Information processing method and device

By retrieving knowledge vectors with high and low similarity in the vector database and determining candidate response information in combination with large models, the problem of high complexity of large language models is solved, and efficient information replies on devices with limited hardware performance are achieved.

CN119088944APending Publication Date: 2024-12-06LENOVO (BEIJING) LTD
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202411322460.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-20
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The large language model has high complexity in information replies and cannot be applied to electronic devices with limited hardware performance, making it difficult to accurately determine the reply information corresponding to the user input information.

Method used

By retrieving knowledge vectors with high and low similarity from the vector database, determining candidate response information in combination with the big model, and determining target response information based on the candidate response information, the requirements for hardware performance are reduced.

Benefits of technology

It realizes efficient determination of reply information on devices with limited hardware performance, reduces complexity, and improves the accuracy and efficiency of information processing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119088944A_ABST
    Figure CN119088944A_ABST
Patent Text Reader

Abstract

The invention discloses an information processing method and device, and the method comprises the steps: searching at least one first knowledge vector and at least one second knowledge vector from a vector database according to prompt information inputted by a user, and if the vector similarity between the first knowledge vector and the prompt information exceeds a first threshold value, determining that the prompt information is not a first knowledge vector; the vector similarity between the second knowledge vector and the prompt information is lower than a second threshold value, and the first threshold value is greater than the second threshold value; based on the prompt information, the at least one first knowledge vector and the at least one second knowledge vector, at least one piece of candidate reply information of the prompt information is determined by utilizing at least one large model; and determining target reply information suitable for the prompt information based on the at least one candidate reply information.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to an information processing method and device. Background Art

[0002] The application scenarios of question answering or information response based on large models such as large language models are becoming increasingly widespread.

[0003] In order to accurately determine the reply information corresponding to the user input information, the large model needs to be retrained or the model size of the large model has high requirements, which makes the complexity of information reasoning using the large model high and cannot be applied to some electronic devices with limited hardware performance. Therefore, how to reduce the complexity of determining the reply information by the large model is a technical problem that technicians in this field need to solve. Summary of the invention

[0004] On the one hand, the present application provides an information processing method, comprising:

[0005] According to the prompt information input by the user, at least one first knowledge vector and at least one second knowledge vector are retrieved from the vector database, wherein the vector similarity between the first knowledge vector and the prompt information exceeds a first threshold, the vector similarity between the second knowledge vector and the prompt information is lower than a second threshold, and the first threshold is greater than the second threshold;

[0006] Determine at least one candidate response information for the prompt information using at least one large model based on the prompt information, at least one first knowledge vector and at least one second knowledge vector;

[0007] Based on the at least one candidate reply information, target reply information suitable for the prompt information is determined.

[0008] In a possible implementation, determining at least one candidate reply information for the prompt information by using at least one large model based on the prompt information, at least one first knowledge vector, and at least one second knowledge vector includes:

[0009] Determine first candidate response information for the prompt information using a large model based on the prompt information and at least one first knowledge vector;

[0010] Based on the prompt information and at least one second knowledge vector, the large model is used to determine second candidate response information for the prompt information.

[0011] In a possible implementation, the determining at least one candidate reply information for the prompt information by using at least one large model based on the prompt information, at least one first knowledge vector and at least one second knowledge vector further includes:

[0012] Based on the prompt information, at least one first knowledge vector and at least one second knowledge vector, the large model is used to determine third candidate reply information for the prompt information.

[0013] In yet another possible implementation, determining, based on the at least one candidate reply information, target reply information suitable for the prompt information includes:

[0014] Constructing a query statement, the query statement including a target condition that the target reply information suitable for the prompt information needs to satisfy;

[0015] Based on the query statement, the large model is used to determine target reply information that meets the target condition in the at least one candidate reply information.

[0016] In yet another possible implementation, the information processing method further includes:

[0017] If the target reply information cannot be determined based on the at least one candidate reply information, retrieving a third knowledge vector matching the candidate reply information from the vector database;

[0018] Based on the third knowledge vector, at least one first knowledge vector and at least one second knowledge vector that match the candidate reply information, at least one large model is used to determine the target reply information that is suitable for the prompt information.

[0019] In yet another possible implementation, the information processing method further includes:

[0020] If the target reply information is determined, outputting the target reply information and information about the reasoning process for determining the target reply information to a user interaction interface;

[0021] If the target reply information cannot be determined, an unanswerable reminder and reasoning process information that the target reply information cannot be determined are output to the user interaction interface.

[0022] In yet another possible implementation, retrieving at least one first knowledge vector and at least one second knowledge vector from a vector database according to prompt information input by a user includes:

[0023] According to the prompt information input by the user, at least one first knowledge vector matching the prompt information is retrieved from a vector database;

[0024] Constructing a negative prompt of the prompt information, wherein the negative prompt is used to request to obtain reply information corresponding to prompt content opposite to the prompt information;

[0025] At least one second knowledge vector matching the negative prompt is retrieved from the vector database.

[0026] In another possible implementation, the determining, based on the prompt information, at least one first knowledge vector and at least one second knowledge vector, at least one large model of at least one candidate response information for the prompt information comprises: if the total number of the first knowledge vectors and the second knowledge vectors retrieved from the vector database is not less than a set number threshold, determining, based on the prompt information, at least one first knowledge vector and at least one second knowledge vector, at least one large model of at least one candidate response information for the prompt information;

[0027] The information processing method further includes:

[0028] If the total number is lower than the set number threshold, the at least one large model is used to determine at least one sub-prompt information required to process the prompt information step by step, and based on the at least one sub-prompt information, the at least one large model is used to determine the target reply information suitable for the prompt information.

[0029] In yet another possible implementation, the information processing method further includes:

[0030] Based on the current processing link of the prompt information, the processing result of the processing link is output to the user interaction interface, and the processing link includes: retrieving at least one first knowledge vector and at least one second knowledge vector from the vector database; and determining at least one candidate reply information.

[0031] On the other hand, the present application also provides an information processing device, comprising:

[0032] a knowledge retrieval unit, configured to retrieve at least one first knowledge vector and at least one second knowledge vector from a vector database according to prompt information input by a user, wherein a vector similarity between the first knowledge vector and the prompt information exceeds a first threshold, a vector similarity between the second knowledge vector and the prompt information is lower than a second threshold, and the first threshold is greater than the second threshold;

[0033] a candidate determination unit, configured to determine at least one candidate reply information of the prompt information by using at least one large model based on the prompt information, at least one first knowledge vector and at least one second knowledge vector;

[0034] A target determination unit is used to determine target reply information suitable for the prompt information based on the at least one candidate reply information. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. Throughout the accompanying drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and the originals and elements are not necessarily drawn to scale.

[0036] Figure 1 A flowchart of the information processing method provided in this application;

[0037] Figure 2 A schematic diagram of another flow chart of the information processing method provided by the present application;

[0038] Figure 3 A schematic diagram of another flow chart of the information processing method provided by the present application;

[0039] Figure 4 A schematic diagram of another flow chart of the information processing method provided by the present application;

[0040] Figure 5 A schematic diagram of the structure of an information processing device provided by the present application;

[0041] Figure 6 A schematic diagram of the composition architecture of an electronic device provided in this application. DETAILED DESCRIPTION

[0042] The embodiments of the present application are described below in conjunction with the drawings in the embodiments of the present application. The terms used in the implementation mode of the present application are only used to explain the specific embodiments of the present application, and are not intended to limit the present application. It is known to those skilled in the art that with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0043] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and need not be used to describe a specific order or sequential order. It should be understood that the terms used in this way can be interchangeable under appropriate circumstances, which is only to describe the distinction mode adopted by the objects of the same attributes when describing in the embodiments of the present application. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, so that the process, method, system, product or equipment comprising a series of units need not be limited to those units, but may include other units that are not clearly listed or inherent to these processes, methods, products or equipment.

[0044] like Figure 1, showing a flow chart of the information processing method provided by an embodiment of the present application, this embodiment can be applied to an electronic device, which may be a mobile phone, a tablet computer, a laptop computer or a server, etc., without limitation.

[0045] The method of this embodiment may include:

[0046] S101, according to prompt information input by a user, retrieve at least one first knowledge vector and at least one second knowledge vector from a vector database.

[0047] The prompt information is used to describe the response result that the user expects to obtain. The prompt information can guide the large model to produce the expected result. For example, the prompt information can be a question description, an instruction, or interactive information, etc., without limitation.

[0048] The vector database can store knowledge vectors of multiple knowledge. Depending on the actual application scenario, the types of knowledge vectors involved in the vector database will also vary, which can be set as needed and are not limited to this.

[0049] In the present application, the first knowledge vector is a knowledge vector in the vector database whose similarity with the vector of the prompt information exceeds a first threshold. The second knowledge vector is a knowledge vector in the vector database whose similarity with the vector of the prompt information is lower than a second threshold. The first threshold is greater than the second threshold.

[0050] For example, the information vector of the prompt information can be determined first, and based on the vector distance between each knowledge vector in the vector database and the information vector, the similarity between the knowledge vector and the information vector of the prompt information can be determined. On this basis, at least one first knowledge vector whose similarity with the information vector of the prompt information is greater than a first threshold can be determined, and at least one second knowledge vector whose similarity with the information vector of the prompt information is less than a second threshold can be determined. In this case, the first knowledge vector is a knowledge vector whose vector distance with the prompt information is less than a first distance value (also called positively correlated knowledge), and the second knowledge vector is a knowledge vector whose vector distance with the prompt information is greater than a second distance value, and the first distance value is less than the second distance value (also called negatively correlated knowledge).

[0051] For another example, for the first knowledge vector, at least one first knowledge vector matching the prompt information can be directly retrieved from the vector database based on the prompt information input by the user. The specific retrieval method can be referred to the above description. In order to retrieve the second knowledge vector, the present application can also first construct a negative prompt of the prompt information. The negative prompt is used to request a reply information corresponding to the prompt content opposite to the prompt information. Accordingly, at least one second knowledge vector matching the negative prompt can be retrieved from the vector database.

[0052] It is understandable that since the response information required for the negative prompt is the response information corresponding to the prompt content opposite to the prompt information, the knowledge vector matched with the negative prompt is the knowledge vector used to exclude the response information that does not correspond to the prompt information. For example, the prompt information is: Who did Jia Baoyu marry? Then, the constructed negative prompt can be: Jia Baoyu cannot marry whom.

[0053] Of course, there may be other ways to retrieve the first knowledge vector and the second knowledge vector, and this application does not limit this.

[0054] S102: Determine at least one candidate reply information of the prompt information by using at least one large model based on the prompt information, at least one first knowledge vector and at least one second knowledge vector.

[0055] For example, only one large model may be used to determine at least one candidate reply message for the prompt message.

[0056] As another example, multiple large models are used to determine at least one candidate response information for the prompt information, wherein each large model determines at least one candidate response information for the prompt information based on at least two types of information among the prompt information, at least one knowledge vector and at least one second knowledge vector.

[0057] In the present application, the reply information will be different based on the type of prompt information. For example, if the prompt information is a question, then the reply information will be the answer to the question; if the prompt information is the chat conversation content entered by the user, then the reply information will be the interactive content for the chat conversation content.

[0058] For the sake of distinction, in the present application, the reply information of the prompt information inferred in step S102 is referred to as candidate reply information, and the reply information of the prompt information finally determined later is referred to as target reply information.

[0059] It can be understood that the knowledge represented by the first knowledge vector is knowledge that is highly correlated with the prompt information, that is, the knowledge required for reference to determine the reply information corresponding to the prompt information (also known as positively correlated knowledge). The knowledge represented by the second knowledge vector is relatively less correlated with the prompt information, and can also be used to exclude knowledge that is unlikely to be the reply result corresponding to the prompt information. Based on this, the large model in this application combines the first knowledge vector and the second knowledge vector, two different types of knowledge, to more comprehensively and accurately determine the possible candidate reply information for the prompt information.

[0060] In the present application, the specific model type of the large model may not be restricted. For example, the large model may be a large language model or other deep learning model.

[0061] It should be noted that the big model mentioned in this application is a name for a model that uses deep learning technology or other artificial intelligence technology. With the continuous development of artificial intelligence and other technologies, the name of the big model can also be replaced by the model name corresponding to other models that can achieve information reasoning. The big model includes but is not limited to a big language model, a deep learning model or a machine learning model.

[0062] S103: Determine target reply information suitable for the prompt information based on the at least one candidate reply information.

[0063] In the present application, for the sake of easy distinction, the reply information of the prompt information finally determined is referred to as target reply information.

[0064] There are many specific implementation methods for determining the target reply information.

[0065] For example, when at least one large model determines the at least one candidate reply information, if the confidence corresponding to each candidate reply information is determined, then the candidate reply information with the highest confidence that exceeds the set confidence threshold can be determined as the target reply information.

[0066] For another example, the target reply information can be determined from the at least one candidate reply information using a large model in combination with at least one candidate reply information. Taking a possible implementation as an example, the present application can construct an inquiry statement, which includes the target condition that the target reply information suitable for the prompt information needs to meet. Accordingly, based on the inquiry statement, the target reply information that meets the target condition in the at least one candidate reply information is determined using a large model (such as any one of the at least one large model).

[0067] The target condition can be set as needed. For example, the target condition can be that the reply information suitable for being the prompt information has the highest degree of suitability and the degree of suitability is higher than a set degree threshold. Accordingly, the constructed query statement can prompt the large model to determine the candidate reply information suitable for being the prompt information from at least one candidate reply information, which has the highest degree of suitability and the degree of suitability is higher than the set degree threshold as the target reply information.

[0068] For another example, instead of using a large model, at least one candidate reply message may be integrated to determine a new reply message that is suitable for the prompt message as the target reply message.

[0069] Of course, the target reply information may also be determined through other implementation methods, which are not limited.

[0070] From the above content, it can be seen that the first knowledge vector retrieved from the vector database based on the prompt information input by the user in this application is the knowledge with a greater correlation with the prompt information, that is, the knowledge required to refer to determine the reply information corresponding to the prompt information, and the knowledge represented by the retrieved second knowledge vector has a relatively small correlation with the prompt information, that is, it can be used to exclude knowledge that cannot be the reply result corresponding to the prompt information. Based on this, the present application combines the knowledge required to determine the reply information corresponding to the prompt information and the knowledge used to exclude the reply information that cannot be used as the prompt information, and uses at least one large model to more comprehensively analyze the possible candidate reply information of the prompt information, and naturally can more accurately determine the candidate reply information that may be used as the prompt information, so that the target reply information suitable for the prompt information can be accurately determined based on each candidate reply information. Moreover, since the solution of the present application only requires a large model to infer the knowledge vectors of two dimensions, and does not require complex disassembly of the prompt information and step-by-step reasoning of the disassembled problems, and does not require complex operations such as retraining the large model, it can naturally reduce the complexity of determining the reply information, thereby reducing the hardware performance requirements for electronic devices.

[0071] In the present application, there may be many possibilities for the specific implementation of using at least one large model to determine at least one candidate reply information. Several possible implementations are described below as examples.

[0072] In a possible implementation, based on the prompt information and at least one first knowledge vector, a large model is used to determine first candidate response information for the prompt information. And based on the prompt information and at least one second knowledge vector, the large model is used to determine second candidate response information for the prompt information.

[0073] In this implementation, the same large model is used to infer the reply information of the prompt information, respectively for reference knowledge used to determine the reply information and reference knowledge used to exclude information that is not likely to be a reply. Thus, the first candidate reply information and the second candidate reply information are obtained through two inferences from two dimensions of different knowledge.

[0074] Furthermore, the present application can also use the large model to determine third candidate reply information for the prompt information based on the prompt information, at least one first knowledge vector and at least one second knowledge vector. Figure 2 Provide detailed explanation.

[0075] like Figure 2 , shows another flow chart of the information processing method provided in the embodiment of the present application. The method of this embodiment may include:

[0076] S201, retrieving at least one first knowledge vector and at least one second knowledge vector from a vector database according to prompt information input by a user.

[0077] Among them, the vector similarity between the first knowledge vector and the prompt information exceeds a first threshold, the vector similarity between the second knowledge vector and the prompt information is lower than a second threshold, and the first threshold is greater than the second threshold.

[0078] The specific implementation of this step can refer to the relevant introduction of the previous embodiment, which will not be repeated here.

[0079] S202: Based on the prompt information and at least one first knowledge vector, determine first candidate reply information for the prompt information using a large model.

[0080] Among them, since at least one first knowledge vector represents the knowledge related to replying to the prompt information in the vector database, the large model can refer to the at least one first knowledge vector to infer the reply information of the prompt information. For the sake of distinction, the reply information of the prompt information determined only by referring to the at least one first knowledge vector is called the first candidate reply information.

[0081] It is understandable that if the first knowledge vector does not exist in the vector database, then in step S202, the large model can determine the first candidate reply information only based on the prompt information.

[0082] S203: Based on the prompt information and at least one second knowledge vector, determine second candidate reply information for the prompt information using a large model.

[0083] As mentioned above, the second knowledge vector is used to exclude reference knowledge that cannot be used as a response to the prompt information. Therefore, at least one first knowledge vector of the large model can effectively exclude relevant information that cannot be a response to the prompt information, thereby accurately determining the possible response information to the prompt information.

[0084] For ease of distinction, the reply information determined by the large model with reference to at least one second knowledge vector is called second candidate reply information, and correspondingly, the reply information determined by the candidate with reference to at least one first knowledge vector and at least one second knowledge vector is called third candidate reply information.

[0085] S204: Based on the prompt information, at least one first knowledge vector and at least one second knowledge vector, determine third candidate reply information for the prompt information using the large model.

[0086] It can be understood that the large model combines the knowledge vector that can determine the prompt information and the knowledge vector used to exclude the replies that cannot serve as the prompt information, which can further improve the accuracy of determining the reply information of the prompt information.

[0087] It should be noted that the order of the above steps S202 to S204 is not limited to Figure 3 As shown, in actual application, the sequence of these three steps can be adjusted, or the three steps can be executed synchronously, and there is no restriction on this.

[0088] S205: Determine target reply information suitable for the prompt information based on the first candidate reply information, the second candidate reply information, and the third candidate reply information.

[0089] The specific implementation of determining the target reply information based on the three candidate reply information may refer to the related introduction of the previous embodiment.

[0090] For example, if the large model outputs the confidences of the first candidate reply information, the second candidate reply information, and the third candidate reply information respectively, the target reply information with the highest confidence and a confidence exceeding a set confidence threshold can be determined from these three candidate replies.

[0091] For another example, the present application may also construct a query statement, which includes a target condition that the target reply information suitable for the prompt information needs to meet. For example, in this embodiment, the query statement may be which candidate reply information among the currently determined candidate reply information has an accuracy that meets the target condition, and the target condition is the highest accuracy and the accuracy exceeds a set threshold. Accordingly, based on the query statement, the target reply information that meets the target condition is selected from the three candidate reply information using the large model.

[0092] Among them, the big model determines the target reply information from the three candidate reply information based on the inquiry statement. The big model can combine the reasoning process of the three candidate reply information and the context information related to the reasoning process to determine the accuracy of the three candidate reply information as reply information of the prompt information, and determine the target reply information whose accuracy meets the target conditions.

[0093] In the present application, not only will the first knowledge vector and the second knowledge vector be combined alone to determine the candidate reply information of the prompt information, but the two knowledge vectors will also be combined to determine the candidate reply information of the prompt information, so as to more comprehensively analyze the various candidate reply information that may serve as the prompt information. On this basis, by combining these three types of candidate reply information, the target reply information suitable for the prompt information can be more accurately determined.

[0094] It is understandable that if the accuracy of at least one candidate reply information determined by at least one large model is relatively low, then it may be impossible to determine the target reply information suitable for the prompt information by combining at least one candidate reply information.

[0095] On this basis, in order to determine the target reply information suitable for the prompt information, when the target reply information cannot be determined based on the at least one candidate reply information, for each candidate reply information, the present application will also retrieve a third knowledge vector matching the candidate reply information from the vector database. On this basis, based on the third knowledge vector matching each candidate reply information, at least one first knowledge vector and at least one second knowledge vector, at least one large model is used to determine the target reply information suitable for the prompt information.

[0096] Among them, the third knowledge vector matching the candidate reply information can be a knowledge vector whose vector similarity with the reply vector of the candidate reply information exceeds the first threshold. Of course, it can also be considered as a knowledge vector whose vector distance with the reply vector of the candidate reply information is lower than the first distance value.

[0097] It is understandable that for any candidate reply information, at least one third knowledge vector may be retrieved from the vector database, or the third knowledge vector may not be retrieved. Based on this, the present application can combine the retrieval results of the third knowledge vector corresponding to each candidate reply information, at least one first knowledge vector and at least one second knowledge vector, and use at least one large model to infer the target reply information.

[0098] Among them, the target reply information is determined by combining the third knowledge vector matched by each candidate reply information, at least one first knowledge vector and at least one second knowledge vector, using at least one large model. This can be achieved by combining these three types of knowledge vectors to re-infer the reply information of the prompt information so as to finally determine the target reply information. The specific implementation method may not be restricted.

[0099] For example, in a possible implementation, for each candidate reply information, the search result of the third knowledge vector corresponding to the candidate reply information and the at least one first knowledge vector determined most recently are determined as the at least one first knowledge vector currently retrieved, and the search result of the third knowledge vector and the at least one second knowledge vector determined most recently are determined as the at least one second knowledge vector currently retrieved. Then, the operation of combining at least one first knowledge vector and at least one second knowledge vector and determining at least one candidate reply information of the prompt information using at least one large model is returned.

[0100] Accordingly, based on at least one candidate reply information currently obtained, the target reply information of the prompt information is determined. If the target reply information is determined, or the number of words in the candidate reply information cumulatively output by at least one large model for the prompt information reaches the set threshold, the reasoning is determined to be ended, otherwise, the operation of retrieving the third knowledge vector from the vector database for each candidate reply information determined in this round of iteration is returned until the target reply information is determined, or the number of words in the reply information cumulatively output by the large model for the prompt information reaches the set number threshold.

[0101] When determining the end of the reasoning, if the target reply information can be determined, the target reply information can be output; if the target reply information cannot be determined, it can be determined that the reply information of the prompt information cannot be obtained.

[0102] To facilitate understanding of the implementation process of determining target reply information through continuous iteration, a possible implementation method is described below as an example.

[0103] like Figure 3 , shows another implementation flow diagram of the information processing method provided in an embodiment of the present application. The method of this embodiment may include:

[0104] S301, according to prompt information input by the user, retrieve at least one first knowledge vector and at least one second knowledge vector from a vector database.

[0105] Wherein, the vector similarity between the first knowledge vector and the prompt information exceeds a first threshold, the vector similarity between the second knowledge vector and the prompt information is lower than a second threshold, and the first threshold is greater than the second threshold;

[0106] S302: Based on the prompt information and at least one first knowledge vector, determine first candidate reply information for the prompt information using a large model.

[0107] S303: Based on the prompt information and at least one second knowledge vector, determine second candidate reply information for the prompt information using the large model.

[0108] S304: Based on the prompt information, at least one first knowledge vector and at least one second knowledge vector, determine third candidate reply information for the prompt information using the large model.

[0109] The above steps can be referred to the related introduction of the previous embodiment, which will not be repeated here.

[0110] S305: Construct a query statement.

[0111] The query statement includes a target condition that the target reply information that the prompt information is suitable for needs to meet. For example, the target reply information that meets the target condition is the reply information with the highest accuracy and exceeding a set threshold.

[0112] S306, based on the query statement, using the large model to determine the target reply information that meets the target condition among the first candidate reply information, the second candidate reply information and the third candidate reply information.

[0113] It should be noted that, for ease of understanding, this embodiment takes an implementation method of determining the target reply information from the first candidate reply information, the second candidate reply information and the third candidate reply information as an example, but determining the target reply information from these three candidate reply information by the other methods mentioned above is also applicable to this embodiment and will not be repeated here.

[0114] S307: If the large model determines the target reply information, the target reply information is output to the user interaction interface.

[0115] S308, if the large model determines that there is no target reply information that meets the target condition in the at least one candidate reply information, determine whether the number of words in each candidate reply information cumulatively determined by the large model before the current moment exceeds the set number threshold, if not, execute step S309; ​​if yes, execute step S311.

[0116] Among them, the number of words in each candidate reply information cumulatively determined by the large model refers to the total number of words (or participles) contained in all the first candidate reply information, second candidate reply information and third candidate reply information determined in each round during the iteration process of the large model.

[0117] It can be understood that if the total number of words contained in the candidate reply information determined by the large model exceeds the set threshold, it means that the large model has performed relatively sufficient reasoning. On this basis, if the target reply information cannot be determined, it can be understood that the large model cannot infer the reply information of the prompt information.

[0118] The set quantity threshold may be set based on experience or actual needs, and there is no restriction on this.

[0119] S309: for any one of the first candidate reply information, the second candidate reply information and the third candidate reply information, retrieve a third knowledge vector matching the candidate reply information from the vector database.

[0120] S310, take the retrieval results of the third knowledge vector corresponding to each candidate reply information and at least one first knowledge vector determined most recently as at least one first knowledge vector currently retrieved, take the retrieval results of the third knowledge vector corresponding to each candidate reply information and at least one second knowledge vector determined most recently as at least one second knowledge vector currently retrieved, and return to execute steps S302 to S306.

[0121] Among them, for each candidate reply information in the first candidate reply information, the second candidate reply information and the third candidate reply information, the search result of the third knowledge vector corresponding to the candidate reply information may be that the third knowledge vector is not retrieved, or each third knowledge vector is detected. The search result of the third knowledge vector detected from the vector database based on the candidate reply information can be used to assist the large model in determining whether the candidate reply information inferred last time is accurate and the degree of accuracy, so as to facilitate the large model to further reasonably infer the reply information of the prompt information.

[0122] It is understandable that if step S310 is performed for the first time, the most recently determined first knowledge vector is actually at least one first knowledge vector retrieved from the data vector library, and similarly, the most recently determined at least one second knowledge vector is at least one second knowledge vector retrieved from the data vector library. If step S310 is not performed for the first time, the most recently determined first knowledge vector actually includes the search results of at least one first knowledge vector retrieved from the data vector library and the third knowledge vector corresponding to each candidate reply information retrieved before the current moment.

[0123] It can be understood that the retrieval results of the third knowledge vector corresponding to each candidate reply information are also used as the information contained in at least one first knowledge vector and at least one second knowledge vector. After returning to execute steps S302 and S306, the large model will not only consider the first knowledge vector and the second knowledge vector retrieved from the vector database based on the prompt information, but also combine the matching results of each candidate reply information inferred in the previous iteration process with the knowledge vector in the vector database to further infer the reply information suitable for the prompt information, so as to more reasonably and accurately determine the reply information of the prompt information.

[0124] S311, outputting an unanswerable reminder to the user interaction interface.

[0125] The unanswerable reminder is used to prompt that the prompt information cannot determine the reply information. For example, the unanswerable reminder may be: the answer cannot be determined or the prompt information cannot be answered.

[0126] In a possible implementation, in any of the above embodiments of the present application, if the target reply information is determined, the target reply information can also be output to the user interaction interface so that the user can know the target reply information obtained for the prompt information. Furthermore, in order to enable the user to intuitively understand the specific implementation process of obtaining the target reply information with the help of the large model, the present application can also output the reasoning process information for determining the target reply information while outputting the target reply information. Among them, the reasoning process information can include at least one of the specific steps for inferring the target reply information with the help of the large model and the referenced information.

[0127] For example, the reasoning process information may include: first combining a first knowledge vector and a second knowledge vector that match the prompt information retrieved from a vector database, using a large model to determine at least one candidate reply information, and then determining the target reply information based on the at least one candidate reply information.

[0128] For ease of understanding, Figure 2 Taking the shown embodiment as an example, the reasoning process information may be: after respectively inferring the first candidate reply information and the second candidate reply information by combining the positively correlated knowledge and the negatively correlated knowledge retrieved from the vector database, and inferring the third candidate reply information by combining the positively correlated knowledge and the negatively correlated knowledge, the target reply information is determined by combining these three candidate reply information.

[0129] And for Figure 3 For the embodiment shown, the reasoning process information includes the above Figure 2 In addition to the relevant reasoning process information in the embodiment, it can also include: when the target reply information cannot be determined, based on the retrieval results of the third knowledge vector matching each candidate reply information in the vector database, the candidate reply information and the target reply information are repeatedly iterated and inferred with the positively correlated knowledge and negatively correlated knowledge of the prompt information.

[0130] Similarly, if the target reply information cannot be determined, an unanswerable reminder can be output to the user interface, and the reasoning process information of the target reply information cannot be determined can also be output. The reasoning process information of the target reply information cannot be determined can include the implementation process of reasoning the target reply information with the help of a large model, and can also include information such as the specific reasoning steps that lead to the inability to reason out the target reply information.

[0131] It is understandable that considering that it may take a certain amount of time to reply to the prompt information with the help of the large model reasoning, during the waiting process, the user cannot determine whether the application used to determine the reply information has an abnormal operation or the large model is still inferring, thus affecting the user experience. Based on this, the present application can also output the processing results of the processing link to the user interaction interface based on the current processing link of the prompt information. On this basis, the user can know the processing link in which the prompt information is currently being processed, as well as the processing results obtained in the current processing link based on the processing results of the processing link, and can also determine that the application is still running normally.

[0132] The processing step may include: retrieving at least one first knowledge vector and at least one second knowledge vector from a vector database; and determining at least one candidate reply information. The processing step may also include a processing step of determining target reply information.

[0133] It can be understood that determining at least one candidate reply information involves determining the first candidate reply information, the second candidate reply information and the third candidate reply information. Therefore, the processing link of determining at least one candidate reply information can be further subdivided into three processing links: determining the first candidate reply information of the prompt information, determining the second candidate reply information of the prompt information, and determining the third candidate reply information of the prompt information.

[0134] Of course, the processing link of processing prompt information also involves constructing query statements and iterative loops, etc. The processing link of processing prompt information may also include corresponding processing links, which will not be repeated here.

[0135] Among them, the processing result obtained by the processing link is the result obtained by the electronic device executing the processing link.

[0136] For example, assuming that the processing link currently processing the prompt information is to retrieve at least one first knowledge vector and a second knowledge vector from a vector database, then when the retrieval of the first knowledge vector and the second knowledge vector is completed, the retrieval results can be output. The retrieval results can represent the retrieval status of the first knowledge vector and the second knowledge vector, such as whether the first knowledge vector is retrieved, whether the second knowledge vector is retrieved, the number of the retrieved first knowledge vectors and the second knowledge vectors, etc.

[0137] It can be understood that in the above embodiments, if the first knowledge vector and the second knowledge vector cannot be retrieved from the vector database based on the prompt information, or the number of the retrieved first knowledge vectors and the second knowledge vectors is small, then based on the retrieved first knowledge vectors and the second knowledge vectors, it may not be possible to accurately determine the target reply information suitable for the prompt information.

[0138] On this basis, if the total number of first knowledge vectors and second knowledge vectors retrieved from the vector database is not less than the set quantity threshold, the present application can determine at least one candidate reply information for the prompt information based on the prompt information, at least one first knowledge vector and at least one second knowledge vector using at least one large model, and perform an operation of determining the target reply information based on the at least one candidate reply information.

[0139] If the total number of the retrieved first knowledge vectors and the second knowledge vectors is lower than the set number threshold, the present application may use at least one large model to determine at least one sub-prompt information that needs to be processed step by step to process the prompt information. Then, based on the at least one sub-prompt information, the at least one large model is used to determine the target reply information suitable for the prompt information. That is to say, when the total number of the retrieved first knowledge vectors and the second knowledge vectors is small, the present application may disassemble the prompt information into at least one sub-prompt information, and process each sub-prompt information in turn with the help of the large model, so as to achieve the purpose of disassembling the prompt information and processing it step by step.

[0140] For example, when the prompt information is a question, the question can be broken down into at least one sub-question, and then each sub-question is processed in turn, and finally the target response information is determined.

[0141] In order to facilitate understanding of the specific implementation of distributed processing of the prompt information, the information processing method of the present application is introduced below by taking one implementation as an example.

[0142] like Figure 4 , shows another implementation flow diagram of the information processing method provided in an embodiment of the present application. The method of this embodiment may include:

[0143] S401, according to prompt information input by the user, retrieve at least one first knowledge vector and at least one second knowledge vector from a vector database.

[0144] S402: If the total number of first knowledge vectors and second knowledge vectors retrieved from the vector database is not less than the set quantity threshold, at least one candidate reply information of the prompt information is determined using at least one large model based on the prompt information, at least one first knowledge vector and at least one second knowledge vector, and step S403 is executed.

[0145] S403: Determine target reply information suitable for the prompt information based on at least one candidate reply information.

[0146] Wherein, when the total number of the retrieved first knowledge vectors and the second knowledge vectors is not less than the set number threshold, the specific implementation of determining the target reply information can refer to the previous Figures 1 to 3 The relevant introduction of the embodiments will not be repeated here.

[0147] S404: If the total number of the retrieved first knowledge vectors and the second knowledge vectors is lower than a set number threshold, use the large model to determine at least one sub-prompt information required to process the prompt information in steps.

[0148] The purpose of using the big model to split the prompt information into at least one sub-prompt information is to process each of the split sub-prompt information step by step in a Chain of Thought manner. The specific process of splitting at least one sub-prompt information may not be restricted.

[0149] For example, assuming the prompt information is "Xiao Ming bought three boxes of pencils, and he also bought two boxes of pens. Each box of pencils has 5 pencils, and each box of pens has 3 pens. How many pens does he have now?", it can be divided into two sub-prompt information: Sub-prompt information 1 is: Xiao Ming bought three boxes of pencils, each box has 5 pencils, how many pencils did Xiao Ming buy; and sub-prompt information 2, Xiao Ming bought two boxes of pens, each box has 3 pens, how many pens did Xiao Ming buy.

[0150] Of course, in actual applications, the prompt information can be further split into sub-prompt information 3: how many pens did Xiao Ming buy in total, and the reply result of the sub-prompt information 3 is the reply result required by the prompt information.

[0151] S405: For each sub-prompt information, retrieve at least one fourth knowledge vector and at least one fifth knowledge vector related to the sub-prompt information from the vector database.

[0152] The similarity between the fourth knowledge vector and the vector of the sub-prompt information exceeds a first threshold, and the similarity between the fifth knowledge vector and the vector of the sub-prompt information is lower than a second threshold.

[0153] It can be understood that the specific implementation process of retrieving the fourth knowledge vector based on the sub-prompt information can be similar to the previous implementation process of retrieving the first knowledge vector based on the prompt information, and the implementation process of retrieving the fourth knowledge vector based on the sub-prompt information can be similar to the previous process of retrieving the second knowledge vector based on the prompt information. For details, please refer to the previous introduction and no further details will be given.

[0154] It should be noted that, in this embodiment, each sub-prompt information may be processed in sequence according to the sequence of the decomposed at least one sub-prompt information. Of course, if each sub-prompt information is independent of each other, each sub-prompt information may also be processed in parallel, and there is no limitation to this.

[0155] S406: Based on the sub-prompt information, at least one fourth knowledge vector and at least one fifth knowledge vector, use at least one large model to determine at least one candidate sub-reply information of the sub-prompt information.

[0156] For the sake of distinction, the candidate reply information of the sub-prompt information determined by using the large model is referred to as candidate sub-reply information.

[0157] It can be understood that the specific implementation process of using at least one large model to determine at least one candidate sub-reply information can be similar to the specific implementation process of using at least one large model to determine at least one candidate reply information. For details, please refer to the relevant introduction above.

[0158] For example, step S406 may include: based on the sub-prompt information and at least one fourth knowledge vector, using the big model to determine the first candidate sub-reply information of the sub-prompt information; based on the sub-prompt information and at least one fifth knowledge vector, using the big model to determine the second candidate sub-reply information of the sub-prompt information; and, based on the sub-prompt information, at least one fourth knowledge vector and at least one fifth knowledge vector, using the big model to determine the third candidate sub-reply information of the sub-prompt information.

[0159] S407: Determine target sub-reply information suitable for the sub-prompt information based on at least one candidate sub-reply information of the sub-prompt information.

[0160] Among them, the implementation process of determining whether the sub-prompt information is suitable for the target sub-reply information based on at least one candidate sub-reply information is similar to the implementation process of determining the target reply information based on at least one candidate reply information. For details, please refer to the relevant introduction above.

[0161] For example, a sub-query statement may be constructed, the sub-query statement including a sub-target condition satisfied by the target sub-reply information, for example, the sub-target condition may be that the degree of suitability for the sub-prompt information is the highest and the degree of suitability is greater than a set degree threshold. Accordingly, based on at least one candidate sub-reply information of the sub-prompt information and the sub-query statement, the target sub-reply information satisfying the target condition is determined from the at least one candidate sub-reply information using the large model.

[0162] S408, based on the target sub-reply information corresponding to each sub-prompt information, use the large model to determine the target reply information that is suitable for the prompt information.

[0163] For example, by synthesizing the target sub-response information corresponding to each sub-prompt information, the target response information corresponding to the prompt information can be derived using the large model. Take the prompt information including "How many pens did Xiao Ming buy in total" as an example. The prompt information is split into sub-prompt information 1 and sub-prompt information 2. Then the target sub-response information based on sub-prompt information 1 can determine that Xiao Ming bought 15 pencils; based on the target sub-response information of sub-prompt information 2, it can be determined that Xiao Ming bought 6 pens. On this basis, the total number of pens Xiao Ming bought is the total number of pencils and pens Xiao Ming bought, that is, 15+6=21.

[0164] In particular, if the last sub-prompt information among the multiple sub-prompt information split from the prompt information is the same as the reply information expected by the prompt information, then the target sub-reply information corresponding to the last sub-prompt information may be determined, and the target sub-reply information may be determined as the target reply information suitable for the prompt information. For example, taking the prompt information including "How many pens did Xiao Ming buy in total" split into sub-prompt information 1, sub-prompt information 2, and sub-prompt information 3 as an example, the target sub-reply information of sub-prompt information 3 is that Xiao Ming bought a total of 21 pens, then the target sub-reply information corresponding to the sub-prompt information may be determined as the target reply information suitable for the prompt information.

[0165] Corresponding to an information processing method of the present application, the present application also provides an information processing device.

[0166] like Figure 5 , shows a schematic diagram of a composition structure of an information processing device provided by the present application. The device of this embodiment may include:

[0167] A knowledge retrieval unit 501 is used to retrieve at least one first knowledge vector and at least one second knowledge vector from a vector database according to prompt information input by a user, wherein the vector similarity between the first knowledge vector and the prompt information exceeds a first threshold, the vector similarity between the second knowledge vector and the prompt information is lower than a second threshold, and the first threshold is greater than the second threshold;

[0168] A candidate determination unit 502, configured to determine at least one candidate reply information of the prompt information by using at least one large model based on the prompt information, at least one first knowledge vector and at least one second knowledge vector;

[0169] The target determination unit 503 is used to determine target reply information suitable for the prompt information based on the at least one candidate reply information.

[0170] In a possible implementation, the candidate determination unit includes:

[0171] A first candidate determination subunit, configured to determine first candidate reply information for the prompt information by using a large model based on the prompt information and at least one first knowledge vector;

[0172] The second candidate determination subunit is used to determine the second candidate reply information of the prompt information by using the large model based on the prompt information and at least one second knowledge vector.

[0173] Furthermore, the candidate determination unit further includes:

[0174] The third candidate determination subunit is used to determine the third candidate reply information of the prompt information by using the large model based on the prompt information, at least one first knowledge vector and at least one second knowledge vector.

[0175] In yet another possible implementation, the target determination unit includes:

[0176] A sentence construction subunit, used to construct an inquiry sentence, wherein the inquiry sentence includes a target condition that the target reply information suitable for the prompt information needs to satisfy;

[0177] The target determination subunit is used to determine the target response information that meets the target condition in the at least one candidate response information based on the query statement and using the large model.

[0178] In yet another possible implementation, the device further includes:

[0179] a vector retrieval unit, configured to retrieve a third knowledge vector matching the candidate reply information from the vector database if the target reply information cannot be determined based on the at least one candidate reply information;

[0180] A reply determination unit is used to determine target reply information suitable for the prompt information using at least one large model based on a third knowledge vector matched by the candidate reply information, at least one first knowledge vector and at least one second knowledge vector.

[0181] In yet another possible implementation, the device further includes:

[0182] A first output unit, configured to output the target reply information and information on a reasoning process for determining the target reply information to a user interaction interface if the target reply information is determined;

[0183] The second output unit is used to output an unanswerable reminder and reasoning process information of being unable to determine the target reply information to the user interaction interface if the target reply information cannot be determined.

[0184] In yet another possible implementation, the knowledge retrieval unit includes:

[0185] A first retrieval subunit, configured to retrieve at least one first knowledge vector matching the prompt information from a vector database according to the prompt information input by the user;

[0186] A prompt construction subunit, used for constructing a negative prompt of the prompt information, wherein the negative prompt is used for requesting to obtain reply information corresponding to the prompt content opposite to the prompt information;

[0187] The second retrieval subunit is used to retrieve at least one second knowledge vector matching the negative prompt from the vector database.

[0188] In another possible implementation, the candidate determination unit is specifically configured to determine at least one candidate reply information of the prompt information by using at least one large model based on the prompt information, at least one first knowledge vector and at least one second knowledge vector if the total number of the first knowledge vectors and the second knowledge vectors retrieved from the vector database is not less than a set number threshold;

[0189] The device also includes: a distribution processing unit, which is used to use the at least one large model to determine at least one sub-prompt information required to process the prompt information in steps if the total number is lower than the set number threshold, and based on the at least one sub-prompt information, use the at least one large model to determine the target reply information suitable for the prompt information.

[0190] In yet another possible implementation, the device further includes:

[0191] A link information output unit is used to output the processing result of the processing link to the user interaction interface based on the current processing link of the prompt information, and the processing link includes: retrieving at least one first knowledge vector and at least one second knowledge vector from the vector database; and determining at least one candidate reply information.

[0192] The present application also provides an electronic device. Figure 6 , which shows a schematic diagram of a composition structure of the electronic device, and the electronic device at least includes a processor 601 and a memory 602;

[0193] The processor 601 is configured to execute the information processing method described in any one of the above embodiments;

[0194] The memory 602 is used to store the program required by the processor to perform operations

[0195] It can be understood that the electronic device may further include a display unit 603 and an input unit 604 .

[0196] Of course, the electronic device may also have Figure 6 There may be more or fewer components, without limitation.

[0197] A computer program product is also provided in an embodiment of the present application, including computer-readable instructions. When the computer-readable instructions are executed on an electronic device, the electronic device implements any information processing method provided in the embodiment of the present application.

[0198] A computer-readable storage medium is also provided in an embodiment of the present application. The storage medium carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any information processing method provided in the embodiment of the present application.

[0199] It should also be noted that the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed over multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the drawings of the device embodiments provided by the present application, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines.

[0200] Through the description of the above implementation mode, the technicians in the field can clearly understand that the present application can be implemented by means of software plus necessary general hardware, and of course, it can also be implemented by special hardware including special integrated circuits, special CPUs, special memories, special components, etc. In general, all functions completed by computer programs can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be various, such as analog circuits, digital circuits or special circuits. However, for the present application, software program implementation is a better implementation mode in more cases. Based on such an understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a readable storage medium, such as a computer floppy disk, a U disk, a mobile hard disk, a ROM, a RAM, a disk or an optical disk, etc., including a number of instructions to enable a computer device (which can be a personal computer, a training device, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0201] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product.

[0202] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from a website site, a computer, a training device, or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, training device, or data center. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device, a data center, etc. that includes one or more available media integrations. The available medium may be a magnetic medium, (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)), etc.

Claims

1. An information processing method, comprising: According to the prompt information input by the user, at least one first knowledge vector and at least one second knowledge vector are retrieved from the vector database, wherein the vector similarity between the first knowledge vector and the prompt information exceeds a first threshold, the vector similarity between the second knowledge vector and the prompt information is lower than a second threshold, and the first threshold is greater than the second threshold; Determine at least one candidate response information for the prompt information using at least one large model based on the prompt information, at least one first knowledge vector and at least one second knowledge vector; Based on the at least one candidate reply information, target reply information suitable for the prompt information is determined.

2. The information processing method according to claim 1, wherein the determining, based on the prompt information, at least one first knowledge vector and at least one second knowledge vector, at least one large model of at least one candidate reply information for the prompt information comprises: Determine first candidate response information for the prompt information using a large model based on the prompt information and at least one first knowledge vector; Based on the prompt information and at least one second knowledge vector, the large model is used to determine second candidate response information for the prompt information.

3. The information processing method according to claim 2, wherein the determining, based on the prompt information, at least one first knowledge vector and at least one second knowledge vector, at least one candidate reply information of the prompt information by using at least one large model further comprises: Based on the prompt information, at least one first knowledge vector and at least one second knowledge vector, the large model is used to determine third candidate reply information for the prompt information.

4. The information processing method according to claim 1, wherein determining the target reply information suitable for the prompt information based on the at least one candidate reply information comprises: Constructing a query statement, the query statement including a target condition that the target reply information suitable for the prompt information needs to satisfy; Based on the query statement, the large model is used to determine target reply information that meets the target condition in the at least one candidate reply information.

5. The information processing method according to claim 1, further comprising: If the target reply information cannot be determined based on the at least one candidate reply information, retrieving a third knowledge vector matching the candidate reply information from the vector database; Based on the third knowledge vector, at least one first knowledge vector and at least one second knowledge vector that match the candidate reply information, at least one large model is used to determine the target reply information that is suitable for the prompt information.

6. The information processing method according to claim 1, further comprising: If the target reply information is determined, outputting the target reply information and information about the reasoning process for determining the target reply information to a user interaction interface; If the target reply information cannot be determined, an unanswerable reminder and reasoning process information that the target reply information cannot be determined are output to the user interaction interface.

7. The information processing method according to claim 1, wherein the retrieving at least one first knowledge vector and at least one second knowledge vector from the vector database according to the prompt information input by the user comprises: According to the prompt information input by the user, at least one first knowledge vector matching the prompt information is retrieved from a vector database; Constructing a negative prompt of the prompt information, wherein the negative prompt is used to request to obtain reply information corresponding to prompt content opposite to the prompt information; At least one second knowledge vector matching the negative prompt is retrieved from the vector database.

8. The information processing method according to claim 1, wherein the determining, based on the prompt information, at least one first knowledge vector and at least one second knowledge vector, at least one large model of at least one candidate reply information of the prompt information comprises: If the total number of the first knowledge vectors and the second knowledge vectors retrieved from the vector database is not less than a set number threshold, based on the prompt information, at least one first knowledge vector and at least one second knowledge vector, at least one large model is used to determine at least one candidate response information for the prompt information; The information processing method based on the large model also includes: If the total number is lower than the set number threshold, the at least one large model is used to determine at least one sub-prompt information required to process the prompt information step by step, and based on the at least one sub-prompt information, the at least one large model is used to determine the target reply information suitable for the prompt information.

9. The information processing method according to claim 1, further comprising: Based on the current processing link of the prompt information, the processing result of the processing link is output to the user interaction interface, and the processing link includes: retrieving at least one first knowledge vector and at least one second knowledge vector from the vector database; and determining at least one candidate reply information.

10. An information processing device, comprising: a knowledge retrieval unit, configured to retrieve at least one first knowledge vector and at least one second knowledge vector from a vector database according to prompt information input by a user, wherein a vector similarity between the first knowledge vector and the prompt information exceeds a first threshold, a vector similarity between the second knowledge vector and the prompt information is lower than a second threshold, and the first threshold is greater than the second threshold; a candidate determination unit, configured to determine at least one candidate reply information of the prompt information by using at least one large model based on the prompt information, at least one first knowledge vector and at least one second knowledge vector; A target determination unit is used to determine target reply information suitable for the prompt information based on the at least one candidate reply information.

Citation Information

Patent Citations

  • Question and answer processing method and device, equipment and medium

    CN111382255A

  • Question and answer data query method, device and equipment and storage medium

    CN114003688A

  • Information processing method, device and equipment

    CN118093804A

  • Question answering device

    JP2021124824A

  • Method for generating learning model for predicting semiconductor device structure and apparatus for predicting semiconductor device structure

    KR1020240173448A