header of the segmented document in the form of a title, thereby reconstructing the direct relationship between the document content and the information of the chapter to which it belongs. Finally, the transformed document slices are vectorized and stored in a vector database. During the retrieval process, for questions involving both chapter content and text content, since each document slice contains the information of the chapter to which it belongs, the relevant document slices can be accurately retrieved. The slicing technology proposed by the present invention solves the problem of low accuracy in chapter-related retrieval problems.

[0054] (3) The present invention also designs a dynamic routing mechanism. Different from the traditional dynamic routing that is limited to only identifying the intent of user query input and routing it to a specific knowledge base for document retrieval, this mechanism not only accurately routes user queries to a specific knowledge base for efficient retrieval based on the intent recognition results, but also adaptively selects the output content format that best meets the user's query requirements, and combines customized content templates to generate user-satisfied reply content. It supports reply content in various forms such as plain text, structured tables, comparative analysis of text content, and text mixed with pictures, greatly improving the efficiency of user information acquisition and user experience. More advanced is that the dynamic routing mechanism can also analyze from multiple angles based on the user's background information (such as identity, preferences, etc.) and query intent, combined with relevant document slices, to provide users with personalized reply content. This means that even if users with different backgrounds conduct the same query, they can get replies that better meet their needs, realizing the generation of "one thousand faces for one thousand people" reply content.

[0055] (4) The present invention also proposes a hybrid retrieval scheme that integrates multiple retrieval methods. The scheme first trains a specific named entity recognition model through business corpus and constructs a business keyword library. A keyword importance scoring model is used to assign a corresponding weight score to each keyword. For keywords that co-occur in user query statements and document slices in the knowledge base, the keyword correlation scores of the query statements and each document slice in the knowledge base are calculated. At the same time, the similarity scores of user queries and knowledge base document slices based on semantic text vector embedding and text matching algorithms based on statistical word frequency information (such as BM25, TF-IDF, etc.) are also calculated. By adjusting the weight coefficients of different methods, the document slice similarity scores obtained by each retrieval method are accumulated to obtain a final similarity score, and several document slices with the highest similarity scores are selected as relevant documents for the user query. In addition, according to business rules and prior experience, the weight coefficients of different retrieval methods can be adjusted in real time to further optimize the document retrieval effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 An execution flow chart of a commercial real estate question-and-answer content generation system based on retrieval enhancement technology provided by an embodiment of the present invention;

[0057] Figure 2 A schematic diagram of a business logic relationship knowledge graph provided by an embodiment of the present invention;

[0058] Figure 3 An operation flow chart of the knowledge question and answer module provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to more clearly illustrate the embodiments of the present invention, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other accompanying drawings and other implementation methods can be obtained based on these accompanying drawings without creative work.

[0060] The present invention provides a commercial real estate question and answer content generation system based on retrieval enhancement technology, which mainly includes the following modules.

[0061] Raw data processing module: Utilizes technologies such as OCR (Optical Character Recognition), multimodal large models, etc. to efficiently parse and process multi-source heterogeneous data such as documents, PPTs, pictures, audio and video collected in the commercial real estate field into text content, and completes the preprocessing of raw data after manual verification.

[0062] Text content slicing module: The main function is to slice the parsed text content. Traditional slicing schemes are mainly based on text length, punctuation or text semantics. However, for texts containing directory chapters, such as books, papers and other texts, the directory information contains highly concentrated semantic information. If the traditional slicing scheme is used, on the one hand, for the slices cut in the directory chapters, the slice content has no substantial content and cannot provide useful information for model generation. On the other hand, the slice content that does not contain directory information loses extremely important text structure information. Therefore, the present invention proposes and implements a new slicing method for unstructured documents. By detecting the chapter structure of unstructured documents, the slicing is first performed according to the chapters, and the documents in the chapters are subjected to conventional slicing operations. Then, the chapter information of the slice is added to the header of the slice document in the form of a title, and the direct connection between the document content and the chapter information is re-established. Since each slice contains the chapter information, the text information contained in the entire slice is more sufficient, which significantly improves the accuracy of retrieval.

[0063] Knowledge base construction module: After the sliced ​​documents are processed, they are classified according to their content, and then the slices are vectorized and stored in the vector database (knowledge base); and the knowledge content of each different knowledge base is briefly described so that user questions can be routed to the appropriate knowledge base for retrieval according to user intentions, thereby improving retrieval efficiency and accuracy.

[0064] Query rewriting module: Combined with the historical conversation information between the user and the system, the current user input query is disambiguated to ensure that the user input query is rewritten into instructions with clear and unambiguous semantics. At the same time, combined with the business keyword annotation library, if the user's input query contains business-related keywords, the user's input query will be further supplemented with business keyword annotations.

[0065] Problem planning and decomposition module: By setting reasonable prompt word templates, the big model is guided to determine whether the rewritten user query is a complex task instruction. If it is a simple task instruction, no modification will be made to it. If the current rewritten user query is a complex task instruction, several business entity nodes related to the instruction and the operational relationships between the corresponding entities will be matched from the pre-built business logic knowledge graph, and this series of information will be input into the big model. By using the task decomposition and planning capabilities of the big model, the original instruction will eventually be decomposed into a series of simple and logically clear sub-tasks. According to the business rule map obtained by summarizing previous business, the execution logic of each sub-task is comprehensively planned. If it is judged according to the business rules that there is a mutual dependence between the sub-tasks, the sub-tasks will be executed serially, otherwise the sub-tasks will be executed in parallel to improve the work efficiency of the entire system.

[0066] Intent routing module: A small number of manually annotated intent recognition training samples are used as seed data, synthetic data is constructed as a training set, and the distributed training framework Deepspeed is used for model fine-tuning training to obtain a commercial real estate-specific intent recognition model that can accurately recognize the intent of disassembled user task instructions. Through the model's intent recognition of user instructions, the user instructions are routed to the specific content generation submodule downstream of the system, and retrieved from the knowledge base that meets the user's intention. At the same time, according to different intent recognition results, specific format templates will be selected from the content template library for use by subsequent content generation modules.

[0067] Hybrid retrieval module: First, a specific named entity recognition model is trained through business corpus, and a business keyword library is constructed. Using the keyword importance scoring model, a corresponding weight score is assigned to each keyword. For keywords that co-occur in the user query statement and the document slices in the knowledge base, the keyword correlation score between the query statement and each document slice in the knowledge base is calculated. The business keyword retrieval score is equal to the sum of the scores of each business keyword contained in the user query input text, which can be expressed as the following formula:

[0068] ;

[0069] in, Indicates the importance score of keywords. Represent the user's input query text and a slice document in the knowledge base respectively; Indicates the business keyword library, Represents the words in the user query contained in the business keyword library; represents a keyword importance scoring function extracted based on business rules. The stronger the business attribute of the keyword satisfies the function, the higher the output score is;

[0070] At the same time, the hybrid retrieval module also calculates the similarity scores between user queries and knowledge base document slices based on semantic text vector embedding and text matching algorithms based on statistical word frequency information (such as BM25, TF-IDF, etc.). The similarity score calculation method based on text embedding vector is as follows:

[0071] ;

[0072] in, Represents user query With Slice Document The similarity score of represents a word embedding model used to convert the text description of sliced ​​documents and user queries into semantic vectors; represents the two-norm;

[0073] Taking BM25 as an example, the similarity score of the text sparse vector representation based on statistical word frequency can be calculated as follows:

[0074] ;

[0075] Where t represents the term in the query sentence input by the user; Indicates that word t is in the slice document The frequency of occurrence in is a hyperparameter for adjusting the degree of word frequency saturation, and its value range is usually between [1.2, 2.0]. b is a hyperparameter for adjusting the effect of document length on the score, and its value is usually 0.75. For slice documents Length, represents the average length of all documents in the corpus; is the inverse document frequency of word t, which is calculated as follows:

[0076] ;

[0077] N and They represent the total number of document slices in the knowledge base and the number of words t contained in the slices respectively.

[0078] A set of weight coefficients are learned on the constructed business training data, and the document slice similarity scores obtained by each retrieval method are weighted and added together to obtain the final similarity score. , and select several document slices with the highest scores as relevant documents for the user's query. The specific calculation method is as follows:

[0079] ;

[0080] in, Query for users With Slice Document Similarity score represented by sparse vector of texts by counting word frequencies; Indicates the maximum similarity score based on statistical word frequency; Represents the minimum similarity score based on statistical word frequency; , , They are the similarity score based on text embedding, the similarity score based on statistical word frequency, and the weight coefficient based on keyword importance score; Indicates the maximum value of the keyword importance score; Indicates the minimum value of the keyword importance score.

[0081] In addition, the weight coefficients of different retrieval methods can be adjusted dynamically and in real time according to business rules and prior experience to further optimize the document retrieval effect.

[0082] Content generation module: The content generation module includes four sub-modules: knowledge questions and answers, intelligent retrieval, case matching, and general questions and answers. The knowledge questions and answers mainly provide detailed answers to questions about business knowledge, rules and regulations, document regulations, etc. in business scenarios; intelligent retrieval focuses on the information display of commercial plazas and merchant brands in the real estate field, and conducts comparative analysis from multiple dimensions based on their specific content; case matching focuses on finding matching excellent solution cases for the pain points that arise in the operation of commercial plazas, enriching the operational ideas of managers; general questions and answers are used to answer user questions about general scenarios that are not related to the business. In addition to referring to relevant documents retrieved based on user queries, the reply content of each module will also take into account the user's background information (such as identity, preferences, etc.), and provide users with customized question replies and analysis suggestions, realizing the generation of "one thousand faces for one thousand people" reply content.

[0083] In addition, the technical solution proposed in the present invention also includes a module for processing and saving the user's historical conversation information (information saving module), a module for filtering the relevance of retrieved document slices (slice filtering module), etc. In summary, the technical solution of the present invention has higher accuracy and user experience.

[0084] Based on the technical solution of the present invention, the following case scenario is used to illustrate the implementation process of the present invention in practical application. The specific application implementation plan is as follows:

[0085] Figure 1 This is an implementation case flow of the present invention. When a user enters a query on the interactive terminal, the specific execution flow is as follows:

[0086] 1. The interactive front end receives user input and passes it to the query rewriting module to rewrite the user input in combination with the user's historical conversation information, expressing the original user input as clear and specific task instructions as possible; at the same time, if the user's input contains business keywords, while rewriting the user input, it will also combine with the business keyword library to supplement the keyword annotation information in the input. For example, the user input question in the previous round was "Introduce milk tea brand A and milk tea brand B", and the current user input is "Compare the operating status of the two in the northern region". The query rewriting module will combine the previous round of conversation history and modify the user input of this round to "Compare the operating status of milk tea brand A and milk tea brand B in the northern region of a certain group";

[0087] 2. According to the result of question rewriting, the big model is called to first determine whether the current instruction is a simple question instruction containing only a single intent. If so, the instruction will not be disassembled. If the result is that the current instruction is a complex intent instruction, the instruction will be matched with the pre-built business logic knowledge graph to obtain several business entity nodes related to the instruction and the operational relationship between the corresponding entities. This series of information is input into the big model, and the task disassembly and planning capabilities of the big model are used to disassemble the original instruction into a series of simple and logically clear subtasks. According to the business rule map, the execution logic of each subtask is comprehensively planned. If it is determined according to the business rules that there is a mutual dependence between the subtasks, the subtasks are executed serially, otherwise the subtasks are executed in parallel to improve the work efficiency of the entire system. Specifically, Figure 2 As shown in the figure, taking the commercial real estate business scenario as an example, a brief business logic knowledge graph is constructed, which includes business entities such as personnel, brands, operations and investment promotion. Each entity corresponds to a series of business logics, and the edge relationships between entities correspond to the operational relationships between entities. For example, personnel can perform investment promotion analysis and brand analysis, and the specific execution actions of investment promotion are to attract related brands.

[0088] 3. The fine-tuned intent recognition model determines the intent of the task instructions entering the intent routing module, and routes the task instructions to the appropriate sub-modules in the subsequent knowledge question and answer / intelligent retrieval / case matching / general question and answer based on the determined intent. The subsequent process is introduced using knowledge question and answer as an example. The specific flow chart is as follows: Figure 3 As shown;

[0089] 3-1. After the decomposed sub-questions enter the knowledge question-answering module, they will be searched in a specific knowledge base according to the intent recognition results. The search method is an innovative search method that combines vector search, full-text search and keyword search proposed in this invention. For example, if a user enters "how to improve the efficiency of investment promotion", the user query will be routed to the "investment promotion business" knowledge base for search. If the user enters "what are the measures to increase the flow of customers in shopping malls", it will be routed to the "operation business" knowledge base;

[0090] 3-2. Use the re-ranking model to re-score and sort the relevance of the retrieved text slices to the user instructions, and retain the top-ranked text slices. For example, for a user input of "What are the measures to increase the customer flow of the shopping mall?", hundreds of relevant document slices are retrieved from the "operation business" knowledge base. After re-scoring these text slices using the re-ranking model, the top 20 relevant document slices are retained;

[0091] 3-3. According to the business keyword rules, compare the user instructions with the text slice content, filter out the text slices that do not contain the phrases corresponding to the user instruction keywords, and obtain the final document slice retrieval results. According to the user input "What are the measures to increase the flow of customers in the mall" and the business keyword rules, the user instruction keywords "operation" and "customer development" are obtained. For the document slices retained in the above step 3-2, if they do not contain relevant keywords, they are filtered;

[0092] 3-4. Based on the results of intent recognition in the previous steps, select a customized template that matches the user's intent from the content template, combine the user background information obtained from the database, refer to the final retrieved text slice, generate the reply content by the large model, and save the current conversation information to the historical conversation information. For example, for the same user input "What are the measures to increase the flow of customers in the mall", if the user's identity information is found to be a front-line operation staff, specific measures will be given in combination with the relevant document slice content. If the user's identity information is found to be the general manager of a single store, relevant suggestions will be given from the perspective of management of operation staff in combination with the document content;

[0093] 3-5. Call the model again to determine whether the generated reply content clearly answers the user's question, and give different status codes according to the judgment results to record the final results;

[0094] 3-6. Finally, in the question recommendation module, the current user question and generated content will be referred to, and 3 related questions that the user may be interested in will be recommended to the user and displayed to the user on the front end. For example, if the user inputs "What are the measures to increase the flow of customers in the mall", the user may continue to be recommended related questions such as "How to improve customer stickiness" and "How to increase the turnover of the mall".

[0095] The above process implements the general process of the text content generation method based on retrieval enhancement generation technology of the present invention to reply to user queries. The specific technical implementation is only an example, and other optional implementation methods or alternative methods should not deviate from the core content of the present invention.

[0096] The present invention proposes a system for realizing intelligent question and answer of business knowledge and business knowledge retrieval comparison based on a multimodal retrieval enhancement generation method constructed in a modular way, which significantly improves the efficiency and accuracy of knowledge retrieval on the basis of the existing RAG technical method. Specifically, in the data preprocessing stage, a multimodal model and OCR technology are used to parse and extract multi-source heterogeneous data such as text, PPT, pictures, audio and video, and a proprietary domain business knowledge base is constructed based on this, which greatly improves the knowledge breadth and coverage of the business knowledge base; the large language model is used to understand natural language and to disassemble and plan complex tasks, and through the careful design of the prompt word project, the model is guided to optimize and rewrite the original query in combination with the user's current query and historical dialogue information, and decompose the complex query into several single-purpose sub-queries; a new dynamic routing mechanism is adopted, which can efficiently route the query to the task-specific knowledge base for retrieval, and at the same time realize a customized output mode in combination with the user's background information; the large language model is supervised and fine-tuned by using the corpus data of the proprietary domain, so that the generated content of the model is closer to the business scenario and the wording is more professional. In summary, the present invention proposes a retrieval enhanced content generation method that is more efficient, more accurate, and also greatly improves the user experience.

[0097] The innovative features of the present invention are as follows:

[0098] 1. The present invention discloses an improved scheme for intelligent question-answering of business knowledge and comparison of business data analysis based on retrieval generation enhancement. In this scheme, a new method for disassembling and planning user question-answering tasks is proposed. The method mainly includes two steps: rewriting and disassembling the user task. First, by utilizing the high generalization and understanding ability of the large language model for natural language, combined with the user's multi-round historical dialogue information, by setting a reasonable prompt word template, the current user input statement is semantically disambiguated and rewritten into a task instruction with clear semantics and clear intentions. Further, the rewritten task instruction is matched with the pre-constructed business logic knowledge graph, and the business entities and operational relationships between entities that may be contained in the instruction are disassembled. This series of information is input into the large model. With the help of the task planning ability of the large model, the task instruction is split into a series of simple tasks with inherent business logic, and the subsequent task flow is executed serially / parallel according to the execution logic relationship of each subtask.

[0099] 2. The present invention proposes and implements a new slicing method for unstructured documents. Traditional slicing methods are based on punctuation marks, text length or paragraph-based methods. For documents containing chapter structures, traditional document slicing schemes often break the important connection between directory structure information and detailed content, and cannot well retrieve documents related to issues that contain both chapter content and text content. The document slicing method based on chapter structure proposed by the present invention detects the chapter structure of unstructured documents, and first slices them according to chapters when slicing, and then performs conventional slicing operations on the documents in the chapters, and then adds the chapter information where the slices are located in the form of titles to the header of the sliced ​​document, and re-establishes a direct connection between the document content and the chapter information where they are located. Finally, the transformed slices are vectorized and stored in a vector database operation. When searching for issues containing chapter content and text content, since each slice contains the chapter information where it is located, the relevant document slices can be accurately retrieved. The slicing method proposed by the present invention increases the chapter information contained in each slice, solving the problem of inaccurate retrieval of chapter-related issues.

[0100] 3. The present invention designs a novel dynamic routing mechanism, which is different from the traditional dynamic routing that only performs intent recognition on the user's query input, and then routes the query to a specific knowledge base for document retrieval to obtain relevant document slices, and then hands it over to the content generation module to obtain the final text response output. The dynamic routing mechanism proposed this time not only routes the user query to a specific knowledge base for retrieval according to the intent recognition result, improving the retrieval accuracy and efficiency, but also adaptively determines the output content format that best meets the user's query, and presents the final reply content in different forms such as plain text, structured tables, supplemented by comparative analysis of text content, and text mixed with pictures, thereby improving the user's efficiency in obtaining information and the user experience. Furthermore, the dynamic routing mechanism will also give targeted replies from different perspectives and analysis angles based on the provided user background information (such as identity, preferences, etc.) and user intent, combined with the relevant document slices of the user's query. Users with different background information will also get different reply contents that are more in line with the user's own needs when performing the same query input, realizing the generation of "thousands of faces" reply content.

[0101] 4. The present invention proposes a hybrid retrieval scheme that integrates multiple retrieval methods. First, a specific named entity recognition model is trained in the business corpus and a business keyword library is constructed. The keyword importance scoring model is used to assign corresponding weight scores to each keyword. For the keyword results of the co-occurrence of the user query statement and the document slices in the knowledge base, the keyword correlation scores of the query statement and each document slice in the knowledge base are calculated. In addition, the similarity scores of the user query and the knowledge base document slices based on semantic-based text vector embedding and the text matching algorithm based on statistical word frequency information (such as BM25, TF-IDF, etc.) are calculated at the same time. Through different weight coefficients, the similarity scores of each method are integrated and accumulated to obtain the final similarity score, and several document slices with the highest scores obtained by retrieval are selected as the relevant information of the user query. At the same time, according to business rules and prior experience, the weight coefficients of different retrieval methods can be adjusted in real time to further optimize the document retrieval effect.

[0102] The above description is only a detailed description of the preferred embodiments and principles of the present invention. For ordinary technicians in this field, according to the ideas provided by the present invention, there will be changes in the specific implementation methods, and these changes should also be regarded as the protection scope of the present invention.

Claims

1. A commercial real estate question-answering content generation system based on retrieval enhancement technology, characterized in that: include: The raw data processing module is used to parse and process the collected multi-source heterogeneous data in the field of commercial real estate into text content; The text content slicing module is used to slice the parsed text content; The knowledge base construction module is used to classify the sliced ​​documents according to the corresponding document content, then vectorize the document slices and store them in different knowledge bases; at the same time, briefly describe the knowledge content of each different knowledge base; The query rewriting module is used to perform reference disambiguation on the current user input query and rewrite it into an instruction with clear and explicit semantics; The question planning and decomposition module is used to guide the large model to determine whether the rewritten user query is a complex task instruction and decompose it by setting a reasonable prompt word template; The intention routing module is used to identify the intention of the disassembled user task instructions, obtain the intention recognition results, and search from the knowledge base that meets the user's intention; The hybrid retrieval module is used to obtain the comprehensive similarity score of documents by building a business keyword library and designing scoring rules based on the combination of semantic vector similarity and text matching method based on statistical word frequency information; at the same time, several document slices with the highest similarity scores are selected as relevant documents for user queries; The content generation module is used to provide users with customized responses and analysis suggestions based on the relevant documents retrieved from the user's input query, taking into account the user's background information; The problem planning and decomposition module specifically includes the following processes: If the current rewritten user query is a simple task instruction, no modification is made; if the current rewritten user query is a complex task instruction, several business entity nodes related to the instruction and the operation relationships between the corresponding entities are matched from the pre-built business logic knowledge graph and input into the big model; By utilizing the task decomposition and planning capabilities of the large model, the original instructions are ultimately decomposed into several subtasks; Based on the business rule map obtained from the summary of previous business, the execution logic of each subtask is comprehensively planned; if it is determined according to the business rule map that there is a mutual dependency between subtasks, the subtasks are executed serially, otherwise the subtasks are executed in parallel; The intention routing module specifically includes the following processes: We use manually annotated intent recognition training samples as seed data, construct synthetic data as training sets, and use the distributed training framework Deepspeed to fine-tune the model, thus obtaining a commercial real estate-specific intent recognition model for intent recognition of disassembled user task instructions. The user's task instructions are identified through the proprietary intention recognition model in the commercial real estate field, and the user's task instructions are routed to the specific content generation submodule downstream of the system, and retrieved from the knowledge base that meets the user's intention; at the same time, according to different intention recognition results, a specific format template is selected from the content template library for use by the subsequent content generation module; The hybrid retrieval module specifically includes the following processes: S1, train a specific named entity recognition model through business corpus and build a business keyword library; use the keyword importance scoring model to assign a corresponding weight score to each keyword; for the keywords that co-occur in the user query statement and the document slices in the knowledge base, calculate the keyword correlation score between the query statement and each document slice in the knowledge base; the business keyword retrieval score is equal to the sum of the scores of each business keyword contained in the user query input text, which is specifically expressed as the following formula: Among them, Keyword Score represents the keyword importance score, q and d represent the user's input query text and a slice document in the knowledge base respectively; Set kw Represents the business keyword library, t i Represents the words in the user query contained in the business keyword library; Rule(·) represents the keyword importance scoring function extracted based on the business rules. The stronger the business attribute of the keyword satisfied by the function, the higher the output score; S2, calculates the similarity score between the user query and the knowledge base document slice based on the semantic text vector embedding and the text matching method based on statistical word frequency information; the similarity score calculation method based on the text embedding vector is as follows: Where Similarity Score(q,d) represents the similarity score between user query q and slice document d, Embed(·) represents the word embedding model used to convert the text description of slice document and user query into semantic vector; ‖·‖ represents the bi-norm; S3, learns a set of weight coefficients on the constructed business training data, and adds the document slice similarity scores obtained by each retrieval method according to the weights to obtain the final similarity score FinalScore, and selects several document slices with the highest scores as the relevant documents for the user query. The specific calculation method is as follows: Among them, BMScore(q,d) is the similarity score between the user query q and the slice document d represented by the text sparse vector of the statistical word frequency; BMScore max Indicates the maximum similarity score based on statistical word frequency; BMScore min Represents the minimum similarity score based on statistical word frequency; α1, α2, α3 are the weight coefficients of similarity score based on text embedding, similarity score based on statistical word frequency and similarity score based on keyword importance score respectively; Keyword Score max Indicates the maximum value of the keyword importance score; Keyword Score min Indicates the minimum value of the keyword importance score.

2. The commercial real estate question and answer content generation system based on retrieval enhancement technology according to claim 1 is characterized in that: In the original data processing module, the multi-source heterogeneous data includes documents, PPTs, pictures, and audio and video.

3. The commercial real estate question-and-answer content generation system based on retrieval enhancement technology according to claim 1 is characterized in that: The text content cutting module specifically includes the following processes: By detecting the chapter structure of unstructured documents, the initial slicing is first performed according to chapters, and then the documents within the chapters are subjected to regular slicing operations. Subsequently, the chapter information of the slice is added to the header of the sliced ​​document in the form of a title, thereby re-establishing a direct connection between the document content and the corresponding chapter information.

4. The commercial real estate question-and-answer content generation system based on retrieval enhancement technology according to claim 1 is characterized in that: In the query rewriting module, if the user's input query contains business-related keywords, the business keyword annotation library will be combined to supplement the user's input query with business keyword annotations.

5. The commercial real estate question and answer content generation system based on retrieval enhancement technology according to claim 1 is characterized in that: The content generation module includes a knowledge question and answer submodule, an intelligent retrieval submodule, a case matching submodule and a general question and answer submodule; The knowledge question and answer submodule is used to provide detailed answers to questions about business knowledge, rules and regulations, and document regulations within the business scenario; The intelligent search submodule is used to display information about commercial plazas and merchant brands in the real estate field, and to conduct comparative analysis from multiple dimensions based on specific content; The case matching submodule is used to find corresponding excellent solution cases for the pain points that occur in the operation of the commercial plaza; The general question and answer submodule is used to answer questions from users about general scenarios that are not related to the business.

6. The commercial real estate question and answer content generation system based on retrieval enhancement technology according to claim 1 is characterized in that: Also includes: Information storage module, used to process and store the user's historical conversation information; The slice filtering module is used to perform relevance filtering on the retrieved document slices.

Images