A dynamic large language model adaptation method, device, medium and equipment

Abstract

The application relates to a dynamic large language model adaptation method and device, a medium and equipment. The method comprises the following steps: receiving input information of a user by a unified chat interface; processing the input information to obtain a keyword or a type of a file in the input information; obtaining static features and dynamic features of the user, fusing the static features and the dynamic features of the user to obtain a user portrait vector; when the keyword or the type of the file meets a certain preset condition, adapting a large language model related to a field corresponding to the preset condition to the user; when the keyword or the type of the file does not meet any preset condition, calculating the discrete degree of the user portrait vector and a training set of each large language model, and adapting the large language model with the smallest discrete degree to the user. The user does not need to manually switch the interfaces of different large language models, which helps to improve operation efficiency and continuity of intent understanding, makes the output result more in line with expectations, can dynamically select the adapted large language model according to the industry, region, knowledge background and the like of the user, improves response accuracy in a professional field, and improves the accuracy of the output result.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a method, apparatus, medium and device for adapting dynamic large language models. Background Technology

[0002] Large Language Models (LLMs) have made significant progress in the field of natural language processing. For example, large language models such as GPT-4 and LLaMA have achieved great success in multiple domains and are gradually being applied to specific fields such as finance, healthcare, and education. However, different large language models have their own areas of expertise. For example, Wenxin Yiyan 4.0 specializes in Chinese government and enterprise scenarios and the healthcare field; DeepSeek-R1 specializes in STEM fields (mathematics / code optimization); and Claude3 specializes in the legal field and long text processing.

[0003] Existing large language model dialogue systems have the following technical problems:

[0004] Current large language models employ static model matching, typically triggering predefined model calls through keyword matching (such as "law" or "finance"). For instance, when the keyword "contract" is detected, a large language model with expertise in the legal field is invoked. Static model matching cannot dynamically select an appropriate large language model based on the user's industry, region, and knowledge background, resulting in insufficient accuracy in professional domain responses.

[0005] The parameters of existing large language models are fixed, and general parameter configurations (such as Temperature and Top-P) are difficult to meet the needs of different scenarios (such as legal rigor vs. creative divergence).

[0006] When users submit questions from different domains, they may need to manually switch between different model interfaces, resulting in fragmented interaction, low operational efficiency, and impaired understanding of intent. Summary of the Invention

[0007] To overcome the problems existing in related technologies, this application provides a dynamic large language model adaptation method, apparatus, medium and device.

[0008] According to a first aspect of this application, a dynamic large language model adaptation method is provided, comprising:

[0009] The system receives user input information via a unified chat interface, and the input information includes text and / or files.

[0010] The input information is processed to obtain keywords or file types from the input information;

[0011] Obtain the static and dynamic features of the user, and fuse the static and dynamic features of the user to obtain the user profile vector;

[0012] When the keyword or file type meets a certain preset condition, the user is adapted to a domain-related large language model corresponding to the preset condition.

[0013] When the type of the keyword or file does not meet any preset condition, the dispersion of the user profile vector and the training set of each large language model is calculated, and the user is matched with the large language model with the smallest dispersion.

[0014] In some embodiments of this application, based on the foregoing scheme, the processing of the input information includes:

[0015] One or more of the following operations: text cleaning, file parsing, and character recognition.

[0016] In some embodiments of this application, based on the foregoing scheme, the step of obtaining the user's static and dynamic features, and fusing the user's static and dynamic features to obtain a user profile vector includes:

[0017] Extract user industry and geographic information, and vectorize them to generate static features;

[0018] Dynamically model user conversation history and uploaded files, and vectorize them to generate dynamic features.

[0019] The dynamic and static features are fused together to form a high-dimensional vector;

[0020] The high-dimensional vector is reduced in dimensionality to a user profile vector of a specified dimension.

[0021] In some embodiments of this application, based on the foregoing scheme, the dynamic large language model adaptation method further includes:

[0022] Send a request to the adapted large language model, the request including user input information and parameters of the large language model;

[0023] The user is responded to using the output of the adapted large language model.

[0024] The system receives feedback from the user, enabling it to optimize the parameters of the large language model based on this feedback.

[0025] In some embodiments of this application, based on the foregoing scheme, the dynamic large language model adaptation method further includes:

[0026] Based on the confidence level of the output results, the parameters of the adapted large language model are dynamically optimized.

[0027] In some embodiments of this application, based on the foregoing scheme, the dynamic optimization of the parameters of the adapted large language model includes:

[0028] When the confidence level is greater than or equal to the first preset threshold, the corresponding parameters are adjusted according to the technical field to which the large language model belongs.

[0029] In some embodiments of this application, based on the foregoing scheme, the dynamic optimization of the parameters of the adapted large language model further includes:

[0030] When the confidence level is less than the second preset threshold, a manual review is requested.

[0031] In some embodiments of this application, based on the foregoing scheme, after processing the input information, the method further includes:

[0032] The language type of the input information is detected, and the corresponding language processing pipeline is invoked.

[0033] According to another aspect of this application, a dynamic large language model adaptation device is provided, comprising:

[0034] A receiving module is used to receive user input information from a unified chat interface, the input information including text and / or files;

[0035] The processing module is used to process the input information and obtain the keywords or file types in the input information;

[0036] The user profile module is used to acquire the static and dynamic features of the user, and fuse the static and dynamic features of the user to obtain a user profile vector.

[0037] An adaptation module is used to adapt a domain-related large language model to the user when the type of the keyword or file meets a certain preset condition; and to calculate the dispersion of the user profile vector and the training set of each large language model when the type of the keyword or file does not meet any preset condition, and to adapt the user to the large language model with the smallest dispersion.

[0038] In some embodiments of this application, based on the foregoing scheme, the portrait module is further used for:

[0039] Extract user industry and geographic information, and vectorize them to generate static features;

[0040] Dynamically model user conversation history and uploaded files, and vectorize them to generate dynamic features.

[0041] The dynamic and static features are fused together to form a high-dimensional vector;

[0042] The high-dimensional vector is reduced in dimensionality to a user profile vector of a specified dimension.

[0043] In some embodiments of this application, based on the foregoing scheme, the dynamic large language model adaptation device further includes:

[0044] The output module is used to output results using an adapted large language model and respond to the user.

[0045] An optimization module is used to receive feedback from the user so that the system can optimize the model parameters of the large language model based on the feedback.

[0046] In some embodiments of this application, based on the foregoing scheme, the optimization module is further configured to dynamically optimize the parameters of the adapted large language model according to the confidence level of the output result.

[0047] In some embodiments of this application, based on the foregoing scheme, the optimization module is further configured to adjust the corresponding parameters according to the technical field to which the large language model belongs when the confidence level is greater than or equal to a first preset threshold; and to request manual review when the confidence level is less than a second preset threshold.

[0048] According to another aspect of this application, a computer-readable storage medium is provided, on which a computer program is stored, wherein the computer program, when executed, implements the steps of a dynamic large language model adaptation method.

[0049] According to another aspect of this application, a computer device is provided, including a processor, a memory, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of a dynamic large language model adaptation method.

[0050] This application receives user input information, including text and / or files, through a unified chat interface; processes the input information to obtain keywords or file types; acquires the user's static and dynamic features, and fuses these features to obtain a user profile vector; when the keyword or file type meets a preset condition, it adapts the user to a domain-related large language model corresponding to the preset condition; when the keyword or file type does not meet any preset condition, it calculates the dispersion of the user profile vector with the training sets of each large language model, and adapts the user to the large language model with the smallest dispersion. This allows users to call different large language models without manually switching between different model interfaces, improving operational efficiency and the continuity of intent understanding, making the output results more consistent with expectations; it can dynamically select the appropriate large language model based on the user's industry, region, knowledge background, etc., improving the response accuracy in professional fields and the accuracy of the output results.

[0051] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0052] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0053] Figure 1 This is a flowchart illustrating a dynamic large language model adaptation method according to an exemplary embodiment.

[0054] Figure 2 This is a block diagram illustrating a dynamic large language model adaptation device according to an exemplary embodiment.

[0055] Figure 3 This is a block diagram illustrating a dynamic large language model adaptation device according to an exemplary embodiment.

[0056] Figure 4 This is a block diagram illustrating a computer device according to an exemplary embodiment. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0058] To address the problems existing in the prior art, this application provides a dynamic large language model adaptation method, apparatus, medium, and device. Figure 1 This is a flowchart illustrating a dynamic large language model adaptation method according to an exemplary embodiment. (See reference) Figure 1 The dynamic large language model adaptation method provided in this application includes at least steps S11-S15.

[0059] Step S11: Receive user input information through a unified chat interface. The input information includes text and / or files.

[0060] The system provides users with a unified chat interface, which can connect to multiple large language models in the backend. Users input information of interest through this unified chat interface. User input can be text or uploaded files. Uploaded files can be PDF files or image files. Users can also upload files while inputting text.

[0061] Step S12: Process the input information to obtain keywords or file types from the input information.

[0062] Process user input to obtain keywords from the input or the file type of the file uploaded by the user.

[0063] In one exemplary embodiment, processing the input information includes:

[0064] One or more of the following operations: text cleaning, file parsing, and character recognition.

[0065] For example, Spacy (an open-source Python library for natural language processing that supports multiple languages ​​and provides efficient syntactic analysis and other natural language processing functions such as part-of-speech tagging and named entity recognition) can be used to clean user-input text, removing stop words and extracting entities; PyPDF2 (a pure Python open-source library specifically designed for processing PDF files, requiring no external programs or libraries to read, modify, merge, and split PDF files) can be used to parse user-uploaded PDF files; and OCR (Optical Character Recognition) can be used to recognize user-uploaded image files and convert them into text data. By processing input information, various types of user-input data can be parsed and converted, enabling natural language processing and a better understanding of user intent.

[0066] Step S13: Obtain the user's static and dynamic features, and fuse the user's static and dynamic features to obtain the user profile vector.

[0067] Different users have different industries, positions, and locations, so their expression habits and language will also differ. Different expressions can have the same meaning. The same words may also express different meanings in different contexts.

[0068] The dynamic large language model adaptation method provided in this application obtains the static and dynamic features of users, fuses the static and dynamic features of users to obtain user profile vectors, and can more accurately analyze user intentions and user characteristics based on user profile vectors.

[0069] Step S14: When the keyword or file type meets a certain preset condition, adapt the user to the domain-related large language model corresponding to the preset condition. Specifically, adapt the large language model to the user based on its capability level in the professional domain. Prioritize large language models that have achieved commercial-grade standards in the target domain benchmark tests and whose architecture includes domain-specific modules.

[0070] For example, in an exemplary embodiment, in the model matching decision, decision conditions for large language models related to different domains are set, along with parameter sets corresponding to each large language model. The model matching decision is as follows:

[0071] **Legal Field**:

[0072] If "Article XX" in text or file type == "Contract":

[0073] We selected a large language model relevant to the legal field, with the following parameter set: Temperature = 0.2, Top-P = 0.8.

[0074] **Financial Sector**:

[0075] If the uploaded file contains a "cash flow statement" or the conversation contains a "risk assessment":

[0076] We selected a large language model relevant to the financial field, with the following parameter set: Temperature = 0.3, FrequencyPenalty = 0.7.

[0077] **Education Sector**:

[0078] If the dialogue contains "teaching syllabus" or the uploaded file is a PPT:

[0079] We selected a large language model relevant to the education field, with the following parameter set: Temperature = 0.7, Top-K = 50.

[0080] If the user's input text includes "Article XX", or the file type uploaded by the user is "contract", it can be determined that the user's area of ​​interest is the legal field, and a large language model related to the legal field can be selected for the user;

[0081] If a user's uploaded file includes a "cash flow statement" or the conversation contains "risk assessment", it can be determined that the user's area of ​​interest is the financial field, and a large language model related to the financial field can be selected for the user.

[0082] If the user uploads a file of type "PPT" or the conversation contains "teaching syllabus", it can be determined that the user's area of ​​interest is education, and a large language model related to education can be selected for the user.

[0083] In the above embodiments, the large language model matching decision only involves three domain-related large language model adaptation conditions. In practical applications, the system can connect more domain-related large language models and set adaptation conditions for each large language model.

[0084] Step S15: When the type of keyword or file does not meet any preset condition, calculate the dispersion of the user profile vector and the training set of each decision model, and adapt the user to the large language model with the smallest dispersion.

[0085] When the keywords or file types in the user input information do not meet any preset conditions after processing, although it is not possible to select the corresponding domain-related large language model according to the preset conditions, the user can be matched with the large language model with the smallest dispersion by calculating the dispersion between the user profile vector and the training set of each large language model.

[0086] The dispersion can be expressed as: Among them, P input For the word frequency distribution of the input text, for example, "liability for breach of contract" appears 5 times and "contract law" appears 3 times; P domain The word frequency distribution of the training set for a large language model; D KL The KL divergence value measures the difference between the word frequency distribution of the user input file and the word frequency distribution of the large language model training set. The smaller the divergence value, the closer the input content is to the domain related to the training set.

[0087] For example, if a user enters: "Article 52 of the Contract Law on Liability for Breach of Contract", P input ={Breach of contract liability: 0.5, Contract law: 0.3, Article 52: 0.2}, Word frequency distribution P of a large language model related to the legal field. domain ={Liability for breach of contract: 0.4, Clause: 0.3, ...}, Word frequency distribution P of a large language model related to the financial field. domain ={Risk: 0.6, Return: 0.2, ...}, through dispersion calculation, the user input has the lowest dispersion in the large language model related to the legal field, thus adapting the large language model related to the legal field to the user.

[0088] The dynamic large language model adaptation method provided in this application integrates multiple large language models related to different domains. Users can input information through a unified chat interface and obtain feedback results from the large language models. Even if users need to query questions in different domains during the chat, they can do so through the unified chat interface. Users do not need to manually switch between different model interfaces, which helps improve operational efficiency and the continuity of intent understanding, making the output results more consistent with expectations. In addition, the adapted large language model can be dynamically selected based on the user's industry, region, knowledge background, etc., improving the response accuracy in professional fields and the accuracy of the output results.

[0089] In an exemplary embodiment, step S13 involves acquiring the user's static and dynamic features, fusing the static and dynamic features to obtain a user profile vector, including:

[0090] Extract user industry and geographic information, and vectorize them to generate static features;

[0091] Dynamically model user conversation history and uploaded files, and vectorize them to generate dynamic features.

[0092] The dynamic and static features are fused together to form a high-dimensional vector;

[0093] The high-dimensional vector is reduced in dimensionality to a user profile vector of a specified dimension.

[0094] It can extract industry information such as user's industry and job title from user's registration information, obtain user's geographic information from user's IP address, and generate static features by vectorization.

[0095] Dynamic behavior modeling is performed based on the user's conversation history and uploaded files. For example, features are extracted from the most recent 5 rounds of conversation, including TF-IDF (term frequency–inverse document frequency) keywords for the current session, and the top 10 keywords are determined based on their weights. Uploaded files are analyzed; if the file is a financial report, features such as return on assets and debt-to-equity ratio are extracted; if the file is a contract, features such as penalties for breach of contract and liability are identified. The extracted features are vectorized to generate dynamic features. Static and dynamic features can comprehensively represent the user's state, and different users can generate different static or dynamic features.

[0096] Static and dynamic features are concatenated and fused into a high-dimensional vector, such as a 1024-dimensional vector. Then, the 1024-dimensional vector is reduced in dimensionality to a user profile vector of a specified dimension (128 dimensions).

[0097] By fusing static and dynamic features, multi-source heterogeneous data can be integrated. Dimensionality reduction of high-dimensional vectors reduces computational overhead and avoids overfitting. This results in user profile vectors, which comprehensively represent user states. Based on these vectors, user intent can be understood more accurately, improving response precision.

[0098] In one exemplary embodiment, the dynamic large language model adaptation method further includes:

[0099] Send a request to the adapted large language model, the request including user input information and parameters of the large language model;

[0100] Use the adapted large language model to output the results and respond to the user;

[0101] Receive user feedback so that the system can optimize the parameters of the large language model based on the feedback.

[0102] The system adapts a large language model to the user based on keywords or file type, or based on the dispersion of the user profile vector and the training sets of various decision models. After this adaptation, the user input is submitted to the adapted large language model, and a unified chat interface sends a request to the model, including the input information and the model's parameters. Different parameter sets are adapted for different large language models. For example, the parameter set for a law-related large language model is: Temperature = 0.2, Top-P = 0.8 (Top-P, a sampling strategy based on cumulative probability, selects one word set with a cumulative probability of P), and Presence Penalty = 1.5 (Presence Penalty, an existence penalty to increase the diversity of generated text). The parameter set for an education-related large language model is: Temperature = 0.7, Top-K = 50 (a probability-based word selection strategy, selecting one of the top K most likely words as the output). The parameter sets for different domain-related large language models are pre-set by the system to ensure that the corresponding large language model outputs the desired results. For example, in the legal field, a large language model is expected to output highly deterministic results, with a temperature parameter of 0.2 in its parameter set. In the educational field, however, a large language model is expected to output more diverse results, with a temperature parameter of 0.7 in its parameter set.

[0103] The system sends user requests to an adapted large language model, outputs results using the adapted large language model, and responds to the user. In this application, the model source (e.g., "Generated by Legal Large Language Model v3.1") is displayed at the top of the output results, along with a confidence level (e.g., "Professional Match: 92%)" for user reference. After responding to the user, the system can also receive user feedback, allowing it to optimize the large language model parameters based on the feedback, thereby improving the accuracy of the output results. For example, in the education field, if the temperature parameter of the large language model is set to 0.7, but user feedback indicates that the model output results are inaccurate, the system can set the temperature parameter to 0.6 to improve the accuracy of the large language model's output results.

[0104] In one exemplary embodiment, the dynamic large language model adaptation method further includes:

[0105] Based on the confidence level of the output results, the parameters of the adapted large language model are dynamically optimized.

[0106] In addition to optimizing the parameters of the large language model based on user feedback, the system can also dynamically optimize the parameters of the large language model based on the confidence level of the output results, thereby improving service quality.

[0107] In one exemplary embodiment, dynamically optimizing the parameters of the adapted large language model includes:

[0108] When the confidence level is greater than or equal to the first preset threshold, the corresponding parameters are adjusted according to the technical field to which the large language model belongs. When the technical field requires high determinism in the output results, the temperature parameter is adjusted; when the technical field requires diversity in the output results, the Top-K parameter is adjusted.

[0109] For example, when the confidence level is greater than 0.8, in the legal field, to make the output results more certain, the temperature parameter is reduced to further suppress the randomness of the output results of the large language model.

[0110] It can be expressed as: T = 0.5 × (1 - C) + 0.1 × C, where T is the temperature value and C is the confidence level. The higher the confidence level, the lower the temperature.

[0111] In the field of education, the Top-K parameters are adjusted to ensure greater diversity in output results. This can be represented as K = 50 × e 2C Where K is the Top-K parameter value. The higher the confidence level, the more output results are generated.

[0112] In one exemplary embodiment, dynamically optimizing the parameters of the adapted large language model further includes:

[0113] When the confidence level is less than the second preset threshold, a manual review is requested.

[0114] When the confidence level of the input result of the large language model is less than the preset threshold, such as less than 0.3, the output result has a high degree of uncertainty and requires manual intervention to determine whether the parameter setting is incorrect or the wrong large language model is matched.

[0115] In one exemplary embodiment, after processing the input information, the method further includes:

[0116] Detect the language type of the input information and invoke the corresponding language processing pipeline.

[0117] Multiple large language models can be pre-categorized by language and divided into different pipelines. The system will then invoke the corresponding language processing pipeline based on the language type of the input text or uploaded file. For example, if the user's input is detected as Chinese, the Chinese processing pipeline will be invoked, and a suitable domain-related large language model will be matched to the user based on the user's input keywords, file type, or user profile vector. If the user's input is detected as Arabic, the less commonly spoken language processing pipeline will be invoked, and a suitable domain-related large language model will be matched to the user based on the user's input keywords, file type, or user profile vector. This enables the provision of corresponding services to users who speak different languages.

[0118] Figure 2 This is a block diagram illustrating a dynamic large language model adaptation device according to an exemplary embodiment. (Reference) Figure 2 The dynamic large language model adaptation device includes a receiving module 201, a processing module 202, a portrait module 203, and an adaptation module 204.

[0119] The receiving module 201 is configured to receive user input information via a unified chat interface, including text and / or files.

[0120] The processing module 202 is configured to process input information and obtain keywords or file types from the input information.

[0121] The profiling module 203 is configured to acquire the static and dynamic features of the user, and fuse the static and dynamic features to obtain a user profile vector.

[0122] The adaptation module 204 is configured to adapt the user to a domain-related large language model corresponding to the preset condition when the type of the keyword or file meets a certain preset condition; and to calculate the dispersion of the user profile vector and the training set of each large language model when the type of the keyword or file does not meet any preset condition, and to adapt the user to the large language model with the smallest dispersion.

[0123] In one exemplary embodiment, the image module 203 is further configured to:

[0124] Extract user industry and geographic information, and vectorize them to generate static features;

[0125] Dynamically model user conversation history and uploaded files, and vectorize them to generate dynamic features.

[0126] Dynamic and static features are fused into a high-dimensional vector;

[0127] Dimensionality reduction is performed on high-dimensional vectors to reduce them to user profile vectors of a specified dimension.

[0128] Figure 3 This is a block diagram illustrating a dynamic large language model adaptation device according to an exemplary embodiment. Referring to the image, the dynamic large language model adaptation device further includes:

[0129] Output module 301 is used to output results using the adapted large language model and respond to the user;

[0130] Optimization module 302 is used to receive user feedback and enable the system to optimize the parameters of the large language model based on the feedback.

[0131] In an exemplary embodiment, the optimization module 302 is further configured to: adjust the corresponding parameters according to the technical field to which the large language model belongs when the confidence level is greater than or equal to a first preset threshold; and request manual review when the confidence level is less than a second preset threshold.

[0132] Figure 4 This is a block diagram illustrating a computer device 400 for dynamic large language model adaptation according to an exemplary embodiment. For example, the computer device 400 may be provided as a server. (Refer to...) Figure 4 The computer device 400 includes a processor 401, the number of which can be set to one or more as needed. The computer device 400 also includes a memory 402 for storing instructions executable by the processor 401, such as application programs. The number of memories can be set to one or more as needed. The stored application programs can be one or more. The processor 401 is configured to execute instructions to perform the aforementioned dynamic large language model adaptation method.

[0133] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus (devices), or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. Computer storage media include volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data), including but not limited to RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible by a computer. Furthermore, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and can include any information delivery medium.

[0134] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0135] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes

[0136] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0137] In this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0138] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0139] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if these modifications and variations fall within the scope of the claims of this application and their equivalents, the intent of this application also includes these modifications and variations.

Claims

1. A dynamic large language model adaptation method, characterized in that, include: The system receives user input information via a unified chat interface, and the input information includes text and / or files. The input information is processed to obtain keywords or file types from the input information; Obtain the static and dynamic features of the user, and fuse the static and dynamic features of the user to obtain the user profile vector; When the keyword or file type meets a certain preset condition, the user is adapted to a domain-related large language model corresponding to the preset condition. When the type of the keyword or file does not meet any preset condition, the dispersion of the user profile vector and the training set of each large language model is calculated, and the user is matched with the large language model with the smallest dispersion.

2. The dynamic large language model adaptation method as described in claim 1, characterized in that, The processing of the input information includes: One or more of the following operations: text cleaning, file parsing, and character recognition.

3. The dynamic large language model adaptation method as described in claim 1, characterized in that, The process of acquiring the user's static and dynamic features, and fusing these features to obtain a user profile vector, includes: Extract user industry and geographic information, and vectorize them to generate static features; Dynamically model user conversation history and uploaded files, and vectorize them to generate dynamic features. The dynamic and static features are fused together to form a high-dimensional vector; The high-dimensional vector is reduced in dimensionality to a user profile vector of a specified dimension.

4. The dynamic large language model adaptation method as described in claim 1, characterized in that, Also includes: Send a request to the adapted large language model, the request including user input information and parameters of the large language model; The user is responded to using the output of the adapted large language model. The system receives feedback from the user, enabling it to optimize the parameters of the large language model based on this feedback.

5. The dynamic large language model adaptation method as described in claim 4, characterized in that, Also includes: Based on the confidence level of the output results, the parameters of the adapted large language model are dynamically optimized.

6. The dynamic large language model adaptation method as described in claim 4, characterized in that, The dynamic optimization of the parameters of the adapted large language model includes: When the confidence level is greater than or equal to the first preset threshold, the corresponding parameters are adjusted according to the technical field to which the large language model belongs.

7. The dynamic large language model adaptation method as described in claim 6, characterized in that, The dynamic optimization of the parameters of the adapted large language model also includes: When the confidence level is less than the second preset threshold, a manual review is requested.

8. The dynamic large language model adaptation method as described in claim 1, characterized in that, After processing the input information, the method further includes: The language type of the input information is detected, and the corresponding language processing pipeline is invoked.

9. A dynamic large language model adaptation device, characterized in that, include: A receiving module is used to receive user input information from a unified chat interface, the input information including text and / or files; The processing module is used to process the input information and obtain the keywords or file types in the input information; The user profile module is used to acquire the static and dynamic features of the user, and fuse the static and dynamic features of the user to obtain a user profile vector. An adaptation module is used to adapt a domain-related large language model to the user when the type of the keyword or file meets a certain preset condition; and to calculate the dispersion of the user profile vector and the training set of each large language model when the type of the keyword or file does not meet any preset condition, and to adapt the user to the large language model with the smallest dispersion.

10. The dynamic large language model adaptation device as described in claim 9, characterized in that, The portrait module is also used for: Extract user industry and geographic information, and vectorize them to generate static features; Dynamically model user conversation history and uploaded files, and vectorize them to generate dynamic features. The dynamic and static features are fused together to form a high-dimensional vector; The high-dimensional vector is reduced in dimensionality to a user profile vector of a specified dimension.

11. The dynamic large language model adaptation device as described in claim 9, characterized in that, Also includes: The output module is used to output results using an adapted large language model and respond to the user. An optimization module is used to receive feedback from the user so that the system can optimize the parameters of the large language model based on the feedback.

12. The dynamic large language model adaptation device as described in claim 10, characterized in that, The optimization module is also used for: Based on the confidence level of the output results, the parameters of the adapted large language model are dynamically optimized.

13. The dynamic large language model adaptation device as described in claim 12, characterized in that, The optimization module is also used to adjust the corresponding parameters according to the technical field to which the large language model belongs when the confidence level is greater than or equal to the first preset threshold; and to request manual review when the confidence level is less than the second preset threshold.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the steps of the method as described in any one of claims 1-8.

15. A computer device comprising a processor, a memory, and a computer program stored in the memory, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-8.

Images