A large model question and answer device based on a local knowledge base

Abstract

The application relates to the technical field of intelligent management, and particularly discloses a large model question and answer device based on a local knowledge base, which comprises a knowledge base establishing module, a query and retrieval module and a model interaction module.The knowledge base establishing module constructs a local knowledge base based on acquired domain knowledge and dynamically updates the knowledge base.The query and retrieval module fuses a user role intelligent judgment and an intelligent analysis mechanism, and outputs actual retrieval results based on scene-cause double-dimensional analysis.The model interaction module outputs answers according to actual retrieval results, and provides a feedback function to optimize the answers.Through constructing the local knowledge base and dynamically updating the knowledge base, fusing the user role intelligent judgment and the intelligent analysis mechanism, outputting the actual retrieval results, using the target question and answer large model to output answers according to the actual retrieval results, and providing the feedback function to optimize the answers, the user demand individualization adaptation can be effectively realized, the question and answer intelligent level is improved, the device can adapt to a complex and dynamic working environment, and efficient decision support can be provided for enterprises.

Description

Technical Field

[0001] This invention relates to the field of intelligent management technology, and in particular to a large-scale question-and-answer device based on a local knowledge base. Background Technology

[0002] In recent years, with the rapid development of artificial intelligence technology, large-scale models have demonstrated enormous potential in fields such as intelligent question answering due to their powerful language understanding and generation capabilities. Meanwhile, in terms of industrial digital transformation, industrial cloud platforms and industrial internet platforms, by integrating various data resources in industrial production, provide strong support for enterprise production, management, and decision-making. For example, industrial cloud platforms enable cloud storage of knowledge such as process documents and equipment parameters, while industrial internet platforms collect real-time equipment operating data through IoT technology.

[0003] However, when using existing large models for question-and-answer interaction, challenges arise such as the highly specialized and complex nature of domain knowledge, the difficulty for large models to accurately understand and process it, the dynamic changes in industrial production processes, the rapid updates of knowledge, the lag in updating existing knowledge bases, and the diversity of user roles with significant differences in information needs and query intentions among different roles, resulting in a lack of accurate user role judgment capabilities and an inability to provide personalized question-and-answer services.

[0004] Therefore, this invention provides a large-scale question-answering device based on a local knowledge base to address the pain points of industrial cloud platforms and industrial internet platforms in knowledge management and intelligent question answering, effectively realize personalized adaptation to user needs, and improve the level of question-answering intelligence, enabling it to adapt to complex and dynamic industrial environments and provide enterprises with efficient decision support. Summary of the Invention

[0005] This invention provides a large-scale question-answering device based on a local knowledge base. It constructs and dynamically updates a local knowledge base; integrates intelligent user role judgment and intelligent analysis mechanisms; and accurately outputs customized actual search results based on a scenario-causal dual-dimensional analysis. Then, it utilizes a target question-answering model built on artificial intelligence to output answers based on the actual search results and provides feedback to optimize the answers. This effectively achieves personalized adaptation to user needs and improves the level of question-answering intelligence, enabling it to adapt to complex and dynamic industrial environments and provide efficient decision support for enterprises.

[0006] This invention provides a large-scale question-answering device based on a local knowledge base, comprising:

[0007] Knowledge base building module: used to build a local knowledge base based on domain knowledge crawled from multi-source heterogeneous data sources, and to realize dynamic updates of the knowledge base;

[0008] Query and retrieval module: It integrates intelligent judgment and intelligent analysis mechanism of user role, and accurately outputs customized actual search results based on scenario-cause dual-dimensional analysis;

[0009] Model Interaction Module: This module is used by the target question-answering model to output the actual search answer based on the actual search results input, and to provide user feedback to optimize the search answer.

[0010] Preferably, the knowledge base building module includes:

[0011] Data acquisition and cleaning unit: used to automatically extract domain knowledge data from multi-source heterogeneous data sources, and perform data cleaning to obtain target construction data;

[0012] Knowledge base construction unit: used to generate a local knowledge base after performing semantic processing, vectorization processing and storage on the target construction data;

[0013] Knowledge base update unit: This unit is used to analyze the update monitoring status of domain knowledge from various data sources, determine and execute a knowledge base update strategy, and update the local knowledge base.

[0014] Preferably, the knowledge base update unit includes:

[0015] Monitoring subunit: Used to monitor changes in domain knowledge from various data sources in real time using a preset monitoring interface, and to obtain changed knowledge items by using a set identification technology;

[0016] Item filtering sub-unit: used to determine real-time change items and timed change items by analyzing the short and long-term query frequency of change items and assessing content sensitivity;

[0017] Strategy generation subunit: used to summarize and organize all real-time change items, as well as timed change items and their corresponding actual change intervals, to generate knowledge base update strategies;

[0018] Update sub-unit: Used to immediately synchronize real-time changed items to the local knowledge base according to the knowledge base update strategy, triggering the preset incremental update process;

[0019] For items that are changed on a scheduled basis, add them to the task queue according to the actual change interval and synchronize them to the local knowledge base on a regular basis.

[0020] And after the knowledge entries are updated, an integrity check is performed on the knowledge base.

[0021] Preferably, the item filtering subunit includes:

[0022] Frequency Analysis Block: Used to obtain the historical query frequency of each changed knowledge item within a preset short-term sliding window and a preset long-term sliding window, and match the corresponding short-term query frequency level and long-term query frequency level;

[0023] Mark change knowledge entries with high query frequency in either the short-term or long-term query frequency levels as real-time change entries.

[0024] Change knowledge items whose short-term query frequency level and long-term query frequency level are both not high query frequency are considered sensitive selection items.

[0025] Sensitivity Analysis Block: Used to perform sensitivity analysis on sensitive selected items using a content sensitivity assessment mechanism to obtain an item sensitivity score;

[0026] Sensitive selection items whose sensitivity scores exceed the preset sensitivity threshold are marked as items to be changed in real time.

[0027] Sensitive selection items whose sensitivity scores do not exceed the preset sensitivity threshold are marked as items to be changed periodically;

[0028] Interval Analysis Block: This block is used to adjust the baseline change interval by combining historical query frequency with entry sensitivity score, thereby obtaining the actual change interval for the corresponding timed change entry.

[0029] Preferably, the query and retrieval module includes:

[0030] Query unit: Used to perform semantic analysis and processing on the query content of the target user to obtain query vector blocks;

[0031] From the local knowledge base, text vector blocks whose vector similarity to the query vector block exceeds a set vector similarity threshold are used as the initial retrieval vector block output;

[0032] Role Judgment Unit: Used to obtain the historical user query records of the current target user within a preset time period, and extract all query key text features of each round of historical queries and the corresponding answer key text features of historical answers from the historical user query records;

[0033] Analyze the number of similar features of the answer key text features that exceed the set text similarity threshold with each query key text feature, and label the corresponding text similarity as the reference similarity.

[0034] A query-answer roundup matrix is ​​constructed by using the number of similar features of each query key text feature and the mean of all corresponding reference similarities as matrix elements;

[0035] The query-answer polling matrix is ​​calculated to obtain the polling performance vector value;

[0036] By assigning time decay weights to the polling performance vector values ​​of each round of historical queries, and then summing and normalizing them, a comprehensive role judgment value is obtained.

[0037] Based on the comprehensive role judgment value, the query level of the current target user is determined;

[0038] Normal analysis unit: When the target user's query level is normal query, the obtained initial search vector block is output as the actual search result;

[0039] Specialized Analysis Unit: When the target user's query level is a specialized query, it performs causal reasoning and scenario association analysis based on the initial retrieval vector block and captures specialized retrieval vector blocks from the local knowledge base.

[0040] Then, the initial search vector block and all the obtained specialized search vector blocks are output as the actual search results.

[0041] Preferably, the specialized analysis unit includes:

[0042] Scene analysis subunit: used to obtain the relevant application scenarios for each initial retrieval vector block, as well as the corresponding scene association system;

[0043] When there are multiple initial retrieval vector blocks, the initial retrieval vector blocks with scene similarity exceeding a set scene similarity threshold are clustered to obtain a vector block set;

[0044] Obtain and analyze the historical retrieval frequency and historical retrieval correction frequency for each relevant application scenario corresponding to the current vector block set to obtain the retrieval verification score;

[0045] The application scenario with the highest retrieval verification score is taken as the comprehensive application scenario of the current vector block set, and the corresponding scenario association system is taken as the comprehensive scenario association system.

[0046] Causal analysis subunit: used to combine the corresponding query content of each initial retrieval vector block and use the set causal discovery algorithm to determine the corresponding single causal link;

[0047] When a set of vector blocks exists, the relationship attributes of the single causal links of all initially retrieved vector blocks in the same set are compared and adjusted, duplicate links are deleted, and causal links are merged to generate a set of causal links.

[0048] Output sub-unit: Used to capture the corresponding associated vector blocks of each scene-related knowledge system and the comprehensive scene-related knowledge system from the local knowledge base, and output them as professional vector blocks;

[0049] Capture the corresponding associated vector blocks for each single causal link and set of causal links from the local knowledge base, and combine all causal links as a specialized vector block for output.

[0050] Preferably, the causal analysis subunit further includes:

[0051] Behavior Analysis Block: Used to obtain the historical query behavior paths of the current target user, and to determine the reference behavior path by analyzing the fit between the query nodes of each historical query behavior path and all initial retrieval vector blocks in the current vector block set;

[0052] Attribute unification block: Used to unify the representation of a single causal chain within the same set of vector blocks where relational attributes differ but are essentially related;

[0053] Duplicate removal block: Used to perform deduplication on identical single causal chains that exist in the same set of vector blocks, ultimately retaining only one of them;

[0054] Set generation block: Used to combine all the corresponding single causal chains of the vector block set after processing by attribute unification block and duplicate removal block, with reference behavior path, and input to a pre-established link generation model to obtain set causal links.

[0055] Preferably, the module interaction module includes:

[0056] Answer generation unit: This unit generates the actual search answer based on the input search results from the pre-deployed target question-answering model.

[0057] Answer correction unit: Used to provide user feedback functionality, allowing users to mark incorrect answers and trigger model retraining.

[0058] The beneficial effects of this invention compared to existing technologies are as follows: By constructing a local knowledge base and dynamically updating it; integrating intelligent judgment and analysis mechanisms for user roles; accurately outputting customized actual search results based on scenario-causal dual-dimensional analysis; and then using a target question-and-answer model built on artificial intelligence to output answers based on actual search results, as well as providing feedback functions to optimize answers, this invention can effectively achieve personalized adaptation to user needs and improve the level of question-and-answer intelligence, enabling it to adapt to complex and dynamic industrial environments, providing efficient decision support for enterprises, promoting the exploration of artificial intelligence application scenarios, and laying a good foundation for the formation of replicable and scalable demonstration application scenarios.

[0059] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0060] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0062] Figure 1 This is a schematic diagram of a large-scale question-answering device based on a local knowledge base, as described in an embodiment of the present invention. Detailed Implementation

[0063] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention. Example 1:

[0064] This invention provides a large-scale question-answering device based on a local knowledge base, with reference to... Figure 1 ,include:

[0065] Knowledge base building module: used to build a local knowledge base based on domain knowledge crawled from multi-source heterogeneous data sources, and to realize dynamic updates of the knowledge base;

[0066] Query and retrieval module: It integrates intelligent judgment and intelligent analysis mechanism of user role, and accurately outputs customized actual search results based on scenario-cause dual-dimensional analysis;

[0067] Model Interaction Module: This module is used by the target question-answering model to output the actual search answer based on the actual search results input, and to provide user feedback to optimize the search answer.

[0068] In this embodiment, multi-source heterogeneous data sources refer to multiple data sources that can acquire domain knowledge with different data differences (including but not limited to data structures (such as structured, semi-structured, and unstructured) and data formats (such as text, images, videos, JSON, XML, etc.)), such as enterprise document systems, databases, API interfaces, etc.; local knowledge bases refer to a collection of specific domain knowledge built on local storage devices, storing structured or semi-structured knowledge (such as product information stored in tabular form, industry reports stored in document form, etc.), and the stored knowledge has been vectorized for easy access and processing by computers.

[0069] In this embodiment, intelligent user role judgment refers to judging the user's query level; intelligent analysis mechanism refers to triggering professional / general dual-track analysis based on the user's query level, with the professional path introducing scenario clustering verification and dynamic integration of causal chains to adapt to different demand scenarios; scenario-causal dual-dimensional analysis refers to causal reasoning and scenario association analysis; the target question-answering big model refers to the actual search answer generated by the big model question-answering device based on the actual search results input in the local knowledge base; the actual search answer is the final answer generated by the target question-answering big model based on the actual search results input and directly presented to the target user, presented in natural language, clearly and accurately answering the user's query question; the user feedback function is an interaction mechanism provided by the model interaction module for users, allowing target users to evaluate and provide feedback on the generated actual search answer.

[0070] The beneficial effects of the above technologies are as follows: by building a local knowledge base and dynamically updating the knowledge base; by integrating intelligent judgment and intelligent analysis mechanisms for user roles; by accurately outputting customized actual search results based on scenario-causal dual-dimensional analysis; and by using a target question-and-answer model based on artificial intelligence to output answers based on actual search results, as well as providing feedback functions to optimize answers, it is possible to effectively achieve personalized adaptation to user needs and improve the level of question-and-answer intelligence, enabling it to adapt to complex and dynamic industrial environments and provide efficient decision support for enterprises. Example 2:

[0071] This invention provides a large-scale question-answering device based on a local knowledge base, wherein the knowledge base building module includes:

[0072] Data acquisition and cleaning unit: used to automatically extract domain knowledge data from multi-source heterogeneous data sources, and perform data cleaning to obtain target construction data;

[0073] Knowledge base construction unit: used to generate a local knowledge base after performing semantic processing, vectorization processing and storage on the target construction data;

[0074] Knowledge base update unit: This unit is used to analyze the update monitoring status of domain knowledge from various data sources, determine and execute a knowledge base update strategy, and update the local knowledge base.

[0075] In this embodiment, multi-source heterogeneous data sources refer to various data sources capable of acquiring domain knowledge, such as enterprise document systems, industry standard documents, databases, API interfaces, etc.; target construction data refers to the data obtained after cleaning the crawled domain knowledge data (such as using regular expressions and NLP algorithms to clean noisy data); semantic processing specifically refers to using natural language processing (NLP) technology to perform a series of operations on the target construction data to understand and extract the semantic information of the text, such as part-of-speech tagging, named entity recognition, semantic annotation, etc.; vectorization processing refers to converting the semantically processed text fragment data into vector form.

[0076] In this embodiment, the local knowledge base refers to a collection of knowledge in a specific domain built on a local storage device. It stores structured or semi-structured knowledge (such as product information stored in tabular form, industry reports stored in document form, etc.), and the stored knowledge has been vectorized for easy access and processing by computers. For example, there is a knowledge base building process: first, the crawled domain knowledge is processed by text embedding based on natural language processing algorithms, that is, long text is encoded; then, an m×n vector stack is constructed and segmented to ensure the performance of subsequent queries; finally, it is stored in a vector database through vector embedding to form a local knowledge base.

[0077] The purpose of establishing the local knowledge base is to provide a knowledge foundation for the large-scale question-answering device. Specifically, it combines the local knowledge base with the large-scale question-answering capabilities based on retrieval-enhanced generation technology. When a user asks a question, the device can quickly retrieve relevant information from the local knowledge base and generate an accurate answer using the capabilities of the large-scale model. The specific question-answering process is as follows: When a user asks a question, the device first converts the user's question into vector form (the vector conversion process is consistent with the vectorization processing of knowledge in the local knowledge base, such as using word embedding models or encoding methods); then, it performs vector similarity and association analysis in the vector database of the local knowledge base to obtain knowledge retrieval vectors; finally, it inputs the knowledge retrieval vectors into the large-scale question-answering model, which combines the knowledge retrieval vectors to generate an accurate answer and return it to the user.

[0078] In this embodiment, the update monitoring status refers to the results obtained by monitoring the domain knowledge update status of each data source, including whether the data source has been updated, the updated content, and the update frequency, etc.; the knowledge base update strategy is used to clarify which data source updates need to be immediately synchronized to the local knowledge base, which can be synchronized periodically, and the specific time interval for synchronization, etc.

[0079] The beneficial effects of the above technologies are as follows: by automatically extracting and cleaning domain knowledge data from multi-source heterogeneous data sources, high-quality target construction data is obtained, providing a reliable data foundation for the construction of a local knowledge base; a local knowledge base is generated based on semantic processing and vectorization of the target construction data, and dynamic updates of the local knowledge base are achieved by analyzing the domain knowledge update monitoring status of various data sources, which can effectively improve the timeliness and accuracy of the knowledge in the base, and help to generate more accurate and user-friendly answers by combining the powerful generation capabilities of large models. Example 3:

[0080] This invention provides a large-scale question-answering device based on a local knowledge base, wherein the knowledge base updating unit includes:

[0081] Monitoring subunit: Used to monitor changes in domain knowledge from various data sources in real time using a preset monitoring interface, and to obtain changed knowledge items by using a set identification technology;

[0082] Item filtering sub-unit: used to determine real-time change items and timed change items by analyzing the short and long-term query frequency of change items and assessing content sensitivity;

[0083] Strategy generation subunit: used to summarize and organize all real-time change items, as well as timed change items and their corresponding actual change intervals, to generate knowledge base update strategies;

[0084] Update sub-unit: Used to immediately synchronize real-time changed items to the local knowledge base according to the knowledge base update strategy, triggering the preset incremental update process;

[0085] For items that are changed on a scheduled basis, add them to the task queue according to the actual change interval and synchronize them to the local knowledge base on a regular basis.

[0086] And after the knowledge entries are updated, an integrity check is performed on the knowledge base.

[0087] In this embodiment, the preset monitoring interface refers to the interface used to monitor in real time whether there are changes in knowledge content in each data source; the set identification technology refers to the pre-set technology to accurately identify changed knowledge entries from the data source, such as hash verification, timestamp comparison, or version number matching technology; the changed knowledge entry refers to the knowledge content in the data source that has changed, and the change method includes addition, deletion, and modification; the short-term and long-term query frequency refers to the frequency at which the changed knowledge entry is queried in the short term and the long term; and the content sensitivity assessment refers to the process of judging the importance of the content of the changed knowledge entry.

[0088] In this embodiment, real-time change entries refer to change knowledge entries that are determined to need to be updated to the local knowledge base immediately; timed change entries refer to change knowledge entries that need to be updated to the local knowledge base periodically at certain time intervals; and actual change interval refers to the time interval at which timed change entries actually change in the local knowledge base.

[0089] In this embodiment, the knowledge base update strategy is a set of update rules and plans comprehensively formulated based on information such as real-time changed items, timed changed items, and corresponding actual change intervals. It clarifies which items need to be updated immediately, which items need to be updated on a timed basis, and the specific time interval for timed updates. The pre-set incremental update process refers to a pre-defined update process used to handle the updates of changed knowledge items, such as updating only the changed knowledge items. The integrity check is a verification operation performed on the local knowledge base after the knowledge item update is completed. The core check is whether the updated knowledge base is complete and whether there is any data loss, damage, or inconsistency.

[0090] The beneficial effects of the above technologies are as follows: by monitoring changes in domain knowledge from various data sources in real time and accurately acquiring changed knowledge items with the help of set identification technology; by determining whether to change changed knowledge items in real time or at regular intervals, and by adopting different change methods based on the determination results, knowledge update strategies can be generated, which can effectively avoid unnecessary frequent updates, rationally allocate system resources, and improve the overall efficiency of knowledge base updates. Example 4:

[0091] This invention provides a large-scale question-answering device based on a local knowledge base, wherein the item filtering subunit includes:

[0092] Frequency Analysis Block: Used to obtain the historical query frequency of each changed knowledge item within a preset short-term sliding window and a preset long-term sliding window, and match the corresponding short-term query frequency level and long-term query frequency level;

[0093] Mark change knowledge entries with high query frequency in either the short-term or long-term query frequency levels as real-time change entries.

[0094] Change knowledge items whose short-term query frequency level and long-term query frequency level are both not high query frequency are considered sensitive selection items.

[0095] Sensitivity Analysis Block: Used to perform sensitivity analysis on sensitive selected items using a content sensitivity assessment mechanism to obtain an item sensitivity score;

[0096] Sensitive selection items whose sensitivity scores exceed the preset sensitivity threshold are marked as items to be changed in real time.

[0097] Sensitive selection items whose sensitivity scores do not exceed the preset sensitivity threshold are marked as items to be changed periodically;

[0098] Interval Analysis Block: This block is used to adjust the baseline change interval by combining historical query frequency with entry sensitivity score, thereby obtaining the actual change interval for the corresponding timed change entry.

[0099] In this embodiment, the preset short-term sliding window is a pre-set short time interval used to perform short-term statistics and analysis on the historical query frequency of changed knowledge items, such as the most recent week or day; the preset long-term sliding window is a pre-set long time interval used to perform long-term statistics and analysis on the historical query frequency of changed knowledge items, such as the most recent month; the historical query frequency refers to the number of times a certain changed knowledge item has been queried by a user in the past period of time (determined by the preset short-term or long-term sliding window).

[0100] In this embodiment, the short-term query frequency level is the query frequency level selected from a preset short-term frequency level mapping table, using historical query frequencies within a preset short-term sliding window as the matching condition. The short-term query frequency level includes three levels: high, medium, and low. The preset short-term frequency level mapping table consists of the correspondence between different historical query frequency ranges and their corresponding short-term query frequency levels (high, medium, low). For example, within a preset short-term sliding window (such as the most recent week), queries of 0-10 times are considered a low query frequency level, 11-50 times a medium query frequency level, and 51 times or more a high query frequency level. Similarly, the long-term query frequency level is the query frequency level selected from a preset long-term frequency level mapping table, using historical query frequencies within a preset long-term sliding window as the matching condition. The long-term query frequency level includes three levels: high, medium, and low. The preset long-term frequency level mapping table consists of the correspondence between different historical query frequency ranges and their corresponding long-term query frequency levels (high, medium, low). For example, if the preset long-term sliding window is the most recent month, queries of 0-50 times are defined as a low query frequency level, 51-200 times as a medium query frequency level, and so on. 201 queries or more is considered a medium query frequency level, while 201 queries or more is considered a high query frequency level.

[0101] In this embodiment, real-time change items refer to knowledge items that need to be changed or updated immediately. Specifically, they refer to change knowledge items with a high query frequency in either the short-term or long-term query frequency level, as well as sensitive selection items whose sensitivity scores exceed a preset sensitivity threshold. Sensitive selection items refer to change knowledge items whose short-term and long-term query frequency levels are not both high.

[0102] In this embodiment, the content sensitivity assessment mechanism specifically involves: evaluating the importance of selected sensitive items using preset important indicators, normalizing the obtained important indicator values, and then weighting and averaging them to obtain an item sensitivity score. The preset important indicators include the rate of content change, the proportion of sensitive information (such as sensitive information related to personal privacy, trade secrets, and national security), and the number of application areas, etc. The weight values ​​assigned to the indicator values ​​of each preset important indicator are obtained by solving the matrix constructed after pairwise comparison and scoring using the analytic hierarchy process, and the values ​​range from (0, 1).

[0103] In this embodiment, for example, the important evaluation index values ​​of the preset important index of the sensitive selection item 1 are respectively and Then the sensitivity score of sensitive selection item 1 is equal to ,in, This represents the weight value assigned to the indicator value of the important indicator of the j-th preset item.

[0104] In this embodiment, the item sensitivity score is a quantitative value obtained after analyzing the importance of sensitive selected items using a content sensitivity assessment mechanism. The value range is (0, 1), and the higher the score, the more important the item. The preset sensitivity threshold is a pre-set sensitivity score limit value. When the item sensitivity score of a sensitive selected item exceeds this threshold, the item will be marked as a real-time change item. Conversely, if the item sensitivity score does not exceed the threshold, it will be marked as a timed change item. The baseline change interval is a pre-set initial item change time interval. The actual change interval is obtained by adjusting the baseline change interval based on the historical query frequency and the item sensitivity score.

[0105] In this embodiment, for example, the historical query frequency of item 1, which is changed periodically, is 20 times, which is between 11 and 50 times, and is considered a medium query frequency level. The item sensitivity score is 0.6, which is less than the preset sensitivity threshold of 0.8.

[0106] At this point, the formula for calculating the actual change interval of timed change item 1 is as follows:

[0107] ;

[0108] in, This represents the actual change interval for timed change item 1; This is represented as the baseline change interval; This represents the weight of the impact of item sensitivity on the adjustment of the actual change interval, with a value range of (0, 1). The value represents the weight of the influence of query frequency on the actual change interval, with a range of (0, 1); e represents the base of the natural logarithm, with a value of 2.7; the weights assigned to item sensitivity and query frequency are obtained by solving the matrix constructed after pairwise comparison and scoring using the analytic hierarchy process.

[0109] The beneficial effects of the above technologies are as follows: by comprehensively considering the frequency of historical queries and the sensitivity of content, it is possible to accurately determine whether to make real-time or timed changes to knowledge items, which can help avoid unnecessary frequent updates, rationally allocate system resources, and improve question-and-answer efficiency. Example 5:

[0110] This invention provides a large-scale question-answering device based on a local knowledge base, wherein the query and retrieval module includes:

[0111] Query unit: Used to perform semantic analysis and processing on the query content of the target user to obtain query vector blocks;

[0112] From the local knowledge base, text vector blocks whose vector similarity to the query vector block exceeds a set vector similarity threshold are used as the initial retrieval vector block output;

[0113] Role Judgment Unit: Used to obtain the historical user query records of the current target user within a preset time period, and extract all query key text features of each round of historical queries and the corresponding answer key text features of historical answers from the historical user query records;

[0114] Analyze the number of similar features of the answer key text features that exceed the set text similarity threshold with each query key text feature, and label the corresponding text similarity as the reference similarity.

[0115] A query-answer roundup matrix is ​​constructed by using the number of similar features of each query key text feature and the mean of all corresponding reference similarities as matrix elements;

[0116] The query-answer polling matrix is ​​calculated to obtain the polling performance vector value;

[0117] By assigning time decay weights to the polling performance vector values ​​of each round of historical queries, and then summing and normalizing them, a comprehensive role judgment value is obtained.

[0118] Based on the comprehensive role judgment value, the query level of the current target user is determined;

[0119] Normal analysis unit: When the target user's query level is normal query, the obtained initial search vector block is output as the actual search result;

[0120] Specialized Analysis Unit: When the target user's query level is a specialized query, it performs causal reasoning and scenario association analysis based on the initial retrieval vector block and captures specialized retrieval vector blocks from the local knowledge base.

[0121] Then, the initial search vector block and all the obtained specialized search vector blocks are output as the actual search results.

[0122] In this embodiment, the target user refers to the user currently using the large-model question-answering device based on the local knowledge base to make a query, such as managers, technicians, operators, etc.; the query content refers to the query question raised by the target user to the question-answering device; the query vector block refers to the vector representation of the query content after semantic analysis and processing; the vector similarity refers to the degree of similarity between the query vector block and the text vector blocks in the local knowledge base, which is usually calculated using the cosine similarity algorithm; the set vector similarity threshold is a pre-set similarity threshold used to determine whether the query vector block is sufficiently similar to the text vector blocks in the local knowledge base (referring to each text entry stored in the local knowledge base that has been converted into a vector representation after semantic analysis processing) in order to filter out text vector blocks that are sufficiently relevant to the query content; for example, if there is query vector block 1 and text vector block 1, the vector similarity is calculated using the cosine similarity algorithm, and the vector similarity is 0.85, while the currently set vector similarity threshold is 0.8; at this time, text vector block 1 is output as the initial retrieval vector block.

[0123] In this embodiment, the preset time period is a pre-defined time interval used to collect and analyze the target user's historical query records within that time range; historical user query records refer to all query records left by the target user when using the question-and-answer device within the preset time period.

[0124] In this embodiment, the key text features of the query refer to the text features extracted from historical user query records that can represent the core meaning of the query content, including but not limited to keywords, entities, and phrases (phrases with specific meanings or expressing specific intentions); the key text features of the answer refer to the text features extracted from the corresponding historical answers of each round of historical queries that represent the core meaning of the answer, including but not limited to key information and professional terms.

[0125] In this embodiment, for example, if there is a query for "how to optimize the manufacturing process of automobile engines to improve production efficiency", the corresponding key text features of the query include: keywords: automobile engine, manufacturing process, production efficiency; entity: automobile engine; phrases: optimize manufacturing process, improve production efficiency.

[0126] The historical answer was "Optimizing the automotive engine manufacturing process can improve production efficiency by introducing automated equipment, rationally arranging processes, and strengthening employee training." The key textual features of the corresponding answer include: key information: introducing automated equipment, rationally arranging processes, and strengthening employee training; professional terms: automated equipment, process arrangement, and employee training.

[0127] In this embodiment, the number of similar features refers to the number of answer key text features whose text similarity to each query key text feature exceeds a set text similarity threshold. Text similarity refers to the degree of similarity between query key text features and answer key text features, typically calculated using a cosine similarity algorithm. The set text similarity threshold is a pre-defined similarity threshold used to determine whether query key text features and answer key text features are sufficiently similar to filter out answer content that is sufficiently relevant to the query content. Reference similarity refers to the text similarity corresponding to each answer key text feature whose similarity to each query key text feature exceeds the set text similarity threshold.

[0128] In this embodiment, the query-answer round-robin matrix refers to a matrix constructed using the number of similar features of each query key text feature and the mean of all corresponding reference similarities as matrix elements, which is used to reflect the correlation between any query and answer in each round of historical queries within a preset time period; the round-robin performance vector value refers to the feature vector value obtained after performing feature value decomposition on the query-answer round-robin matrix, which is used to measure the strength of the correlation between the query and the answer.

[0129] In this embodiment, the comprehensive role judgment value is obtained by assigning time decay weights (determined using an exponential decay function) to the polling performance vector values ​​of each round of historical queries, summing them up to obtain the initial role judgment value, and then performing normalization processing (using methods such as minimum-maximum normalization and Z-score normalization).

[0130] The initial character judgment value is represented as follows: ;in, This represents the initial character assessment value; This represents the polling performance vector value of the corresponding query-answer polling matrix for the i-th round of historical queries within a preset time period, where i = 1, 2, 3, ..., n; n represents the total number of historical rounds of queries performed by the target user within the preset time period; This represents the time decay weight assigned to the polling performance vector value of the i-th round of historical queries within a preset time period; This represents the time interval between the i-th round of historical queries within a preset time period and the current time.

[0131] In this embodiment, the query level includes two levels: ordinary query and professional query. The professional retrieval vector block refers to the text vector block that the professional analysis unit captures from the local knowledge base after performing causal reasoning and scenario association analysis based on the initial retrieval vector block when the target user's query level is determined to be a professional query. This vector block is highly relevant to the query content, has professional depth, and multi-dimensional technical details. The actual retrieval result refers to the final retrieval result output by the ordinary analysis unit or the professional analysis unit according to the target user's query level. When the query level is ordinary query, the actual retrieval result is the initial retrieval vector block. When the query level is professional query, the actual retrieval result includes the initial retrieval vector block and all professional retrieval vector blocks.

[0132] The beneficial effects of the above technologies are as follows: by judging the query level of target users and adopting different search strategies based on the judgment results, the search needs of different users in different scenarios can be met; for users with professional query levels, through causal reasoning and scenario association analysis, more in-depth information related to the query topic can be captured, thereby providing users with more professional, comprehensive and accurate search results. Example 6:

[0133] This invention provides a large-scale question-answering device based on a local knowledge base, wherein the specialized analysis unit includes:

[0134] Scene analysis subunit: used to obtain the relevant application scenarios for each initial retrieval vector block, as well as the corresponding scene association system;

[0135] When there are multiple initial retrieval vector blocks, the initial retrieval vector blocks with scene similarity exceeding a set scene similarity threshold are clustered to obtain a vector block set;

[0136] Obtain and analyze the historical retrieval frequency and historical retrieval correction frequency for each relevant application scenario corresponding to the current vector block set to obtain the retrieval verification score;

[0137] The application scenario with the highest retrieval verification score is taken as the comprehensive application scenario of the current vector block set, and the corresponding scenario association system is taken as the comprehensive scenario association system.

[0138] Causal analysis subunit: used to combine the corresponding query content of each initial retrieval vector block and use the set causal discovery algorithm to determine the corresponding single causal link;

[0139] When a set of vector blocks exists, the relationship attributes of the single causal links of all initially retrieved vector blocks in the same set are compared and adjusted, duplicate links are deleted, and causal links are merged to generate a set of causal links.

[0140] Output sub-unit: Used to capture the corresponding associated vector blocks of each scene-related knowledge system and the comprehensive scene-related knowledge system from the local knowledge base, and output them as professional vector blocks;

[0141] Capture the corresponding associated vector blocks for each single causal link and set of causal links from the local knowledge base, and combine all causal links as a specialized vector block for output.

[0142] In this embodiment, the relevant application scenario refers to the application scenario in which the initial retrieval vector block is identified and labeled using a scenario classification model in advance, such as equipment maintenance and production process optimization. The steps for establishing the scenario classification model are as follows: First, data from various application scenarios are collected, and the collected data is preprocessed and features are extracted (for example, in the equipment maintenance scenario, extracted features include equipment type, maintenance history, fault records, maintenance costs, etc.; in the production process optimization scenario, extracted features include production links, production efficiency, quality control data, resource utilization, etc.). Then, the model is trained using a preset classification algorithm (such as decision tree, support vector machine, etc.) to establish the scenario classification model. The scenario association system refers to the relationship between various components, links, or elements within the current relevant application scenario. For example, in the equipment maintenance scenario, the scenario association system refers to the association between equipment and maintenance personnel, the association between equipment and maintenance plans, and the association between equipment and fault types.

[0143] In this embodiment, for example, there is a scenario association system for equipment maintenance scenario 1: the association between equipment and maintenance personnel. The corresponding system architecture is equipment-maintenance personnel mapping table-maintenance personnel information acquisition. The maintenance personnel mapping table records the maintenance personnel information corresponding to each equipment, including maintenance personnel ID, name, contact information, etc.

[0144] In this embodiment, the scene similarity threshold is a preset numerical standard used to determine whether the related application scenarios of two initial retrieval vector blocks are sufficiently similar. The vector block set refers to the set of vector blocks formed by clustering multiple initial retrieval vector blocks whose scene similarity is higher than the set scene similarity threshold. The determination of scene similarity is achieved by using cosine similarity (or other similarity algorithms, such as Jaccard similarity, Euclidean distance, etc.) to calculate the similarity of the scene association system between related application scenarios. Specifically, it refers to comparing whether the components, links, or elements in the scene association system of two application scenarios are similar, and whether the association relationship between them is consistent or similar.

[0145] In this embodiment, historical retrieval frequency refers to the number of times a relevant application scenario has been retrieved within a past period; historical retrieval correction frequency refers to the number of times a user has corrected or adjusted the retrieval results for a certain relevant application scenario within a past period; the calculation steps for the retrieval verification score are as follows: first, calculate the ratio of the historical retrieval frequency of the current relevant application scenario to the average historical retrieval frequency of all relevant application scenarios within the same vector block set, i.e., the frequency ratio; then, combine the historical retrieval correction frequency of the current relevant application scenario with an exponential function (the reason for choosing an exponential function is that it can cause the correction decay index to decrease exponentially as the historical retrieval correction frequency increases, more reasonably reflecting the impact of correction frequency on the retrieval verification score) to determine the correction decay index (for example, if there are relevant application scenarios). Historical search correction frequency is The corresponding corrected decay exponent is Finally, the frequency ratios mentioned above are added to the corrected attenuation index, and the result is the retrieval verification score. The comprehensive related application scenario refers to the application scenario with the highest retrieval verification score among all related application scenarios in the vector block set. The comprehensive scenario association system refers to the corresponding scenario association system of the comprehensive related application scenario as the comprehensive scenario association system.

[0146] In this embodiment, the set causal discovery algorithm refers to a pre-set algorithm used to infer the causal structure between variables from the corresponding initial retrieval vector block by combining the query content, such as the PC algorithm; a single causal link refers to a specific causal relationship inferred from the initial retrieval vector block based on the set causal discovery algorithm, such as the decrease in the maintenance frequency of device A leading to a decrease in production efficiency.

[0147] In this embodiment, relational attribute comparison and adjustment refers to comparing the relational attributes between the single causal links of all initially retrieved vector blocks within the same vector block set. Different single causal links may involve different causal subjects, objects, or causal action modes. Set causal links refer to higher-level causal links obtained by integrating and summarizing all single causal links within the same vector block set after attribute unification and duplicate deletion. Related vector blocks refer to system units (such as components, links, or elements) or link nodes of specific causal links (single causal links and set causal links of vector block sets) that are closely related to a specific scenario-related knowledge system (referring to the corresponding scenario-related system of related application scenarios and vector block sets). Professional vector blocks refer to text vector blocks captured from the local knowledge base that are highly relevant to the query content, have professional depth, and multi-dimensional technical details.

[0148] The beneficial effects of the above technologies are as follows: for users with professional query levels, causal reasoning and scenario association analysis can capture more in-depth information related to the query topic, providing users with more professional, comprehensive and accurate search results, thereby improving the level of question answering intelligence and adapting to more complex and varied query needs. Example 7:

[0149] This invention provides a large-scale question-answering device based on a local knowledge base, wherein the causal analysis subunit further includes:

[0150] Behavior Analysis Block: Used to obtain the historical query behavior paths of the current target user, and to determine the reference behavior path by analyzing the fit between the query nodes of each historical query behavior path and all initial retrieval vector blocks in the current vector block set;

[0151] Attribute unification block: Used to unify the representation of a single causal chain within the same set of vector blocks where relational attributes differ but are essentially related;

[0152] Duplicate removal block: Used to perform deduplication on identical single causal chains that exist in the same set of vector blocks, ultimately retaining only one of them;

[0153] Set generation block: Used to combine all the corresponding single causal chains of the vector block set after processing by attribute unification block and duplicate removal block, with reference behavior path, and input to a pre-established link generation model to obtain set causal links.

[0154] In this embodiment, the historical query behavior path refers to the path formed by a series of query operations performed by the target user in chronological order over a period of time. It connects the continuous query behaviors of each round of historical queries by the target user in chronological order to construct the query behavior path. The query behavior path can be represented as a sequence of query nodes, each node representing a specific query. A query node refers to each independent query operation in the historical query behavior path, representing a specific query intent or need of the user. For example, there might be a historical query behavior path like "Equipment Fault Alarm Query" → "Equipment Maintenance Record Query" → "Spare Parts Inventory Status Query" → "Maintenance Team Dispatch Query"; the corresponding query nodes are "Equipment Fault Alarm," "Equipment Maintenance Record Query," "Spare Parts Inventory Status," and "Maintenance Team Dispatch."

[0155] In this embodiment, the steps for determining the fit between the query node and the initial retrieval vector block are as follows: First, perform semantic analysis on the query node to extract key features (such as keywords, entity names, and behavioral intent); then, calculate the semantic similarity between the key features and the initial retrieval vector block (for example, use the cosine similarity algorithm to calculate the semantic similarity, with a value range of (0, 1)); next, if the semantic similarity between the query node and an initial retrieval vector block exceeds a set node similarity threshold (a pre-set value (such as 0.8) used to determine the matching degree between the query node and the vector block), then the query node is considered to be compatible with the initial retrieval vector block; otherwise, it is considered not compatible.

[0156] In this embodiment, the reference behavior path refers to the behavior path with the highest path fit score (most relevant or representative to the current query) selected from the target user's historical query behavior paths by analyzing the fit between the query nodes of each historical query behavior path and all initial retrieval vector blocks in the current vector block set. The path fit score is obtained by directly multiplying the number of initial retrieval vector blocks fitted to each query node within the path by the average semantic similarity (the average semantic similarity of all corresponding initial retrieval vector blocks fitted to the query node), and then summing the results. For example, in historical query behavior path 1, each query node... and The number of matching vector blocks with the initially retrieved vector block in the current vector block set are 3, 5, 5, and 4 respectively; the corresponding average semantic similarity scores are 0.81, 0.9, 0.88, and 0.82 respectively; at this time, the path matching score of historical query behavior path 1 is 3. 0.81+5 0.9+5 0.88+4 0.82 = 14.61.

[0157] In this embodiment, relational attributes refer to features that describe the relationship between two or more entities, such as cause-effect attributes; unified expression means unifying the description of single causal chains within the same vector block set that have different relational attributes but are essentially the same; for example, if there are two single causal links that are "factor A causes result B" and "factor A triggers result B", they can be unified into the expression "factor A causes / triggers result B".

[0158] In this embodiment, the steps for establishing the link generation model are as follows: First, a large amount of causal chain data is collected from the local knowledge base or historical query logs (covering various scenarios and needs that the target user may query, and the collected causal chain data is cleaned (noise, duplicates and irrelevant information are removed to ensure data quality and accuracy), segmented, vectorized, and labeled and classified (e.g., divided according to relationship type (such as cause-effect, condition-conclusion, etc.)); then, key features (such as entities, relationships, context, etc.) are extracted from each causal chain, and then the extracted features are converted into vector representations using an embedding model (such as Word2Vec, BERT, etc.); then, a suitable similarity measurement method (such as cosine similarity, Jaccard similarity, etc.) is selected to calculate the similarity between causal chains, and a similarity matrix is ​​constructed, and the selected clustering algorithm (such as K-means, hierarchical clustering, DBSCAN, etc.) is used to cluster similar causal chains; finally, the clustering results are used to train a neural network.

[0159] In this embodiment, the set causal link refers to a higher-level causal link obtained by integrating and summarizing all single causal links within the same vector block set after attribute unification and duplicate removal processing, reflecting the common characteristics and patterns among multiple single causal links.

[0160] The beneficial effects of the above technologies are as follows: by analyzing users' historical query behavior preferences, and by comparing and adjusting the relational attributes of single causal chains in the same vector block set, deleting duplicate links, and merging causal links, the readability and understandability of causal analysis can be improved, and interference from redundant information can be avoided, thereby enhancing the intelligence and personalization of question answering. Example 8:

[0161] This invention provides a large-scale question-answering device based on a local knowledge base, wherein the module interaction module includes:

[0162] Answer generation unit: This unit generates the actual search answer based on the input search results from the pre-deployed target question-answering model.

[0163] Answer correction unit: Used to provide user feedback functionality, allowing users to mark incorrect answers and trigger model retraining.

[0164] In this embodiment, the target question-answering large model refers to a large language model built based on pre-training of massive text data and combined with deep learning technology. It has learned rich language knowledge, semantic understanding and generation capabilities, and is used to generate actual search answers in the large model question-answering device based on the input actual search results. For example, it is a Transformer architecture model after being fine-tuned with data from a specific domain. The actual search results are a set of information related to the target user's query obtained by the query retrieval module after a series of processing.

[0165] In this embodiment, the actual search answer is the final answer generated by the target question-answering model based on the actual search results input, and directly presented to the target user. It is presented in natural language, clearly and accurately answering the user's query question. The user feedback function is an interactive mechanism provided by the model interaction module, allowing the target user to evaluate and provide feedback on the generated actual search answer. Specifically, when the user believes that the answer is incorrect, inaccurate, or incomplete, they can mark the answer as incorrect through this function. Model retraining is the process of retraining the target question-answering model under the trigger of the user feedback function. That is, when the target user marks the answer as incorrect and triggers model retraining, the question-answering device will collect user feedback information, as well as related query records, actual search results, and incorrect answers, as new training samples. Then, these training samples are used to fine-tune or retrain the target question-answering model to optimize the model's parameters and performance.

[0166] The beneficial effects of the above technologies are as follows: by utilizing the target question-answering model to generate accurate and clear actual search answers based on the actual search results, it effectively ensures that user queries can receive high-quality responses and improves the accuracy of question-answering; by introducing a user feedback function, users can evaluate and provide feedback on the answers, and when an answer is marked as incorrect, the model is retrained, which can effectively achieve continuous optimization of model performance, help reduce the occurrence of incorrect answers, and further improve the reliability of question-answering.

[0167] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A large-scale question-answering device based on a local knowledge base, characterized in that, include: Knowledge base building module: used to build a local knowledge base based on domain knowledge crawled from multi-source heterogeneous data sources, and to realize dynamic updates of the knowledge base; Query and retrieval module: It integrates intelligent judgment and intelligent analysis mechanism of user role, and accurately outputs customized actual search results based on scenario-cause dual-dimensional analysis; Model Interaction Module: This module is used by the target question-answering model to output the actual search answer based on the actual search results input, and to provide user feedback to optimize the search answer. The query and retrieval module includes: Query unit: Used to perform semantic analysis and processing on the query content of the target user to obtain query vector blocks; From the local knowledge base, text vector blocks whose vector similarity to the query vector block exceeds a set vector similarity threshold are used as the initial retrieval vector block output; Role Judgment Unit: Used to obtain the historical user query records of the current target user within a preset time period, and extract all query key text features of each round of historical queries and the corresponding answer key text features of historical answers from the historical user query records; Analyze the number of similar features of the answer key text features that exceed the set text similarity threshold with each query key text feature, and label the corresponding text similarity as the reference similarity. A query-answer roundup matrix is ​​constructed by using the number of similar features of each query key text feature and the mean of all corresponding reference similarities as matrix elements; The query-answer polling matrix is ​​calculated to obtain the polling performance vector value; By assigning time decay weights to the polling performance vector values ​​of each round of historical queries, and then summing and normalizing them, a comprehensive role judgment value is obtained. Based on the comprehensive role judgment value, the query level of the current target user is determined; Normal analysis unit: When the target user's query level is normal query, the obtained initial search vector block is output as the actual search result; Specialized Analysis Unit: When the target user's query level is a specialized query, it performs causal reasoning and scenario association analysis based on the initial retrieval vector block and captures specialized retrieval vector blocks from the local knowledge base. Then, the initial search vector block and all the obtained specialized search vector blocks are output as the actual search results.

2. The large-scale question-answering device based on a local knowledge base according to claim 1, characterized in that, The knowledge base building module includes: Data acquisition and cleaning unit: used to automatically extract domain knowledge data from multi-source heterogeneous data sources, and perform data cleaning to obtain target construction data; Knowledge base construction unit: used to generate a local knowledge base after performing semantic processing, vectorization processing and storage on the target construction data; Knowledge base update unit: This unit is used to analyze the update monitoring status of domain knowledge from various data sources, determine and execute a knowledge base update strategy, and update the local knowledge base.

3. The large-scale question-answering device based on a local knowledge base according to claim 2, characterized in that, The knowledge base update unit includes: Monitoring subunit: Used to monitor changes in domain knowledge from various data sources in real time using a preset monitoring interface, and to obtain changed knowledge items by using a set identification technology; Item filtering sub-unit: used to determine real-time change items and timed change items by analyzing the short and long-term query frequency of change items and assessing content sensitivity; Strategy generation subunit: used to summarize and organize all real-time change items, as well as timed change items and their corresponding actual change intervals, to generate knowledge base update strategies; Update sub-unit: Used to immediately synchronize real-time changed items to the local knowledge base according to the knowledge base update strategy, triggering the preset incremental update process; For items that are changed on a scheduled basis, add them to the task queue according to the actual change interval and synchronize them to the local knowledge base on a regular basis. And after the knowledge entries are updated, an integrity check is performed on the knowledge base.

4. The large-scale question-answering device based on a local knowledge base according to claim 3, characterized in that, The item filtering subunit includes: Frequency Analysis Block: Used to obtain the historical query frequency of each changed knowledge item within a preset short-term sliding window and a preset long-term sliding window, and match the corresponding short-term query frequency level and long-term query frequency level; Mark change knowledge entries with high query frequency in either the short-term or long-term query frequency levels as real-time change entries. Change knowledge items whose short-term query frequency level and long-term query frequency level are both not high query frequency are considered sensitive selection items. Sensitivity Analysis Block: Used to perform sensitivity analysis on sensitive selected items using a content sensitivity assessment mechanism to obtain an item sensitivity score; Sensitive selection items whose sensitivity scores exceed the preset sensitivity threshold are marked as items to be changed in real time. Sensitive selection items whose sensitivity scores do not exceed the preset sensitivity threshold are marked as items to be changed periodically; Interval Analysis Block: This block is used to adjust the baseline change interval by combining historical query frequency with entry sensitivity score, thereby obtaining the actual change interval for the corresponding timed change entry.

5. The large-scale question-answering device based on a local knowledge base according to claim 1, characterized in that, The specialized analysis unit includes: Scene analysis subunit: used to obtain the relevant application scenarios for each initial retrieval vector block, as well as the corresponding scene association system; When there are multiple initial retrieval vector blocks, the initial retrieval vector blocks with scene similarity exceeding a set scene similarity threshold are clustered to obtain a vector block set; Obtain and analyze the historical retrieval frequency and historical retrieval correction frequency for each relevant application scenario corresponding to the current vector block set to obtain the retrieval verification score; The application scenario with the highest retrieval verification score is taken as the comprehensive application scenario of the current vector block set, and the corresponding scenario association system is taken as the comprehensive scenario association system. Causal analysis subunit: used to combine the corresponding query content of each initial retrieval vector block and use the set causal discovery algorithm to determine the corresponding single causal link; When a set of vector blocks exists, the relationship attributes of the single causal links of all initially retrieved vector blocks in the same set are compared and adjusted, duplicate links are deleted, and causal links are merged to generate a set of causal links. Output sub-unit: Used to capture the corresponding associated vector blocks of each scene-related knowledge system and the comprehensive scene-related knowledge system from the local knowledge base, and output them as professional vector blocks; Capture the corresponding associated vector blocks for each single causal link and set of causal links from the local knowledge base, and combine all causal links as a specialized vector block for output.

6. The large-scale question-answering device based on a local knowledge base according to claim 5, characterized in that, The causal analysis subunit further includes: Behavior Analysis Block: Used to obtain the historical query behavior paths of the current target user, and to determine the reference behavior path by analyzing the fit between the query nodes of each historical query behavior path and all initial retrieval vector blocks in the current vector block set; Attribute unification block: Used to unify the representation of a single causal chain within the same set of vector blocks where relational attributes differ but are essentially related; Duplicate removal block: Used to perform deduplication on identical single causal chains that exist in the same set of vector blocks, ultimately retaining only one of them; Set generation block: Used to combine all the corresponding single causal chains of the vector block set after processing by attribute unification block and duplicate removal block, with reference behavior path, and input to a pre-established link generation model to obtain set causal links.

7. The large-scale question-answering device based on a local knowledge base according to claim 1, characterized in that, The model interaction module includes: Answer generation unit: This unit generates the actual search answer based on the input search results from the pre-deployed target question-answering model. Answer correction unit: Used to provide user feedback functionality, allowing users to mark incorrect answers and trigger model retraining.

Images