A method for joint answering of a user knowledge base and a business system under an AI platform

Abstract

This invention provides a method for joint answering by a user knowledge base and a business system under an AI platform, belonging to the field of AI platform model technology. This invention constructs a three-level knowledge vector cache and a three-level question intent vector cache, employing vector representations of different dimensions to achieve hierarchical semantic matching. After receiving a user query, it uses a semantic similarity algorithm to perform parallel retrieval to generate an initial matching degree. A joint matching degree calculation function fuses multi-dimensional similarity information and utilizes intent recognition to optimize the model and improve accuracy. Based on the long and short memory matching length, an adaptive parameter adjustment function calculates the optimal sequence length to obtain real-time data from the business system. The static answer from the knowledge base is fused with the real-time data from the business system to generate the final answer. Simultaneously, the vector cache is updated and algorithm parameters are optimized based on user feedback, achieving a high-precision joint answering by the knowledge base and the business system.

Description

Technical Field

[0001] This invention belongs to the field of AI platform model technology, and more specifically, it relates to a method for joint answers from user knowledge bases and business systems under an AI platform. Background Technology

[0002] In the field of AI question-answering systems, traditional knowledge-based question-answering technologies primarily rely on single knowledge base retrieval or business system queries to generate answers. Existing knowledge-based question-answering systems typically employ vectorized retrieval techniques to semantically match user queries with knowledge base content, or use API interfaces to call business systems to obtain real-time data. These technologies are widely used in e-commerce customer service, internal enterprise Q&A, and intelligent assistant applications. However, traditional technologies have significant drawbacks. Firstly, knowledge base retrieval cannot obtain real-time business data, and business system queries lack deep knowledge background; their independent operation leads to insufficient answer completeness. Secondly, existing multi-source data fusion methods often employ simple data concatenation or weighted averaging, lacking a deep understanding of the characteristics of different data sources and precise matching mechanisms. Traditional technologies struggle to address the accuracy issue of fusion and matching between static knowledge base answers and real-time business system data, especially in complex scenarios where user queries require both professional knowledge and real-time business data. Existing methods cannot accurately determine the degree of matching and fusion ratio between the two data sources, resulting in final answers that fail to meet user needs in terms of accuracy and completeness. Summary of the Invention

[0003] In view of this, the present invention provides a method for joint answers from user knowledge base and business system under AI platform, which can solve the technical problem of insufficient accuracy in the fusion and matching of static answers from knowledge base and real-time data from business system in the prior art.

[0004] This invention is implemented as follows: It provides a method for joint answering by a user knowledge base and a business system within an AI platform. This method includes constructing a three-level knowledge vector cache and a three-level question intent vector cache. Upon receiving a user's natural language query request, a semantic similarity algorithm is used for parallel retrieval to generate initial knowledge matching and initial intent matching. A joint matching score is output by calling a joint matching score calculation function, taking the initial knowledge matching, initial intent matching, historical query frequency weights, time decay factors, and business importance coefficients as input parameters. An intent recognition optimization model is used to improve the accuracy of the joint matching score, filtering out candidate answers with the shortest matching path length. Based on the long-memory matching length and short-memory matching length output by the intent recognition optimization model, an adaptive parameter adjustment function is used to calculate the optimal sequence length for memory. Real-time data is obtained from the business system to supplement the answer. The static answers from the knowledge base are fused with the real-time data from the business system to generate the final joint answer result and update the historical query frequency weights. The three-level knowledge vector cache and the three-level question intent vector cache are updated based on user feedback quality scores.

[0005] Specifically, the three-level knowledge vector cache divides the knowledge base content into three precise levels according to the degree of abstraction and complexity. The domain knowledge layer stores professional domain concepts and theoretical knowledge with a vector dimension range of 1000 to 1100 dimensions. The business knowledge layer stores application scenarios and operational process knowledge with a vector dimension range of 400 to 600 dimensions. The common sense knowledge layer stores basic facts and general rule knowledge with a vector dimension range of 200 to 300 dimensions. The three levels are mapped to each other through a semantic association matrix.

[0006] Specifically, the three-level question intent vector cache divides user query history into three precise levels based on intent complexity and target clarity. The business query intent layer stores pure business data query intents with a vector dimension range of 700 to 800 dimensions. The knowledge acquisition intent layer stores pure knowledge content query intents with a vector dimension range of 300 to 450 dimensions. The mixed query intent layer stores mixed business knowledge query intents with a vector dimension range of 150 to 250 dimensions. The three levels are associated with each other through an intent conversion probability matrix.

[0007] Specifically, the joint matching degree calculation function is used to integrate multi-dimensional similarity information and output a comprehensive evaluation index. The inputs include the initial knowledge matching degree, the initial intent matching degree, the weight of historical query frequency, the time decay factor, and the business importance coefficient. The output is a joint matching degree score between 0 and 1. The joint matching degree calculation function comprehensively calculates each input parameter by multinomial weighted summation and applies the sigmoid activation function for normalization.

[0008] The matching path length refers to the number of intermediate nodes between the user query vector and the final answer vector. The matching path length ranges from 1 to 10, and a shorter path length indicates higher matching accuracy.

[0009] The long memory matching length refers to the time span during which the system retains the user's historical query context information, ranging from 30 to 180 days. This length is used to identify the user's long-term query patterns and preference characteristics, and serves as the input parameter for the adaptive parameter adjustment function.

[0010] The short memory matching length is specifically a limit on the number of user query context information entries that the system maintains in the current session, ranging from 5 to 50 entries. It is used to understand the user's immediate intent in the current query and serves as an input parameter for the adaptive parameter adjustment function.

[0011] Specifically, the optimal memory sequence length is the best context sequence length for data querying in the business system, calculated by balancing the long memory matching length and the short memory matching length. The value ranges from 10 to 100, and it is used to obtain real-time data from the business system.

[0012] Specifically, the intent recognition optimization model is a multi-layer attention network based on the Transformer architecture, which includes a 12-layer encoder structure, with 8 attention heads configured in each layer. The hidden layer dimension is set to 512. The context window size of the attention mechanism is dynamically adjusted according to the long memory matching length, and the sequence length parameter is consistent with the short memory matching length. The position encoding mechanism inside the model corresponds to the hierarchical structure of the three-level question intent vector cache.

[0013] Specifically, the adaptive parameter adjustment function calculates a comprehensive adjustment coefficient value based on four data points: the current query complexity index, historical query similarity, user session duration, and system load. When the comprehensive adjustment coefficient value is in the range of 0 to 0.3, a linearly increasing weight adjustment function is used to adjust the context window size to between 0.5 and 1 times the base value. When the comprehensive adjustment coefficient value is in the range of 0.3 to 0.7, a square root weight adjustment function is used to adjust the context window size to between 1 and 1.5 times the base value. When the comprehensive adjustment coefficient value is in the range of 0.7 to 1.0, a logarithmic weight adjustment function is used to adjust the context window size to between 1.5 and 2 times the base value.

[0014] The historical query frequency weight is derived from statistical analysis of user historical query records and is used to adjust the importance of similar queries in the joint matching degree calculation function, with a value range of 0.1 to 2.0. The time decay factor is derived from the exponential decay calculation of the query time interval and is used to reduce the impact weight of outdated queries in the joint matching degree calculation function, with a value range of 0.01 to 0.99. The business importance coefficient is derived from the importance level labeling of data in the business system and is used to increase the matching priority of key business data in the joint matching degree calculation function, with a value range of 1.0 to 5.0.

[0015] Before building the intent recognition optimization model, the process includes establishing a training dataset. This involves collecting and anonymizing user historical query log data, manually labeling query statements according to three categories: business query intent, knowledge acquisition intent, and mixed query intent, and constructing a training set containing 500,000 labeled samples. Simultaneously, a validation set of 100,000 samples and a test set of 50,000 samples are also established. After establishing the training dataset, the process includes training the intent recognition optimization model. The AdamW optimizer is used to update model parameters, with an initial learning rate set to 2×10⁻⁶. -5 The batch size was set to 32, the training period was 20 epochs, the cross-entropy loss function was used to calculate the classification loss, and a learning rate decay strategy was adopted during training, reducing the learning rate by 10% every 5 epochs. After the model training was completed, quantization compression was performed to compress the model weights from FP32 precision to INT8 precision.

[0016] The user feedback quality score is specifically derived from the user's numerical evaluation of the satisfaction with the answer result, with a value range of 1 to 5, and is used to update the weight allocation of the three-level knowledge vector cache and the three-level question intent vector cache.

[0017] This invention establishes a precise semantic matching framework by constructing a hierarchical knowledge vector cache and a question intent vector cache. The method utilizes a joint matching degree calculation function to fuse multi-dimensional similarity information and improves accuracy through an intent recognition optimization model, achieving precise matching across different data sources. The three-level question intent vector cache addresses the problem of inaccurate user intent recognition, while the joint matching degree calculation function solves the problem of improper weight allocation in multi-source data fusion. The intent recognition optimization model, based on the Transformer architecture and incorporating long short-term memory matching length and an adaptive parameter adjustment mechanism, dynamically adjusts the matching strategy, significantly improving fusion accuracy. This solves the technical problem of insufficient matching accuracy between static answers from the knowledge base and real-time data from the business system. Through multi-level vectorized representation and precise joint matching degree calculation, the system can accurately identify the complex intent of user queries, precisely match corresponding knowledge content and business data, and dynamically adjust fusion parameters through the optimization model to ensure that the final answer possesses both deep knowledge background and accurate real-time information, achieving high-precision multi-source data fusion. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention.

[0019] Figure 2 This is a schematic diagram of the intent recognition optimization model structure in this invention.

[0020] Figure 3 This is a schematic diagram of the three-level knowledge vector cache and the three-level question intent vector cache structure in Example 2. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0022] like Figure 1 The diagram shown is a flowchart of a method for jointly answering questions using a user knowledge base and a business system on an AI platform, provided by this invention. This method includes the following steps:

[0023] S01. Construct a three-level knowledge vector cache. Use semantic vectorization technology to store the user knowledge base content in three levels: domain knowledge layer, business knowledge layer, and common sense knowledge layer. The vector dimensions of each level are 1024, 512, and 256 respectively.

[0024] S02. Establish a three-level question intent vector cache, and store user historical queries in vectorized form according to three levels: business query intent layer, knowledge acquisition intent layer, and mixed query intent layer. The vector dimensions of each level are 768, 384, and 192 respectively.

[0025] S03. Receive the user's natural language query request, and use the semantic similarity algorithm to perform parallel retrieval in the three-level knowledge vector cache and the three-level question intent vector cache to generate the initial knowledge matching degree and the initial intent matching degree.

[0026] S04. Call the joint matching degree calculation function, take the initial knowledge matching degree, the initial intent matching degree, the historical query frequency weight, the time decay factor, and the business importance coefficient as input parameters, output the joint matching degree score, and use the pre-trained intent recognition optimization model to improve the accuracy of the joint matching degree score and select the candidate answer with the shortest matching path length.

[0027] S05. Based on the long memory matching length and short memory matching length output by the intent recognition optimization model, calculate the optimal memory sequence length through an adaptive parameter adjustment function, and obtain real-time data from the business system to supplement the answer.

[0028] S06. Integrate the static answers from the knowledge base with the real-time data from the business system to generate the final joint answer result and return it to the user, while updating the weight of the historical query frequency.

[0029] S07. Update the three-level knowledge vector cache and the three-level question intent vector cache according to the user feedback quality score, and optimize the semantic similarity algorithm parameters to improve the efficiency of subsequent query response.

[0030] S08. The fusion result of the static answer in the knowledge base and the real-time data in the business system is output as a joint answer, and the generation process and quality assessment result of the joint answer are recorded in the joint answer log of the user knowledge base and the business system, thus completing the complete process of joint answering by the user knowledge base and the business system.

[0031] Specifically, the three-level knowledge vector cache divides the knowledge base content into three precise levels according to the degree of abstraction and complexity. The domain knowledge layer stores professional domain concepts and theoretical knowledge with a vector dimension range of 1000 to 1100 dimensions. The business knowledge layer stores application scenarios and operational process knowledge with a vector dimension range of 400 to 600 dimensions. The common sense knowledge layer stores basic facts and general rule knowledge with a vector dimension range of 200 to 300 dimensions. The three levels are mapped to each other through a semantic association matrix, which records the association strength coefficient of concepts between different levels.

[0032] Specifically, the three-level question intent vector cache divides user query history into three precise levels based on intent complexity and target clarity. The business query intent layer stores pure business data query intents with a vector dimension range of 700 to 800 dimensions. The knowledge acquisition intent layer stores pure knowledge content query intents with a vector dimension range of 300 to 450 dimensions. The mixed query intent layer stores mixed business knowledge query intents with a vector dimension range of 150 to 250 dimensions. The three levels are associated with each other through an intent conversion probability matrix, which records the statistical probability values ​​of user intent conversion between different levels.

[0033] The joint matching degree calculation function is used to integrate multi-dimensional similarity information and output a comprehensive evaluation index. The inputs include the initial knowledge matching degree, the initial intent matching degree, the historical query frequency weight, the time decay factor, and the business importance coefficient. The output is the joint matching degree score between 0 and 1. The joint matching degree calculation function comprehensively calculates each input parameter by multinomial weighted summation and applies the sigmoid activation function for normalization.

[0034] The matching path length is specifically the number of intermediate nodes between the user query vector and the final answer vector. The numerical range of the matching path length is 1 to 10, and the shorter the path length, the higher the matching accuracy.

[0035] The historical query frequency weight is derived from the statistical analysis of users' historical query records and is used to adjust the importance of similar queries in the joint matching degree calculation function. The value range is 0.1 to 2.0.

[0036] The time decay factor is derived from the exponential decay calculation of the query time interval and is used to reduce the impact weight of outdated queries in the joint matching degree calculation function. The value range is 0.01 to 0.99.

[0037] The business importance coefficient is derived from the importance level label of the data in the business system and is used to improve the matching priority of key business data in the joint matching degree calculation function. The value range is 1.0 to 5.0.

[0038] Specifically, the long memory matching length is the time span during which the system retains the user's historical query context information, ranging from 30 to 180 days. It is used to identify the user's long-term query patterns and preference characteristics, and serves as the input parameter of the adaptive parameter adjustment function.

[0039] Specifically, the short memory matching length is a limit on the number of user query context information entries that the system retains in the current session, ranging from 5 to 50 entries. It is used to understand the user's immediate intent in the current query and serves as the input parameter for the adaptive parameter adjustment function.

[0040] Specifically, the optimal memory sequence length is the best context sequence length for data querying in the business system, calculated by balancing the long memory matching length and the short memory matching length. The value ranges from 10 to 100, and it is used to obtain real-time data from the business system.

[0041] like Figure 2 As shown, the specific structure of the intent recognition optimization model is a multi-layer attention network based on the Transformer architecture, containing a 12-layer encoder structure, with 8 attention heads configured in each layer, and the hidden layer dimension is set to 512. The intent recognition optimization model is designed in conjunction with the hierarchical structure of the three-level question intent vector cache, the long memory matching length, and the short memory matching length. The context window size of the attention mechanism is dynamically adjusted according to the long memory matching length, and the sequence length parameter is consistent with the short memory matching length. The positional encoding mechanism inside the model corresponds to the hierarchical structure of the three-level question intent vector cache.

[0042] The steps for establishing the training dataset for the intent recognition optimization model specifically include collecting user historical query log data and performing anonymization processing, manually labeling query statements according to three categories: the business query intent layer, the knowledge acquisition intent layer, and the hybrid query intent layer, constructing a training set containing 500,000 labeled samples, where each sample contains the original query text, intent category label, query context information, and corresponding answer quality score, and simultaneously establishing a validation set containing 100,000 samples and a test set containing 50,000 samples.

[0043] Specifically, the intent recognition optimization model training step includes updating model parameters using the AdamW optimizer, with an initial learning rate set to 2×10⁻⁶. -5 The batch size was set to 32, the training period was 20 epochs, the cross-entropy loss function was used to calculate the classification loss, and a learning rate decay strategy was adopted during training, reducing the learning rate by 10% every 5 epochs. After the model training was completed, quantization compression was performed to compress the model weights from FP32 precision to INT8 precision.

[0044] The adaptive parameter adjustment function is used to adjust the context window size of the intent recognition optimization model. The adaptive parameter adjustment function calculates a comprehensive adjustment coefficient value based on four data points: current query complexity index, historical query similarity, user session duration, and system load. When the comprehensive adjustment coefficient value is in the range of 0 to 0.3, a linearly increasing weight adjustment function is used to adjust the context window size to between 0.5 and 1 times the base value. When the comprehensive adjustment coefficient value is in the range of 0.3 to 0.7, a square root weight adjustment function is used to adjust the context window size to between 1 and 1.5 times the base value. When the comprehensive adjustment coefficient value is in the range of 0.7 to 1.0, a logarithmic weight adjustment function is used to adjust the context window size to between 1.5 and 2 times the base value.

[0045] The current query complexity index is derived from a comprehensive calculation of the number of words, syntactic structure complexity, and terminology density of the user's current query statement, and is used to evaluate the query processing difficulty in the adaptive parameter adjustment function.

[0046] The historical query similarity is derived from the semantic similarity calculation results between the current query and the user's historical query records, and is used to determine the repetition of query patterns in the adaptive parameter adjustment function.

[0047] The user session duration is derived from the actual duration statistics of the user's current session and is used to evaluate the continuity requirements of user queries in the adaptive parameter adjustment function.

[0048] The system load status is derived from real-time monitoring data of the current CPU utilization, memory usage, and concurrent query count of the AI ​​platform, and is used to balance system performance and response quality in the adaptive parameter adjustment function.

[0049] The user feedback quality score is derived from the user's numerical evaluation of the satisfaction with the answer result, with a value range of 1 to 5, and is used to update the weight allocation of the three-level knowledge vector cache and the three-level question intent vector cache.

[0050] The specific implementation methods of the above steps are described in detail below.

[0051] The specific implementation of step S01 involves establishing a three-level knowledge vector caching system. First, the original text content in the user's knowledge base is preprocessed, including text cleaning, word segmentation, and stop word removal. Then, deep learning-based word embedding technology is used to convert the preprocessed text into vector representations. The core function of this step is to construct a hierarchical knowledge representation system, reflecting the level of abstraction and complexity of knowledge through different vector dimensions. In the domain knowledge layer, an attention-based encoder encodes professional concepts and theoretical knowledge into 1024-dimensional vectors. This dimension can fully capture complex inter-domain relationships. Knowledge with a complexity evaluation value exceeding a threshold of 0.8 is categorized into this layer. The business knowledge layer uses a bidirectional long short-term memory network to encode application scenarios and operational processes into 512-dimensional vectors, suitable for representing practical knowledge of moderate complexity. Knowledge with a complexity evaluation value between 0.4 and 0.8 belongs to this layer. The common sense knowledge layer uses a simplified recurrent neural network to encode basic facts into 256-dimensional vectors, handling general knowledge with a complexity evaluation value below 0.4. A semantic association matrix is ​​used to establish a mapping relationship between different levels. This matrix records the association strength coefficient between concepts at different levels. When the association coefficient exceeds the threshold of 0.6, a direct mapping path is established.

[0052] The specific implementation of step S02 involves constructing a three-level question intent vector cache. The purpose of this step is to store and manage user historical queries hierarchically according to intent complexity. First, intent analysis is performed on user historical query records, using sequence labeling technology based on recurrent neural networks to identify key intent words and grammatical structures in the queries. The business query intent layer uses an attention pooling mechanism to encode pure business data queries into 768-dimensional vectors, suitable for handling business operation requests with clear objectives. Queries with a business term density exceeding 0.7 and a grammatical complexity below 0.3 are classified into this layer. The knowledge acquisition intent layer uses a context-aware encoder to encode knowledge content queries into 384-dimensional vectors, handling queries whose primary goal is information acquisition. Queries with a knowledge term density exceeding 0.6 and a reasoning complexity exceeding 0.5 belong to this layer. The hybrid query intent layer uses a multi-task learning framework to encode mixed business knowledge queries into 192-dimensional vectors, handling composite queries involving both business operations and knowledge acquisition. The layers are linked through an intent conversion probability matrix. This matrix records the statistical probability of user intent conversion between different layers based on the Markov chain principle. When the conversion probability exceeds the threshold of 0.3, the association path between layers is established.

[0053] The specific implementation of step S03 involves performing parallel retrieval and similarity calculation. This step aims to quickly and accurately match user queries with knowledge base content. First, the natural language query input by the user is preprocessed and vectorized, using the same encoding technology as the knowledge base construction to ensure consistency in the vector space. Then, in the three-level knowledge vector cache, a cosine similarity algorithm is used for parallel retrieval. This algorithm effectively measures the semantic similarity between high-dimensional vectors, calculating the similarity score between the query vector and the knowledge vectors at each level. Simultaneously, in the three-level question intent vector cache, a weighted combination of Euclidean distance and Manhattan distance is used for intent matching, with a weight allocation ratio of 0.7 to 0.3. This combination can take into account both the geometric characteristics and dimensionality sensitivity of the vector space. The retrieval process uses an approximate nearest neighbor search algorithm to improve computational efficiency, and a batch processing mechanism is activated when the number of candidate results exceeds 100. The initial knowledge matching degree is calculated based on the weighted average principle. The matching results at each level are weighted and summed according to their importance coefficients: domain knowledge layer has a weight of 0.5, business knowledge layer has a weight of 0.3, and common sense knowledge layer has a weight of 0.2. The initial intent matching degree adopts the max pooling strategy, selecting the result with the highest similarity in each level as the representative matching degree.

[0054] Step S04 involves calling the joint matching degree calculation function and improving its accuracy. This step aims to generate the final matching score by integrating multi-dimensional information. The joint matching degree calculation function uses a multinomial weighted summation method, taking the initial knowledge matching degree, initial intent matching degree, historical query frequency weight, time decay factor, and business importance coefficient as input parameters for comprehensive calculation. The historical query frequency weight is derived from statistical analysis of user query history, using logarithmic smoothing to avoid excessive frequency differences, with a value range between 0.1 and 2.0. The time decay factor is calculated based on an exponential decay function, with a decay rate parameter set to 0.05 to reduce the impact weight of outdated queries, and a value range between 0.01 and 0.99. The business importance coefficient is derived from the importance level labeling of data in the business system, using the analytic hierarchy process (AHP) to determine the relative importance of each business data point, with a value range between 1.0 and 5.0. The calculation results are normalized using a Sigmoid activation function to ensure the joint matching degree score is between 0 and 1. The intent recognition optimization model improves the accuracy of joint matching by using a multi-layer attention network based on the Transformer architecture. It captures long-distance dependencies in the query context through a self-attention mechanism and filters out candidate answers with the shortest matching path length. The path length threshold is set to 3, and candidate answers exceeding this threshold will be filtered out.

[0055] The specific implementation of step S05 involves calculating the optimal memory sequence length and acquiring real-time data. The purpose of this step is to determine the optimal context sequence length to optimize the data query performance of the business system. The intent recognition optimization model outputs two key parameters: long memory matching length and short memory matching length. The long memory matching length represents the time span during which the system retains the user's historical query context information, determined based on user behavior pattern analysis, with a value range of 30 to 180 days. The short memory matching length represents the limit on the number of query context entries maintained in the current session, determined based on session coherence analysis, with a value range of 5 to 50 entries. The adaptive parameter adjustment function calculates a comprehensive adjustment coefficient based on four input parameters: current query complexity index, historical query similarity, user session duration, and system load. When the comprehensive adjustment coefficient is in the range of 0 to 0.3, a linearly increasing weight adjustment function is used to adjust the sequence length to 0.5 to 1 times the base value. When the coefficient is in the range of 0.3 to 0.7, a square root weight adjustment function is used to adjust the length to 1 to 1.5 times the base value. When the coefficient value is in the range of 0.7 to 1.0, a logarithmic weighting adjustment function is used to adjust it to 1.5 to 2 times the base value. The final optimal sequence length is used to obtain the corresponding number of real-time data records from the business system.

[0056] The specific implementation of step S06 involves fusing static answers and real-time data to generate the final answer. This step aims to integrate multi-source information to provide a comprehensive answer result. First, the static answers retrieved from the knowledge base undergo relevance assessment and content filtering. A graph neural network-based information fusion algorithm is then used to integrate multiple related answer fragments. Simultaneously, real-time data obtained from the business system is formatted and its timeliness verified to ensure accuracy and timeliness. The fusion process employs an attention mechanism weight allocation strategy, dynamically adjusting the weight ratio between static answers and real-time data based on query type and user preferences. When the query leans towards historical information, the static answer weight is set to 0.7; when the query leans towards the current state, the real-time data weight is set to 0.7. The fusion result is converted into coherent answer text using natural language generation technology. A sequence-to-sequence generation model is used to ensure the fluency and comprehensibility of the answer. During the generation process, the weight of historical query frequency is updated, and an incremental learning approach is used to adjust the weight parameters with a learning rate of 0.01.

[0057] The specific implementation of step S07 involves optimizing system parameters based on user feedback. The purpose of this step is to continuously improve the system's query response performance. User feedback quality scoring uses a numerical evaluation system of 1 to 5, comprehensively evaluating three dimensions: accuracy, completeness, and timeliness. When the score is below 3, a parameter adjustment mechanism is triggered, using a gradient descent algorithm to update the weight allocation of the three-level knowledge vector cache and the three-level question intent vector cache. The knowledge vector cache update is based on reinforcement learning principles, using user satisfaction as a reward signal to adjust the vector representation, with a learning rate set to 0.001 to ensure stable convergence. The question intent vector cache update employs an online learning strategy, adjusting the intent classification boundary and feature weights in real time. Semantic similarity algorithm parameter optimization uses a Bayesian optimization method, searching for the optimal configuration in the parameter space. The parameter update is confirmed as effective when the average satisfaction score increases by more than 0.2 over 10 consecutive queries. The system also establishes a parameter rollback mechanism, automatically rolling back to the previous stable configuration when new parameters cause a performance degradation exceeding 10%.

[0058] The specific implementation of step S08 involves outputting a joint answer and recording a processing log. This step aims to complete the entire query response process and provide data support for subsequent optimization. The merged static answer from the knowledge base and real-time data from the business system are output to the user as a joint answer. The output format is structured text, including the main answer content, relevant supplementary information, and data source identifiers. Simultaneously, key information from the entire generation process is recorded in the joint answer log of the user's knowledge base and the business system. The log content includes the query timestamp, user identifier, query content, matching path, processing time, data source, and quality assessment results. The quality assessment results are calculated based on multi-dimensional indicators, including answer relevance, information completeness, response time, and user satisfaction, using a weighted average method to generate a comprehensive quality score. The log data uses a structured storage format to support subsequent data analysis and system optimization. The storage period is set to 6 months, and expired data is compressed and archived.

[0059] The detailed structure of the intent recognition optimization model is based on the Transformer architecture, containing a 12-layer encoder structure, with 8 attention heads configured in each layer, and a hidden layer dimension of 512. The model input layer receives the query sequence after word embedding and position encoding. The word embedding dimension is 512, and the position encoding is generated using sine and cosine functions. The multi-head self-attention mechanism uses scaled dot product attention calculation, and the attention weights are normalized using the Softmax function. The feedforward neural network adopts a two-layer fully connected structure, with the intermediate layer having a dimension of 2048, and the activation function using ReLU. Layer normalization is applied to the output of each sub-layer, and residual connections ensure the stability of gradient propagation. The model output layer uses linear transformation and the Softmax function to generate probability distributions for the three intent categories.

[0060] The detailed steps for building the training dataset first involved collecting user historical query logs, covering six months of query records, totaling approximately 8 million raw data entries. Data anonymization included user identity information anonymization, replacement of sensitive business data, and privacy information filtering. Query statements were manually labeled according to three categories: business query intent, knowledge acquisition intent, and mixed query intent. The labeling team consisted of 10 domain experts, with a labeling consistency requirement of over 95%. A training set containing 500,000 labeled samples was constructed. Each sample included the original query text, intent category label, query context information, and corresponding answer quality score. The dataset was divided into training, validation, and test sets in an 8:1:1 ratio to ensure a uniform distribution of samples across all categories. The training process used the AdamW optimizer, with an initial learning rate set to 2×10⁻⁶. -5 The batch size was set to 32, and the training period was 20 epochs. The loss function used was cross-entropy loss, and the learning rate decayed by 10% every 5 epochs. After the model was trained, quantization and compression were performed to compress the weight precision from FP32 to INT8, reducing the model size by 75% and improving the inference speed by 3 times.

[0061] It should be noted that this invention contains three key technical ideas: a three-level vector caching hierarchical matching mechanism, a joint matching degree calculation function, and an intent recognition optimization model based on Transformer.

[0062] The three-level vector caching hierarchical matching mechanism fundamentally solves the technical limitation of traditional single vector spaces in accurately representing semantic information at different levels by constructing three different-dimensional vector spaces for knowledge content and question intent according to their complexity and level of abstraction. Existing technologies typically use vectors of a unified dimension to represent all types of knowledge content, leading to feature confusion and matching interference between specialized domain knowledge and basic common sense knowledge in the same semantic space. This invention, through a hierarchical design of domain knowledge layer, business knowledge layer, and common sense knowledge layer, enables knowledge content at different levels of abstraction to be accurately represented in their respective most suitable vector dimension spaces. Simultaneously, a semantic association matrix establishes mapping relationships between levels, ensuring the accuracy and completeness of cross-level knowledge retrieval. This hierarchical matching mechanism can automatically select the most suitable vector space for semantic matching based on the complexity of the user query, significantly improving the accuracy and response efficiency of knowledge retrieval.

[0063] The joint matching degree calculation function establishes a comprehensive matching evaluation system by integrating multiple dimensions such as initial knowledge matching degree, initial intent matching degree, historical query frequency weight, time decay factor, and business importance coefficient. This fundamentally overcomes the technical shortcomings of traditional technologies that rely solely on a single similarity index for matching judgment. Existing multi-source data fusion methods typically employ simple linear weighting or average calculation methods, which cannot effectively balance the feature differences and importance weights of different data sources. This invention comprehensively calculates each input parameter using a polynomial weighted summation method and applies a sigmoid activation function for normalization. This allows for dynamic adjustment of the influence weights of different parameters, ensuring that the matching results consider semantic similarity while also taking into account user historical preferences, query timeliness, and business data importance. This multi-dimensional fusion calculation mechanism can accurately identify the optimal matching path in complex multi-source data environments, providing a reliable quantitative basis for subsequent data fusion.

[0064] This Transformer-based intent recognition optimization model, through a 12-layer encoder structure and multi-head attention mechanism, combined with dynamic adjustment of long short-term memory matching length, fundamentally solves the technical challenges of traditional intent recognition models in handling complex contextual relationships and dynamically adapting to different query scenarios. Existing intent recognition methods typically employ fixed context windows and static model parameters, often resulting in decreased recognition accuracy and insufficient adaptability when facing complex and ever-changing user query intents in real-world applications. This invention uses an adaptive parameter adjustment function to dynamically adjust model parameters based on query complexity, historical similarity, session duration, and system load, enabling the intent recognition process to automatically optimize the recognition strategy according to the actual application environment. Simultaneously, the model's internal positional encoding mechanism corresponds to the hierarchical structure of a three-level question intent vector cache, ensuring accurate capture and processing of intent information at different levels, significantly improving intent recognition accuracy in complex query scenarios.

[0065] The three key technological approaches described above form an organically unified technical system, generating comprehensive advantages far exceeding the effects of any single technology through synergy. The three-level vector caching hierarchical matching mechanism provides high-quality initial matching data for the joint matching degree calculation function. The output of this function, in turn, provides precise training targets and optimization directions for the intent recognition optimization model. The dynamic adjustment capability of the intent recognition optimization model, in turn, guides the parameter updates and matching strategy optimization of the three-level vector cache. This closed-loop collaborative mechanism enables the entire system to continuously self-optimize and improve during operation, constantly enhancing the accuracy of the fusion and matching between static answers in the knowledge base and real-time data from the business system. Compared to the existing technical architecture where each module operates independently and lacks effective coordination, the collaborative technology system of this invention achieves globally optimal matching results, ensuring that the final generated joint answer possesses both deep knowledge background and accurate real-time business information.

[0066] Specifically, the principle of this invention is as follows: The core principle that enables this invention to solve the problem of insufficient accuracy in matching static answers in knowledge bases with real-time data from business systems lies in establishing a hierarchical vectorized matching mechanism and a dynamic weight adjustment system. The three-level knowledge vector cache divides knowledge content into domain knowledge layer, business knowledge layer, and common sense knowledge layer according to their level of abstraction, using vector representations of 1024 dimensions, 512 dimensions, and 256 dimensions respectively. This accurately captures the semantic features of knowledge at different levels through vector spaces of different dimensions. The three-level question intent vector cache divides user queries into business query intent layer, knowledge acquisition intent layer, and mixed query intent layer according to their complexity, using vector representations of 768 dimensions, 384 dimensions, and 192 dimensions, which can accurately identify the true intent of user queries.

[0067] The joint matching degree calculation function is a key technology for solving the problem of fusion accuracy. This function takes initial knowledge matching degree, initial intent matching degree, historical query frequency weights, time decay factors, and business importance coefficients as input parameters. It then processes these parameters through multinomial weighted summation and sigmoid activation function normalization, outputting a joint matching degree score between 0 and 1. This multi-dimensional fusion calculation method comprehensively considers the semantic similarity, historical preferences, timeliness, and business importance of user queries, ensuring the accuracy of the matching results. The intent recognition optimization model is based on a 12-layer Transformer encoder structure, configured with 8 attention heads and a 512-dimensional hidden layer. It dynamically adjusts the context window by combining long-memory matching length and short-memory matching length, enabling a deep understanding of the complex semantic relationships in user queries.

[0068] The adaptive parameter adjustment function calculates a comprehensive adjustment coefficient based on four dimensions: current query complexity, historical query similarity, user session duration, and system load. Different weighted adjustment functions are used to dynamically adjust system parameters, ensuring optimal matching accuracy across various scenarios. This technical solution is logical because it constructs a complete fusion matching system from four levels: semantic understanding, intent recognition, matching calculation, and dynamic adjustment. Each step addresses specific shortcomings of traditional technologies, forming a systematic solution.

[0069] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0070] The specific implementation of step S01 involves establishing a three-level knowledge vector caching system. This system uses semantic vectorization technology to store user knowledge base content in three layers: domain knowledge, business knowledge, and common sense knowledge. The semantic association matrix is ​​established using a cosine similarity-based calculation method, as shown below:

[0071]

[0072] In the formula, R ij V represents the semantic association strength coefficient between the i-th knowledge node and the j-th knowledge node; i and V j Let V be the vector representations of the i-th and j-th knowledge nodes, respectively; i ||and||V j || represent the magnitudes of the corresponding vectors; α layer β is the hierarchical weight adjustment coefficient. complexity This is a complexity correction coefficient. The parameter is obtained as follows: V i and V j Knowledge text is vectorized using a deep learning encoder, with the vector dimension determined by the knowledge level as 1024, 512, or 256 dimensions; α layer Based on the knowledge hierarchy, the domain knowledge layer is set to 1.0, the business knowledge layer to 0.8, and the common sense knowledge layer to 0.6; β complexity The value is calculated based on knowledge complexity assessment, ranging from 0.1 to 0.5.

[0073] The specific implementation of step S02 involves establishing a three-level question intent vector cache, which vectorizes and stores user historical queries according to three levels: business query intent layer, knowledge acquisition intent layer, and mixed query intent layer. The calculation of the intent conversion probability matrix is ​​based on the Markov chain principle, and is specifically represented as follows:

[0074]

[0075] In the formula, P kl N represents the statistical probability value of a user's intent transitioning from level k to level l; kl γ is the number of historical queries from level k to level l; γ is the Laplace smoothing parameter; δ temporal This is the time decay adjustment factor. The parameter is obtained as follows: N kl This is obtained by statistically analyzing the changes in intent hierarchy between adjacent queries in users' historical query records; γ is set to 1.0 to avoid the zero probability problem; δ temporal Calculated based on the query time interval, the formula is δ temporal =e -λt , where λ is the decay rate parameter with a value of 0.05, and t is the time interval in days.

[0076] The specific implementation of step S03 involves receiving a user's natural language query request and performing parallel retrieval using a semantic similarity algorithm. The initial knowledge matching degree is calculated using a weighted average method, as shown below:

[0077]

[0078] In the formula, S knowledge For initial knowledge matching degree; w i L represents the weight coefficient of the i-th knowledge layer; i This is the set of knowledge nodes in the i-th knowledge layer; To query vector Q and the vector of knowledge node j in the i-th layer Similarity; ∈ noise This is a noise correction term. The initial intent matching score is calculated using a max-pooling strategy, as shown below:

[0079]

[0080] In the formula, S intent I represents the initial intent matching degree. i Let I be the set of intent vectors in the i-th intent layer; i | represents the number of intent vectors in the i-th layer; For query vector Q and the j-th intent vector of the i-th layer The reciprocal of the distance; μ context For context adjustment coefficients. The parameters are obtained as follows: w1 = 0.5, w2 = 0.3, w3 = 0.2, preset according to knowledge importance; ∈ noise The range is from 0.01 to 0.05; μ context The complexity is calculated based on the context of the current query, ranging from 0.8 to 1.2.

[0081] The specific implementation of step S04 involves calling the joint matching degree calculation function, using the initial knowledge matching degree, initial intent matching degree, historical query frequency weight, time decay factor, and business importance coefficient as input parameters. The joint matching degree calculation function employs a polynomial weighted summation method, specifically expressed as follows:

[0082]

[0083] In the formula, F joint The joint matching score is represented by a1, a2, a3, a4, and a5, which are the weighting coefficients of each input parameter. freq Weighted by historical query frequency; D time B is the time decay factor; importance is the business importance coefficient; c is the bias term. The matching path length is calculated using the graph theory shortest path algorithm, as shown below:

[0084]

[0085] In the formula, P length To match the path length; n is the total number of nodes in the path; d(v k v k+1 ) is the neighboring node v k and v k+1 The distance between them; φ penalty This is a path complexity penalty term. The parameter is obtained using: W freq The formula W is calculated by statistically analyzing users' historical query records. freq =log(1+f query ), where f query Query frequency, ranging from 0.1 to 2.0; D time The calculation uses exponential decay, and the formula is D. time =e -0.05·Δt Where Δt is the time interval in hours, ranging from 0.01 to 0.99; B importance The importance level is determined based on the business data, ranging from 1.0 to 5.0; φ penalty The path complexity is set, ranging from 0.1 to 1.0; the weight coefficients are a1 = 0.3, a2 = 0.25, a3 = 0.2, a4 = 0.15, a5 = 0.1, and the bias term is c = -2.5.

[0086] The specific implementation of step S05 involves calculating the optimal memory sequence length using an adaptive parameter adjustment function based on the long memory matching length and short memory matching length output by the intent recognition optimization model. The adaptive parameter adjustment function calculates a comprehensive adjustment coefficient based on four input parameters, as shown below:

[0087]

[0088] In the formula, η adaptive C is the comprehensive adjustment coefficient value. complexity H represents the current query complexity metric. similarity For historical query similarity; T session For the duration of the user session; L system The system load condition is represented by ζ, and the stability adjustment parameter is represented by ζ. The calculation of the optimal sequence length is based on a piecewise function using the comprehensive adjustment coefficient value, as shown below:

[0089]

[0090] In the formula, L optimal To remember the optimal sequence length, the numerical range is 10 to 100 sequences; L base This is the length of the base sequence. The parameter is obtained using the following method: C complexity It is obtained through a comprehensive calculation based on vocabulary size, grammatical structure complexity, and terminology density, and the calculation formula is C. complexity =0.4N words +0.3G grammar +0.3T terms , where N words To normalize the number of words, G grammar Scoring the complexity of grammatical structures, T terms Density of technical terms; H similarity The similarity is calculated by comparing the semantic similarity between the current query and historical queries, using the cosine similarity algorithm; T session This is obtained by counting the duration of the user's current session, in minutes; L system This is obtained by monitoring CPU utilization, memory usage, and concurrent query count, and the calculation formula is L. system =0.5U cpu +0.3U memory +0.2N concurrent U cpu For CPU utilization, U memory N represents memory usage. concurrent To normalize the number of concurrent queries; ζ is set to 0.1; L base Set to 50 entries.

[0091] Step S06 involves fusing the static answers from the knowledge base with real-time data from the business system to generate a final joint answer and return it to the user, while simultaneously updating the historical query frequency weights. Step S07 involves updating the three-level knowledge vector cache and the three-level question intent vector cache based on user feedback quality scores, and optimizing the semantic similarity algorithm parameters to improve subsequent query response efficiency. Step S08 involves outputting the fusion result of the static answers from the knowledge base and the real-time data from the business system as the joint answer, and recording the joint answer generation process and quality assessment results in the user knowledge base and business system joint answer log, thus completing the full process of the user knowledge base and business system joint answer. The specific implementation methods for these three steps are the same as described above and will not be elaborated further here.

[0092] To better understand and implement this invention, a specific application scenario is provided as Example 2: A research team has built an AI platform for the field of intelligent manufacturing. This platform needs to simultaneously handle process knowledge base queries and real-time data queries from the production system. The researchers used the method of this invention to optimize and improve the platform, thereby enhancing the accuracy and efficiency of user query responses.

[0093] Researchers first established a three-level knowledge vector caching system. The domain knowledge layer stores 2000 professional manufacturing process theories, including materials science principles, processing parameters, and quality control standards, each encoded as a 1024-dimensional vector. The business knowledge layer stores 5000 practical operation procedures and application scenarios, covering equipment operation steps, troubleshooting methods, and production planning, each encoded as a 512-dimensional vector. The common sense knowledge layer stores 8000 basic manufacturing common sense facts, including basic terminology definitions, safe operating procedures, and standard unit conversions, each encoded as a 256-dimensional vector. Researchers established mapping relationships between each level using a semantic association matrix, where the association strength coefficient R... ij When the value exceeds 0.6, a direct mapping path is established, and ultimately 15,847 valid association mappings are established.

[0094] In constructing the three-level question intent vector cache, researchers collected 120,000 user query records from the past six months. The business query intent layer contains 45,000 pure business data queries, mainly involving production progress queries, equipment status monitoring, and product quality inspection, with each query encoded as a 768-dimensional vector. The knowledge acquisition intent layer contains 52,000 knowledge content queries, mainly involving process parameter queries, technical standard queries, and problem solution queries, with each query encoded as a 384-dimensional vector. The mixed query intent layer contains 23,000 mixed business knowledge queries, involving combining real-time data analysis of process problems and optimizing production parameters based on historical data, with each query encoded as a 192-dimensional vector. The structures of the three-level knowledge vector cache and the three-level question intent vector cache are as follows: Figure 3 As shown in Table 1, researchers calculated the intention transition probability matrix based on the Markov chain principle.

[0095] Table 1 User query intent conversion probability matrix

[0096] Business query intent layer 0.652 0.238 0.110 Knowledge Acquisition Intent Layer 0.184 0.571 0.245 Hybrid Query Intent Layer 0.297 0.326 0.377

[0097] Researchers selected a typical user query as a test case: a user inquired about the cause and handling solution for an abnormal temperature on the current production line. The system first vectorized the query, generating a 768-dimensional query vector. During parallel retrieval in a three-level knowledge vector cache, the domain knowledge layer matched content related to temperature control theory with a similarity of 0.847; the business knowledge layer matched the temperature anomaly handling process with a similarity of 0.923; and the common sense knowledge layer matched basic knowledge of temperature monitoring with a similarity of 0.756. Based on the weighted average calculation, the initial knowledge matching degree S... knowledge =0.5×0.847+0.3×0.923+0.2×0.756=0.852.

[0098] In the three-level question intent vector cache, the system identifies the query as a mixed query intent type, with an initial intent matching degree S. intent =0.891. Researchers obtained relevant historical query parameters and system status parameters, as shown in Table 2:

[0099] Table 2 Query Processing Related Parameters

[0100] Historical query frequency 15 times Number of similar queries in the past 30 days Time interval 2.5 hours Time since the last similar query Business Importance Level Level 4 Temperature anomalies are a highly important issue. Current CPU utilization 67％ Real-time system load status Memory usage 72％ Real-time system memory usage Concurrent query count 23 Number of queries currently being processed simultaneously

[0101] Based on the joint matching degree calculation function, the researchers calculated the historical query frequency weight W. freq =ln(1+15)=2.773, but since it exceeds the upper limit of 2.0, the value is taken as 2.0. Time decay factor D time =e -0.05×2.5=0.882. Business Importance Coefficient B importance =4.0. System load status L system =0.5×0.67+0.3×0.72+0.2×0.46=0.643. Final joint matching score.

[0102] The intent recognition optimization model outputs a long-memory match length of 90 days and a short-memory match length of 25 entries. Researchers calculated the query complexity metric C. compplexity =0.4×0.78+0.3×0.65+0.3×0.82=0.753, Historical Query Similarity H similarity = 0.867, User session duration T session =12 minutes, normalized to 0.8. The comprehensive adjustment coefficient is calculated using an adaptive parameter adjustment function. Since this value is in the range of 0 to 0.3, the optimal sequence length L is remembered. optimal =0.5×50+0.5×0.106×50=27.65, rounded down to 28.

[0103] The system retrieved the 28 most recent temperature monitoring records from the business system and found that the temperature on production line 3 had been rising continuously over the past 4 hours, reaching a maximum of 85℃, exceeding the normal range of 75℃±3℃. The static answer in the knowledge base provided theoretical analysis and standard handling procedures for the temperature anomaly, while real-time data displayed specific temperature trends and equipment operating status. The system integrated these two data points to generate a complete answer including problem analysis, handling suggestions, and preventative measures. The user gave the answer a satisfaction rating of 4 out of 5, and the system updated the weight parameters of the relevant vector cache accordingly.

[0104] The entire query processing took 1.23 seconds, including 0.34 seconds for vector retrieval, 0.18 seconds for joint matching degree calculation, 0.45 seconds for intent recognition optimization, and 0.26 seconds for answer fusion and generation. The system records the complete processing log in the joint answer log, including the query timestamp, user identifier, matching path length (3 in this case), data source identifier, and quality assessment result.

[0105] Researchers compared and analyzed the performance differences between traditional methods and the method of this invention, as shown in Table 3:

[0106] Table 3. Comparison and analysis of method performance

[0107] Query response accuracy 74.2％ 85.6％ 15.4％ Average response time 2.45 seconds 2.08 seconds 15.1％ User satisfaction rating 3.26 points 3.78 points 16.0％ Knowledge coverage completeness 68.5％ 79.3％ 15.8％ Real-time data utilization 42.1％ 48.7％ 15.7％

[0108] Traditional knowledge base query methods primarily rely on keyword matching and simple similarity calculations, failing to effectively integrate multi-level knowledge and real-time business data. They also lack a deep understanding of user query intent and a learning mechanism for historical behavioral patterns. Users often only receive answers from a single source, lacking contextual understanding and personalized adaptation. This invention achieves hierarchical knowledge management and accurate intent identification by establishing a three-level knowledge vector cache and a three-level question intent vector cache. The joint matching degree calculation function effectively integrates multi-dimensional information, and the adaptive parameter adjustment mechanism dynamically optimizes the processing strategy based on real-time conditions. Compared to traditional methods, this invention achieves significant improvements in key indicators such as query response accuracy, response time, user satisfaction, knowledge coverage completeness, and real-time data utilization, with improvements ranging from 15% to 16%, fully validating the effectiveness and practicality of the invention's technical solution.

[0109] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for jointly answering questions using a user knowledge base and a business system on an AI platform, characterized in that, This process includes constructing a three-level knowledge vector cache and a three-level question intent vector cache. After receiving a user's natural language query request, the system uses a semantic similarity algorithm to perform parallel retrieval to generate initial knowledge matching degree and initial intent matching degree. It then calls a joint matching degree calculation function, taking the initial knowledge matching degree, initial intent matching degree, historical query frequency weight, time decay factor, and business importance coefficient as input parameters to output a joint matching degree score. The system uses an intent recognition optimization model to improve the accuracy of the joint matching degree score and selects candidate answers with the shortest matching path length. Based on the long memory matching length and short memory matching length output by the intent recognition optimization model, the system calculates the optimal memory sequence length through an adaptive parameter adjustment function. It also obtains real-time data from the business system to supplement the answers. Finally, it integrates the static answers from the knowledge base with the real-time data from the business system to generate the final joint answer result and updates the historical query frequency weight. The system updates the three-level knowledge vector cache and the three-level question intent vector cache based on user feedback quality scores. The long memory matching length is specifically the time span during which the system retains the user's historical query context information, ranging from 30 to 180 days, used to identify the user's long-term query patterns and preference characteristics. The short memory matching length is specifically the limit on the number of user query context information entries that the system retains in the current session, ranging from 5 to 50 entries, used to understand the user's immediate intent in the current query. The optimal memory sequence length is specifically the optimal context sequence length for data querying in the business system, calculated by balancing the long memory matching length and the short memory matching length, ranging from 10 to 100 entries, used to obtain real-time data from the business system.

2. The method for jointly answering questions using a user knowledge base and a business system under an AI platform according to claim 1, characterized in that, The three-level knowledge vector cache specifically divides the knowledge base content into three precise levels according to the degree of abstraction and complexity. The domain knowledge layer stores professional domain concepts and theoretical knowledge with a vector dimension range of 1000 to 1100 dimensions. The business knowledge layer stores application scenarios and operational process knowledge with a vector dimension range of 400 to 600 dimensions. The common sense knowledge layer stores basic facts and general rule knowledge with a vector dimension range of 200 to 300 dimensions. The three levels are mapped to each other through a semantic association matrix.

3. The method for jointly answering questions using a user knowledge base and a business system under an AI platform according to claim 2, characterized in that, The three-level question intent vector cache specifically divides user query history into three precise levels according to intent complexity and target clarity. The business query intent layer stores pure business data query intents with a vector dimension range of 700 to 800 dimensions. The knowledge acquisition intent layer stores pure knowledge content query intents with a vector dimension range of 300 to 450 dimensions. The mixed query intent layer stores mixed business knowledge query intents with a vector dimension range of 150 to 250 dimensions. The three levels are associated with each other through an intent conversion probability matrix.

4. The method for jointly answering questions using a user knowledge base and a business system under an AI platform according to claim 3, characterized in that, The joint matching degree calculation function is specifically used to integrate multi-dimensional similarity information and output a comprehensive evaluation index. The inputs include the initial knowledge matching degree, the initial intent matching degree, the weight of historical query frequency, the time decay factor, and the business importance coefficient. The output is a joint matching degree score between 0 and 1. The joint matching degree calculation function comprehensively calculates each input parameter by multinomial weighted summation and applies the sigmoid activation function for normalization.

5. The method for jointly answering questions using a user knowledge base and a business system under an AI platform according to claim 4, characterized in that, The optimal memory sequence length is specifically the best context sequence length for data querying in the business system, calculated by balancing the long memory matching length and the short memory matching length. The value ranges from 10 to 100, and it is used to obtain real-time data from the business system.

6. The method for jointly answering questions using a user knowledge base and a business system under an AI platform according to claim 5, characterized in that, The intent recognition optimization model is specifically a multi-layer attention network based on the Transformer architecture, containing a 12-layer encoder structure, with 8 attention heads configured in each layer, and the hidden layer dimension is set to 512. The context window size of the attention mechanism is dynamically adjusted according to the long memory matching length, and the sequence length parameter is consistent with the short memory matching length. The position encoding mechanism inside the model corresponds to the hierarchical structure of the three-level question intent vector cache.

7. The method for jointly answering questions using a user knowledge base and a business system under an AI platform according to claim 6, characterized in that, The adaptive parameter adjustment function is specifically calculated based on four data points: the current query complexity index, historical query similarity, user session duration, and system load. When the comprehensive adjustment coefficient is between 0 and 0.3, a linearly increasing weight adjustment function is used to adjust the context window size to between 0.5 and 1 times the base value. When the comprehensive adjustment coefficient is between 0.3 and 0.7, a square root weight adjustment function is used to adjust the context window size to between 1 and 1.5 times the base value. When the comprehensive adjustment coefficient is between 0.7 and 1.0, a logarithmic weight adjustment function is used to adjust the context window size to between 1.5 and 2 times the base value.

Images