GLM-based knowledge recommendation seat assistant method

Abstract

The invention discloses a GLM-based knowledge recommendation seat assistant method. The method comprises the following steps: S1, firstly, downloading a knowledge file to a specified directory of a knowledge recommendation seat assistant through FTP / HTTP; s2, after the text processing system reads the knowledge file, through text cleaning, forming a cleaned knowledge file and storing the cleaned knowledge file; s3, the knowledge files in the knowledge folder are processed through the knowledge slices, and a vector storage Embedding layer and a Faise index are formed; s4, carrying out vector search on the basis of an Embedding layer, a Fiss index and Fiss, splicing texts of search results to obtain context, and then filling a prompt template to obtain a prompt of an LLM (Logical Language Model); and S5, the questions are fed back through GLM + Faiiss, and corresponding answers are obtained. According to the method and the system, the problem that a large amount of time is consumed and a customer waits for answering due to the fact that an operator seat needs to click a document or other third-party platforms to search corresponding professional knowledge or data before, during and after communication with the user at present is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of software monitoring, and in particular to a method for recommending an agent assistant based on GLM. Background Art

[0002] With the rapid development of big data and artificial intelligence technologies, large language models, as an important technology in the field of natural language processing, have attracted widespread attention. Large language models can understand and generate natural language by learning from large amounts of text data, thereby achieving a variety of tasks such as machine translation, question-answering systems, and text summarization. Large language models have achieved remarkable results in many natural language processing tasks, even surpassing traditional methods. With the maturity of natural language processing technology, people have discovered that there is huge value hidden behind the massive amount of calls, which has attracted more and more people to study text data.

[0003] Especially for customer service hotlines, when agents handle problems online when users call the hotline, they need to quickly respond to the content involved in the problem and quickly provide results to users. This field covers too many detailed knowledge points and is too complicated. It is difficult for employees who have not been working in this field for less than 1 year to quickly respond to user content. Waiting for too long will easily cause dissatisfaction. In order to ensure that practitioners have matching professional knowledge recommended to agents before and during telephone communication, it is necessary to combine the professional knowledge in the enterprise knowledge base. Since there are about 5,000 complaint tickets to be handled every day, and dozens of tickets are assigned to each agent, the content requirements of each ticket are inconsistent. Before communicating with users each time, agents need to query a lot of information, such as what is the package type, or what are the types of outages, what is the corresponding speech or processing plan, and they need to understand the situation from different channels or third-party platforms before communicating with users, which will consume a lot of manpower and time. This will lead to problems such as long waiting time for users and reduced user experience. Summary of the invention

[0004] The purpose of the present invention is to solve the problems in the prior art and to propose a method for recommending an agent assistant based on knowledge of GLM.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for recommending an agent assistant based on knowledge of GLM comprises the following steps:

[0007] S1, first download the knowledge file to the directory specified by the knowledge recommendation agent assistant via FTP / HTTP;

[0008] S2. After the text processing system reads the knowledge file, it cleans the text to form a cleaned knowledge file and stores it;

[0009] S3. Process the knowledge files in the knowledge folder through knowledge slicing to form a vector storage Embedding layer and Faiss index;

[0010] S4. Based on the Embedding layer and Faiss index, Faiss performs vector search, concatenates the text of the search results to obtain the context, and then fills the prompt template to obtain the LLM prompt;

[0011] S5. Feedback the questions through GLM+Faiss and get the corresponding answers.

[0012] Preferably, in S1, the knowledge file includes knowledge from the traditional knowledge base in the vertical industry knowledge file.

[0013] Preferably, in S2, text cleaning specifically includes first processing the data of the enterprise knowledge base, then correcting the standard business knowledge through manual sampling, and establishing the Langchain knowledge base.

[0014] Preferably, S3 specifically includes:

[0015] Slice the cleaned knowledge and stop slicing when encountering punctuation marks;

[0016] Combine each slice and the original text into Documents;

[0017] Use the Embedding model to perform Embedding operations on each Document generated after slicing and generate vector storage;

[0018] Generate Faiss index for the stored vector using Faiss.

[0019] Preferably, S4 specifically includes:

[0020] First, use the Embedding model to embed the user's question to generate a question vector;

[0021] Then use Faiss to search the question vector in the Faiss index and vector stored previously to obtain the Top K Documents;

[0022] The Documents obtained by Faiss search are stitched together to form a Context, and a Prompt is added, which is sent to LLM to obtain the LLM Prompt.

[0023] Preferably, S5 also includes:

[0024] Use the TF-IDF method to match the feedback results in the information security database to filter out illegal information and sensitive data;

[0025] The content filtered by security information and Faiss search results are mapped to the original knowledge link and packaged into an answer and returned to the end user.

[0026] Preferably, in S2, the slicing step is a number of words as a segment.

[0027] Compared with the prior art, the present invention has the following advantages:

[0028] The present invention first processes the data of the enterprise knowledge base through knowledge cleaning, and then corrects the standard business knowledge through manual sampling, thereby completing the establishment of the Langchain knowledge base. Then, the Langchain knowledge base is connected to the GLM through Embedding, and the low-dimensional knowledge data is upgraded to make the general knowledge more refined and accurate. When the agent communicates with the user, the question vector is searched in the previously stored Faiss index and Embedding vector through Faiss technology, and the search results plus the corresponding Prompt are sent to GLM to obtain the answer of professional knowledge, and the answer is pushed to the agent and user communication interface, forming an advanced agent assistant application. In summary, the present invention uses this method to solve the problem that the current operator agent needs to click on documents or other third-party platforms to find the corresponding professional knowledge or information before, during, and after communicating with the user, which consumes a lot of time and makes the customer wait for the answer. It is based on the fields of statistics and natural language processing, and has the characteristics of reliability, strong modeling, and high accuracy. The whole process requires a small amount of manual operation and does not rely on the operator to provide training materials, thereby saving a lot of manpower and time costs for the operator. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A flowchart of a method for recommending an agent assistant based on knowledge in GLM proposed by the present invention;

[0030] Figure 2 A Langchain framework flow chart of a method for constructing a knowledge recommendation agent assistant based on GLM proposed by the present invention;

[0031] Figure 3 This is a knowledge search flow chart of a method for a knowledge recommendation agent assistant based on GLM proposed by the present invention. DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0033] Embedding: Embedding or embedding layer is a bridge between the two worlds. Through it, text, pictures, voices, and videos in the real world can be converted into a language that computers can recognize and use, and no information is lost during the conversion process.

[0034] Langchain: LangChain is an open source framework that allows developers working on artificial intelligence to combine large language models such as GPT-4 with external computation and data sources. The framework is currently available in the form of Python or JavaScript packages.

[0035] GLM model: Generalized Linear Models is an extension of the linear model. It establishes the relationship between the mathematical expectation of the response variable and the linear combination of the predictor variables through the connection function. Its characteristics are that it does not forcibly change the natural measurement of the data, and the data can have nonlinear and non-constant variance structure. It is a development of the linear model in studying the non-normal distribution of the response value and the concise and direct linear transformation of the nonlinear model.

[0036] Faiss: Faiss (Facebook AI Similarity Search) is an open source library for clustering and similarity search. It provides efficient similarity search and clustering for dense vectors and is an approximate nearest neighbor search library.

[0037] LLM model: LLM (Language Model with Lattice) is a model used for natural language processing, mainly used to transcribe audio signals into text

[0038] Reference Figure 1-3 , a method for recommending an agent assistant based on knowledge of GLM, comprising the following steps:

[0039] S1, first download the knowledge file to the directory specified by the knowledge recommendation agent assistant through FTP / HTTP; the knowledge file includes the knowledge from the traditional knowledge base in the vertical industry knowledge file.

[0040] S2. After the text processing system reads the knowledge file, it cleans the text to form a cleaned knowledge file and stores it.

[0041] Text cleaning specifically includes first processing the data of the enterprise knowledge base, then correcting the standard business knowledge through manual sampling, and establishing the Langchain knowledge base;

[0042] S3. Process the knowledge files in the knowledge folder through knowledge slicing to form a vector storage Embedding layer and Faiss index.

[0043] Specifically include:

[0044] Slice the cleaned knowledge, and take a certain number of words as a segment. For example, based on experience, take 20 words as a segment, and stop slicing when encountering punctuation marks;

[0045] Combine each slice and the original text into Documents;

[0046] Use the Embedding model (embedded model, usually refers to the model used to generate embedded vectors) to perform Embedding operations on each Document generated after slicing to generate vector storage;

[0047] Use Faiss technology to generate Faiss indexes for stored vectors. Through the above steps S1-S3, vertical industry knowledge has been constructed into knowledge in the Langchain framework in a completely new way.

[0048] S4. Based on the Embedding layer and Faiss index, Faiss performs vector search, concatenates the text of the search results to get the context, and then fills the prompt template to get the LLM prompt. Specifically,

[0049] When a user asks a question, the question is first embedded using the Embedding model to generate a question vector.

[0050] Then use Faiss technology to search the question vector in the previously stored Faiss index and vector to obtain the Top K Documents, that is, the files within the ranking K;

[0051] The Documents obtained by Faiss search are stitched together to form a Context, and a Prompt is added, which is sent to LLM to obtain the prompt of LLM.

[0052] S5. Feedback the questions through GLM+Faiss and get the corresponding answers.

[0053] Specifically, it also includes using the TF-IDF method to match the feedback results in the information security database to filter out illegal information and sensitive data;

[0054] The content filtered by security information and Faiss search results are mapped to the original knowledge link and packaged into an answer and returned to the end user.

[0055] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for recommending agent assistants based on knowledge in GLM, characterized in that: The following steps are involved: S1, first download the knowledge file via FTP / HTTP to the directory specified by the knowledge recommendation agent assistant; S2. After the text processing system reads the knowledge file, it cleans the text to form a cleaned knowledge file and stores it; S3. Process the knowledge files in the knowledge folder through knowledge slicing to form a vector storage Embedding layer and Faiss index; S4. Based on the stored Embedding layer and Faiss index, Faiss performs a vector search, concatenates the text of the search results to obtain the context, and then fills the prompt template to obtain the LLM prompt; S5. Feedback on questions through GLM+Faiss and get corresponding answers.

2. The method for recommending an agent assistant based on knowledge in GLM according to claim 1, characterized in that: In S1, the knowledge files include the knowledge in the vertical industry knowledge files that comes from the traditional knowledge base.

3. The method for recommending an agent assistant based on knowledge in GLM according to claim 1, characterized in that: In S2, text cleaning specifically includes first processing the data of the enterprise knowledge base, then correcting the standard business knowledge through manual sampling, and establishing the Langchain knowledge base.

4. The method for recommending an agent assistant based on knowledge in GLM according to claim 1, characterized in that: S3 specifically includes: Slice the cleaned knowledge and stop slicing when encountering punctuation marks; Combine each slice and the original text into Documents; Use the Embedding model to perform Embedding operations on each Document generated after slicing and generate vector storage; Generate Faiss index for the stored vector using Faiss.

5. The method for recommending an agent assistant based on knowledge in GLM according to claim 1, characterized in that: S4 specifically includes: First, the user's question is embedded using the Embedding model to generate a question vector; Then use Faiss to search the question vector in the previously stored Faiss index and vector to obtain the top K documents; The Documents obtained by the Faiss search are spliced ​​into a Context, and a Prompt is added. This is sent to the LLM to obtain the LLM Prompt.

6. The method for recommending an agent assistant based on knowledge in GLM according to claim 5, characterized in that: S5 also includes, Use the TF-IDF method to match the feedback results in the information security database to filter out illegal information and sensitive data; The content filtered by security information and Faiss search results are mapped to the original knowledge link and packaged into an answer and returned to the end user.

7. The method for recommending an agent assistant based on knowledge in GLM according to claim 4, characterized in that: In S2, the slicing step takes several words as a segment.

Images