Information processing method and apparatus, computing device cluster, and storage medium

Abstract

The present application provides an information processing method, an information processing apparatus, a computing device cluster, and a computer-readable storage medium. The method comprises: acquiring a plurality of texts to be stored, wherein said plurality of texts include a first text and a second text; using a first embedding model to process the first text so as to obtain a first result; using a second embedding model to process the second text so as to obtain a second result, wherein the number of dimensions of the first embedding model and the number of dimensions of the second embedding model are different; and storing the first result and the second result into a knowledge base. The present application can save storage resources and avoid the excessively large number of dimensions of vectors generated by embedding models in the process of constructing a knowledge base.

Description

Information processing method and device, computing device cluster, and storage medium

[0001] This application claims priority to the Chinese patent application No. 202411345464.6, filed on September 25, 2024, and entitled “Information processing method and device, computing device cluster, and storage medium”, the entire content of which is incorporated herein by reference. TECHNICAL FIELD

[0002] The present application relates to the field of artificial intelligence, and in particular to an information processing method, an information processing device, a computing device cluster, and a computer-readable storage medium. BACKGROUND

[0003] In the art, there is a technology of converting a segment of text containing certain knowledge information into a vector by embedding a model, so as to construct a knowledge base by the vectors. However, in the prior art, the same embedding model is used to embed all segments of all texts in the same way, and vectors of the same dimension are obtained. Such an embedding method is prone to cause the problem of embedding some segments by using an embedding model with too high dimension, thereby obtaining vectors with too high dimension, resulting in waste of storage resources for storing the knowledge base. Therefore, there is an urgent need in the art for a knowledge base construction technology capable of saving storage resources and avoiding too high dimension of vectors generated by an embedding model. SUMMARY

[0004] To this end, the present application is committed to providing an information processing method, an information processing device, a computing device cluster, and a computer-readable storage medium, which can save storage resources and avoid too high dimension of vectors generated by an embedding model in the process of constructing a knowledge base.

[0005] In one aspect, the present application provides an information processing method, comprising: obtaining a plurality of to-be-stored texts, the plurality of to-be-stored texts comprising a first text and a second text; processing the first text by using a first embedding model to obtain a first result; processing the second text by using a second embedding model to obtain a second result, the dimension number of the first embedding model and the second embedding model being different; and storing the first result and the second result to a knowledge base.

[0006] According to the present aspect, different texts are processed by embedding models with different dimension numbers, which is beneficial to selecting an embedding model with a suitable dimension number for embedding operation according to the characteristics of the text itself, so as to obtain a vector with a suitable dimension number for constructing a knowledge base. This is beneficial to avoiding the use of an embedding model with too large dimension number for embedding operation on a relatively short text or a text that does not need to be represented by a large number of dimensions, thereby causing the dimension number of the embedded vector to be too large and resulting in waste of computing and storage resources.

[0007] In a possible implementation of the present application, the first text is processed by using the first embedding model to obtain a first result, including: dividing the first text into multiple segments according to the characteristics of the first text; embedding the multiple segments by using the first embedding model to obtain multiple first vectors as the first result. Wherein, the second text is processed by using the second embedding model to obtain a second result, including: dividing the second text into multiple segments according to the characteristics of the second text; embedding the multiple segments by using the second embedding model to obtain multiple second vectors as the second result.

[0008] According to the present implementation, different ways are used to divide and embed the segments according to different characteristics of the text, so that the vectors obtained by embedding have different dimensions, which can make the number of dimensions of the vectors obtained by embedding match the characteristics of the text and the length of the segments, avoiding the problem of waste of storage resources caused by using the same embedding model to embed the segments of all texts.

[0009] In a possible implementation of the present application, the characteristics of the first text or the characteristics of the second text include one or more of the following: white paper type, product manual type, and question and answer type.

[0010] According to the present implementation, white papers, product manuals, and questions and answers are common texts that usually contain a lot of knowledge information. Dividing the knowledge texts into these three types is conducive to further processing (including segment division and embedding) according to different characteristics of the text, so that the constructed knowledge base has more knowledge information and the vectors therein have appropriate dimensions.

[0011] In a possible implementation of the present application, the first text is divided into multiple segments according to the characteristics of the first text, including: if the first text belongs to the white paper type, the first text is divided into multiple segments by using a first way, and the first way includes dividing according to a first fixed length.

[0012] According to the present implementation, since the paragraphs of the text of the white paper type are long and often have long narrative and explanatory sentences, segmenting the text of the white paper type according to a fixed length can make appropriate division according to the characteristics of the text of the white paper type, thereby obtaining vectors with appropriate dimensions.

[0013] In a possible implementation of the present application, the first text is divided into multiple segments according to the characteristics of the second text, including: if the second text belongs to the product manual type, the second text is divided into multiple segments by using a second way, and the second way includes dividing according to a second fixed length smaller than the first fixed length.

[0014] According to the present implementation, since the product specification type text paragraph is shorter than the white paper type text paragraph, the product specification type text is divided according to the second fixed length smaller than the first fixed length, which can be suitable for the characteristics of the type of text, so as to obtain a vector with a suitable dimension.

[0015] In a possible implementation of the present application, the first text is divided into multiple segments according to the characteristics of the first text, including: if the first text belongs to the question and answer type, the first text is divided into multiple segments by using a first mode, wherein the first mode includes dividing according to the paragraph.

[0016] According to the present implementation, since the paragraph of the question and answer type text is short and usually consists of short questions and answers, the question and answer type text can be divided according to the paragraph, so as to divide the segments with a suitable length according to the characteristics of the question and answer type text, thereby obtaining a vector with a suitable dimension.

[0017] In a possible implementation of the present application, the method further includes: embedding the inquiry sentence of the user by using the first embedding model to obtain a first retrieval vector; embedding the inquiry sentence by using the second embedding model to obtain a second retrieval vector; and retrieving in the knowledge base through the first retrieval vector and the second retrieval vector to obtain multiple retrieval results.

[0018] According to the present implementation, when retrieving the knowledge base constructed based on the above-mentioned mode, the inquiry sentence is embedded by using different embedding models used when constructing the knowledge base to obtain retrieval vectors with different dimensions, so as to retrieve in the vector database with different dimensions through the retrieval vectors with different dimensions, which can fully utilize the knowledge contained in different types of vectors, perform comprehensive retrieval, fully extract the knowledge information contained in the vectors with different dimensions, and obtain more optimized retrieval results, while saving the storage resources required for storing the knowledge base.

[0019] In a possible implementation of the present application, the retrieving in the knowledge base through the first retrieval vector and the second retrieval vector to obtain multiple retrieval results includes: retrieving in the knowledge base through the first retrieval vector to obtain X first retrieval results; retrieving in the knowledge base through the second retrieval vector to obtain X second retrieval results; rearranging the X first retrieval results and the X second retrieval results according to the similarity; and selecting the first Y retrieval results from high to low as the multiple retrieval results.

[0020] According to the present implementation, the search results obtained from the knowledge base are rearranged according to the similarity, and the several search results with the highest similarity are selected as the final results, so that the search results as accurate as possible and meeting the user's requirements can be obtained. The search results meeting the requirements most can be filtered out by rearranging according to the unified similarity standard, and efficient matching is realized.

[0021] In a possible implementation of the present application, the method further comprises: constructing a prompt word according to the query statement and the plurality of search results; inputting the query statement and the prompt word into the large language model to obtain an answer to the query statement.

[0022] According to the present implementation, by searching and constructing the prompt word in the above manner, the quality of the answer can be ensured, while the vector dimension of the short text segment is reduced, so that the database storage space required by the question and answer method is smaller. In addition, by constructing the prompt word for the plurality of search results ranked high, and inputting the prompt word and the query statement into the large language model to obtain the answer, important and key information can be obtained through sufficient background knowledge, so that the quality and accuracy of the answer are improved.

[0023] In a possible implementation of the present application, the constructing of the prompt word according to the query statement and the plurality of search results comprises: decomposing the query statement into a plurality of questions, each question corresponding to one search result of the plurality of search results; combining each question with the corresponding search result; and taking the plurality of combinations of the plurality of questions and the plurality of search results as the prompt word.

[0024] According to the present implementation, by decomposing the query statement into a plurality of questions, and combining the questions and the answers into a plurality of combinations, and taking the plurality of combinations as the prompt word, the user can be prompted from different aspects and directions about the answer content that the user wants to obtain, and the most comprehensive and accurate answer result can be obtained.

[0025] In another aspect, the present application also provides an information processing apparatus, comprising: an acquisition module configured to acquire a plurality of to-be-stored texts, the plurality of to-be-stored texts comprising a first text and a second text; a first processing module configured to process the first text using a first embedding model to obtain a first result; a second processing module configured to process the second text using a second embedding model to obtain a second result, the number of dimensions of the first embedding model and the second embedding model being different; and a storage module configured to store the first result and the second result to a knowledge base.

[0026] In a possible implementation of the present application, the first processing module is further configured to: divide the first text into a plurality of segments according to a characteristic of the first text; and embed the plurality of segments respectively by using the first embedding model to obtain a plurality of first vectors as the first result. The second processing module is further configured to: divide the second text into a plurality of segments according to a characteristic of the second text; and embed the plurality of segments respectively by using the second embedding model to obtain a plurality of second vectors as the second result.

[0027] In a possible implementation of the present application, the characteristic of the first text or the characteristic of the second text comprises one or more of the following: a white paper type, a product manual type, and a question and answer type.

[0028] In a possible implementation of the present application, the first processing module is further configured to: if the first text belongs to the white paper type, divide the first text into a plurality of segments by using a first manner, and the first manner comprises dividing according to a first fixed length.

[0029] In a possible implementation of the present application, the second processing module is further configured to: if the second text belongs to the product manual type, divide the second text into a plurality of segments by using a second manner, and the second manner comprises dividing according to a second fixed length smaller than the first fixed length.

[0030] In a possible implementation of the present application, the first processing module is further configured to: if the first text belongs to the question and answer type, divide the first text into a plurality of segments by using a first manner, and the first manner comprises dividing according to a paragraph.

[0031] In a possible implementation of the present application, the apparatus further comprises: a first embedding module configured to embed the inquiry sentence of the user by using the first embedding model to obtain a first retrieval vector; a second embedding module configured to embed the inquiry sentence by using the second embedding model to obtain a second retrieval vector; and a retrieval module configured to perform retrieval in the knowledge base by using the first retrieval vector and the second retrieval vector to obtain a plurality of retrieval results.

[0032] In a possible implementation of the present application, the retrieval module is further configured to: perform retrieval in the knowledge base by using the first retrieval vector to obtain X first retrieval results; perform retrieval in the knowledge base by using the second retrieval vector to obtain X second retrieval results; re-arrange the X first retrieval results and the X second retrieval results according to a similarity; and select a first Y retrieval results with a highest similarity to a lowest similarity as the plurality of retrieval results.

[0033] In a possible implementation of the present application, the device further comprises a constructing module configured to construct the prompt word according to the query sentence and the plurality of search results; and an inputting module configured to input the query sentence and the prompt word into the large language model to obtain an answer to the query sentence.

[0034] In a possible implementation of the present application, the constructing module is further configured to: decompose the query sentence into a plurality of questions, each question corresponding to one of the plurality of search results; combine each question with the corresponding search result; and combine the plurality of questions with the plurality of search results as the prompt word.

[0035] In another aspect, the present application also provides a computing device cluster comprising at least one computing device, each computing device comprising a processor and a memory; the processor of the at least one computing device is configured to execute instructions stored in the memory of the at least one computing device to cause the computing device cluster to perform the information processing method described above.

[0036] In another aspect, the present application provides a computer-readable storage medium storing a computer program, the computer program being configured to perform the information processing method described above.

[0037] In another aspect, the present application provides a computer program product comprising program code, which, when the computer runs the computer program product, causes the computer to implement the information processing method described above.

[0038] Any of the information processing devices, computing device clusters, computer-readable storage media or computer program products provided above are configured to perform the knowledge base construction method provided above, and thus can achieve the beneficial effects of the corresponding solutions of the corresponding method provided above, which will not be described here again. BRIEF DESCRIPTION OF DRAWINGS

[0039] Hereinafter, a specific embodiment of the present application will be described in detail with reference to the accompanying drawings, in which:

[0040] FIG. 1 shows an architectural schematic diagram of an information processing method according to an embodiment of the present application;

[0041] FIG. 2 shows a flow schematic diagram of an information processing method according to an embodiment of the present application;

[0042] FIG. 3 shows a flow schematic diagram of an information processing method according to another embodiment of the present application;

[0043] FIG. 4 shows a flow schematic diagram of an information processing method according to another embodiment of the present application;

[0044] FIG. 5 shows a schematic diagram of a prompt word in the information processing method according to the embodiment of FIG. 4;

[0045] FIG. 6 shows a structural schematic diagram of an information processing apparatus according to an embodiment of the present application;

[0046] FIG. 7 shows a structural schematic diagram of a computing device according to an embodiment of the present application;

[0047] FIG. 8 shows a structural schematic diagram of a computing device cluster according to an embodiment of the present application;

[0048] FIG. 9 shows a structural schematic diagram of a computing device cluster according to another embodiment of the present application. DETAILED DESCRIPTION

[0049] In order to make the concept and thought of the present application more clearly understood by those skilled in the art, the present application is described in detail below in conjunction with specific embodiments. It should be understood that the embodiments given herein are only a part of all the embodiments that the present application can have. Those skilled in the art can make improvements, modifications, or replacements to part or all of the following embodiments after reading the description of the present application, and these improvements, modifications, or replacements are also included in the scope of protection of the present application.

[0050] In this document, the terms "one", "a", and other similar words are not intended to mean that only one of the described things exists, but that the description is directed to only one of the described things, which can have one or more. In this document, the terms "comprise", "include", and other similar words are intended to mean logical relationships, and cannot be regarded as indicating spatial structural relationships. For example, "A includes B" is intended to mean that B logically belongs to A, and not that B is located inside A in space. In addition, the meaning of the terms "comprise", "include", and other similar words should be regarded as open, rather than closed. For example, "A includes B" is intended to mean that B belongs to A, but B does not necessarily constitute all of A, and A can also include C, D, E, and other elements.

[0051] In this document, the terms "first", "second", and other similar words do not imply any order, number, and importance, but are only used to distinguish different elements. In this document, the terms "embodiment", "the present embodiment", "an embodiment", "one embodiment" do not mean that the description is only applicable to one specific embodiment, but that the description can also be applicable to another one or more embodiments. Those skilled in the art should understand that any description made herein for a certain embodiment can be replaced, combined, or otherwise combined with the description of another one or more embodiments, and the new embodiments resulting from the replacement, combination, or other combination are easily thought of by those skilled in the art and are within the scope of protection of the present application.

[0052] Retrieval-augmented generation (RAG) refers to the optimization of large language model outputs to enable referencing an authoritative knowledge base beyond the training data sources before generating responses. Large language models (LLMs) are trained on vast amounts of data, using billions of parameters to generate raw outputs for tasks such as answering questions, translating languages, and completing sentences. Building on the already powerful capabilities of LLMs, RAG extends them to access internal knowledge bases specific to a domain or organization, all without the need to retrain the model. This is an economically efficient way to improve LLM outputs, keeping them relevant, accurate, and practical in various contexts.

[0053] Embedding refers to the representation of high-dimensional discrete features in a relatively low-dimensional continuous vector space. In RAG systems, document segment embedding is performed to convert text into computable vectors for subsequent similarity calculation and retrieval. Text embedding relies on a text embedding model. Embedding models are specifically designed to learn how to convert raw data (such as text, images, or other categories of data) into dense vectors (i.e., embeddings). This process involves learning the underlying structure and features of the data, allowing it to be represented in a form more suitable for machine learning models. That is, a text embedding model can preserve the semantic features of the text, which may include vocabulary, syntax, semantics, sentiment, emotion, topic, context, etc. Thus, semantically similar text segments can be found during retrieval. For example, embedding the sentence "Your text string goes here" using the text-embedding-ada-002 embedding model generates a 1536-dimensional vector, resulting in the following: "-0.006929283495992422, -0.005336422007530928, … -4547132266452536e-05, -0.024047505110502243". This is an array of length 1536. This vector contains all the features of the sentence, including vocabulary and syntax, which can be stored in a vector database for subsequent semantic search.

[0054] In some technologies in the field, low-dimensional embedding models do not perform well on long text, so to achieve good embedding results, high-dimensional embedding models are used regardless of the size of the document segment, resulting in high-dimensional vectors after embedding and large storage space; the similarity calculation is large during retrieval, and the retrieval speed is slow.

[0055] To this end, in some embodiments of the present application, a plurality of different embedding models are selected, and the length range of the text segment suitable for use is calculated according to the dimension data thereof; when the text segment is embedded in the RAG system, the embedding model with the lowest dimension suitable for use is selected according to the length of the text for embedding, and the embedding dimension is reduced on the premise of ensuring the representation ability of the embedding model. The embedding dimension required for a shorter text is not very high because the shorter text has fewer features. High dimension is used for long text (a large section of description in a white paper document), and low dimension is used for short text (such as the question and answer pairs in most FAQ documents). The embedding dimension of each text segment is reduced as much as possible, so that the storage resource consumption is significantly reduced, and the retrieval speed of the medium and low dimension vector library is greatly improved.

[0056] FIG. 1 shows an architecture schematic diagram of a knowledge base construction, information retrieval and question answering method according to an embodiment of the present application.

[0057] As shown in FIG. 1, after the RAG system loads the document, the corresponding segmentation method is selected according to the document type identification, the FAQ (frequently-asked questions) type is segmented by question paragraph, that is, each question and answer is segmented into a segment, and other documents are segmented by text token length. After segmentation, the corresponding embedding model is selected according to the text segment length, the vector of the corresponding dimension is generated after embedding, and then stored in the vector database. In other words, the long document is segmented into smaller text segments, such as fixed token segmentation, chapter segmentation, etc. Different segmentation strategies are used for different types of documents. Then, according to the token length of the text segment to be embedded, the most suitable and lowest dimension model is selected from the configured multiple embedding models for embedding, the vector is generated and stored in the corresponding vector database. As shown in the figure, there are three vectors, namely the first dimension vector, the second dimension vector and the third dimension vector, and there are three corresponding vector databases. However, those skilled in the art should know that the number of vectors and vector databases in the technical solution of the present application can be any suitable number, which is not limited by the present application. The user can determine the selection and number of embedding models according to the characteristics of the document to be embedded, and each embedding model corresponds to a vector database. When the user has a query request, the user input query statement is received, and the configured M (for example, 3) embedding models are used for embedding respectively to form the corresponding multiple vectors, and the top K related text segments are retrieved from the corresponding dimension vector database, and then the top N most similar segments are selected as background knowledge from the KxM segments through re-ranking. Then, the top N retrieval segments are used as background knowledge together with the query statement to construct a prompt word. In other words, when the user of the RAG system asks a question, the query statement is first embedded using embedding models of different dimensions, and then the relevant segments are retrieved from the vector database according to the embedded vectors. All segments retrieved from different vector databases are merged and sorted with high precision, and the top several segments after reordering and the query statement are used to construct a prompt word. Finally, the prompt word is sent to the large model for summary and answer.

[0058] FIG. 2 shows a flowchart of an information processing method according to an embodiment of the present application.

[0059] According to the present embodiment, the information processing method includes steps S210 to S240, which are described in detail below.

[0060] S210, a plurality of to-be-stored texts are obtained, and the plurality of to-be-stored texts include a first text and a second text.

[0061] In this embodiment, the text can refer to the literal content or collection of literal content recording knowledge and information. For example, the text can be a document, an article, a paper, a question and answer, a white paper, a specification, a manual, a book, etc. In this embodiment, the text to be stored can refer to the text waiting to be stored in the knowledge base. Generally, the text to be stored is the text originally obtained, which has various contents and formats and cannot be directly used to build a knowledge base and needs to be processed before being stored in the knowledge base so as to facilitate query or retrieval operations, etc. In this embodiment, the first and second texts can refer to any two texts in the text to be stored, the contents of which are different. Although only the first and second texts are given in this embodiment, those skilled in the art should know that the technical solution of this embodiment can also be processed for the third, fourth and more texts, and the first and second texts here are only illustrative.

[0062] S220, processing the first text by using the first embedding model to obtain a first result.

[0063] S230, processing the second text by using the second embedding model to obtain a second result, the number of dimensions of the first embedding model and the second embedding model being different.

[0064] In this embodiment, embedding can refer to the representation of high-dimensional discrete features mapped into a relatively low-dimensional continuous vector space. In the RAG system, the embedding of the document fragment is to convert the text into a computable vector, which is convenient for subsequent similarity calculation and retrieval. The embedding of the text needs to rely on the text embedding model. The embedding model is specially designed to learn how to convert the original data (such as text, pictures or other categories of data) into a dense vector (i.e. embedding). This process involves learning the underlying structure and features of the data, so as to be able to represent these original data in a form more suitable for processing by machine learning models. The text embedding model can preserve the semantic features of the text, which can include vocabulary, syntax, semantics, sentiment, emotion, theme, context, etc., so that semantically similar text fragments can be found during retrieval.

[0065] In this embodiment, the result obtained by processing the text by using the embedding model can refer to the process of inputting the text or its fragment into the embedding model and converting it into a corresponding vector through the calculation of the embedding model. The result obtained by the embedding model processing can be the vector directly obtained after the embedding calculation, or the result obtained by further calculating or processing the vector.

[0066] In this embodiment, the first and second embedding models have different numbers of dimensions, which can refer to different numbers of dimensions of the vectors generated by the first and second embedding models, and thus different vectors, i.e., different results. In some embodiments, the vector generated by the first embedding model is taken as one processing result, the vector generated by the second embedding model is taken as another processing result, and so on. Each processing result can correspond to a vector database, and the vector database can store vectors of a specific number of dimensions. The vector databases collectively constitute the knowledge base.

[0067] As an example, to process the first text using the first embedding model to obtain a first result, the first text can be divided into multiple segments according to the characteristics of the first text; then the first embedding model is used to embed the multiple segments respectively to obtain multiple first vectors as the first result. In this example, to process the second text using the second embedding model to obtain a second result, the second text can be divided into multiple segments according to the characteristics of the second text; then the second embedding model is used to embed the multiple segments respectively to obtain multiple second vectors as the second result.

[0068] In this example, the characteristics of the text can refer to the information structure or the writing style of the text. For example, according to the length characteristics of the text, the text can be divided into long text, medium text, and short text; according to the language characteristics of the text, the text can be divided into Chinese text, English text, etc.; according to the content characteristics of the text, the text can be divided into introduction text, description text, and reasoning text. For example, the characteristics of the text can be that the text belongs to a question and answer (also referred to as Q&A) type composed of multiple questions and multiple answers, or a long text (such as a book) type composed of multiple long paragraphs, or a short text (such as a network article) type composed of multiple short paragraphs. In this example, a segment can refer to a long text divided into multiple short text paragraphs according to certain rules.

[0069] In this example, dividing the text into multiple segments according to the characteristics of the text can refer to determining the division method of the text according to the characteristics of the text, and then dividing the text into multiple segments according to the determined division method. For example, for a text with long paragraphs, the text is divided into long segments, and for a text with short paragraphs, the text is divided into short segments.

[0070] In this example, the embedding model is used to embed multiple segments to obtain multiple vectors, which can mean that multiple segments divided from a certain text are input into an embedding model corresponding to the text, and multiple vectors of corresponding dimensions are obtained through the calculation of the embedding model. In this embodiment, the dimension number of the embedding model is adapted to the type of the text and the length of the segments divided therefrom. For a shorter text segment, the dimension number of the corresponding embedding model is usually smaller, and for a longer text segment, the dimension number of the corresponding embedding model is usually larger.

[0071] As an example, the characteristics of the first text or the characteristics of the second text include one or more of the following: white paper type, product manual type, and question and answer type.

[0072] In this example, the white paper type text has the characteristic of longer paragraphs, the product manual type text has shorter paragraphs, and the question and answer type text is composed of questions and answers, and the paragraphs of the text are the shortest. Different division methods can be used for the text according to the characteristics of the text.

[0073] As an example, in order to divide the first text into multiple segments according to the characteristics of the first text, when the first text belongs to the white paper type, the first text can be divided into multiple segments by using a first method, and the first method includes division according to a first fixed length. As an example, in order to divide the second text into multiple segments according to the characteristics of the second text, when the second text belongs to the product manual type, the second text can be divided into multiple segments by using a second method, and the second method includes division according to a second fixed length smaller than the first fixed length. As an example, in order to divide the first text into multiple segments according to the characteristics of the first text, when the first text belongs to the question and answer type, the first text can be divided into multiple segments by using the first method, and the first method includes division according to paragraphs.

[0074] In this example, since the paragraphs of the white paper type text are the longest, the first method, i.e., division according to a longer first fixed length, can be used for division. Since the paragraphs of the product manual type text are in the middle, the second method, i.e., division according to a shorter second fixed length, can be used for division. Since the paragraphs of the question and answer type text are the shortest, the third method, i.e., division according to paragraphs, can be used for division.

[0075] Specifically, three segmentation strategies can be adopted: 1) segmenting a segment every 1000 tokens with 200 tokens overlap, which is suitable for documents with long paragraph descriptions, and a large number of paragraphs are longer than 500 words, such as white papers; 2) segmenting a segment every 500 tokens with 100 tokens overlap, which is suitable for documents with moderate semantic description length, and most paragraphs are within 500 words, such as product description documents, and each feature is introduced in a small paragraph of text; and 3) segmenting by paragraph, which is suitable for FAQ documents, and the length of each question and answer is generally not high, and most are within 200 words. The judgment of the three types of documents is made by manual judgment according to the classification standard when importing the document.

[0076] As an example, each embedding model can have a different number of dimensions. Taking three embedding models as an example, the first embedding model can have 1 to 1000 dimensions, the second embedding model can have 1000 to 3000 dimensions, and the third embedding model can have 3000 to 5000 dimensions.

[0077] In this example, the first embedding model with 1 to 1000 dimensions can refer to the number of dimensions of the vector obtained after embedding, the second embedding model with 1000 to 3000 dimensions can refer to the number of dimensions of the vector obtained after embedding, and the third embedding model with 3000 to 5000 dimensions can refer to the number of dimensions of the vector obtained after embedding.

[0078] Specifically, three embedding models with different dimensions can be selected: BAAI / bge-base-zh (768 dimensions), BAAI / bge-m3 (1024 dimensions), and openai / text-embedding-3-large (3072 dimensions). The text segments with less than 200 tokens, 200-500 tokens, and 500-1000 tokens are embedded according to their token lengths and stored in the corresponding dimension vector database. Therefore, the embedding process of the text segment is fully automatic and does not require manual selection of the embedding model. In this way, compared to using a high-dimensional (3072-dimensional) embedding model for all samples, if the number of text segments in each of the three interval ranges is 1 / 3, the storage space after vectorization can be saved by (1 / 3) x (3072-1024) / 3072 + (1 / 3) x (3072-768) / 3072 = 47%. The amount of calculation of vector similarity calculation in the retrieval process is also reduced by 47%, thereby improving the retrieval efficiency and problem answering speed.

[0079] S240, storing the first result and the second result to the knowledge base.

[0080] In the embodiment, the knowledge base can refer to a set of interrelated knowledge pieces in computer memory storage, organization, management and use for solving one or some domain problems, which include theoretical knowledge, factual data, heuristic knowledge obtained from expert experience, such as definitions, theorems and operation rules related to the domain and common sense knowledge. The knowledge base construction can refer to the creation and improvement process of the knowledge base, including collecting knowledge of a specific domain, organizing and storing according to certain rules, and thus building a knowledge base that can be accessed or retrieved.

[0081] In the embodiment, storing the first result and the second result to the knowledge base can refer to storing the vector obtained by the first embedding model and the vector obtained by the second embedding model into different vector databases, and the knowledge base is composed of multiple vector databases. In the embodiment, the vector database can refer to a database storing vectors, and the data stored in such a database are all vectors of a certain type.

[0082] FIG. 3 shows a flowchart of an information processing method according to an embodiment of the present application.

[0083] According to the embodiment, the information processing method includes steps S310 to S370, which are described in detail below.

[0084] S310, a plurality of to-be-stored texts are obtained, and the plurality of to-be-stored texts include a first text and a second text.

[0085] S320, the first text is processed using a first embedding model to obtain a first result.

[0086] S330, the second text is processed using a second embedding model to obtain a second result, and the number of dimensions of the first embedding model and the second embedding model are different.

[0087] S340, the first result and the second result are stored to a knowledge base.

[0088] For details of S310 to S340, refer to the detailed description of S210 to S240 in the embodiment of FIG. 2 above, which will not be repeated here.

[0089] S350, the user's inquiry sentence is embedded using the first embedding model to obtain a first retrieval vector.

[0090] In the embodiment, the query statement can refer to a search field input by the user for querying information, such as a keyword, a question, or the like. For example, the query statement can refer to a statement input by the user for indicating which aspects of information the user wants to query or which aspects of answers the user wants to obtain. For example, the user wants to know the relevant information about "three-layer networking", and the query statement can be "what is three-layer networking" or "what advantages does three-layer networking have", or the like. In the embodiment, the first search vector can refer to a vector that can be queried in the knowledge base (containing the vector database corresponding to the first search vector) and is converted from the query statement of the user and can be processed by the vector database to obtain relevant knowledge information.

[0091] In the embodiment, the first embedding model is used to embed the query statement of the user to obtain the first search vector, which can refer to embedding the query statement by the first embedding model with a specific number of dimensions to obtain the first search vector with a specific number of dimensions. The number of dimensions of the search vector corresponds to the number of dimensions of the embedding model, and thus corresponds to the number of dimensions of the vector in the vector database embedded by the corresponding embedding model in the knowledge base.

[0092] S360, embedding the query statement by using the second embedding model to obtain the second search vector.

[0093] In the embodiment, the second search vector can refer to a vector with a different number of dimensions from the first search vector. The knowledge base has a vector database corresponding to the second search vector, which has the same number of dimensions as the second search vector, and thus has the same number of dimensions as the second embedding model.

[0094] S370, searching in the knowledge base by using the first search vector and the second search vector to obtain a plurality of search results.

[0095] In the embodiment, the search result can refer to those vectors that are most similar to the search vector obtained by searching and matching in the knowledge base by using the search vector. In the embodiment, the plurality of search results obtained by searching in the knowledge base by using the first search vector and the second search vector can refer to the plurality of search results obtained by searching in the vector database corresponding to the first search vector in the knowledge base by using the first search vector and searching in the vector database corresponding to the second search vector in the knowledge base by using the second search vector. The plurality of search results can be directly obtained by searching the search vector, or can be obtained by further computing or processing the search results directly obtained by searching the search vector.

[0096] As an example, in order to obtain a plurality of retrieval results by searching the knowledge base through the first retrieval vector and the second retrieval vector, the first retrieval vector can be used to search the knowledge base first to obtain X first retrieval results; then the second retrieval vector is used to search the knowledge base to obtain X second retrieval results; then the X first retrieval results and the X second retrieval results are rearranged according to the similarity; finally, the top Y retrieval results with the highest similarity are selected as the plurality of retrieval results.

[0097] In the present example, the knowledge base includes a vector database corresponding to the first retrieval vector and the second retrieval vector respectively. Searching the corresponding vector database through the first or second retrieval vector can enable the information to be searched in the vector database with the same number of dimensions according to the number of dimensions of the first or second retrieval vector. In the present example, rearranging the X first retrieval results and the X second retrieval results according to the similarity can mean arranging all the retrieval results (i.e. the retrieved vectors) in the vector database according to the numerical value of the similarity with the retrieval vector, with the most similar retrieval results arranged at the top and the least similar retrieval results arranged at the bottom. In the present example, selecting the top Y retrieval results with the highest similarity as the plurality of retrieval results can mean that after all the retrieval results are arranged according to the similarity, the top Y retrieval results arranged at the top are the most similar Y retrieval results, which are taken as the final retrieval results.

[0098] Specifically, the query of the user can be embedded by three embedding models respectively to search the top 5 text segments in the corresponding vector database, and 3x5=15 segments can be obtained after merging. Then, a re-ranking model (such as bge-reranker-base) can be used to select the top 3 from the 15 segments as the reference knowledge.

[0099] FIG. 4 shows a flowchart of an information processing method according to an embodiment of the present application.

[0100] According to the present embodiment, the information processing method includes steps S410 to S490, which will be described in detail below.

[0101] S410, a plurality of to-be-stored texts are obtained, and the plurality of to-be-stored texts include a first text and a second text.

[0102] S420, the first text is processed by using a first embedding model to obtain a first result.

[0103] S430, the second text is processed by using a second embedding model to obtain a second result, and the number of dimensions of the first embedding model and the second embedding model is different.

[0104] S440, store the first result and the second result to the knowledge base.

[0105] For details of S410 to S440, refer to the detailed description of S210 to S240 in the embodiment of FIG. 2 above, which will not be repeated here.

[0106] S450, embedding the inquiry sentence of the user by using the first embedding model to obtain a first retrieval vector.

[0107] S460, embedding the inquiry sentence by using the second embedding model to obtain a second retrieval vector.

[0108] S470, retrieving in the knowledge base by using the first retrieval vector and the second retrieval vector to obtain a plurality of retrieval results.

[0109] For details of S450 to S470, refer to the detailed description of S350 to S370 in the embodiment of FIG. 3 above, which will not be repeated here. S480, constructing a prompt word according to the inquiry sentence and the plurality of retrieval results.

[0110] In this embodiment, the prompt word can refer to a special language input used to guide the AI model to perform a specific task, such as generating text, image or performing other types of creation. The design and selection of the prompt word have a crucial impact on the output results of the AI model, because it directly determines how the AI understands and responds to the user's instructions. In this embodiment, constructing a prompt word according to the inquiry sentence and the plurality of retrieval results can refer to combining the inquiry sentence and the plurality of results obtained by retrieval to form a prompt word that can prompt the AI model what kind of answer it needs, so as to help the user get more accurate and comprehensive answer content.

[0111] S490, inputting the inquiry sentence and the prompt word into a large language model to obtain an answer to the inquiry sentence.

[0112] In this embodiment, the large language model can refer to a deep learning model trained using a large amount of text data, which can generate natural language text or understand the meaning of language text. The large language model can handle various natural language tasks such as text classification, question answering, dialogue, etc., and is an important way to artificial intelligence. Currently, large language models use similar architectures and pre-training objectives as small models, and the difference between small models is to increase the model size, training data, and computing resources. In this embodiment, the query sentence and the prompt word are input into the large language model to obtain the answer to the query sentence. This can mean that the query sentence and the prompt word constructed by the query sentence and the retrieval result before the query sentence are input into the large language model as input data of the large language model, and the answer corresponding to the user query sentence is obtained through the calculation and processing of the large language model. Such an answer has more accurate and comprehensive characteristics compared to the answer obtained by simply inputting the query sentence, which can better answer the user's questions and improve the user's information acquisition efficiency.

[0113] As an example, in order to construct a prompt word according to a query sentence and multiple retrieval results, the query sentence can be first decomposed into multiple questions, each question corresponding to one of the multiple retrieval results; then, each question and the corresponding retrieval result are combined; finally, the multiple combinations of multiple questions and multiple retrieval results are used as the prompt word.

[0114] In this example, decomposing the query sentence into multiple questions can mean that according to the content of each retrieval result, a question about one aspect of the knowledge information involved in the query sentence is used as a sub-question to form a combination of sub-questions and sub-answers with the retrieval result. Each retrieval result forms a combination, and multiple retrieval results form multiple combinations of sub-questions and sub-answers, thereby serving as a prompt word.

[0115] Specifically, based on the Top3 reference knowledge and the user's query sentence, a prompt word (Prompt) similar to that shown in FIG. 5 is constructed, and the constructed prompt word is sent to a large model interface (an interface of a self-built open source large model or a commercial large model).

[0116] Based on the method embodiment described in the foregoing FIG. 2, the embodiment of the present application further provides an information processing device, and a structure diagram thereof is shown in FIG. 6. The device is used to execute each step in the foregoing FIG. 2.

[0117] According to the embodiment, the information processing apparatus includes an obtaining module 610, a first processing module 620, a second processing module 630, and a storage module 640. The obtaining module 610 is configured to obtain a plurality of to-be-stored texts, the plurality of to-be-stored texts including a first text and a second text. The first processing module 620 is configured to process the first text by using a first embedding model to obtain a first result. The second processing module 630 is configured to process the second text by using a second embedding model to obtain a second result, the first embedding model and the second embedding model being different in dimension number. The storage module 640 is configured to store the first result and the second result to a knowledge base.

[0118] It should be noted that the information processing apparatus 600 provided by the embodiment shown in FIG. 6, when performing the information processing method, is only an example of the division of the above functional modules, and in actual application, the above functions can be completed by different functional modules according to the needs, that is, the internal structure of the apparatus is divided into different functional modules to complete all or part of the functions described above. In addition, the information processing apparatus 600 provided by the above embodiment and the information processing method embodiment shown in FIG. 2 respectively belong to the same concept, and the specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0119] The present application also provides a computing device 700. As shown in FIG. 7, the computing device 700 includes a bus 702, a processor 704, a memory 706, and a communication interface 708. The processor 704, the memory 706, and the communication interface 708 communicate through the bus 702. The computing device 700 can be a server or a terminal device. It should be understood that the present application does not limit the number of processors and memories in the computing device 700.

[0120] The bus 702 can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one line is shown in FIG. 7, but it does not mean that there is only one bus or only one type of bus. The bus 702 can include a path for transmitting information between various components (e.g., the memory 706, the processor 704, the communication interface 708) of the computing device 700.

[0121] The processor 704 can include any one or more of a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor (MP), or a digital signal processor (DSP), among other processors.

[0122] The memory 706 can include volatile memory, such as random access memory (RAM), and non-volatile memory, such as read-only memory (ROM), floppy disks, or hard disks.

[0123] The memory 706 stores executable program code that the processor 704 executes to implement the functions of the aforementioned obtaining module, first processing module, second processing module, and storage module, respectively, to implement the information processing method. That is, the memory 706 stores instructions for executing the information processing method.

[0124] The communication interface 708 uses a transceiver module, such as but not limited to a network interface card or a transceiver, to enable communication between the computing device 700 and other devices or communication networks.

[0125] Embodiments of the present disclosure also provide a computing device cluster. The computing device cluster includes at least one computing device. The computing device can be a server, such as a central server, an edge server, or a local server in a local data center. In some embodiments, the computing device can also be a terminal device, such as a desktop computer, a notebook computer, or a smartphone.

[0126] As shown in FIG. 8, the computing device cluster includes at least one computing device 700. The memory 706 in one or more computing devices 700 in the computing device cluster can store the same instructions for executing the information processing method.

[0127] In some possible implementations, the memory 706 of one or more computing devices 700 in the computing device cluster can also respectively store partial instructions for executing the information processing method. In other words, the combination of one or more computing devices 700 can collectively execute the instructions for executing the information processing method.

[0128] It should be noted that the memories 706 in different computing devices 700 in the computing device cluster can store different instructions for respectively performing part of the functions of the information processing method. That is, the instructions stored in the memories 706 in different computing devices 700 can implement the functions of one or more of the obtaining module, the first processing module, the second processing module and the storing module.

[0129] In some possible implementation manners, one or more computing devices in the computing device cluster can be connected through a network. The network can be a wide area network, a local area network or the like. FIG. 9 shows a possible implementation manner. As shown in FIG. 9, two computing devices 700A and 700B are connected through a network. Specifically, the computing devices are connected to the network through the communication interfaces in the computing devices. In this kind of possible implementation manner, the memory 706 in the computing device 700A stores instructions for performing the functions of the obtaining module and the first processing module. Meanwhile, the memory 706 in the computing device 700B stores instructions for performing the functions of the second processing module and the storing module.

[0130] The connection manner between the computing device cluster shown in FIG. 9 can be that, considering that the information processing method provided in the present application needs to store a large amount of data, it is considered to assign the functions implemented by the second processing module and the storing module to the computing device 700B.

[0131] It should be understood that the functions of the computing device 700A shown in FIG. 9 can also be completed by multiple computing devices 700. Similarly, the functions of the computing device 700B can also be completed by multiple computing devices 700.

[0132] The embodiments of the present application also provide another computing device cluster. The connection relationship between the computing devices in the computing device cluster can be similar to the connection manners of the computing device clusters described with reference to FIG. 8 and FIG. 9. The difference is that the memories 706 in one or more computing devices 700 in the computing device cluster can store the same instructions for performing the information processing method.

[0133] In some possible implementation manners, the memories 706 of one or more computing devices 700 in the computing device cluster can also respectively store part of the instructions for performing the information processing method. In other words, the combination of one or more computing devices 700 can collectively execute the instructions for performing the information processing method.

[0134] The embodiments of the present application further provide a computer program product containing instructions. The computer program product can be software or program product containing instructions, which can run on a computing device or be stored in any available medium. When the computer program product runs on at least one computing device, the at least one computing device is caused to perform the information processing method.

[0135] The embodiments of the present application can also be a computer readable storage medium, which stores computer program instructions. When the computer program instructions are run by a processor, the processor is caused to perform the steps in the knowledge base construction method according to various embodiments of the present application described above in the specification.

[0136] The computer readable storage medium can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium may, for example, include but is not limited to an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or apparatus, or any combination thereof. More specific examples (non-exhaustive list) of the readable storage medium include an electrical connection having one or more wires, a portable disc, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0137] The concepts, principles and ideas of the present application are described in detail above in combination with the specific implementation (including embodiments and examples). Those skilled in the art should understand that the implementation of the present application is not limited to the above forms. After reading the present application document, those skilled in the art can make any possible improvements, replacements and equivalents to the steps, methods, devices and components in the above embodiments. These improvements, replacements and equivalents should be considered to fall within the scope of the present application. The protection scope of the present application is only subject to the claims.

Claims

1. An information processing method characterized by comprising: The method comprises: acquiring a plurality of to-be-stored texts, the plurality of to-be-stored texts comprising a first text and a second text; processing the first text by using a first embedding model to obtain a first result; processing the second text by using a second embedding model to obtain a second result, the first embedding model and the second embedding model being different in dimension number; storing the first result and the second result to a knowledge base.

2. The method of claim 1, wherein, The processing of the first text by using the first embedding model to obtain the first result comprises: dividing the first text into a plurality of segments according to a characteristic of the first text; embedding the plurality of segments by using the first embedding model to obtain a plurality of first vectors as the first result; The processing of the second text by using the second embedding model to obtain the second result comprises: dividing the second text into a plurality of segments according to a characteristic of the second text; embedding the plurality of segments by using the second embedding model to obtain a plurality of second vectors as the second result.

3. The method of claim 2, wherein, The characteristic of the first text or the characteristic of the second text comprises one or more of the following: white paper type, product manual type, and question-and-answer type.

4. The method of claim 3, wherein, The dividing of the first text into a plurality of segments according to the characteristic of the first text comprises: if the first text belongs to the white paper type, dividing the first text into a plurality of segments by using a first mode, the first mode comprising dividing according to a first fixed length.

5. The method of claim 4, wherein, The dividing of the first text into a plurality of segments according to the characteristic of the first text comprises: if the second text belongs to the product manual type, dividing the second text into a plurality of segments by using a second mode, the second mode comprising dividing according to a second fixed length smaller than the first fixed length.

6. The method of claim 3, wherein, The dividing of the first text into a plurality of segments according to the characteristic of the first text comprises: if the first text belongs to the question-and-answer type, dividing the first text into a plurality of segments by using a first mode, the first mode comprising dividing according to a paragraph.

7. The method according to any one of claims 1 to 6, characterized in that, The method further comprises: embedding a query sentence of a user by using the first embedding model to obtain a first retrieval vector; embedding the query sentence by using the second embedding model to obtain a second retrieval vector; retrieving in the knowledge base by using the first retrieval vector and the second retrieval vector to obtain a plurality of retrieval results.

8. The method of claim 7, wherein, The retrieving in the knowledge base by using the first retrieval vector and the second retrieval vector to obtain the plurality of retrieval results comprises: retrieving in the knowledge base by using the first retrieval vector to obtain X first retrieval results; retrieving in the knowledge base by using the second retrieval vector to obtain X second retrieval results; rearranging the X first retrieval results and the X second retrieval results according to similarity; selecting a first Y retrieval results from high to low in similarity as the plurality of retrieval results.

9. The method of claim 7, wherein, The method further comprises: constructing a prompt word according to the query sentence and the plurality of retrieval results; inputting the query sentence and the prompt word into a large language model to obtain an answer for the query sentence.

10. The method of claim 9, wherein, The constructing the prompt word according to the query sentence and the plurality of retrieval results comprises: decomposing the query sentence into a plurality of questions, each of the questions corresponding to one of the plurality of retrieval results; combining each of the questions with the corresponding retrieval result; using the plurality of questions and the plurality of retrieval results as the prompt word.

11. An information processing apparatus comprising: The apparatus comprises: an acquisition module configured to acquire a plurality of to-be-stored texts, the plurality of to-be-stored texts comprising a first text and a second text; a first processing module configured to process the first text using a first embedding model to obtain a first result; a second processing module configured to process the second text using a second embedding model to obtain a second result, the first embedding model and the second embedding model having different dimension numbers; a storage module configured to store the first result and the second result to a knowledge base.

12. The apparatus of claim 11, wherein, The first processing module is further configured to: divide the first text into a plurality of segments according to a characteristic of the first text; embed the plurality of segments using the first embedding model to obtain a plurality of first vectors as the first result; The second processing module is further configured to: divide the second text into a plurality of segments according to a characteristic of the second text; embed the plurality of segments using the second embedding model to obtain a plurality of second vectors as the second result.

13. The apparatus of claim 12, wherein, The characteristic of the first text or the characteristic of the second text comprises one or more of the following: a white paper type, a product manual type, and a question-and-answer type.

14. The apparatus of claim 13, wherein, The first processing module is further configured to: if the first text belongs to the white paper type, divide the first text into a plurality of segments using a first manner, the first manner comprising dividing according to a first fixed length.

15. The apparatus of claim 14, wherein, The second processing module is further configured to: if the second text belongs to the product manual type, divide the second text into a plurality of segments using a second manner, the second manner comprising dividing according to a second fixed length smaller than the first fixed length.

16. The apparatus of claim 13, wherein, The first processing module is further configured to: if the first text belongs to the question-and-answer type, divide the first text into a plurality of segments using a first manner, the first manner comprising dividing according to a paragraph.

17. The apparatus of any one of claims 11 to 16, wherein, The apparatus further comprises: a first embedding module configured to embed a query sentence of a user using the first embedding model to obtain a first retrieval vector; a second embedding module configured to embed the query sentence using the second embedding model to obtain a second retrieval vector; a retrieval module configured to perform retrieval in the knowledge base through the first retrieval vector and the second retrieval vector to obtain a plurality of retrieval results.

18. The apparatus of claim 17, wherein, The retrieval module is further configured to: perform retrieval in the knowledge base through the first retrieval vector to obtain X first retrieval results; perform retrieval in the knowledge base through the second retrieval vector to obtain X second retrieval results; rearrange the X first search results and the X second search results according to the similarity; select the first Y search results with the highest similarity as the plurality of search results.

19. The apparatus of claim 17, wherein, The device further comprises: a constructing module configured to construct a prompt word according to the query sentence and the plurality of search results; an inputting module configured to input the query sentence and the prompt word into a large language model to obtain an answer to the query sentence.

20. The apparatus of claim 19, wherein, The constructing module is further configured to: decompose the query sentence into a plurality of questions, each of which corresponds to one of the plurality of search results; combine each of the questions with the corresponding search result; use the plurality of combinations of the plurality of questions and the plurality of search results as the prompt word.

21. A cluster of computing devices, characterized in that, The at least one computing device each includes a processor and a memory; the processor of the at least one computing device is configured to execute instructions stored in the memory of the at least one computing device, so that the computing device cluster executes the information processing method according to any one of claims 1 to 10.

22. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program for executing the information processing method according to any one of claims 1 to 10. The computer readable storage medium stores a computer program for executing the information processing method according to any one of claims 1 to 10.

Images