Method for generating question reply through retrieval enhancement based on iteration mode

Abstract

The invention discloses a method for generating question reply through retrieval enhancement based on an iteration mode. The method comprises the following steps: S1, taking user input as a retrieval text for retrieval; s2, summarizing the large model to obtain a primary information summary; s3, sorting the content retrieved and returned in the step S1 into a recall fragment set; s4, summarizing the initial information in the S2 as a retrieval text to perform the retrieval; s5, the big model perfects the previous information summary by using the content returned by the current retrieval to obtain the current information summary; s6, sorting the content returned by the retrieval into a recall fragment set, identifying whether a new fragment exists or not, if not, entering step S8, and otherwise, entering step S7; s7, performing current retrieval by taking the current information summary as a retrieval text, and entering S5; and S8, replying user input according to the information summary. The RAG accuracy is improved based on an iteration mode, more perfect information is obtained through iteration retrieval, and the model reply quality is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method for generating question responses based on retrieval enhancement in an iterative manner. Background Art

[0002] RAG is the abbreviation of Retrieval Augmented Generation. It obtains certain text fragments through retrieval and provides them to a large model (also called a large language model, English: Large Language Model, abbreviated LLM) as supplementary information for generating answers, thereby reducing the probability of hallucinations when the large model is generated.

[0003] The prerequisite of RAG is to store the data in a database (called a knowledge base). A common method is to divide several articles into multiple segments (shards), extract the shard features through an embedding model, vectorize them, and store them in a vector database. The vector database contains the shard vectors and the shard text content.

[0004] There are two common methods for shard retrieval, namely semantic retrieval (vector retrieval) and full-text retrieval (keyword retrieval). Semantic retrieval is to vectorize the question and query the vector library for similar vectors, thereby recalling (i.e. taking out) a certain number of shards. Full-text retrieval is to extract the words in the question and then fuzzily match the text content of the shards based on the extracted words, thereby recalling a certain number of shards. Both methods will also be used simultaneously (hybrid retrieval) to return a certain number of shards with the highest search scores. The text content of these shards can be used as reference information for the generation of the large model.

[0005] Most of the existing RAG systems on the market obtain answers to questions through a single round of search enhancement generation, or obtain a long text answer through multiple rounds of search combined with generation. This method is essentially a single round of search enhancement generation, which splices multiple short answers into a long answer. Single round search is difficult to solve the problem of processing the reference part in the document and the problem that the required information is distributed in multiple fragments and is difficult to recall.

[0006] Therefore, in order to solve these problems, it is necessary to provide a method for generating question responses based on iterative retrieval enhancement to overcome the shortcomings of the existing technology. Summary of the invention

[0007] The purpose of the present invention is to avoid the shortcomings of the prior art and provide a method for generating question responses based on retrieval enhancement in an iterative manner, thereby improving the accuracy of RAG in an iterative manner.

[0008] The purpose of the present invention is achieved through the following technical measures.

[0009] A method for generating question responses based on retrieval enhancement in an iterative manner is provided, which is performed by the following steps:

[0010] S1. Use the user input as the search text for retrieval, and retrieve a certain amount of information fragments according to the user input question;

[0011] S2. The large model summarizes the information in the content returned by S1 that is helpful for answering the user's input question and obtains the initial information summary;

[0012] S3. Arrange the content returned by S1 into the recall shard set;

[0013] S4. Use the initial information summary in S2 as the search text for this search;

[0014] S5. The large model uses the information in the content returned by this search that is useful for answering the user's input question to improve the previous information summary and obtain the current information summary;

[0015] S6. Arrange the content returned by this search into the recalled shard set, and identify whether there are new shards. If not, proceed to step S8, otherwise proceed to S7;

[0016] S7. Perform this search based on the current information summary as the search text and enter S5;

[0017] S8. Respond to the user input based on the current information summary.

[0018] Preferably, the above-mentioned method of generating question responses through retrieval enhancement based on iteration, S5 specifically comprises: the large model summarizes the information in the content returned by this retrieval that is helpful for answering the question input by the user to obtain this information summary; and then uses this information summary to improve the previous information summary to obtain the current information summary.

[0019] Another preferred method of generating question responses based on iterative retrieval enhancement is as follows: S5 is specifically: the large model superimposes the information in the content returned by this retrieval that is helpful for answering the question input by the user with the previous information summary to obtain the current information summary.

[0020] Preferably, in the above-mentioned method of generating question responses based on iterative retrieval enhancement, in S6, the content returned by this retrieval is sorted into a collection of recalled shards, which is to perform a union processing of the historical recalled shards in the process and the current recalled shards.

[0021] Preferably, in the above-mentioned method of generating question responses based on retrieval enhancement in an iterative manner, identifying whether there are new fragments in S6 specifically involves identifying the difference between the currently recalled fragments and the historically recalled fragments.

[0022] Preferably, in the above-mentioned method of generating question responses through retrieval enhancement based on iteration, S2 or S5 is performed by a large model, using the following prompt words:

[0023] Please complete the current integration information based on user questions, current integration information and known information, so that you can answer user questions completely and clearly based on the current integration information in the future;

[0024] If the current integration information is empty, rewrite the current integration information based on the known information;

[0025] If the known information contains data, the specific data is displayed in the integrated information;

[0026] User Question:

[0027] {{userInput}}

[0028] Current integration information:

[0029] {{summary}}

[0030] Known information:

[0031] {{chunks}}.

[0032] Preferably, in the above-mentioned method for generating question responses based on iterative retrieval enhancement, S8 summarizes and responds to user input through a large model based on the current information, using the following prompt words:

[0033] Answer questions based on current integrated information;

[0034] When answering questions, if there is data in the knowledge base, display the specific data in the answer;

[0035] User Question:

[0036] {{userInput}}

[0037] Current integration information:

[0038] {{summary}};

[0039] The content contained in {{}} is filled in dynamically during implementation. {{userInput}} corresponds to the content entered by the user, {{summary}} corresponds to the information summarized in the current information, and {{chunks}} corresponds to the recalled shard information.

[0040] Large Language Model (LLM), also known as Large Language Model, is an artificial intelligence model designed to understand and generate human language. They are trained on large amounts of text data and can perform a wide range of tasks, including text summarization, translation, sentiment analysis, and more. LLMs are characterized by their large size, which helps them learn complex patterns in language data.

[0041] The method for generating question replies based on iterative retrieval enhancement of the present invention is to extract and improve the information required for replying to the user input according to the user input and the recalled text in each iteration through multiple iterations, and whether the current recall has new fragments is used as a sign of whether to end the iteration, and finally the improved information is used as the supplementary information of the RAG, so as to achieve the purpose of improving the accuracy of the reply. Through multiple searches, the information can be gradually improved, so as to obtain the partial information of the document reference and the information missing from a single search. At the same time, it avoids the problem that too many fragments are recalled at one time, resulting in a large model input exceeding the upper limit of the model window, and too few fragments are recalled at one time, resulting in incomplete information.

[0042] The timing of the end of iteration is the core logic of iteration. The present invention proposes to use whether there are new fragments in the current recall as a sign of whether to end the iteration, so as to avoid the problem of insufficient iteration depth and low flexibility caused by limiting the number of iterations, and also avoid the problem of too many iterations caused by constantly raising sub-questions for iteration. The method of the present invention can solve the problem of missing information in the quoted part of the article, reduce the negative impact caused by the inappropriate number of recall fragments, improve the completeness of the recall information, thereby improving the quality of the answer, and return high-quality answers by using refined information. DETAILED DESCRIPTION

[0043] The present invention will be further described with reference to the following examples.

[0044] Example 1.

[0045] A method for generating question responses based on iterative retrieval enhancement, which improves the accuracy of RAG in an iterative manner. The core idea is to obtain more complete information through iterative retrieval, so as to improve the quality of model responses.

[0046] The method for generating question responses based on iterative retrieval enhancement in this embodiment is performed through the following steps:

[0047] S1. Use the user input as the search text for searching, and retrieve a certain amount of information fragments according to the user input question; S1 searches according to the user input, using the conventional search enhancement RAG process, which will not be repeated here.

[0048] S2. The big model summarizes the information in the content returned by S1 that is helpful for answering the user's input question and obtains the initial information summary. The content returned by S1 is the recalled shard information. Summarization is to let the big model refine the information retrieved by S1 in combination with the user's question, retain only the information that is useful for answering the user's question, and eliminate the information that is useless for answering the user's question.

[0049] S3. Arrange the content returned by S1 into the recall shard set. Step S1 is the first search based on the user input question. The previous recall shard set is empty. S3 directly arranges the content returned by S1 into the recall shard set.

[0050] S4. Use the initial information summary in S2 as the search text for this search. S2 only retains the initial information summary that is useful for answering user questions. Using this as the search text for searching through a large model can improve the accuracy of the search.

[0051] S5. The large model uses the information in the content returned by this search that is helpful in answering the question input by the user to improve the previous information summary and obtain the current information summary.

[0052] S6. Arrange the content returned by this search into the recalled shard set, and identify whether there are new shards. If not, proceed to step S8, otherwise proceed to S7.

[0053] S7. Perform this search based on the current information summary as the search text and enter S5.

[0054] S8. Respond to the user input based on the current information summary.

[0055] S5 can adopt this method: the large model summarizes the information in the content returned by this search that is helpful for answering the questions input by the user to obtain this information summary; then use this information summary to improve the previous information summary to obtain the current information summary.

[0056] S5 can also adopt this method: the large model superimposes the information in the content returned by this retrieval that is helpful in answering the user's input question with the previous information summary to obtain the current information summary.

[0057] In S6, the content returned by this search is organized into a collection of recalled shards, which is to perform a union process on the historical recalled shards in the process and the current recalled shards, and identify the difference between the current recalled shards and the historical recalled shards.

[0058] In the method of generating question responses based on iterative retrieval enhancement in this embodiment, S2 or S5 is performed through a large model, and the following prompt words may be used:

[0059] Please complete the current integration information based on user questions, current integration information and known information, so that you can answer user questions completely and clearly based on the current integration information in the future;

[0060] If the current integration information is empty, rewrite the current integration information based on the known information;

[0061] If the known information contains data, the specific data is displayed in the integrated information;

[0062] User Question:

[0063] {{userInput}}

[0064] Current integration information:

[0065] {{summary}}

[0066] Known information:

[0067] {{chunks}}.

[0068] In the method of generating question responses based on iterative retrieval enhancement in this embodiment, S8 summarizes the responses to user inputs based on the current information through a large model, and the following prompt words may be used:

[0069] Answer questions based on current integrated information;

[0070] When answering questions, if there is data in the knowledge base, display the specific data in the answer;

[0071] User Question:

[0072] {{userInput}}

[0073] Current integration information:

[0074] {{summary}};

[0075] The content contained in {{}} is filled in dynamically during implementation. {{userInput}} corresponds to the content entered by the user, {{summary}} corresponds to the information summarized in the current information, and {{chunks}} corresponds to the recalled shard information.

[0076] Whether it is semantic retrieval or full-text retrieval, the chapter number, keywords, etc. contained in the quoted part in the summary information may trigger the retrieval conditions, thereby recalling the fragments of the quoted part and further improving the information. The summary information of this method removes irrelevant information, so searching based on the summary information can increase the retrieval weight of the effective chapter information. Avoid the fragments of the retrieval being occupied by fragments of irrelevant information. The search enhancement generation does not require the source of supplementary information. In the end, the supplementary information and questions are input into the large model through prompt words for answer generation.

[0077] The method for generating question replies based on iterative retrieval enhancement of the present invention is to extract and improve the information required for replying to the user input according to the user input and the recalled text in each iteration through multiple iterations, and whether the current recall has new fragments is used as a sign of whether to end the iteration, and finally the improved information is used as the supplementary information of the RAG, so as to achieve the purpose of improving the accuracy of the reply. Through multiple searches, the information can be gradually improved, so as to obtain the partial information of the document reference and the information missing from a single search. At the same time, it avoids the problem that too many fragments are recalled at one time, resulting in a large model input exceeding the upper limit of the model window, and too few fragments are recalled at one time, resulting in incomplete information.

[0078] The timing of the end of iteration is the core logic of iteration. The present invention proposes to use whether there are new fragments in the current recall as a sign of whether to end the iteration, so as to avoid the problem of insufficient iteration depth and low flexibility caused by limiting the number of iterations, and also avoid the problem of too many iterations caused by constantly raising sub-questions for iteration. The method of the present invention can solve the problem of missing information in the quoted part of the article, reduce the negative impact caused by the inappropriate number of recall fragments, improve the completeness of the recall information, thereby improving the quality of the answer, and return high-quality answers by using refined information.

[0079] Example 2.

[0080] A method for generating question responses based on iterative retrieval enhancement of the present invention is described through a specific example.

[0081] The following is a comparison of the results of using the traditional one-shot RAG and the method of generating question responses based on iterative retrieval enhancement using the present invention for the same data set (365 use cases) and the same knowledge base.

[0082] In this embodiment, the traditional method of generating a question response by RAG is as follows:

[0083] S1. Use user input as search text for searching;

[0084] S2. Send the search return content plus user input to the big model;

[0085] S3. The large model answers the questions entered by the user based on the information of the retrieved content.

[0086] The search results may contain "Except for the situation in Chapter xxx" and other references to articles that were not searched and returned, resulting in the large model answer results containing "Except for the situation in Chapter xxx" and other content.

[0087] The test knowledge base contains 28 fire protection laws and regulations. The documents entered in the knowledge base cannot cover all the problems in the test set. The judgment results use a large model (GPT-4o) as a referee, and the results are judged through manual inspection.

[0088] The referee's prompts are as follows:

[0089] <Character>

[0090] You are an answer judge, please judge whether the user's answer is similar to the reference answer.

[0091] <Requirements>

[0092] First output the judgment basis, then output whether it is similar. If it is similar, output "True", otherwise output "False" <other requirements>

[0093] 1. Ignore the order of the answers, as long as the meaning is similar, it is correct.

[0094] 2. The reference answer has a specific value and needs to correspond to the value to be correct

[0095] 3. Ignore the format of the answer.

[0096] The format for inputting answers to referees is as follows:

[0097] Question: {{question}}

[0098] Reference answer: {{answer}}

[0099] User answer: {{assistant_answer}}

[0100] The contents contained in {{}} are filled in dynamically during implementation, which are user questions, reference answers and answers returned by RAG.

[0101] The correct answer rate of the method of the present invention is about 0.6411, and the correct answer rate of the traditional RAG using one search is about 0.5096.

[0102] The same comparative test was conducted multiple times with different models, and the difference in correct answer rate between the iterative RAG method of the present invention and the traditional RAG of one-time retrieval was between 1% and 14%. The stronger the model capability, the more obvious the advantage of the iterative RAG of the present invention.

[0103] Example 3.

[0104] A specific example is used to illustrate that the method of generating question answers through retrieval enhancement based on iteration of the present invention has advantages over the common method of generating question answers through retrieval enhancement.

[0105] User Question: What is the clearance for the fire lane?

[0106] Answer to the question: should not be less than 4 meters.

[0107] Steps for enhancing the generation of question responses for general search:

[0108] Step 1: Query the knowledge base based on the user's question and obtain the following partial recall shard content:

[0109] "Code for Fire Protection Design of Buildings GB50016-2014 (2018 Edition)

[0110] 7 Firefighting and rescue facilities

[0111] 7.1 Fire Lane

[0112] Explanation of Article 7.1.8 Paragraphs 1, 2 and 3 of this Article are mandatory. In order to ensure that the fire lanes meet the needs of fire trucks passing and fighting building fires, this Article determines the minimum clear width and clearance height of the fire lanes according to the external dimensions of various fire trucks currently in service in China, in accordance with the single lane and taking into account the needs of rapid passage of fire trucks, and puts forward requirements for the turning radius. For buildings, roads and bridges that require the passage of special fire trucks, the clear width and clearance height of the fire lanes should also be increased according to the actual situation of the fire trucks. Since most of the fire lanes in cities or certain areas currently need to use urban roads or public roads in residential areas, the turning radius of fire trucks is generally large, usually 9m to 12m. Therefore, whether it is a dedicated fire lane or other roads or highways that also serve as fire lanes, they should meet the turning radius requirements of fire trucks. The turning radius can be determined by combining the configuration of local fire trucks and the construction and planning of buildings in the area.

[0113] The road slope determined in this article is the slope required for the safe driving of fire trucks, not the slope of the site for fire trucks to park and carry out fire-fighting operations.

[0114] According to the actual firefighting situation, except for high-rise buildings that need to set up firefighting and rescue operation sites, general buildings can directly use fire lanes to carry out firefighting and rescue operations. Therefore, there should be sufficient distance and clearance between the fire lanes and buildings to avoid tall trees, overhead high-voltage power lines, overhead pipe galleries, etc. that affect firefighting and rescue operations. "

[0115] Step 2: Give the shard content and question as input to the big model for answering.

[0116] Answer: According to Article 7.1.8 of the Code for Fire Protection Design of Buildings GB50016-2014 (2018 edition), for buildings, roads and bridges that require the passage of special fire trucks, the clearance height of the fire truck lane should be increased according to the actual situation of the fire trucks.

[0117] The method steps for generating question responses based on the iterative retrieval enhancement of the present invention are as follows:

[0118] Step 1: Query the knowledge base based on the user's question and obtain the following partial recall shard content:

[0119] "Code for Fire Protection Design of Buildings GB50016-2014 (2018 Edition)

[0120] 7 Firefighting and rescue facilities

[0121] 7.1 Fire Lane

[0122] Explanation of Article 7.1.8 Paragraphs 1, 2 and 3 of this Article are mandatory. In order to ensure that the fire lanes meet the needs of fire trucks passing and fighting building fires, this Article determines the minimum clear width and clearance height of the fire lanes according to the external dimensions of various fire trucks currently in service in China, in accordance with the single lane and taking into account the needs of rapid passage of fire trucks, and puts forward requirements for the turning radius. For buildings, roads and bridges that require the passage of special fire trucks, the clear width and clearance height of the fire lanes should also be increased according to the actual situation of the fire trucks. Since most of the fire lanes in cities or certain areas currently need to use urban roads or public roads in residential areas, the turning radius of fire trucks is generally large, usually 9m to 12m. Therefore, whether it is a dedicated fire lane or other roads or highways that also serve as fire lanes, they should meet the turning radius requirements of fire trucks. The turning radius can be determined by combining the configuration of local fire trucks and the construction and planning of buildings in the area.

[0123] The road slope determined in this article is the slope required for the safe driving of fire trucks, not the slope of the site for fire trucks to park and carry out fire-fighting operations.

[0124] According to the actual firefighting situation, except for high-rise buildings that need to set up firefighting and rescue operation sites, general buildings can directly use fire lanes to carry out firefighting and rescue operations. Therefore, there should be sufficient distance and clearance between the fire lanes and buildings to avoid tall trees, overhead high-voltage power lines, overhead pipe galleries, etc. that affect firefighting and rescue operations. "

[0125] Step 2: The large model summarizes the information of the recalled fragments in step 1 based on the information that helps answer the question "What is the clearance of the fire lane?" and obtains "According to Article 7.1.8 of the Code for Fire Protection Design of Buildings GB50016-2014 (2018 Edition), for buildings, roads and bridges that require the passage of special fire trucks, the clearance height of the fire lane should be increased according to the actual situation of the fire trucks."

[0126] Step 3: Organize the text fragments retrieved in step 1 into the historical fragment collection.

[0127] Step 4: Use the information from step 2 to query the knowledge base and get new shard information:

[0128] "Code for Fire Protection Design of Buildings GB50016-2014 (2018 Edition)

[0129] 7 Firefighting and rescue facilities

[0130] 7.1 Fire Lane

[0131] 7.1.8 Fire lanes shall meet the following requirements:

[0132] 1. The clear width and clearance height of the lane should not be less than 4.0m;

[0133] 2. The turning radius should meet the turning requirements of fire trucks;

[0134] 3. There should be no obstacles such as trees, overhead pipelines, etc. that may hinder the operation of fire trucks between the fire truck lane and the building;

[0135] 4. The edge of the fire truck lane on the side close to the building's exterior wall should not be less than 5m from the building's exterior wall;

[0136] 5 The slope of the fire lane should not be greater than 8%. "

[0137] Step 5. The large model uses the recalled fragment information in step 4 to complete the information summary based on the answer to the question "What is the clearance of the fire lane?" and obtains "According to Article 7.1.8 of the Code for Fire Protection Design of Buildings GB50016-2014 (2018 Edition)", the clearance height and net width of the fire lane should not be less than 4.0 meters. This is to ensure that the fire lane can meet the basic requirements for the passage of fire trucks and the fighting of building fires. At the same time, the specification points out that the fire lane should not be set with obstacles that hinder the operation of fire trucks, such as trees and overhead pipelines, and the edge of the fire lane on the side of the building's exterior wall should not be less than 5 meters from the building's exterior wall. This information can help understand the clearance standards of fire lanes and related design requirements."

[0138] Step 6: Compare the historical shard set with the shards found in step 4. If a new shard is found, the new shard flag is assigned a value of "true". Intersect the historical shard set with the shards found in step 4 to obtain a new historical shard set.

[0139] Step 7: Identify the new shard as "true" and use the information obtained in step 5 to summarize the query knowledge base.

[0140] Step 8: The large model uses the recalled shard information from step 7 to improve the information summary based on the answer to the question "What is the clearance of the fire lane?" Because there is no new shard information, the information summary remains unchanged.

[0141] Step 9: Compare the historical shard set with the shards queried in step 7. If no new shards are found, the new shard identifier is assigned a value of "false".

[0142] Step 10: Identify that the new shard is marked as "false", and give the information summary content and user questions as input parameters to the large model for answering.

[0143] Answer: "According to Article 7.1.8 of the Code for Fire Protection Design of Buildings GB50016-2014 (2018 Edition)", the clear height and clear width of the fire lane should not be less than 4.0 meters."

[0144] The method for generating question replies based on iterative retrieval enhancement of the present invention is to extract and improve the information required for replying to the user input according to the user input and the recalled text in each iteration through multiple iterations, and whether the current recall has new fragments is used as a sign of whether to end the iteration, and finally the improved information is used as the supplementary information of the RAG, so as to achieve the purpose of improving the accuracy of the reply. Through multiple searches, the information can be gradually improved, so as to obtain the partial information of the document reference and the information missing from a single search. At the same time, it avoids the problem that too many fragments are recalled at one time, resulting in a large model input exceeding the upper limit of the model window, and too few fragments are recalled at one time, resulting in incomplete information.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the essence and scope of the technical solution of the present invention.

Claims

1. A method for generating question responses based on iterative retrieval enhancement, characterized in that: Proceed as follows: S1. Use the user input as the search text for retrieval, and retrieve a certain amount of information fragments according to the user input question; S2. The large model summarizes the information in the content returned by S1 that is helpful for answering the user's input question to obtain the initial information summary; S3. Arrange the content returned by S1 into the recall shard set; S4. Use the initial information summary in S2 as the search text for this search; S5. The large model uses the information in the content returned by this search that is useful for answering the user's input question to improve the previous information summary and obtain the current information summary; S6. Arrange the content returned by this search into the recalled shard set, and identify whether there are new shards. If not, proceed to step S8, otherwise proceed to S7; S7. Perform this search based on the current information summary as the search text and enter S5; S8. Respond to the user input based on the current information summary.

2. The method for generating question responses based on iterative retrieval enhancement according to claim 1, characterized in that: S5 is specifically: the large model summarizes the information in the content returned by this search that is helpful for answering the question input by the user to obtain this information summary; and then uses this information summary to improve the previous information summary to obtain the current information summary.

3. The method for generating question responses based on iterative retrieval enhancement according to claim 1, characterized in that: S5 is specifically: the large model superimposes the information in the content returned by this search that is helpful in answering the question input by the user with the previous information summary to obtain the current information summary.

4. The method for generating question responses based on iterative retrieval enhancement according to any one of claims 1 to 3, characterized in that: In S6, the content returned by this search is organized into a collection of recalled shards, which is to combine the historical recalled shards in the process with the current recalled shards.

5. The method for generating question responses based on iterative retrieval enhancement according to claim 4, characterized in that: In S6, identifying whether there is a new shard is specifically identifying the difference between the currently recalled shard and the historically recalled shard.

6. The method for generating question responses based on iterative retrieval enhancement according to claim 5, characterized in that: S2 or S5 is performed using a large model, using the following prompts: Please complete the current integration information based on user questions, current integration information and known information, so that you can answer user questions completely and clearly based on the current integration information in the future; If the current integration information is empty, rewrite the current integration information based on the known information; If the known information contains data, the specific data is displayed in the integrated information; User Question: {{userInput}} Current integration information: {{summary}} Known information: {{chunks}}.

7. The method for generating question responses based on iterative retrieval enhancement according to claim 6, characterized in that: S8 summarizes the information and responds to the user input through a large model, using the following prompt words: Answer questions based on current integrated information; When answering questions, if there is data in the knowledge base, display the specific data in the answer; User Question: {{userInput}} Current integration information: {{summary}}; The content contained in {{}} is filled in dynamically during implementation. {{userInput}} corresponds to the content input by the user, {{summary}} corresponds to the information summarized in the current information, and {{chunks}} corresponds to the recalled shard information.