Retrieval generation type text reply method and device for assisting planner

Abstract

The invention provides a retrieval generation type text reply method and device for assisting a planner. In the method, a server can efficiently retrieve a large amount of historical dialogue data through a retrieval module of a retrieval enhancement generation model so as to retrieve a first type of query results and a second type of query results of professional knowledge information and language style information required when reply reference texts for replying questions proposed by users are generated; a generation module of the enhanced retrieval generation model can be used for generating a reference reply text with higher accuracy according to the first type of query results and the second type of query results, and providing the reference reply text to the target plan, wherein the language style of the reference reply text is consistent with that of the target plan; therefore, the target planner can reply the user based on the reply reference text, so that the business execution efficiency of the planner is improved, and the cost of insurance recommendation is reduced.

Description

Technical Field

[0001] One or more embodiments of the present specification relate to the field of artificial intelligence technology, and more particularly, to a retrieval-generated text reply method and device for assisting planners. Background Art

[0002] At present, in the field of insurance recommendation, personalized and efficient communication is the key to successful sales. With the intensification of market competition and the increasing diversification of customer needs, insurance planners face many challenges when communicating with customers. Customers' questions are often personalized and divergent, which requires planners to spend a lot of time and energy to respond one by one. For example, customers may ask various complex questions about insurance product terms, claims process, investment returns, etc. These questions not only require professional knowledge to answer, but also require targeted analysis and answers based on the actual situation of the customer.

[0003] In actual operation, on the one hand, insurance planners need to spend a lot of time preparing and organizing customer information in order to better understand customer needs. On the other hand, when answering customer questions, insurance planners may need to repeatedly review relevant information or consult other professionals. These problems not only reduce the work efficiency of planners, but also affect the customer experience, which in turn leads to higher costs for insurance recommendations. Summary of the invention

[0004] In view of this, one or more embodiments of this specification provide the following technical solutions:

[0005] According to a first aspect of one or more embodiments of this specification, a retrieval-generated text reply method for assisting planners is proposed, comprising:

[0006] Get the user's question text;

[0007] The question text is input into a preset retrieval enhancement generation model, so as to obtain a first type of query result and a second type of query result through a retrieval module of the retrieval enhancement generation model; the first type of query result is used to provide the retrieval enhancement generation model with professional knowledge information required to reply to the question text, and the second type of query result is used to provide the retrieval enhancement generation model with language style information required to reply to the question text;

[0008] Through the generation module of the retrieval enhancement generation model, a reply reference text is generated according to the first category query results and the second category query results, and the reply reference text is provided to the target planner so that the target planner can reply to the user according to the reply reference text.

[0009] According to a second aspect of one or more embodiments of this specification, a retrieval generative text reply device for assisting planners is proposed, including:

[0010] An acquisition module for acquiring the question text of the user;

[0011] A query module for inputting the question text into a preset retrieval enhanced generation model, so as to query a first type of query result and a second type of query result through the retrieval module of the retrieval enhanced generation model; the first type of query result is used to provide the professional knowledge information required for replying to the question text for the retrieval enhanced generation model, and the second type of query result is used to provide the language style information required for replying to the question text for the retrieval enhanced generation model;

[0012] An execution module for generating a reply reference text according to the first type of query result and the second type of query result through the generation module of the retrieval enhanced generation model, and providing the reply reference text to the target planner for the target planner to reply to the user according to the reply reference text.

[0013] According to a third aspect of one or more embodiments of this specification, an electronic device is proposed, including: a processor; a memory for storing processor-executable instructions; wherein, the processor realizes the steps of the retrieval generative text reply method for assisting planners as described above by running the executable instructions.

[0014] According to a fourth aspect of one or more embodiments of this specification, a computer-readable storage medium is proposed, on which computer instructions are stored, and when the instructions are executed by a processor, the steps of the retrieval generative text reply method for assisting planners as described above are realized.

[0015] According to a fifth aspect of one or more embodiments of this specification, a computer program product is proposed, including computer programs / instructions, and when the computer programs / instructions are executed by a processor, the steps of the retrieval generative text reply method for assisting planners as described above are realized.

[0016] As can be seen from the above embodiments, this specification obtains the problem text of the user and inputs the problem text into a preset retrieval-augmented generation model. Through the retrieval module of the retrieval-augmented generation model, a first type of query result and a second type of query result are obtained. Among them, the first type of query result is used to provide the professional knowledge information required for the retrieval-augmented generation model to generate a reply to the problem text, and the second type of query result is used to provide the language style information required for the retrieval-augmented generation model to generate a reply to the problem text. Furthermore, through the generation module of the retrieval-augmented generation model, according to the first type of query result and the second type of query result, a reply reference text is generated and provided to the target planner for the target planner to reply to the user according to the reply reference text.

[0017] In this method, a first type of query result for providing the professional knowledge information required for the retrieval-augmented generation model to generate a reply to the problem text, and the language style information required for the retrieval-augmented generation model to generate a reply to the problem text can be queried through the retrieval module of the retrieval-augmented generation model, so that the generation module of the retrieval-augmented generation model can generate a reply reference text with a language style consistent with the language style information included in the second type of query result according to the professional knowledge information included in the first type of query result for the target planner as a reference to assist the target planner in replying to the problem, thereby improving the work efficiency of the target planner and further reducing the cost of the target planner for business execution. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic flowchart of a retrieval-based generative text reply method for assisting a planner provided by an exemplary embodiment;

[0019] Figure 2 is a schematic diagram of the acquisition process of the first type of query result and the second type of query result provided by an exemplary embodiment;

[0020] Figure 3 is a schematic diagram of the generation process of the reply reference text provided by an exemplary embodiment;

[0021] Figure 4 is a schematic diagram of the detection process provided by an exemplary embodiment;

[0022] Figure 5 is a schematic structural diagram of a device provided by an exemplary embodiment;

[0023] Figure 6 is a block diagram of a retrieval-based generative text reply device for assisting a planner provided by an exemplary embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this specification will be clearly and completely described below in conjunction with specific embodiments of this specification and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, rather than all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope protected by this specification.

[0025] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this specification are all information and data that have been authorized by the user or fully authorized by all parties. And the collection, use, and processing of relevant data need to comply with relevant laws, regulations, and standards of relevant countries and regions, and corresponding operation entrances are provided for users to choose to authorize or refuse.

[0026] The following will detail the technical solutions provided by each embodiment of this specification in conjunction with the drawings.

[0027] Figure 1 is a schematic flowchart of a retrieval generative text reply method for assisting planners provided by an exemplary embodiment, including:

[0028] S100: Obtain the problem text of the user.

[0029] In this specification, the business platform can obtain the problem text sent by the user during the conversation between the target planner and the user, so as to generate a reply reference text according to the problem text currently sent by the user and the historical conversation record between the target planner and the user, and then provide the reply reference text to the target planner, so that the target planner can reply to the user according to the reply reference text.

[0030] Among them, the above-mentioned problem text can be the problem text directly input by the user through the text box, or the text corresponding to the document (such as PDF, Word, etc.) containing the problem description uploaded by the user, or the text obtained after converting the colloquial problem input by the user into text form by using speech recognition technology, etc.

[0031] In this specification, the execution entity for implementing the retrieval generative text reply method for assisting planners can refer to a specified device such as a server set in the business platform, or terminal devices such as desktop computers and laptop computers. For the sake of description, the following will only take the server as the execution entity as an example to illustrate the retrieval generative text reply method for assisting planners provided by this specification.

[0032] S102: Input the problem text into a preset retrieval-augmented generation model, and query the first type of query results and the second type of query results through the retrieval module of the retrieval-augmented generation model; the first type of query results is used to provide the professional knowledge information required for the retrieval-augmented generation model to reply to the problem text, and the second type of query results is used to provide the language style information required for the retrieval-augmented generation model to reply to the problem text.

[0033] In this specification, after the server obtains the problem text of the user, it can input the problem text into the retrieval module included in a preset retrieval-augmented generation model (Retrieval-Augmented Generation, RAG), so that the retrieval module of the retrieval-augmented generation model can query the first type of query results and the second type of query results from different preset data sets according to the received problem text, specifically as Figure 2 shown.

[0034] Figure 2 is a schematic diagram of the acquisition process of the first type of query results and the second type of query results provided by an exemplary embodiment.

[0035] Combined with Figure 2 it can be seen that the above different data sets may include: a first type of data set containing historical conversation data between the target planner and the users of the business platform, a second type of data set containing historical conversation data between all planners (i.e., all planners of the business platform) and the users of the business platform, and a knowledge base containing the clause information involved in insurance products (i.e., the legal clause details involved in insurance products). Of course, the knowledge base here may also contain professional knowledge information such as the coverage scope, applicable objects, claim settlement conditions, premium calculation rules, launch information of new industry insurance products, financial and wealth planning (portfolio allocation, risk management strategies, characteristics of savings tools, risk-return characteristics of investment products, etc.), and tax planning.

[0036] The above first type of query results can be used to provide the professional knowledge information required for the generation module of the retrieval-augmented generation model to generate a reply to the problem text. The professional knowledge information here may include: dialogue strategy information, clause information involved in the product itself, and financial knowledge information involved in the product, etc.

[0037] The server can input the problem text into a preset retrieval-augmented generation model, and query the historical conversation data matching the problem text from a preset first data set through the retrieval module of the retrieval-augmented generation model, and query the information matching the problem text from the knowledge base as the first type of query results.

[0038] Specifically, in actual application scenarios, the conversations between planners and users can be divided into strategic stages such as: opening remarks, information collection, product recommendation, solution design, and promoting product sales. Different conversation strategies are adopted for different strategic stages. At this time, the server can determine the strategic stage of the conversation between the target planner and the current user based on the historical conversation records between the target planner and the current user and / or the problem text of the user. Furthermore, according to the strategic stage of the conversation between the target planner and the current user, historical conversation data matching the problem text can be retrieved from the first dataset as the first type of query result.

[0039] In addition, in the insurance product recommendation scenario, different insurance products often have a high degree of complexity. Different product terms, details, and exceptions may lead to differences in understanding, resulting in a relatively low accuracy of the generated reply reference text. Based on this, the server can also retrieve clause information matching the problem text from the knowledge base as the first type of query result.

[0040] In addition, to further improve the work efficiency of the target planner, the server can also query historical conversation data matching the problem text from a preset second dataset through the retrieval module of the retrieval-enhanced generation model as the second type of query result.

[0041] The above-mentioned second type of query result can be used to provide the retrieval-enhanced generation model with the language style information required to generate a reply to the problem text. Here, the language style information is used to reflect the language characteristics and habitual expression methods of the target planner when communicating with users. For example: vocabulary selection preferences, sentence structures, interaction patterns (such as: ways of initiating topics, ways of responding to others' opinions, question habits, etc.), emotional colors, idiomatic expressions, etc.

[0042] In addition, to improve the accuracy of retrieving historical conversation data matching the problem text from each preset dataset, the server can also obtain the user's personal comprehensive attribute data through the retrieval-enhanced generation model and, based on the personal comprehensive attribute data, retrieve the first type of query result and the second type of query result matching the problem text from the first dataset.

[0043] In the above content, the personal comprehensive attribute data is used to reflect at least one of the user's asset status, financial goals, health status, and business status already handled.

[0044] Specifically, the server can also predict the user's product demand information through the retrieval-enhanced generation model based on the user's personal comprehensive attribute data. Furthermore, according to the user's product demand information, the first type of query result and the second type of query result matching the problem text can be retrieved from the first dataset.

[0045] It should be noted that in actual application scenarios, the server can also determine the weight coefficient of each historical conversation data included in the first type of dataset and the second type of dataset according to the attribute data of the planner corresponding to the historical conversation data. Furthermore, it can enable the retrieval module of the retrieval-enhanced generation model to query the first type of query result and the second type of query result that match the question text from the first type of dataset and the second type of dataset according to the weight coefficients of each historical conversation data included in the first type of dataset and the second type of dataset.

[0046] Among them, the above-mentioned attribute data is used to reflect the professional ability, work experience, past performance, etc. of the planner. For example: graduation institution, relevant certificates, work experience, professional skills, customer service records, customer retention rate, annual sales volume, number of new customers acquired, etc.

[0047] S104: Through the generation module of the retrieval-enhanced generation model, generate a reply reference text according to the first type of query result and the second type of query result, and provide the reply reference text to the target planner for the target planner to reply to the user according to the reply reference text.

[0048] After the server obtains the first type of query result and the second type of query result, it can input the first type of query result and the second type of query result into the generation module of the retrieval-enhanced generation model, so as to generate a reply reference text through the generation module of the retrieval-enhanced generation model according to the first type of query result and the second type of query result, and provide the reply reference text to the target planner, so that the target planner determines to adopt the reply reference text and sends the reply reference text to the user for business execution, or enables the target planner to perform a rewriting operation on the reply reference text, thereby obtaining a modified reply reference text and sending the modified reply reference text to the user for business execution.

[0049] Specifically, after the server obtains the first type of query result and the second type of query result, it can generate a prompt message according to the question text, the first type of query result and the second type of query result. Furthermore, it can input the prompt message into the generation module of the retrieval-enhanced generation model, so as to generate a reply reference text through the generation module of the retrieval-enhanced generation model according to the prompt message.

[0050] Among them, the above-mentioned prompt message is used to guide the generation module of the retrieval-enhanced generation model to generate a reply reference text that is consistent with the language style information included in the second type of query result according to the professional knowledge information included in the first type of query result.

[0051] For example: Suppose you are an insurance product planner. Please refer to the content in "The First Type of Query Results" and generate a reference response text for the "Problem Text", and the language style of the reference response text needs to be consistent with the following example: "The Second Type of Query Results".

[0052] It should be noted that the above retrieval-enhanced generation model needs to be trained before it can be deployed to the server for business execution. Among them, the training method for the retrieval-enhanced generation model can be: obtaining historical problem texts, and inputting the historical problem texts into the retrieval-enhanced generation model to be trained, so as to query the first type of sample query results and the second type of sample query results that match the historical problem texts through the retrieval module of the retrieval-enhanced generation model, and generating a sample reference response text according to the first type of sample query results and the second type of sample query results through the generation module of the retrieval-enhanced generation model. Furthermore, the retrieval-enhanced generation model can be trained according to the actual adoption situation corresponding to the sample reference response text to obtain the trained retrieval-enhanced generation model.

[0053] Among them, there are various methods for training the retrieval-enhanced generation model according to the actual adoption situation corresponding to the sample reference response text. For example: determining the reward value corresponding to the sample reference response text according to the actual adoption situation corresponding to the sample reference response text and a preset reward function. Among them, if the sample reference response text is adopted, a positive reward is given, otherwise, a negative reward is given. Furthermore, the retrieval-enhanced generation model can be trained according to the reward value corresponding to the sample reference response text.

[0054] Another example: presenting the sample reference response text to multiple planners, and determining the adoption rate corresponding to the sample reference response text according to the adoption situation of the multiple planners for the sample reference response text. Furthermore, the target loss value can be determined according to the adoption rate corresponding to the sample reference response text, and the retrieval-enhanced generation model can be trained with the goal of minimizing the target loss value. Among them, the adoption rate corresponding to the sample reference response text is negatively correlated with the determined target loss value.

[0055] In the actual application scenario, since different target planners often have different evaluation criteria for the quality of the reference response text, and in order to make the reference response text generated by the retrieval-enhanced generation model more in line with the personalized tendency of the target planner, so as to increase the adoption rate of the reference response text and thus improve the efficiency of the target planner, the server can also represent the evaluation criteria of the target planner for the quality of the reference response text through a preset hint feature, so as to guide the retrieval-enhanced generation model to generate a reference response text that more meets the evaluation criteria of the target planner through the hint feature representation, specifically as Figure 3 shown.

[0056] Figure 3 It is a schematic diagram of the generation process of the reply reference text provided by an exemplary embodiment.

[0057] Combined with Figure 3 It can be seen that the server can input the first type of query results, the second type of query results, and the preset hint feature representation into the generation module of the retrieval-augmented generation model to obtain the reply reference text.

[0058] Among them, the above-mentioned hint feature representation is obtained by performing multiple rounds of adjustment on the preset initial hint feature representation according to the adoption situation of the target planner for the historical reply reference text, and the initial hint feature representation here can be randomly generated.

[0059] Specifically, for each round of adjustment, the server can determine the hint feature representation to be adjusted in this round of adjustment, as well as the historical first type of query results and historical second type of query results required in this round of adjustment. Furthermore, the server can input the hint feature representation to be adjusted, the historical first type of query results, and the historical second type of query results into the generation module of the retrieval-augmented generation model to obtain the historical reply reference text, and can adjust the hint feature representation to be adjusted according to the adoption situation of the target planner for the historical reply reference text (such as: directly adopted, adopted after modification, not adopted, etc.) to obtain the hint feature representation.

[0060] Furthermore, in order to improve the accuracy of the reply reference text generated by the retrieval-augmented generation model, before providing the reply reference text to the target planner, the server can also detect the reply reference text generated by the retrieval-augmented generation model, and determine whether to provide the reply reference text to the target planner or regenerate the reply reference text according to the detection result, specifically as Figure 4 shown.

[0061] Figure 4 It is a schematic diagram of the detection process provided by an exemplary embodiment.

[0062] Combined with Figure 4 It can be seen that the server can input the reply reference text into the preset detection model to obtain the detection result of the reply reference text, and when determining that the difference between the language style of the reply reference text and the language style of the target planner meets the preset conditions according to the detection result, provide the reply reference text to the target planner so that the target planner can reply to the user according to the reply reference text.

[0063] Specifically, the server can generate detection prompt information according to the reply reference text and input the detection prompt information into the detection model to obtain the detection result of the reply reference text.

[0064] Among them, the above detection prompt information is used to guide the detection model to detect the reply reference text to obtain a detection result. The detection prompt information here can be, for example: "Reply reference text. This is a reply reference text generated by a large language model in response to the 'question text'. Please give the detection result based on the difference between the language style of this text and the language style of the target planner."

[0065] The above detection result is used to reflect whether the difference between the language style of the reply reference text and the language style of the target planner meets the preset conditions. The preset conditions here can be determined according to actual needs.

[0066] For example: If the classification result used to characterize the language style of the reply reference text is consistent with the target classification result used to characterize the language style of the target planner, it can be regarded as meeting the preset conditions.

[0067] Another example: If the lexical matching degree between the specified type of vocabulary used to reflect the language style of the reply reference text and the specified type of vocabulary used for the language style of the target planner reaches the preset threshold, it can be regarded as meeting the preset conditions.

[0068] In addition, the server can also evaluate the accuracy and quality of the reply reference text through the detection model to obtain an evaluation result, and based on the evaluation result, determine whether to provide the reply reference text to the target planner.

[0069] It can be seen from the above content that the server can repeatedly detect the language style of the reply reference text output by the retrieval augmented generation model through the neural network model to improve the accuracy of the generated reply reference text. However, during the repeated detection process, if it is determined that the reply reference text does not meet the preset conditions, it is also necessary to regenerate the reply reference text through the retrieval augmented generation model, thereby reducing the reply efficiency.

[0070] Based on this, in order to ensure the reply efficiency while improving the reply accuracy, the server can also generate at least one candidate reply reference text through the retrieval augmented generation model according to the first type of query result and the second type of query result, and then can input each candidate reply reference text into a preset detection model so that the detection model selects a reply reference text from each candidate reply reference text.

[0071] Specifically, the server can generate at least one candidate reply reference text through the generation module of the retrieval augmented generation model according to the first type of query result and the second type of query result, and then can generate detection prompt information based on each candidate reply reference text, and input each candidate reply reference text and the detection prompt information into the detection model so that the detection model selects a reply reference text from each candidate reply reference text.

[0072] At this time, the above detection prompt information is used to guide a preset detection model to detect the language style and content accuracy of each candidate reply reference text.

[0073] The above detection prompt information may be, for example: "Each candidate reply reference text". This is each candidate reply reference text generated by a large language model for replying to the "question text". Please screen out at least one candidate reply reference text from each candidate reply reference text and output it according to the difference between the language style of each candidate reply reference text and the language style of the target planner, and the content accuracy of each candidate reply reference text.

[0074] It should be noted that in this specification, after the server provides the reply reference text to the target planner, it can also adjust the model parameters of the retrieval enhanced generation model according to the adoption situation of the target planner for the reply reference text, obtain an adjusted retrieval enhanced generation model, and adjust the prompt feature representation in the above content to obtain an adjusted prompt feature representation. Furthermore, when replying to the question text proposed by the subsequent user, the adjusted retrieval enhanced generation model can generate a reply reference text according to the adjusted prompt feature representation.

[0075] As can be seen from the above content, the server can efficiently retrieve a large amount of historical conversation data through the retrieval module of the retrieval enhanced generation model to retrieve the first type of query results and the second type of query results of the professional knowledge information and language style information required to generate the reply reference text for replying to the question proposed by the user. Furthermore, the generation module of the retrieval enhanced generation model can generate a reference reply text with higher accuracy and a language style consistent with that of the target planner according to the first type of query results and the second type of query results, and provide it to the target planner, so that the target planner can reply to the user based on the reply reference text, thereby improving the business execution efficiency of the planner and reducing the cost of insurance recommendation.

[0076] Figure 5 It is a schematic structural diagram of a device provided by an exemplary embodiment. Please refer to Figure 5, at the hardware level, the device includes a processor 502, an internal bus 504, a network interface 506, a memory 508, and a non-volatile memory 510. Of course, it may also include other hardware required for other functions. One or more embodiments of this specification can be implemented in a software manner. For example, the processor 502 reads the corresponding computer program from the non-volatile memory 510 into the memory 508 and then runs it. Of course, in addition to the software implementation method, one or more embodiments of this specification do not exclude other implementation methods, such as logical devices or a combination of software and hardware, etc. That is to say, the execution subject of the following processing flow is not limited to each logical unit, and can also be hardware or logical devices.

[0077] Please refer to Figure 6 , the retrieval generative text reply device for assisting planners can be applied to a device as shown in Figure 5 to implement the technical solutions of this specification. Among them, the retrieval generative text reply device for assisting planners may include:

[0078] An acquisition module 601, configured to acquire the question text of the user;

[0079] A query module 602, configured to input the question text into a preset retrieval enhanced generation model, so as to query a first type of query result and a second type of query result through the retrieval module of the retrieval enhanced generation model; the first type of query result is used to provide the professional knowledge information required for replying to the question text for the retrieval enhanced generation model, and the second type of query result is used to provide the language style information required for replying to the question text for the retrieval enhanced generation model;

[0080] An execution module 603, configured to generate a reply reference text according to the first type of query result and the second type of query result through the generation module of the retrieval enhanced generation model, and provide the reply reference text to the target planner for the target planner to reply to the user according to the reply reference text.

[0081] Optionally, the query module 602 is specifically configured to input the question text into a preset retrieval enhanced generation model, so that the retrieval module of the retrieval enhanced generation model acquires the personal comprehensive attribute data of the user, and queries a first type of query result and a second type of query result according to the personal comprehensive attribute data, where the personal comprehensive attribute data is used to reflect at least one of the user's asset status, financial goals, health status, and business status already handled.

[0082] Optionally, the query module 602 is specifically configured to input the question text into a preset retrieval-augmented generation model, so as to query a first type of query result and a second type of query result from different preset data sets through the retrieval module of the retrieval-augmented generation model; the data sets include: a first type of data set containing the historical conversation data between the full-time planner and the user, a second type of data set containing the historical conversation data between the target planner and the user, and a knowledge base containing the clause information involved in the insurance product.

[0083] Optionally, the query module 602 is specifically configured to determine the policy stage of the conversation between the target planner and the current user, and query, according to the policy stage, the historical conversation data matching the question text from the first data set as the first type of query result; and query the clause information matching the question text from the knowledge base as the first type of query result.

[0084] Optionally, the execution module 603 is specifically configured to generate a reply reference text according to the first type of query result, the second type of query result, and a preset prompt feature representation through the generation module of the retrieval-augmented generation model; the prompt feature representation is used to reflect the adoption tendency of the target planner for the reply reference text.

[0085] Optionally, the device further includes: a training module 604;

[0086] The training module 604 is specifically configured to adjust the model parameters of the retrieval-augmented generation model according to the adoption situation of the target planner for the reply reference text to obtain an adjusted retrieval-augmented generation model; and adjust the prompt feature representation to obtain an adjusted prompt feature representation.

[0087] Optionally, the execution module 603 is specifically configured to input the reply reference text into a preset detection model to obtain a detection result of the reply reference text, where the detection result is used to reflect whether the difference between the language style of the reply reference text and the language style of the target planner meets a preset condition; if it is determined according to the detection result that the difference between the language style of the reply reference text and the language style of the target planner meets the preset condition, the reply reference text is provided to the target planner, so that the target planner can reply to the user according to the reply reference text.

[0088] Optionally, the execution module 603 is specifically configured to: through the generation module of the retrieval-enhanced generation model, generate at least one candidate reply reference text according to the first type of query result and the second type of query result; generate detection prompt information according to each candidate reply reference text, where the detection prompt information is used to guide a preset detection model to detect the language style and content accuracy of each candidate reply reference text; input each candidate reply reference text and the detection prompt information into the detection model, so that the detection model selects a reply reference text from each candidate reply reference text.

[0089] Based on the same concept as the above method, this specification also provides an electronic device, including: a processor; a memory for storing processor-executable instructions; wherein, the processor realizes the steps of the method as described in any of the above embodiments by running the executable instructions.

[0090] Based on the same concept as the above method, this specification also provides a computer-readable storage medium, on which computer instructions are stored, and when the instructions are executed by a processor, the steps of the method as described in any of the above embodiments are realized.

[0091] Based on the same concept as the above method, this specification also provides a computer program product, including computer programs / instructions, and when the computer programs / instructions are executed by a processor, the steps of the method as described in any of the above embodiments are realized.

Claims

1. A retrieval - augmented text response method for assisting planners, comprising: Obtaining the question text of the user; Inputting the question text into a preset retrieval - augmented generation model, so as to query a first - type query result and a second - type query result through the retrieval module of the retrieval - augmented generation model; The first - type query result is used to provide the professional knowledge information required for the retrieval - augmented generation model to reply to the question text, and the second - type query result is used to provide the language style information required for the retrieval - augmented generation model to reply to the question text; Through the generation module of the retrieval - augmented generation model, generating a reply reference text according to the first - type query result and the second - type query result, and providing the reply reference text to the target planner for the target planner to reply to the user according to the reply reference text.

2. The method according to claim 1, inputting the question text into a preset retrieval - augmented generation model, so as to query a first - type query result and a second - type query result through the retrieval module of the retrieval - augmented generation model, specifically comprising: Inputting the question text into a preset retrieval - augmented generation model, so that the retrieval module of the retrieval - augmented generation model obtains the personal comprehensive attribute data of the user, and according to the personal comprehensive attribute data, querying a first - type query result and a second - type query result, where the personal comprehensive attribute data is used to reflect at least one of the user's asset status, financial goals, health status, and business status already handled.

3. The method according to claim 1, inputting the question text into a preset retrieval - augmented generation model, so as to query a first - type query result and a second - type query result through the retrieval module of the retrieval - augmented generation model, specifically comprising: Inputting the question text into a preset retrieval - augmented generation model, so as to query a first - type query result and a second - type query result from preset different data sets through the retrieval module of the retrieval - augmented generation model; the data sets include: a first - type data set containing the historical dialogue data of all planners and users, a second - type data set containing the historical dialogue data of the target planner and the user, and a knowledge base containing the clause information involved in insurance products.

4. The method according to claim 3, querying the first - type query result from preset different data sets, specifically comprising: Determining the strategy stage of the conversation between the target planner and the current user, and according to the strategy stage, querying the historical dialogue data matching the question text from the first data set as the first - type query result; and Querying the clause information matching the question text from the knowledge base as the first - type query result.

5. The method according to claim 1, generating a reply reference text according to the first - type query result and the second - type query result through the generation module of the retrieval - augmented generation model, specifically comprising: Through the generation module of the retrieval enhanced generation model, a reply reference text is generated according to the first type of query result, the second type of query result, and a preset prompt feature representation; The prompt feature representation is used to reflect the adoption tendency of the target planner for the reply reference text.

6. The method according to claim 5, further comprising: Adjusting the model parameters of the retrieval enhanced generation model according to the adoption situation of the target planner for the reply reference text to obtain an adjusted retrieval enhanced generation model; and adjusting the prompt feature representation to obtain an adjusted prompt feature representation.

7. The method according to claim 1, providing the reply reference text to the target planner for the target planner to reply to the user according to the reply reference text, specifically comprising: Inputting the reply reference text into a preset detection model to obtain a detection result of the reply reference text, where the detection result is used to reflect whether the difference between the language style of the reply reference text and the language style of the target planner meets a preset condition; If it is determined according to the detection result that the difference between the language style of the reply reference text and the language style of the target planner meets the preset condition, the reply reference text is provided to the target planner so that the target planner can reply to the user according to the reply reference text.

8. The method according to claim 1, generating a reply reference text through the generation module of the retrieval enhanced generation model according to the first type of query result and the second type of query result, specifically comprising: Generating at least one candidate reply reference text through the generation module of the retrieval enhanced generation model according to the first type of query result and the second type of query result; Generating detection prompt information according to each candidate reply reference text, where the detection prompt information is used to guide a preset detection model to detect the language style and content accuracy of each candidate reply reference text; Inputting each candidate reply reference text and the detection prompt information into the detection model so that the detection model selects a reply reference text from the candidate reply reference texts.

9. A retrieval generative text reply device for assisting a planner, comprising: An acquisition module for acquiring the question text of the user; A query module for inputting the question text into a preset retrieval enhanced generation model to query a first type of query result and a second type of query result through the retrieval module of the retrieval enhanced generation model; The first type of query result is used to provide the professional knowledge information required for the retrieval enhanced generation model to reply to the question text, and the second type of query result is used to provide the language style information required for the retrieval enhanced generation model to reply to the question text; An execution module for generating a reply reference text through the generation module of the retrieval enhanced generation model according to the first type of query result and the second type of query result, and providing the reply reference text to the target planner for the target planner to reply to the user according to the reply reference text.

10. The device according to claim 9, wherein the query module is specifically configured to input the problem text into a preset retrieval augmented generation model, so as to query, through the retrieval module of the retrieval augmented generation model, a first type of query result and a second type of query result from different preset data sets; the data sets include: A first type of dataset containing the historical conversation data between the full-time planner and the user, a second type of dataset containing the historical conversation data between the target planner and the user, and a knowledge base containing the clause information involved in the insurance product.

11. The device according to claim 10, wherein the query module is specifically configured to determine the policy stage of the conversation between the target planner and the current user, and query, according to the policy stage, the historical conversation data that matches the problem text from the first dataset as the first type of query result; and query the clause information that matches the problem text from the knowledge base as the first type of query result.

12. An electronic device, comprising: Processor; A memory for storing instructions executable by the processor; wherein the processor realizes the steps of the method according to any one of claims 1-8 by running the executable instructions.

13. A computer-readable storage medium, on which computer instructions are stored, and when the instructions are executed by the processor, the steps of the method according to any one of claims 1-8 are realized.

14. A computer program product, including a computer program / instructions, and when the computer program / instructions are executed by the processor, the steps of the method according to any one of claims 1-8 are realized.

Images