Enterprise suitability information retrieval system based on RAG technology

Abstract

An enterprise suitability information retrieval system based on an RAG technology relates to the technical field of information retrieval and comprises a monitoring center which is in communication connection with a document preprocessing module, an index construction module, a retrieval module, a content generation module and a user feedback module. The method comprises the following steps: preprocessing each document in an enterprise to obtain preprocessed document contents; according to the preprocessed document content, a document index is constructed, and the document index comprises key vocabularies and non-key vocabularies; searching a query request of a user according to the document index to obtain an adaptive document; constructing a content generation model, and generating retrieval content based on the content generation model; according to the information retrieval method and device, the retrieval feedback operation generated after the user receives the retrieval content is obtained, the retrieval content is generated according to the retrieval feedback operation and fed back to the user, the retrieval feedback operation comprises sending of the new query request and sending of the retrieval content regeneration instruction, and the accuracy of information retrieval and the quality of content generation are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of information retrieval technology, and in particular to an enterprise adaptability information retrieval system based on RAG technology. Background Art

[0002] The prior art CN111737608A "Enterprise information retrieval result sorting method and device" searches according to keywords input by the user to obtain search results; determines the stage of the activity and the characteristic indicators of each stage of the activity, the characteristic indicators are determined in combination with the user's historical behavior, and the stages of the activity include before the activity, during the activity and after the activity; determines the score of each characteristic indicator and each basic indicator of each search result according to a preset scoring rule; calculates the overall score of each search result according to the score of each characteristic indicator and each basic indicator of each search result and the preset weight; and sorts the search results in order of the overall score from high to low.

[0003] The prior art CN117972222A "Enterprise information retrieval method and device based on artificial intelligence" method includes: performing entity analysis on enterprise information through a preset named entity recognition model, extracting relationships on the results of the entity analysis through a deep learning neural network model, and constructing an enterprise relationship network; training a preset convolutional neural network model according to the user's historical retrieval data, and obtaining the corresponding user retrieval intent based on the enterprise information retrieval request currently sent by the user and the convolutional neural network model; performing feature fusion on the enterprise relationship network and the user's retrieval intent, and inputting the features obtained after feature fusion into a set multi-task learning model to obtain the enterprise retrieval information output by the multi-task learning model.

[0004] With the development of information technology, a large amount of document data has accumulated within enterprises. How to effectively manage and retrieve this information has become an important issue in enterprise management. Although traditional information retrieval technology can meet the needs to a certain extent, its retrieval effect and the quality of generated content are often unsatisfactory when faced with complex queries. Summary of the invention

[0005] In order to solve the above technical problems, the purpose of the present invention is to provide an enterprise adaptability information retrieval system based on RAG technology, including a monitoring center, wherein the monitoring center is communicatively connected with a document preprocessing module, an index building module, a retrieval module, a content generation module and a user feedback module. The system can improve the accuracy of information retrieval and the quality of content generation by combining information retrieval technology and natural language generation technology, thereby better serving the information management needs within the enterprise;

[0006] The document preprocessing module is used to preprocess various documents within the enterprise and obtain the preprocessed document content;

[0007] The index building module is used to build a document index based on the preprocessed document content, and the document index includes key words and non-key words;

[0008] The retrieval module is used to retrieve the user's query request according to the document index and obtain the adapted document;

[0009] The content generation module is used to construct a content generation model and generate search content based on the content generation model;

[0010] The user feedback module is used to obtain a search feedback operation generated by the user after receiving the search content, generate search content according to the search feedback operation and feed it back to the user, wherein the search feedback operation includes sending a new query request and sending a search content regeneration instruction.

[0011] Furthermore, the process of constructing a document index according to the preprocessed document content by the index construction module includes:

[0012] Extract words from the preprocessed document content, perform statistical analysis on several words in each document, obtain the frequency of occurrence of each word in each document, and the frequency of occurrence of each word in each document in all documents, obtain the criticality coefficient of each word in each document based on the frequency of occurrence of each word in each document, compare the criticality coefficient of each word in each document with a preset criticality threshold, mark the words in each document whose criticality coefficient is greater than or equal to the criticality threshold as key words, and mark the words in each document whose criticality coefficient is less than the criticality threshold as non-key words.

[0013] Furthermore, the retrieval module searches the user's query request according to the document index, and the process of obtaining the adapted document includes:

[0014] Extracting words from the user's query request, obtaining the query weight coefficient of each word in the query request, comparing the words in the query request with the key words and non-key words in each document for consistency, screening out documents containing key words or non-key words that are consistent with the words in the query request, and marking the documents as first documents;

[0015] Obtaining the number of key words and non-key words in the first document that are consistent with the words in the query request, and obtaining the retrieval relevance coefficient of each first document according to the number of key words and non-key words and the query weight coefficient of each word in the query request;

[0016] The first ten first documents with the highest retrieval relevance coefficients are screened out, the first ten first documents are sorted according to their retrieval relevance coefficients, a first document retrieval queue is generated, and a historical semantic retrieval is performed on the first document retrieval queue to obtain an adapted document.

[0017] Furthermore, the process of obtaining the query weight coefficient of each word in the query request includes:

[0018] Obtain the total number of documents within the enterprise, compare the terms in the query request with the terms contained in each document within the enterprise for consistency, if there is a term in the document that is consistent with the term in the query request, mark the document as an associated document of the term, obtain the total number of associated documents of the term, and obtain the query weight coefficient of the term based on the total number of associated documents of the term and the total number of documents.

[0019] Furthermore, the process of performing historical semantic retrieval on the first document retrieval queue to obtain the adapted document includes:

[0020] Determine whether the user has a historical query request record. If the user does not have a historical query request record, select a first document with the highest search correlation coefficient in the first document search queue, and mark the first document as an adapted document;

[0021] If the user has a historical query request record, obtain the adaptation document corresponding to the user's historical query request record, extract the key words in the adaptation document, compare the similarity between the key words in the adaptation document and the key words in each first document in the first document retrieval queue, obtain the key word similarity of each first document, screen out the first document with the highest key word similarity, and mark the first document as an adaptation document.

[0022] Furthermore, the content generation module constructs a content generation model, and the process of generating retrieval content based on the content generation model includes:

[0023] Build a content generation model based on deep learning, use word embedding technology to convert the vocabulary of each document within the enterprise into feature vectors, and use the feature vectors as training sets and test sets;

[0024] Setting initial parameters and a loss function of a content generation model, inputting the training set into the content generation model for training until the loss function training is stable, saving the model parameters, testing the content generation model with a test set until it meets preset requirements, and outputting the content generation model;

[0025] The word embedding technology is used to convert the vocabulary of the adapted document into a feature vector, and the feature vector corresponding to the adapted document is input into the content generation model. The retrieval content is generated according to the content generation model and fed back to the user.

[0026] Furthermore, when the user feedback module sends a new query request according to the search feedback operation, the process of generating search content and feeding back to the user includes:

[0027] When the user's retrieval feedback operation is to send a new query request, the vocabulary included in the new query request is obtained, the vocabulary included in the new query request and the vocabulary included in the user's query requests other than the new query request are combined to construct a vocabulary set, the vocabulary set is compared with the key words and non-key words of each first document in the first document retrieval queue, the retrieval relevance coefficient of each first document is obtained, the first document with the highest retrieval relevance coefficient is selected, the vocabulary of the first document and the vocabulary of the adapted document are converted into a feature vector using word embedding technology, the feature vector is input into a content generation model, and the retrieval content is generated according to the content generation model and fed back to the user.

[0028] Furthermore, when the user feedback module sends a search content regeneration instruction according to the search feedback operation, the process of generating the search content and feeding back to the user includes:

[0029] When the user's search feedback operation is to send a search content regeneration instruction, the first document corresponding to the adapted document in the first document search queue is removed, and a historical semantic search is performed on the first document search queue from which the first document corresponding to the adapted document is removed to obtain the adapted document. The vocabulary of the adapted document is converted into a feature vector using word embedding technology, and the feature vector is input into a content generation model. Search content is generated according to the content generation model and fed back to the user.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] 1. Through the preprocessing module, various types of enterprise documents are standardized, including text cleaning, word segmentation, stop word removal, stemming, etc., to ensure the consistency and comparability of document content. According to the preprocessed document content, a document index is constructed. By statistically analyzing the frequency of occurrence of words in each document, key words and non-key words are distinguished to improve the pertinence and accuracy of retrieval. Based on the document index, the user's query requests are efficiently retrieved, the most relevant documents are screened out, and the consistency comparison of key words and non-key words is performed to further improve the relevance of the retrieval results.

[0032] 2. Use deep learning technology to build a content generation model, convert the vocabulary in the document into feature vectors through word embedding technology, train the model to generate high-quality retrieval content, collect user feedback on the retrieval content through the user feedback module, continuously optimize the retrieval results and generate content, and improve the user experience.

[0033] 3. Combine the user's historical query records to perform historical semantic retrieval, generate retrieval content that better meets user needs based on the user's previous behavior and preferences, and adjust the retrieval strategy based on user feedback, such as sending new query requests or regenerating retrieval content, to ensure the dynamic and real-time nature of the retrieval results, significantly improve the accuracy of information retrieval and the quality of content generation, and provide strong support for internal information management and decision support within the enterprise. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a schematic diagram of an enterprise adaptability information retrieval system based on RAG technology according to an embodiment of the present application. DETAILED DESCRIPTION

[0035] The following is a clear and complete description of the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0036] like Figure 1 As shown, the enterprise adaptability information retrieval system based on RAG technology includes a monitoring center, and the monitoring center is communicatively connected with a document preprocessing module, an index building module, a retrieval module, a content generation module and a user feedback module;

[0037] The document preprocessing module is used to preprocess various documents within the enterprise and obtain the preprocessed document content. The preprocessing content includes but is not limited to text cleaning, word segmentation, and stop word removal operations to ensure the standardization of document content;

[0038] The index building module is used to build a document index based on the preprocessed document content, and the document index includes key words and non-key words;

[0039] The retrieval module is used to retrieve the user's query request according to the document index and obtain the adapted document;

[0040] The content generation module is used to construct a content generation model and generate search content based on the content generation model;

[0041] The user feedback module is used to obtain a search feedback operation generated by the user after receiving the search content, generate search content according to the search feedback operation and feed it back to the user, wherein the search feedback operation includes sending a new query request and sending a search content regeneration instruction.

[0042] It should be further noted that, in the specific implementation process, preprocessing various documents within an enterprise is an important step, which helps to improve the effectiveness of subsequent information retrieval and text analysis. The following is the specific preprocessing process, including operations such as text cleaning, word segmentation, and stop word removal;

[0043] Among them, text cleaning refers to removing noise information in the document, such as HTML tags, special characters, numbers, punctuation marks, etc., to retain meaningful text content. If the document is obtained from a web page, HTML tags need to be removed, which can be accomplished by using the BeautifulSoup library in Python. For numbers and special characters, etc., regular expressions are used to remove non-text content such as numbers, URLs, email addresses, dates, etc., and at the same time convert all letters to lowercase, unify the text format, and delete redundant spaces and line breaks;

[0044] Word segmentation refers to splitting a continuous text into individual lexical units, which are called "tokens". Word segmentation is one of the basic steps in natural language processing. In this invention, an n-gram model is constructed through Python and the NLTK library, and the n-gram model is used to determine how to split word boundaries;

[0045] Stop words refer to those words that frequently appear in the text but contribute little to the text content, such as "de", "shi", "zai", etc. Removing stop words can reduce noise and improve the effect of subsequent processing. The steps for removing stop words in this invention include:

[0046] Loading stop words from a predefined stop word list, and then removing the stop words in the text. The stop word list can be downloaded from open source projects or customized according to specific application scenarios.

[0047] By performing the above preprocessing steps on the document, the noise in the text can be effectively reduced, and the accuracy and efficiency of subsequent processing can be improved.

[0048] It should be further noted that, in the specific implementation process, the process of the index construction module constructing a document index based on the preprocessed document content includes:

[0049] Extract words from the preprocessed document content, perform statistical analysis on several words in each document, obtain the frequency of occurrence of each word in each document, and the frequency of occurrence of each word in each document in all documents, obtain the criticality coefficient of each word in each document based on the frequency of occurrence of each word in each document, compare the criticality coefficient of each word in each document with a preset criticality threshold, mark the words in each document whose criticality coefficient is greater than or equal to the criticality threshold as key words, and mark the words in each document whose criticality coefficient is less than the criticality threshold as non-key words.

[0050] It should be further explained that, in the specific implementation process, according to the occurrence frequency of each word in each document and the occurrence frequency of each word in each document in all documents, the calculation formula for obtaining the criticality coefficient of each word in each document is:

[0051]

[0052] Among them, i represents the criticality coefficient of word i, freq i Indicates the frequency of occurrence of word i in the document to which it belongs, freqA i represents the frequency of occurrence of word i in all documents, and δ represents the conversion coefficient.

[0053] It should be further explained that, in the specific implementation process, the retrieval module searches the user's query request according to the document index, and the process of obtaining the adapted document includes:

[0054] Extracting words from the user's query request, obtaining the query weight coefficient of each word in the query request, comparing the words in the query request with the key words and non-key words in each document for consistency, screening out documents containing key words or non-key words that are consistent with the words in the query request, and marking the documents as first documents;

[0055] Obtaining the number of key words and non-key words in the first document that are consistent with the words in the query request, and obtaining the retrieval relevance coefficient of each first document according to the number of key words and non-key words and the query weight coefficient of each word in the query request;

[0056] The first ten first documents with the highest retrieval relevance coefficients are screened out, the first ten first documents are sorted according to their retrieval relevance coefficients, a first document retrieval queue is generated, and a historical semantic retrieval is performed on the first document retrieval queue to obtain an adapted document.

[0057] It should be further explained that, in the specific implementation process, according to the number of key words and the number of non-key words and the query weight coefficient of each word in the query request, the calculation formula for obtaining the retrieval relevance coefficient of each first document is:

[0058]

[0059] Among them, ds represents the retrieval correlation coefficient, N im Indicates the number of key words, N not represents the number of non-key words, ω z represents the query weight coefficient of word z, ω p represents the query weight coefficient of word p, M im represents the vocabulary set in the query request that is consistent with the key words of the first document, M not Indicates the vocabulary set in the query request that is consistent with the non-key vocabulary of the first document.

[0060] It should be further explained that, in a specific implementation process, the process of obtaining the query weight coefficient of each word in the query request includes:

[0061] Obtain the total number of documents within the enterprise, compare the terms in the query request with the terms contained in each document within the enterprise for consistency, if there is a term in the document that is consistent with the term in the query request, mark the document as an associated document of the term, obtain the total number of associated documents of the term, and obtain the query weight coefficient of the term based on the total number of associated documents of the term and the total number of documents.

[0062] It should be further explained that, in the specific implementation process, the calculation formula for obtaining the query weight coefficient of a word according to the total number of associated documents of the word and the total number of documents is:

[0063]

[0064] Among them, ω i It represents the query weight coefficient of word i. As can be seen from the formula, the more documents are associated with word i, the smaller the query weight coefficient of word i is. N represents the total number of documents, which is always greater than one. NS i Represents the total number of associated documents of word i. The denominator is N-1 because when the total number of associated documents associated with word i is NS i When ω is 1, it ensures i is 1.

[0065] It should be further explained that, in a specific implementation process, the process of performing historical semantic retrieval on the first document retrieval queue to obtain the adapted document includes:

[0066] Determine whether the user has a historical query request record. If the user does not have a historical query request record, select a first document with the highest search correlation coefficient in the first document search queue, and mark the first document as an adapted document;

[0067] If the user has a historical query request record, obtain the adaptation document corresponding to the user's historical query request record, extract the key words in the adaptation document, compare the similarity between the key words in the adaptation document and the key words in each first document in the first document retrieval queue, obtain the key word similarity of each first document, screen out the first document with the highest key word similarity, and mark the first document as an adaptation document.

[0068] It should be further explained that, in the specific implementation process, the calculation formula of the key word similarity of the first document is:

[0069]

[0070] Wherein, sim represents the key word similarity, NC represents the number of the same key words in the adaptation document and the first document, NS represents the number of keywords in the adaptation document, and ND represents the number of keywords in the first document.

[0071] It should be further explained that, in a specific implementation process, the content generation module constructs a content generation model, and the process of generating search content based on the content generation model includes:

[0072] Build a content generation model based on deep learning, use word embedding technology to convert the vocabulary of each document within the enterprise into feature vectors, and use the feature vectors as training sets and test sets;

[0073] Setting initial parameters of the content generation model (including hidden layer size, number of layers, etc.) and a loss function (the present invention uses cross entropy loss), inputting the training set into the content generation model for training until the loss function training is stable, and saving the model parameters, testing the content generation model with a test set until it meets the preset requirements, and outputting the content generation model;

[0074] The word embedding technology is used to convert the vocabulary of the adapted document into a feature vector, and the feature vector corresponding to the adapted document is input into the content generation model. The retrieval content is generated according to the content generation model and fed back to the user.

[0075] It should be further explained that, in a specific implementation process, when the user feedback module sends a new query request according to the search feedback operation, the process of generating search content and feeding back to the user includes:

[0076] When the user's retrieval feedback operation is to send a new query request, the vocabulary included in the new query request is obtained, the vocabulary included in the new query request and the vocabulary included in the user's query requests other than the new query request are combined to construct a vocabulary set, the vocabulary set is compared with the key words and non-key words of each first document in the first document retrieval queue, the retrieval relevance coefficient of each first document is obtained, the first document with the highest retrieval relevance coefficient is selected, the vocabulary of the first document and the vocabulary of the adapted document are converted into a feature vector using word embedding technology, the feature vector is input into a content generation model, and the retrieval content is generated according to the content generation model and fed back to the user.

[0077] It should be further explained that, in a specific implementation process, when the user feedback module sends a search content regeneration instruction according to the search feedback operation, the process of generating the search content and feeding back to the user includes:

[0078] When the user's search feedback operation is to send a search content regeneration instruction, the first document corresponding to the adapted document in the first document search queue is removed, and a historical semantic search is performed on the first document search queue from which the first document corresponding to the adapted document is removed to obtain the adapted document. The vocabulary of the adapted document is converted into a feature vector using word embedding technology, and the feature vector is input into a content generation model. Search content is generated according to the content generation model and fed back to the user.

[0079] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. The enterprise adaptability information retrieval system based on RAG technology is characterized by: It includes a monitoring center, which is communicatively connected to a document preprocessing module, an index building module, a retrieval module, a content generation module and a user feedback module; The document preprocessing module is used to preprocess various documents within the enterprise and obtain the preprocessed document content; The index building module is used to build a document index based on the preprocessed document content, and the document index includes key words and non-key words; The retrieval module is used to retrieve the user's query request according to the document index and obtain the adapted document; The content generation module is used to construct a content generation model and generate search content based on the content generation model; The user feedback module is used to obtain a search feedback operation generated by the user after receiving the search content, generate search content according to the search feedback operation and feed it back to the user, wherein the search feedback operation includes sending a new query request and sending a search content regeneration instruction.

2. The enterprise adaptability information retrieval system based on RAG technology according to claim 1 is characterized in that: The process of constructing a document index according to the preprocessed document content by the index construction module includes: Extract words from the preprocessed document content, perform statistical analysis on several words in each document, obtain the frequency of occurrence of each word in each document, and the frequency of occurrence of each word in each document in all documents, obtain the criticality coefficient of each word in each document based on the frequency of occurrence of each word in each document, compare the criticality coefficient of each word in each document with a preset criticality threshold, mark the words in each document whose criticality coefficient is greater than or equal to the criticality threshold as key words, and mark the words in each document whose criticality coefficient is less than the criticality threshold as non-key words.

3. The enterprise adaptability information retrieval system based on RAG technology according to claim 2 is characterized in that: The retrieval module searches for the user's query request according to the document index, and the process of obtaining the adapted document includes: Extracting words from the user's query request, obtaining the query weight coefficient of each word in the query request, comparing the words in the query request with the key words and non-key words in each document for consistency, screening out documents containing key words or non-key words that are consistent with the words in the query request, and marking the documents as first documents; Obtaining the number of key words and non-key words in the first document that are consistent with the words in the query request, and obtaining the retrieval relevance coefficient of each first document according to the number of key words and non-key words and the query weight coefficient of each word in the query request; The first ten first documents with the highest retrieval relevance coefficients are screened out, the first ten first documents are sorted according to their retrieval relevance coefficients, a first document retrieval queue is generated, and a historical semantic retrieval is performed on the first document retrieval queue to obtain an adapted document.

4. The enterprise adaptability information retrieval system based on RAG technology according to claim 3 is characterized in that: The process of obtaining the query weight coefficient of each word in the query request includes: Obtain the total number of documents within the enterprise, compare the terms in the query request with the terms contained in each document within the enterprise for consistency, if there is a term in the document that is consistent with the term in the query request, mark the document as an associated document of the term, obtain the total number of associated documents of the term, and obtain the query weight coefficient of the term based on the total number of associated documents of the term and the total number of documents.

5. The enterprise adaptability information retrieval system based on RAG technology according to claim 4 is characterized in that: The process of performing historical semantic retrieval on the first document retrieval queue to obtain the adapted document includes: Determine whether the user has a historical query request record. If the user does not have a historical query request record, select a first document with the highest search correlation coefficient in the first document search queue, and mark the first document as an adapted document; If the user has a historical query request record, obtain the adaptation document corresponding to the user's historical query request record, extract the key words in the adaptation document, compare the similarity between the key words in the adaptation document and the key words in each first document in the first document retrieval queue, obtain the key word similarity of each first document, screen out the first document with the highest key word similarity, and mark the first document as an adaptation document.

6. The enterprise adaptability information retrieval system based on RAG technology according to claim 5 is characterized in that: The content generation module constructs a content generation model, and the process of generating retrieval content based on the content generation model includes: Build a content generation model based on deep learning, use word embedding technology to convert the vocabulary of each document within the enterprise into feature vectors, and use the feature vectors as training sets and test sets; Setting initial parameters and a loss function of a content generation model, inputting the training set into the content generation model for training until the loss function training is stable, saving the model parameters, testing the content generation model with a test set until it meets preset requirements, and outputting the content generation model; The word embedding technology is used to convert the vocabulary of the adapted document into a feature vector, and the feature vector corresponding to the adapted document is input into the content generation model. The retrieval content is generated according to the content generation model and fed back to the user.

7. The enterprise adaptability information retrieval system based on RAG technology according to claim 6 is characterized in that: When the user feedback module sends a new query request according to the search feedback operation, the process of generating search content and feeding back to the user includes: When the user's retrieval feedback operation is to send a new query request, the vocabulary included in the new query request is obtained, the vocabulary included in the new query request and the vocabulary included in the user's query requests other than the new query request are combined to construct a vocabulary set, the vocabulary set is compared with the key words and non-key words of each first document in the first document retrieval queue, the retrieval relevance coefficient of each first document is obtained, the first document with the highest retrieval relevance coefficient is selected, the vocabulary of the first document and the vocabulary of the adapted document are converted into a feature vector using word embedding technology, the feature vector is input into a content generation model, and the retrieval content is generated according to the content generation model and fed back to the user.

8. The enterprise adaptability information retrieval system based on RAG technology according to claim 7 is characterized in that: When the user feedback module sends a search content regeneration instruction according to the search feedback operation, the process of generating the search content and feeding back to the user includes: When the user's search feedback operation is to send a search content regeneration instruction, the first document corresponding to the adapted document in the first document search queue is removed, and a historical semantic search is performed on the first document search queue from which the first document corresponding to the adapted document is removed to obtain the adapted document. The vocabulary of the adapted document is converted into a feature vector using word embedding technology, and the feature vector is input into a content generation model. Search content is generated according to the content generation model and fed back to the user.

Images