Large model question answering device based on local knowledge base

Abstract

The invention 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 establishment module for establishing the local knowledge base based on acquired domain knowledge and dynamically updating the knowledge base; the query retrieval module is used for fusing a user role intelligent judgment mechanism and an intelligent analysis mechanism and outputting an actual retrieval result based on scene-causal two-dimensional analysis; according to the model interaction module, the target question and answer large model outputs answers according to the actual retrieval result and provides a feedback function to optimize the answers. By constructing a local knowledge base and dynamically updating the knowledge base, fusing a user role intelligent judgment and intelligent analysis mechanism, outputting an actual retrieval result, outputting an answer according to the actual retrieval result by utilizing a target question and answer large model, and providing a feedback function to optimize the answer, personalized adaptation of user requirements can be effectively realized, and the user experience is improved. And the intelligent level of questions and answers is improved, so that the system can adapt to a complex and dynamic working environment, and efficient decision support is provided for enterprises.

Description

Technical Field

[0001] The present invention relates to the field of intelligent management technology, and in particular to a large-model question-answering device based on a local knowledge base. Background Art

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

[0003] However, when applying existing large models for question-and-answer interactions, we face difficulties such as highly specialized and complex domain knowledge, which makes it difficult for large models to accurately understand and process it; dynamic changes in industrial production processes, rapid knowledge updates, and delayed updates of existing knowledge bases; and diverse user roles, with different roles having very different information needs and query intentions. The system lacks the ability to accurately judge user roles and is unable to provide personalized question-and-answer services.

[0004] Therefore, the present invention provides a large-model question-and-answer device based on a local knowledge base to solve the pain points of industrial cloud platforms and industrial Internet platforms in knowledge management and intelligent question-and-answering, effectively realize personalized adaptation of user needs, and improve the intelligence level of question-and-answering, so that it can adapt to complex and dynamic industrial environments and provide enterprises with efficient decision-making support. Summary of the Invention

[0005] The present invention provides a large-model question-and-answer device based on a local knowledge base, which is used to construct a local knowledge base and dynamically update the knowledge base; integrate user role intelligent judgment and intelligent analysis mechanism, accurately output customized actual retrieval results based on scenario-cause-effect two-dimensional analysis, and then use the target question-and-answer large model established based on artificial intelligence to output answers according to the actual retrieval results, and provide feedback function to optimize the answers, which can effectively realize personalized adaptation of user needs and improve the level of question-and-answer intelligence, so that it can adapt to complex and dynamic industrial environments and provide enterprises with efficient decision-making support.

[0006] The present invention provides a large model question-answering device based on a local knowledge base, comprising: Knowledge base building module: used to build a local knowledge base based on domain knowledge captured from multi-source heterogeneous data sources and realize dynamic update of the knowledge base; Query and retrieval module: used to integrate intelligent judgment of user roles and intelligent analysis mechanisms, based on scenario-cause-effect dual-dimensional analysis, to accurately output customized actual search results; Model interaction module: used by the target question-answering model to output the actual retrieval answer based on the actual retrieval results input, and to provide user feedback function to optimize the retrieval answer.

[0007] Preferably, the knowledge base building module includes: Data acquisition and cleaning unit: used to automatically capture domain knowledge data from multiple heterogeneous data sources and perform data cleaning to obtain target construction data; Knowledge base construction unit: used for performing semantic processing, vectorization processing and storage on the target construction data to generate a local knowledge base; Knowledge base updating unit: used to analyze the update monitoring status of the domain knowledge of each data source, determine the knowledge base update strategy and execute it, and update the local knowledge base.

[0008] Preferably, the knowledge base updating unit includes: Monitoring subunit: used to monitor domain knowledge changes of various data sources in real time using preset monitoring interfaces, and to obtain changed knowledge entries by adopting set recognition technology; Item screening subunit: used to determine real-time change items and scheduled change items by performing short-term and long-term query frequency analysis and content sensitivity assessment on the changed knowledge items; Strategy generation subunit: used to summarize and organize all real-time change items, scheduled change items and corresponding actual change intervals, and generate knowledge base update strategies; Update subunit: used to immediately synchronize real-time changed entries to the local knowledge base according to the knowledge base update strategy, triggering the preset incremental update process; For scheduled change items, add them to the task queue according to the actual change interval and synchronize them to the local knowledge base regularly; And when the knowledge entry is updated, the knowledge base is checked for integrity.

[0009] Preferably, the entry screening subunit includes: Frequency analysis block: used to obtain the historical query frequency of each changed knowledge item in the preset short-term sliding window and the preset long-term sliding window, and match the corresponding short-term query frequency level and long-term query frequency level; Marking the changed knowledge items with a high query frequency in the short-term query frequency level or the long-term query frequency level as real-time changed items; The changed knowledge items whose short-term query frequency level and long-term query frequency level are not high query frequency are regarded as sensitive selection items; Sensitivity analysis block: used to perform sensitivity analysis on sensitive selection items using the content sensitivity assessment mechanism to obtain the item sensitivity score; Mark sensitive selection items whose item sensitivity scores exceed the preset sensitivity threshold as real-time change items; Mark sensitive selection items whose item sensitivity scores do not exceed the preset sensitivity threshold as timed change items; Interval analysis block: used to adjust the baseline change interval by combining the historical query frequency with the item sensitivity score to obtain the actual change interval of the corresponding scheduled change item.

[0010] Preferably, the query retrieval module includes: Query unit: used to perform semantic analysis and processing on the target user's query content to obtain query vector blocks; From the local knowledge base, the text vector blocks whose vector similarity with the query vector block exceeds the set vector similarity threshold are matched and output as the initial search vector blocks; 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 key text features of the corresponding historical answers from the historical user query records; Analyze the number of similar features of the answer key text features whose text similarity with each query key text feature exceeds a set text similarity threshold, and mark the corresponding text similarity as a reference similarity; The query-answer polling matrix is established by taking the number of similar features of each query key text feature and the mean of all corresponding reference similarities as matrix elements; Calculating the query-answer polling matrix to obtain a polling performance vector value; By assigning a time-decay weight to the polling performance vector value of each round of historical queries, summing and normalizing them, we can obtain the comprehensive role judgment value. Determine the query level of the current target user based on the comprehensive role judgment value; Common analysis unit: used to output the obtained initial search vector block as the actual search result when the target user's query level is common query; Professional analysis unit: When the target user's query level is professional, it performs causal reasoning and scenario association analysis based on the initial retrieval vector block and captures the professional retrieval vector block from the local knowledge base; Then the initial search vector block and all obtained professional search vector blocks are output as actual search results.

[0011] Preferably, the professional analysis unit includes: Scenario analysis subunit: used to obtain the relevant application scenarios of each initial search vector block and the corresponding scenario association system; When there are multiple initial search vector blocks, the initial search vector blocks whose scene similarity is higher than the set scene similarity threshold are clustered to obtain a vector block set; Obtain and analyze the historical search frequency and historical search correction frequency of each relevant application scenario corresponding to the current vector block set to obtain the search verification score; The relevant application scenario with the highest retrieval verification score is used as the comprehensive relevant application scenario of the current vector block set, and the corresponding scenario association system is used as the comprehensive scenario association system; Causal analysis subunit: used to combine the corresponding query content of each initial search vector block and use the set causal discovery algorithm to determine the corresponding single causal link; When there is a vector block set, the single causal links of all the initially retrieved vector blocks in the same vector block set are compared and adjusted in terms of relational attributes, duplicate links are deleted, and causal links are merged to generate a set causal link. Output subunit: used to capture the corresponding association 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; The corresponding association vector blocks of each single causal link and collective causal link are captured from the local knowledge base, and all causal links are combined and output as a professional vector block.

[0012] Preferably, the cause-effect analysis subunit further includes: Behavior analysis block: used to obtain the historical query behavior path of the current target user and determine the reference behavior path by analyzing the degree of fit between the query node of each historical query behavior path and all the initial search vector blocks in the current vector block set; Attribute Unification Block: It is used to unify the expression form of a single causal chain with different relationship attributes but related essence within the same vector block set; Deduplication block: It is used to perform deduplication operations on identical single causal chains existing in the same vector block set, and finally retain only one of them; Set generation block: It is used to combine all the corresponding single causal chains of the vector block set processed by the attribute unification block and the duplicate removal block with the reference behavior path, input them into the pre-established link generation model, and obtain the set causal links.

[0013] Preferably, the module interaction module includes: Answer generation unit: used to generate actual retrieval answers based on the input actual retrieval results of the pre-deployed target question-answering model; Answer correction unit: used to provide user feedback function. Target users can mark incorrect answers and trigger model retraining.

[0014] Compared with the existing technology, the beneficial effects of the present invention are as follows: by constructing a local knowledge base and dynamically updating the knowledge base; integrating user role intelligent judgment and intelligent analysis mechanism, accurately outputting customized actual search results based on scenario-cause-effect two-dimensional analysis, and then using the target question-answering large model established based on artificial intelligence to output answers according to the actual search results, and providing feedback function to optimize the answers, it can effectively realize personalized adaptation of user needs and improve the level of question-answering intelligence, so that it can adapt to complex and dynamic industrial environments, provide enterprises with efficient decision-making support, promote the exploration of artificial intelligence application scenarios, and lay a good foundation for the formation of replicable and popularizable demonstration application scenarios.

[0015] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures specifically pointed out in this application document.

[0016] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 Schematic diagram of a large-model question-answering device based on a local knowledge base in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention. Embodiment 1:

[0019] The present invention provides a large model question answering device based on a local knowledge base, referring to Figure 1 ,include: Knowledge base building module: used to build a local knowledge base based on domain knowledge captured from multi-source heterogeneous data sources and realize dynamic update of the knowledge base; Query and retrieval module: used to integrate intelligent judgment of user roles and intelligent analysis mechanisms, based on scenario-cause-effect dual-dimensional analysis, to accurately output customized actual search results; Model interaction module: used by the target question-answering model to output the actual retrieval answer based on the actual retrieval results input, and to provide user feedback function to optimize the retrieval answer.

[0020] In this embodiment, multi-source heterogeneous data sources refer to multiple data sources that can obtain domain knowledge with different data differences (including but not limited to data structures (such as structured, semi-structured, unstructured), data formats (such as text, images, videos, JSON, XML, etc.)), such as enterprise document systems, databases, API interfaces, etc.; local knowledge base refers to a collection of specific domain knowledge established on a local storage device, which stores structured or semi-structured knowledge (such as product information stored in table form, industry reports stored in document form, etc.), and the stored knowledge has been vectorized to facilitate rapid computer access and processing.

[0021] In this embodiment, intelligent judgment of user role refers to the judgment of the user's query level; the intelligent analysis mechanism refers to the combination of user query level to trigger professional / general dual-track analysis, and the professional path introduces scenario clustering verification and dynamic integration of causal chains to adapt to different demand scenarios; scenario-causal two-dimensional analysis refers to causal reasoning and scenario correlation analysis; the target question and answer big model refers to the actual retrieval answer generated by the big model question and answer device based on the local knowledge base according to the actual retrieval results input; the actual retrieval answer is the final answer generated by the target question and answer big model according to the actual retrieval results input and presented directly to the target user. It is presented in the form of natural language and clearly and accurately answers the user's query question; the user feedback function is an interactive mechanism provided by the model interaction module for users, allowing target users to evaluate and provide feedback on the generated actual retrieval answer.

[0022] The beneficial effects of the above technologies are: by building a local knowledge base and dynamically updating the knowledge base; integrating intelligent judgment of user roles and intelligent analysis mechanisms, accurately outputting customized actual search results based on scenario-cause-effect two-dimensional analysis, and then using the target question-answering large model established based on artificial intelligence to output answers according to actual search results, and providing feedback functions to optimize answers, it can effectively achieve personalized adaptation to user needs and improve the level of question-answering intelligence, so that it can adapt to complex and dynamic industrial environments and provide enterprises with efficient decision-making support. Example 2:

[0023] The present invention provides a large-model question-answering device based on a local knowledge base, wherein the knowledge base establishment module includes: Data acquisition and cleaning unit: used to automatically capture domain knowledge data from multiple heterogeneous data sources and perform data cleaning to obtain target construction data; Knowledge base construction unit: used for performing semantic processing, vectorization processing and storage on the target construction data to generate a local knowledge base; Knowledge base updating unit: used to analyze the update monitoring status of the domain knowledge of each data source, determine the knowledge base update strategy and execute it, and update the local knowledge base.

[0024] In this embodiment, multi-source heterogeneous data sources refer to various data sources that can obtain domain knowledge, such as enterprise document systems, industry standard files, databases, API interfaces, etc.; target construction data refers to the data obtained after data cleaning (such as regular expressions, NLP algorithms to clean noise data) of the captured domain knowledge data; semantic processing specifically refers to the use of 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 tagging, etc.; vectorization processing refers to the conversion of semantically processed text fragment data into vector form.

[0025] In this embodiment, the local knowledge base refers to a collection of specific domain knowledge established on a local storage device, storing structured or semi-structured knowledge (such as product information stored in table form, industry reports stored in document form, etc.), and the stored knowledge has been vectorized to facilitate rapid computer access and processing. For example, there is a knowledge base establishment process: first, text embedding processing is performed on the captured domain knowledge based on a natural language processing algorithm, that is, long text is encoded. Then, an m×n vector stack is constructed accordingly and sharding is performed to ensure subsequent query performance; finally, the knowledge is stored in a vector database through vector embedding to form a local knowledge base; Among them, the purpose of establishing a local knowledge base is to provide a knowledge basis for the large-model question-answering device, that is, to combine the local knowledge base and the large-model question-answering capabilities based on retrieval enhancement generation technology. When the user asks a question, the question-answering device can quickly retrieve relevant information from the local knowledge base and generate accurate answers in combination with the capabilities of the large model. The specific question-answering process is: when the user asks a question, the question-answering device first converts the user's question into a vector form (the vector conversion process is consistent with the vectorization processing method of the knowledge in the local knowledge base, such as using a word embedding model or encoding method); then, vector similarity analysis and association analysis are performed in the vector database of the local knowledge base to obtain a knowledge retrieval vector; finally, the knowledge retrieval vector is input into the question-answering large model, and the question-answering large model combines the knowledge retrieval vector to generate an accurate answer and return it to the user.

[0026] 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 update content, the update frequency, etc.; the knowledge base update strategy is used to clarify which data source updates need to be synchronized to the local knowledge base immediately, which can be synchronized regularly, and the specific time interval for synchronization, etc.

[0027] The beneficial effects of the above technologies are: by automatically capturing 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 the local knowledge base; based on the semantic processing and vectorization processing of the target construction data, the local knowledge base is generated, and by analyzing the domain knowledge update monitoring status of each data source, the local knowledge base is dynamically updated, which can effectively improve the timeliness and accuracy of the knowledge in the library, and help to combine the powerful generation capabilities of the large model in the future to generate more accurate and more user-friendly answers. Example 3:

[0028] The present invention provides a large model question-answering device based on a local knowledge base, wherein the knowledge base updating unit includes: Monitoring subunit: used to monitor domain knowledge changes of various data sources in real time using preset monitoring interfaces, and to obtain changed knowledge entries by adopting set recognition technology; Item screening subunit: used to determine real-time change items and scheduled change items by performing short-term and long-term query frequency analysis and content sensitivity assessment on the changed knowledge items; Strategy generation subunit: used to summarize and organize all real-time change items, scheduled change items and corresponding actual change intervals, and generate knowledge base update strategies; Update subunit: used to immediately synchronize real-time changed entries to the local knowledge base according to the knowledge base update strategy, triggering the preset incremental update process; For scheduled change items, add them to the task queue according to the actual change interval and synchronize them to the local knowledge base regularly; And when the knowledge entry is updated, the knowledge base is checked for integrity.

[0029] In this embodiment, the preset monitoring interface refers to an interface used to monitor in real time whether there are changes in the knowledge content of each data source; the set identification technology refers to a pre-set technology for accurately identifying 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 that has changed in the data source, and the change methods include addition, deletion and modification; the short-term and long-term query frequency refers to the frequency of querying the changed knowledge entry in the short and long term; content sensitivity assessment refers to the process of judging the importance of the content of the changed knowledge entry.

[0030] In this embodiment, a real-time change entry refers to a changed knowledge entry that is determined to need to be updated to the local knowledge base immediately; a scheduled change entry refers to a changed knowledge entry that needs to be updated to the local knowledge base regularly at certain time intervals; and an actual change interval refers to the time interval in which a scheduled change entry actually changes in the local knowledge base.

[0031] 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 change entries, scheduled change entries, and corresponding actual change intervals, which clearly specifies which entries need to be updated immediately, which entries need to be updated regularly, and the specific time intervals for scheduled updates; the preset incremental update process refers to a pre-set update process used to handle the update of changed knowledge entries, such as only updating some knowledge entries that have changed; the integrity check is a verification operation performed on the local knowledge base after the knowledge entry update is completed, and the core check is whether the updated knowledge base is complete and whether there is any data loss, damage or inconsistency.

[0032] The beneficial effects of the above technology are: by real-time monitoring of domain knowledge changes in various data sources, and with the help of setting recognition technology, accurately obtaining changed knowledge entries; making real-time changes or scheduled changes to changed knowledge entries, and adopting different change methods based on the judgment results, and then generating knowledge update strategies, which can effectively avoid unnecessary frequent updates, reasonably allocate system resources, and improve the overall efficiency of knowledge base updates. Embodiment 4:

[0033] The present invention provides a large-model question-answering device based on a local knowledge base, wherein the item screening subunit includes: Frequency analysis block: used to obtain the historical query frequency of each changed knowledge item in the preset short-term sliding window and the preset long-term sliding window, and match the corresponding short-term query frequency level and long-term query frequency level; Marking the changed knowledge items with a high query frequency in the short-term query frequency level or the long-term query frequency level as real-time changed items; The changed knowledge items whose short-term query frequency level and long-term query frequency level are not high query frequency are regarded as sensitive selection items; Sensitivity analysis block: used to perform sensitivity analysis on sensitive selection items using the content sensitivity assessment mechanism to obtain the item sensitivity score; Mark sensitive selection items whose item sensitivity scores exceed the preset sensitivity threshold as real-time change items; Mark sensitive selection items whose item sensitivity scores do not exceed the preset sensitivity threshold as timed change items; Interval analysis block: used to adjust the baseline change interval by combining the historical query frequency with the item sensitivity score to obtain the actual change interval of the corresponding scheduled change item.

[0034] In this embodiment, the preset short-term sliding window is a preset shorter time interval, which is used for short-term statistics and analysis of the historical query frequency of the changed knowledge items, such as the last week or day; the preset long-term sliding window is a preset longer time interval, which is used for long-term statistics and analysis of the historical query frequency of the changed knowledge items, such as the last month; the historical query frequency refers to the number of times a certain changed knowledge item has been queried by users in the past period of time (determined by the preset short-term or long-term sliding window).

[0035] In this embodiment, the short-term query frequency level is a query frequency level selected from a preset short-term frequency level mapping table based on the historical query frequency within a preset short-term sliding window as a matching condition, wherein the short-term query frequency level includes three levels: high, medium and low; the preset short-term frequency level mapping table is composed of a correspondence between different historical query frequency ranges and corresponding short-term query frequency levels (high, medium and low). For example, within a preset short-term sliding window (such as the most recent week), the number of queries between 0 and 10 is a low query frequency level, 11 to 50 is a medium query frequency level, and 51 and above is a high query frequency level; the long-term query frequency level is a query frequency level selected from a preset long-term frequency level mapping table based on the historical query frequency within a preset long-term sliding window as a matching condition, wherein the long-term query frequency level includes three levels: high, medium and low; the preset long-term frequency level mapping table is composed of a correspondence between different historical query frequency ranges and corresponding long-term query frequency levels (high, medium and low). For example, if the preset long-term sliding window is the most recent month, the number of queries between 0 and 50 is a low query frequency level, 51 to 200 is a medium query frequency level, and 51 and above is a high query frequency level. times is a medium query frequency level, and 201 times and above is a high query frequency level.

[0036] In this embodiment, real-time change entries refer to knowledge entries that need to be changed or updated immediately, specifically change knowledge entries whose short-term query frequency level or long-term query frequency level is a high query frequency, and sensitive selection entries whose entry sensitivity scores exceed a preset sensitivity threshold; sensitive selection entries refer to change knowledge entries whose short-term query frequency level and long-term query frequency level are not high query frequencies.

[0037] In this embodiment, the content sensitivity evaluation mechanism is specifically: using preset item important indicators to evaluate the importance of sensitive selection items, and normalizing the obtained important evaluation indicator values, and then weighted averaging to obtain the item sensitivity score, wherein the preset item important indicators include the content change rate, the proportion of content sensitive information (for example, personal privacy, commercial secrets, national security and other sensitive information) and the number of application fields, etc.; the weight value assigned to the indicator value of each preset item important indicator is obtained by solving the matrix constructed after pairwise comparison and scoring using the hierarchical analysis method, and the value range is (0, 1).

[0038] In this embodiment, for example, there is a sensitive selection item 1, and the important evaluation index values of the preset item important index are respectively and , then the entry sensitivity score of sensitive selection entry 1 is equal to ,in, It represents the weight value assigned to the index value of the important index of the j-th preset item.

[0039] In this embodiment, the entry sensitivity score is a quantitative value obtained after performing an important analysis on the sensitive selection entry using the content sensitivity evaluation mechanism, and the value range is (0, 1). The higher the score, the more important the entry; the preset sensitivity threshold is a pre-set sensitivity score limit value. When the entry sensitivity score of the sensitive selection entry exceeds this threshold, the entry will be marked as a real-time change entry; conversely, if the entry sensitivity score does not exceed the threshold, it will be marked as a timed change entry; the baseline change interval is a pre-set initial entry change time interval; the actual change interval is obtained after adjustment based on the historical query frequency and the entry sensitivity score on the basis of the baseline change interval.

[0040] In this embodiment, for example, the historical query frequency of scheduled change entry 1 is 20 times, which is between 11 and 50 times, and is a medium query frequency level. The entry sensitivity score is 0.6, which is less than the preset sensitivity threshold of 0.8. At this time, the calculation formula for the actual change interval of scheduled change entry 1 is expressed as follows: ; in, Indicates the actual change interval of scheduled change entry 1; It is expressed as the baseline change interval; It is expressed as the weight of the impact of item sensitivity on the adjustment of the actual change interval, and its value range is (0, 1); It is expressed as the weight of the influence of the query frequency status on the adjustment of the actual change interval, and its value range is (0, 1); e is expressed as the base of the natural logarithm, and its value is 2.7; the weights assigned to the item sensitivity and query frequency status are obtained by solving the matrix constructed after pairwise comparison and scoring using the hierarchical analysis method.

[0041] The beneficial effects of the above technology are: by comprehensively considering the historical query frequency and content sensitivity, accurate judgment of real-time or scheduled changes to changed knowledge entries can be achieved, which can help avoid unnecessary frequent updates, reasonably allocate system resources, and improve question-answering efficiency. Example 5:

[0042] The present invention provides a large-model question-answering device based on a local knowledge base, wherein the query retrieval module includes: Query unit: used to perform semantic analysis and processing on the target user's query content to obtain query vector blocks; From the local knowledge base, the text vector blocks whose vector similarity with the query vector block exceeds the set vector similarity threshold are matched and output as the initial search vector blocks; 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 key text features of the corresponding historical answers from the historical user query records; Analyze the number of similar features of the answer key text features whose text similarity with each query key text feature exceeds a set text similarity threshold, and mark the corresponding text similarity as a reference similarity; The query-answer polling matrix is established by taking the number of similar features of each query key text feature and the mean of all corresponding reference similarities as matrix elements; Calculating the query-answer polling matrix to obtain a polling performance vector value; By assigning a time-decay weight to the polling performance vector value of each round of historical queries, summing and normalizing them, we can obtain the comprehensive role judgment value. Determine the query level of the current target user based on the comprehensive role judgment value; Common analysis unit: used to output the obtained initial search vector block as the actual search result when the target user's query level is common query; Professional analysis unit: When the target user's query level is professional, it performs causal reasoning and scenario association analysis based on the initial retrieval vector block and captures the professional retrieval vector block from the local knowledge base; Then the initial search vector block and all obtained professional search vector blocks are output as actual search results.

[0043] In this embodiment, the target user refers to the user who is currently using the large-model question-and-answer device based on the local knowledge base to conduct inquiries, such as managers, technicians, operators, etc.; the query content refers to the query question raised by the target user to the question-and-answer device; the query vector block refers to the vector representation form converted into 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 block 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 and the text vector block in the local knowledge base (referring to each text entry stored in the local knowledge base and converted into a vector representation form after semantic analysis) are sufficiently similar, so as to screen out text vector blocks that are sufficiently relevant to the query content; for example, there are query vector block 1 and text vector block 1, and the vector similarity calculation is performed 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, the text vector block 1 is output as the initial retrieval vector block.

[0044] In this embodiment, the preset time period is a pre-set time interval for collecting and analyzing the historical query records of the target user within the time range; the historical user query records refer to all query records left by the target user when using the question-answering device within the preset time period.

[0045] In this embodiment, the query key text features refer to 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 answer key text features refer to text features extracted from the corresponding historical answers of each round of historical queries that represent the core meaning of the answers, including but not limited to key information and professional terms.

[0046] In this embodiment, for example, there is a query "How to optimize the automobile engine manufacturing process to improve production efficiency", and the corresponding query key text features include: keywords: automobile engine, manufacturing process, production efficiency, entity: automobile engine, phrases: optimize manufacturing process, improve production efficiency; The historical answer is "Optimizing the automobile engine manufacturing process can improve production efficiency by introducing automated equipment, rationally arranging work processes, and strengthening employee training." The corresponding answer key text features include: Key information: introducing automated equipment, rationally arranging work processes, and strengthening employee training. Professional terms: automated equipment, process arrangement, and employee training.

[0047] In this embodiment, the number of similar features refers to the number of answer key text features whose text similarity with each query key text feature exceeds a set text similarity threshold, wherein text similarity refers to the degree of similarity between the query key text feature and the answer key text feature, which is usually calculated using a cosine similarity algorithm; the set text similarity threshold is a pre-set similarity threshold used to determine whether the query key text feature and the answer key text feature are sufficiently similar to screen out answer content that is sufficiently relevant to the query content; the reference similarity refers to the text similarity corresponding to the answer key text feature whose similarity with each query key text feature exceeds the set text similarity threshold.

[0048] In this embodiment, the query-answer polling matrix refers to a matrix constructed by using the number of similar features of each query key text feature and the average of all corresponding reference similarities as matrix elements to reflect the correlation between any query and the answer in each round of historical query by the user within a preset time period; the polling performance vector value refers to the eigenvalue decomposition operation calculated on the query-answer polling matrix, which is used to measure the strength of the correlation between the query and the answer.

[0049] In this embodiment, the comprehensive role judgment value is obtained by assigning a time-decay weight (determined by an exponential decay function) to the polling performance vector value of each round of historical queries, and then normalizing the initial role judgment value obtained after summing them (using methods such as minimum-maximum normalization and Z-score normalization); Among them, the initial role judgment value is expressed as ;in, Indicates the initial role judgment value; The polling performance vector value of the query-answer polling matrix corresponding to the i-th historical query within the preset time period, where i = 1, 2, 3, ..., n; n represents the total number of historical query rounds performed by the target user within the preset time period; It is represented as the time decay weight assigned to the polling performance vector value of the i-th round of historical queries within a preset time period; It is represented as the time interval between the i-th round of historical query and the current moment within the preset time period.

[0050] In this embodiment, the query level includes two levels: general query and professional query; the professional retrieval vector block refers to the text vector block that is highly relevant to the query content and has professional depth and multi-dimensional technical details captured from the local knowledge base after the professional analysis unit performs causal reasoning and scenario association analysis based on the initial retrieval vector block when the query level of the target user is determined to be a professional query; the actual retrieval result refers to the final retrieval result output by the general analysis unit or the professional analysis unit based on the query level of the target user. When the query level is general 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.

[0051] The beneficial effects of the above technology are: by judging the query level of the target user and adopting different retrieval strategies based on the level judgment results, it can meet the retrieval needs of different users in different scenarios; 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 retrieval results. Example 6:

[0052] The present invention provides a large-model question-answering device based on a local knowledge base, wherein the professional analysis unit includes: Scenario analysis subunit: used to obtain the relevant application scenarios of each initial search vector block and the corresponding scenario association system; When there are multiple initial search vector blocks, the initial search vector blocks whose scene similarity is higher than the set scene similarity threshold are clustered to obtain a vector block set; Obtain and analyze the historical search frequency and historical search correction frequency of each relevant application scenario corresponding to the current vector block set to obtain the search verification score; The relevant application scenario with the highest retrieval verification score is used as the comprehensive relevant application scenario of the current vector block set, and the corresponding scenario association system is used as the comprehensive scenario association system; Causal analysis subunit: used to combine the corresponding query content of each initial search vector block and use the set causal discovery algorithm to determine the corresponding single causal link; When there is a vector block set, the single causal links of all the initially retrieved vector blocks in the same vector block set are compared and adjusted in terms of relational attributes, duplicate links are deleted, and causal links are merged to generate a set causal link. Output subunit: used to capture the corresponding association 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; The corresponding association vector blocks of each single causal link and collective causal link are captured from the local knowledge base, and all causal links are combined and output as a professional vector block.

[0053] In this embodiment, the relevant application scenarios refer to the application scenarios that are pre-identified by the scene classification model for the initial retrieval vector blocks, and are marked accordingly, such as equipment maintenance and production process optimization; the steps for establishing the scene classification model are: first, data of various application scenarios are collected, and the collected data are pre-processed and feature extracted (for example, in the equipment maintenance scenario, the extracted features include equipment type, maintenance history, fault record, maintenance cost, etc.; in the production process optimization scenario, the extracted features include production links, production efficiency, quality control data, resource utilization, etc.); then, training is performed in combination with a preset classification algorithm (such as a decision tree, support vector machine, etc.) to establish a scene classification model; the scene association system refers to the relationship between the various components, links or elements in the current relevant application scenario. For example, in the equipment maintenance scenario, the scene association system refers to: the association between equipment and maintenance personnel, the association between equipment and maintenance plans, the association between equipment and fault types, etc.

[0054] In this embodiment, for example, there is a scenario association system for device maintenance scenario 1: the association between equipment and maintenance personnel, and the corresponding system architecture is device-maintenance personnel mapping table-maintenance personnel information acquisition, wherein the maintenance personnel mapping table records the maintenance personnel information corresponding to each device, including maintenance personnel ID, name, contact information, etc.

[0055] In this embodiment, setting a scene similarity threshold is a preset numerical standard used to determine whether the relevant application scenarios of two initial search vector blocks are sufficiently similar; the vector block set refers to a vector block set formed by clustering multiple initial search vector blocks whose scene similarity is higher than the set scene similarity threshold, wherein the scene similarity is determined by using a cosine similarity algorithm (or other similarity algorithms, such as Jaccard similarity, Euclidean distance, etc.) to calculate the similarity of the scene association systems between the relevant application scenarios, specifically referring to comparing whether the components, links or elements in the scene association systems of the two application scenarios are similar, and whether the association relationships between them are consistent or similar.

[0056] In this embodiment, the historical search frequency refers to the number of times the relevant application scenario has been searched in the past period of time; the historical search correction frequency refers to the number of times the user has corrected or adjusted the search results of a certain relevant application scenario in the past period of time; the calculation steps of the search verification score are: first calculate the ratio of the historical search frequency of the current relevant application scenario to the average historical search frequency of all relevant application scenarios in the same vector block set, that is, the frequency ratio; then combine the historical search correction frequency of the current relevant application scenario with the exponential function (the reason for selecting the exponential function is that it can make the correction attenuation index decrease exponentially with the increase of the historical search correction frequency, which more reasonably reflects the impact of the correction frequency on the search verification score) to determine the correction attenuation index (for example, if there is a relevant application scenario The historical search revision frequency is , then the corresponding modified attenuation index is ); Finally, the above frequency ratio is added to the modified attenuation index, and the result is the retrieval verification score; the comprehensive relevant application scenario refers to the application scenario with the highest retrieval verification score among all relevant application scenarios in the vector block set; the comprehensive scenario association system refers to the corresponding scenario association system of the comprehensive relevant application scenarios as the comprehensive scenario association system.

[0057] In this embodiment, the set causal discovery algorithm refers to a pre-set algorithm used to combine the query content and infer the causal structure between variables from the corresponding initial retrieval vector block, 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, for example, a reduced maintenance frequency of equipment A will lead to a decrease in production efficiency.

[0058] 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. Aggregate causal links refer to higher-level causal links obtained by integrating and summarizing all single causal links within the same vector block set that have undergone attribute unification and deduplication processing. Associated vector blocks refer to system units (such as components, links, or elements) of a knowledge system associated with a specific scenario (referring to a scenario association system corresponding to a comprehensive related application scenario of a related application scenario and a vector block set) or link nodes of a specific causal link (referring to a single causal link and a collective causal link of a vector block set), closely related vector blocks. Professional vector blocks refer to text vector blocks captured from a local knowledge base that are highly relevant to the query content and have professional depth and multi-dimensional technical details.

[0059] The beneficial effects of the above technology are: 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, providing users with more professional, comprehensive and accurate retrieval results, thereby improving the level of question and answer intelligence and adapting to more complex and varied query needs. Example 7:

[0060] The present invention provides a large-model question-answering device based on a local knowledge base, wherein the causal analysis subunit further includes: Behavior analysis block: used to obtain the historical query behavior path of the current target user and determine the reference behavior path by analyzing the degree of fit between the query node of each historical query behavior path and all the initial search vector blocks in the current vector block set; Attribute Unification Block: It is used to unify the expression form of a single causal chain with different relationship attributes but related essence within the same vector block set; Deduplication block: It is used to perform deduplication operations on identical single causal chains existing in the same vector block set, and finally retain only one of them; Set generation block: It is used to combine all the corresponding single causal chains of the vector block set processed by the attribute unification block and the duplicate removal block with the reference behavior path, input them into the pre-established link generation model, and obtain the set causal links.

[0061] 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 the past period of time, that is, the continuous query behaviors of each round of historical queries of the target user are connected in chronological order to construct a query behavior path. The query behavior path can be expressed 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 intention or demand of the user; for example, there is a historical query behavior path of "equipment fault alarm query" → "equipment maintenance record query" → "spare parts inventory status query" → "maintenance team scheduling query"; the corresponding query nodes are "equipment fault alarm", "equipment maintenance record query", "spare parts inventory status" and "maintenance team scheduling".

[0062] In this embodiment, the steps for determining the degree of compatibility between a query node and an initial search vector block include: first, performing semantic analysis on the query node to extract key features (such as keywords, entity names, and behavioral intentions); then, calculating the semantic similarity between the key features and the initial search vector block (for example, using a cosine similarity algorithm to calculate the semantic similarity, with a value range of (0, 1)); then, when the semantic similarity between the query node and a certain initial search vector block exceeds a set node similarity threshold (a pre-set value (such as 0.8) used to determine the degree of matching between the query node and the vector block), the query node is considered to be compatible with the initial search vector block; otherwise, it is considered to be incompatible.

[0063] In this embodiment, the reference behavior path refers to the behavior path with the highest path adaptation score (most relevant or most representative to the current query) in the historical query behavior path of the target user, which is screened out by analyzing the degree of adaptation of the query node of each historical query behavior path to all the initial search vector blocks in the current vector block set; wherein the path adaptation score is obtained by directly multiplying the number of initial search vector blocks adapted to each query node in the path by the average semantic similarity (the average semantic similarity of all initial search vector blocks adapted to the query node) and then summing them; for example, there are each query node in the historical query behavior path 1. and , the number of adapted vector blocks in the current vector block set is 3, 5, 5, and 4 respectively; the corresponding average semantic similarities are 0.81, 0.9, 0.88, and 0.82 respectively; at this time, the path adaptation score of historical query behavior path 1 is 3 0.81+5 0.9+5 0.88+4 0.82=14.61.

[0064] In this embodiment, a relationship attribute refers to a feature that describes the relationship between two or more entities, such as a cause-effect attribute; and a unified expression form refers to unifying the description form of a single causal chain within the same vector block set that has different relationship attributes but is essentially the same; for example, if there are two single causal links, namely "factor A causes result B" and "factor A triggers result B", they can be unified into the expression form of "factor A causes / triggers result B".

[0065] In this embodiment, the steps of 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) through word segmentation, vectorization, labeling and classification (for example, division according to relationship type (such as cause-result, condition-conclusion, etc.)); then, key features (such as entities, relationships, context, etc.) are extracted from each causal chain, and then an embedding model (such as Word2Vec, BERT, etc.) is used to convert the extracted features into vector representations; 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 similar causal chains are clustered using a selected clustering algorithm (such as K-means, hierarchical clustering, DBSCAN, etc.); finally, the clustering results are used to train a neural network.

[0066] In this embodiment, the set causal link refers to a higher-level causal link obtained by integrating and summarizing all single causal links in the same vector block set that have undergone attribute unification and duplication removal, reflecting the common characteristics and laws between multiple single causal links.

[0067] The beneficial effects of the above technology are: by analyzing users' historical query behavior preferences, and comparing and adjusting the relationship attributes of single causal chains in the same vector block set, deleting duplicate links, and merging causal links, it can help improve the readability and comprehensibility of causal analysis, avoid interference from redundant information, and thus enhance the intelligence and personalization of question answering. Example 8:

[0068] The present invention provides a large-model question-answering device based on a local knowledge base, wherein the module interaction module includes: Answer generation unit: used to generate actual retrieval answers based on the input actual retrieval results of the pre-deployed target question-answering model; Answer correction unit: used to provide user feedback function. Target users can mark incorrect answers and trigger model retraining.

[0069] In this embodiment, the target question-answering big model refers to a large language model constructed 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 retrieval answers based on the actual input retrieval results in a big model question-answering device based on a local knowledge base, for example, a Transformer architecture model that has been fine-tuned with specific domain data; the actual retrieval results are a set of information related to the target user's query obtained after a series of processing by the query retrieval module.

[0070] In this embodiment, the actual retrieval answer is the final answer generated by the target question and answer big model based on the actual retrieval result input and presented directly to the target user. It is presented in the form of natural language and clearly and accurately answers the user's query question; the user feedback function is an interactive mechanism provided by the model interaction module to the user, allowing the target user to evaluate and provide feedback on the generated actual retrieval answer. Specifically, when the user believes that the answer is wrong, inaccurate or incomplete, the answer is marked as wrong through this function; model retraining is the process of retraining the target question and answer big model under the triggering of the user feedback function, that is, when the target user marks the answer wrong and triggers model retraining, the question and answer device will collect the user's feedback information and related query records, actual retrieval results, wrong answers and other data as new training samples; then, these training samples are used to fine-tune or retrain the target question and answer big model to optimize the parameters and performance of the model.

[0071] The beneficial effects of the above technologies are: by utilizing the target question-answering large model to generate accurate and clear actual retrieval answers based on actual retrieval results, it effectively ensures that user queries can receive high-quality responses, thereby improving the accuracy of question-answering; by introducing the user feedback function, users can evaluate and provide feedback on the answers, and trigger model retraining when the answers are marked as wrong, which can effectively realize the continuous optimization of model performance, help reduce the occurrence of incorrect answers, and further improve the reliability of question-answering.

[0072] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A large model 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 captured from multi-source heterogeneous data sources and realize dynamic update of the knowledge base; Query and retrieval module: used to integrate intelligent judgment of user roles and intelligent analysis mechanisms, based on scenario-cause-effect dual-dimensional analysis, to accurately output customized actual search results; Model interaction module: used by the target question-answering model to output the actual retrieval answer based on the actual retrieval results input, and to provide user feedback function to optimize the retrieval answer.

2. A large model question-answering device based on a local knowledge base according to claim 1, characterized in that: The knowledge base establishment module includes: Data acquisition and cleaning unit: used to automatically capture domain knowledge data from multiple heterogeneous data sources and perform data cleaning to obtain target construction data; Knowledge base construction unit: used for performing semantic processing, vectorization processing and storage on the target construction data to generate a local knowledge base; Knowledge base updating unit: used to analyze the update monitoring status of the domain knowledge of each data source, determine the knowledge base update strategy and execute it, and update the local knowledge base.

3. A large model question-answering device based on a local knowledge base according to claim 2, characterized in that: The knowledge base updating unit includes: Monitoring subunit: used to monitor domain knowledge changes of various data sources in real time using preset monitoring interfaces, and to obtain changed knowledge entries by adopting set recognition technology; Item screening subunit: used to determine real-time change items and scheduled change items by performing short-term and long-term query frequency analysis and content sensitivity assessment on the changed knowledge items; Strategy generation subunit: used to summarize and organize all real-time change items, scheduled change items and corresponding actual change intervals, and generate knowledge base update strategies; Update subunit: used to immediately synchronize real-time changed entries to the local knowledge base according to the knowledge base update strategy, triggering the preset incremental update process; For scheduled change items, add them to the task queue according to the actual change interval and synchronize them to the local knowledge base regularly; And when the knowledge entry is updated, the knowledge base is checked for integrity.

4. A large model question-answering device based on a local knowledge base according to claim 3, characterized in that: The entry screening subunit includes: Frequency analysis block: used to obtain the historical query frequency of each changed knowledge item in the preset short-term sliding window and the preset long-term sliding window, and match the corresponding short-term query frequency level and long-term query frequency level; Marking the changed knowledge items with a high query frequency in the short-term query frequency level or the long-term query frequency level as real-time changed items; The changed knowledge items whose short-term query frequency level and long-term query frequency level are not high query frequency are regarded as sensitive selection items; Sensitivity analysis block: used to perform sensitivity analysis on sensitive selection items using the content sensitivity assessment mechanism to obtain the item sensitivity score; Mark sensitive selection items whose item sensitivity scores exceed the preset sensitivity threshold as real-time change items; Mark sensitive selection items whose item sensitivity scores do not exceed the preset sensitivity threshold as timed change items; Interval analysis block: used to adjust the baseline change interval by combining the historical query frequency with the item sensitivity score to obtain the actual change interval of the corresponding scheduled change item.

5. The large-scale question-answering device based on a local knowledge base according to claim 1, characterized in that: The query retrieval module includes: Query unit: used to perform semantic analysis and processing on the target user's query content to obtain query vector blocks; From the local knowledge base, the text vector blocks whose vector similarity with the query vector block exceeds the set vector similarity threshold are matched and output as the initial search vector blocks; 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 key text features of the corresponding historical answers from the historical user query records; Analyze the number of similar features of the answer key text features whose text similarity with each query key text feature exceeds a set text similarity threshold, and mark the corresponding text similarity as a reference similarity; The query-answer polling matrix is established by taking the number of similar features of each query key text feature and the mean of all corresponding reference similarities as matrix elements; Calculating the query-answer polling matrix to obtain a polling performance vector value; By assigning a time-decay weight to the polling performance vector value of each round of historical queries, summing and normalizing them, we can obtain the comprehensive role judgment value. Determine the query level of the current target user based on the comprehensive role judgment value; Common analysis unit: used to output the obtained initial search vector block as the actual search result when the target user's query level is common query; Professional analysis unit: When the target user's query level is professional, it performs causal reasoning and scenario association analysis based on the initial retrieval vector block and captures the professional retrieval vector block from the local knowledge base; Then the initial search vector block and all obtained professional search vector blocks are output as actual search results.

6. A large model question-answering device based on a local knowledge base according to claim 5, characterized in that: The professional analysis unit includes: Scenario analysis subunit: used to obtain the relevant application scenarios of each initial search vector block and the corresponding scenario association system; When there are multiple initial search vector blocks, the initial search vector blocks whose scene similarity is higher than the set scene similarity threshold are clustered to obtain a vector block set; Obtain and analyze the historical search frequency and historical search correction frequency of each relevant application scenario corresponding to the current vector block set to obtain the search verification score; The relevant application scenario with the highest retrieval verification score is used as the comprehensive relevant application scenario of the current vector block set, and the corresponding scenario association system is used as the comprehensive scenario association system; Causal analysis subunit: used to combine the corresponding query content of each initial search vector block and use the set causal discovery algorithm to determine the corresponding single causal link; When there is a vector block set, the single causal links of all the initially retrieved vector blocks in the same vector block set are compared and adjusted in terms of relational attributes, duplicate links are deleted, and causal links are merged to generate a set causal link. Output subunit: used to capture the corresponding association 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; The corresponding association vector blocks of each single causal link and collective causal link are captured from the local knowledge base, and all causal links are combined and output as a professional vector block.

7. The large-scale question-answering device based on a local knowledge base according to claim 6, characterized in that: The cause-effect analysis subunit further includes: Behavior analysis block: used to obtain the historical query behavior path of the current target user and determine the reference behavior path by analyzing the degree of fit between the query node of each historical query behavior path and all the initial search vector blocks in the current vector block set; Attribute Unification Block: It is used to unify the expression form of a single causal chain with different relationship attributes but related essence within the same vector block set; Deduplication block: It is used to perform deduplication operations on identical single causal chains existing in the same vector block set, and finally retain only one of them; Set generation block: It is used to combine all the corresponding single causal chains of the vector block set processed by the attribute unification block and the duplicate removal block with the reference behavior path, input them into the pre-established link generation model, and obtain the set causal links.

8. The large-scale question-answering device based on a local knowledge base according to claim 1, characterized in that: The module interaction module includes: Answer generation unit: used to generate actual retrieval answers based on the input actual retrieval results of the pre-deployed target question-answering model; Answer correction unit: used to provide user feedback function. Target users can mark incorrect answers and trigger model retraining.

Images