Enterprise-level schedule planning and knowledge base oriented intelligent collaborative question-answering system and method

Abstract

The invention relates to the technical field of computer systems for natural language processing or semantic processing, and discloses an enterprise-level schedule planning and knowledge base-oriented intelligent collaborative question-answering system and an enterprise-level schedule planning and knowledge base-oriented intelligent collaborative question-answering method. The system comprises a vectorization processing module, an RAG knowledge base module and the like. The system converts natural language input of a user and enterprise knowledge data into semantic vectors, and retrieves related enterprise knowledge fragments from a vector database based on semantic similarity. The large language model core module analyzes the user intention, generates a preliminary answer and identifies whether a schedule type operation request is included or not; and if the schedule operation request exists, the system splits the request through the multi-agent cooperation module and distributes the request to the corresponding agent to obtain a task processing result. And finally, semantic consistency fusion is carried out on the preliminary answer and a task processing result through a context fusion module, a comprehensive answer is generated by a large language model and is returned to a user side, and unified intelligent response of enterprise knowledge and schedule service is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of computer systems for natural language processing, and particularly to an intelligent collaborative question-answering system and method for enterprise-level schedule planning and knowledge base. Background Art

[0002] In the prior art, enterprises often use a schedule management system and a knowledge base management system to separately handle time arrangement and information query tasks. Some systems introduce natural language processing or conversational interfaces to improve the user interaction experience. At the same time, the combination of vectorized retrieval and large language models has been preliminarily applied in question-answering systems, enabling semantic-based knowledge invocation and intelligent response. These systems mainly focus on single functions and lack the ability of unified coordination.

[0003] The prior art generally has the following problems: First, the schedule management and knowledge retrieval systems are independent of each other, and users need to switch between multiple platforms, affecting the usage efficiency. Second, it is difficult for intelligent question-answering systems to combine heterogeneous knowledge within the enterprise and dynamic task requests, and they cannot perform structured parsing and task scheduling on natural language inputs containing operation requests. Third, there is a lack of a semantic fusion mechanism between knowledge response results and operation execution results, resulting in fragmented output information and affecting the overall consistency and response accuracy.

[0004] To solve the above problems, there is an urgent need to provide an intelligent collaborative question-answering system for enterprise-level applications. Summary of the Invention

[0005] The present application provides an intelligent collaborative question-answering system and method for enterprise-level schedule planning and knowledge base, which realizes a unified response of enterprise knowledge answering and schedule task execution based on natural language input, thereby significantly improving the integrated intelligent level of information acquisition and operation instruction processing.

[0006] The present application provides an intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base, including: A vectorization processing module, configured to receive a natural language input of a user, convert the natural language input of the user into a first semantic vector, and convert pre-stored enterprise knowledge data into a second semantic vector; A RAG knowledge base module, configured to retrieve enterprise knowledge fragments corresponding to the second semantic vector semantically related to the first semantic vector in a vector database; A large language model core module, configured to receive the natural language input of the user and the enterprise knowledge fragments, perform semantic understanding and task parsing, generate a preliminary answer including knowledge answering content, and identify whether the natural language input of the user contains a schedule operation request; The multi-agent collaboration module is used to disassemble the schedule operation request into one or more subtasks according to the result of task parsing and schedule them to the corresponding functional agents for execution when it is recognized that the natural language input of the user contains a schedule operation request, and at least includes calling the schedule management agent to access enterprise schedule data and generate a task processing result; The context fusion module is used to receive the preliminary answer and the task processing result, and perform context fusion on the two based on the semantic structure consistency strategy; Among them, the large language model core module is also used to receive the result of context fusion, generate a comprehensive reply, and output it to the user terminal to complete the question and answer response.

[0007] The beneficial effects of this application mainly include: (1) It realizes the unified response of enterprise knowledge retrieval and task operation. By using natural language input for both knowledge recall and task parsing, the system can process query requests and instruction requests simultaneously in one interaction process, improving the intelligence level of enterprise information services. (2) It improves the accuracy and context coherence of the question and answer content. With the context fusion module aligning the semantic structures of the preliminary answer and the task processing result, it avoids the problem of the disconnection between knowledge responses and actual operation results in traditional systems, enhancing the overall consistency and practicality of the results. (3) It constructs a semantic-driven multi-agent dynamic collaboration mechanism. The system can automatically disassemble tasks according to the language understanding results and intelligently allocate them to functional components, adapting to the complex and changeable enterprise-level user requirements, and having good scalability and task adaptation capabilities. (4) It significantly improves the user usage efficiency and experience. Users can obtain both knowledge answers and operation execution results through only one natural language input, saving the steps of multi-platform switching and repeated operations, and optimizing the work process of enterprise employees. Description of the Drawings

[0008] Figure 1 is a schematic diagram of an intelligent collaborative question and answer system for enterprise-level schedule planning and knowledge base provided by the first embodiment of this application.

[0009] Figure 2 is a flowchart of an intelligent collaborative question and answer method for enterprise-level schedule planning and knowledge base provided by the second embodiment of this application. Detailed Embodiments

[0010] Many specific details are set forth in the following description in order to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of this application. Therefore, this application is not limited by the specific embodiments disclosed below.

[0011] The first embodiment of this application provides an intelligent collaborative Q&A system for enterprise-level schedule planning and knowledge base. Please refer to Figure 1 , which is a schematic diagram of the first embodiment of this application. The following combines Figure 1 to elaborate in detail on an intelligent collaborative Q&A for enterprise-level schedule planning and knowledge base provided by the first embodiment of this application.

[0012] The intelligent collaborative Q&A system for enterprise-level schedule planning and knowledge base includes a vectorization processing module 101, a RAG knowledge base module 102, a large language model core module 103, a multi-agent collaboration module 104, and a context fusion module 105.

[0013] The vectorization processing module 101 is used to receive the natural language input of the user, convert the natural language input of the user into a first semantic vector, and convert the pre-stored enterprise knowledge data into a second semantic vector.

[0014] The vectorization processing module 101 is used to implement the semantic representation conversion in the intelligent collaborative Q&A system for enterprise-level schedule planning and knowledge base. Its core function is to preprocess and embed and encode the input natural language text and enterprise knowledge data to generate a vector representation with semantic comparability to support subsequent semantic retrieval and intelligent understanding processing. This module consists of two main processing paths, which are respectively responsible for the vectorization tasks of the natural language input of the user and the pre-stored enterprise knowledge data.

[0015] After receiving the natural language input at the user interaction layer, the vectorization processing module 101 first performs language normalization operations on this input, including preprocessing processes such as word segmentation, stop word removal, and word form unification, and sends it into a selected semantic embedding model. This model can be an open-source or self-developed pre-trained language model, such as semantic embedding models based on the Transformer architecture like BERT, Sentence-BERT, bge-large-zh, etc., for generating the first semantic vector corresponding to this user input. This first semantic vector is a high-dimensional dense vector, usually with a dimension between 256 and 1024, and can capture the context semantics and keyword information in the user input. Its data structure supports fast comparison and semantic matching with the enterprise knowledge semantic vectors in the vector database.

[0016] For the processing path of enterprise knowledge data, during the system deployment phase or data update nodes, the vectorization processing module 101 will batch load the enterprise's internal structured data (such as project records, personnel information, and meeting arrangements in the database), semi-structured data (such as JSON-formatted logs and email summaries), and unstructured data (such as PDF reports, Word documents, etc.). Through a unified information extraction interface, the above multi-source data will be converted into text corpora, and after passing through the preprocessing process, they will be sent into the same type of semantic embedding model for encoding. The generated second semantic vectors are used to construct and maintain the vector database for the RAG knowledge base module 102, and the vector entries will be bound with the unique identifiers and metadata information of the original text fragments (such as document source, creation time, and project number to which they belong, etc.) to ensure knowledge positioning and tracking in subsequent recall processes.

[0017] In addition, during the process of constructing the second semantic vectors, the vectorization processing module 101 needs to support the incremental update and batch vector refresh mechanisms. Specifically, the module should have a data change detection and task scheduling mechanism to trigger the re-vectorization operation of some or all documents according to the knowledge base update event, and after completion, write the updated second semantic vectors into the vector database through the interface. This process needs to maintain the vector consistency with the RAG knowledge base module 102 to avoid recall mismatches.

[0018] To ensure the comparability of the two types of vectors, the embedding model parameters used by the vectorization processing module 101 should be kept consistent, or the two models should be jointly trained through a cross-domain training strategy in a shared semantic space, so as to ensure that the first semantic vectors and the second semantic vectors have comparable distances in the same semantic representation space. In a specific embodiment, the system can also adopt a fine-tuning model for different data sources to improve the semantic fit, but the vector output still needs to go through a unified transformation to maintain the vector space compatibility.

[0019] Finally, to support the integration with the large language model core module 103, the vectorization processing module 101 needs to provide a standardized API interface. The output first semantic vectors and second semantic vectors should adopt a consistent vector format and floating-point precision, and additional position information, sentence boundaries, or paragraph context should be added when necessary to support the combined processing of subsequent multi-modal inputs. The module also needs to include an exception detection and fallback mechanism to output recognizable empty vector markers or trigger error logging in case of embedding failure or missing semantic information, ensuring the robustness of the system.

[0020] In practical applications, the user's natural language input and enterprise knowledge data are two core objects in the semantic calculation of the intelligent collaborative Q&A system. To more clearly illustrate their specific meanings and roles, we can use a typical enterprise scenario as an example.

[0021] Assume that an enterprise has deployed the intelligent collaborative question-answering system described in the present invention, and a marketing manager initiates the following voice or text request through the client: "Please help me check the minutes of the promotion meeting on the X series products last month, and arrange a feedback meeting with the design department this Friday afternoon." This sentence is the user's natural language input, which is expressed in a natural and unstructured way and contains two parallel semantic intentions: one is a knowledge query, that is, querying a certain type of historical meeting records; the other is an operation request, that is, arranging a meeting with a specific department. The vectorization processing module receives the input and converts it into a first semantic vector through the embedding model, so that the system can understand its semantic content and perform subsequent knowledge retrieval and task identification.

[0022] Correspondingly, enterprise knowledge data refers to various information resources accumulated by enterprises during their operations. These data usually exist in structured, semi-structured or unstructured forms, covering meeting minutes, project reports, personnel schedules, system documents, business logs, etc. For example, in the above scenario, the enterprise knowledge data in the system may include a Word document titled "Minutes of the X Series March 2024 Promotion Meeting", which records the meeting time, participants, discussion topics and decision results in detail; it may also include a meeting record in the database, numbered PM-0322, with the meeting theme marked as "X Series Quarterly Promotion Report Meeting", the time is March 15, 2024, and the meeting summary is "The focus was on discussing the conversion rate of social platform advertising and offline channel effect feedback."

[0023] During the system deployment or data synchronization phase, these enterprise knowledge data will be parsed by the vectorization processing module and uniformly sent to the semantic embedding model, generating the second semantic vector and storing it in the vector database. In this way, when the user initiates a natural language request, the system can retrieve the knowledge fragment corresponding to the second semantic vector that is most semantically relevant to it in the vector database based on its first semantic vector, such as the above-mentioned meeting minutes document, thereby providing high-quality contextual basis for the subsequent large language model core module to generate knowledge responses.

[0024] Through this mechanism, the user's natural language input is connected to the enterprise knowledge data at the semantic level, so that the system can not only understand the user's needs, but also quickly locate relevant information from the massive enterprise knowledge, and complete the unified coordination of knowledge response and task operation. This two-way mapping capability based on semantic vectors is the basis for achieving efficient enterprise-level question-answering and scheduling capabilities.

[0025] In summary, the vectorization processing module 101 is not only a pre - processing step for semantic computing. The first semantic vector it outputs directly determines the retrieval accuracy of the RAG knowledge base module 102, while the second semantic vector constitutes the semantic representation basis of the entire knowledge system and is the key to realizing semantic - driven intelligent question - answering and operation scheduling. The architecture design of the module should ensure meeting the real - time and security requirements in aspects such as processing speed, vector consistency, model scalability, and interface compatibility in enterprise - level scenarios.

[0026] Furthermore, the vectorization processing module is also used for: After receiving the user's natural - language input, call an embedding model based on the Transformer architecture to convert the user's natural - language input into a first semantic vector, and attach an input timestamp and a context - position index to the first semantic vector to retain the user - language input time - series information; Before converting the pre - stored enterprise knowledge data into a second semantic vector, select a corresponding semantic - modeling strategy according to the format type of the enterprise knowledge data. Among them, for structured data, call the rule - embedding method, and for natural - language documents, call the same embedding model as the first semantic vector to obtain a preliminary second semantic vector; Perform vector normalization processing on the first semantic vector and the preliminary second semantic vector respectively, and calculate the vector - consistency index between them based on cosine similarity in the same semantic space; When the vector - consistency index is less than a preset semantic - similarity threshold, perform a vector - tuning operation on the preliminary second semantic vector to enhance its cross - source semantic - alignment ability with the first semantic vector. The tuned second semantic vector is used for the RAG knowledge base module to retrieve enterprise - knowledge fragments corresponding to the second semantic vector semantically related to the first semantic vector in the vector database.

[0027] In the intelligent collaborative question - answering system for enterprise - level schedule planning and knowledge base, the vectorization processing module not only has the basic ability to convert natural - language semantic representations, but also further undertakes key functions such as the consistency processing of semantic - representation formats between different types of data, cross - modal semantic fusion, and quality control. Especially in enterprise - level application scenarios, in the face of multi - source heterogeneous knowledge data and diverse user - input semantic expressions, it can achieve dynamic, accurate, and unified semantic - embedding results, thereby providing a stable and comparable vector basis for subsequent modules and significantly improving the retrieval accuracy of the RAG knowledge base module.

[0028] During the system startup or running phase, when the user issues a natural language request through the terminal, the vectorization processing module will receive the natural language input immediately. This input may be a single-sentence instruction, such as "Arrange a meeting with Manager Wang", or a composite request, like "Find the sales data report for the second quarter of last year and remind me to submit the meeting materials tomorrow afternoon", or it can also be a paragraph-style information description. After receiving this type of input, the system encodes it by calling an embedding model based on the Transformer architecture (such as ChatGLM, Baichuan, BERT, RoBERTa, or a lightweight language model customized by the enterprise). The encoding process is not a simple word-level or sentence-level mapping, but rather captures semantic elements such as context dependencies, structural relationships, subject-verb collocations, temporal logic, and verb intentions through multiple attention mechanisms, thereby generating a high-dimensional dense representation, which is the first semantic vector.

[0029] To make the generated first semantic vector have temporal interpretability, the system will also inject two additional information dimensions into this semantic representation. One is the input timestamp of the current request, which records the exact time when the semantic request is issued. This timestamp is usually represented in UNIX time format or ISO standard format and is mapped by the embedding model into a vector channel or an additional feature vector through the time embedding mechanism; the other is the context position index, which is mainly enabled in multi-turn dialogue scenarios and is used to indicate the relative position of this input in the entire conversation, such as "Round 3", "Reply to the previous question", "Follow up on the first arrangement". This helps the system understand the context continuity, which is particularly crucial for expressions such as "Schedule another meeting" and "Postpone yesterday's meeting".

[0030] In contrast to the user input processing process, the sources of enterprise knowledge data are extremely diverse. There may be structured business records in the database (such as meeting databases, task tables, customer management systems), semi-structured logs (such as JSON-format OA approval processes, report interface outputs, system monitoring records), and a large amount of unstructured content (such as employee training manuals, historical emails, project summary documents, meeting minutes, PDF white papers, PPT briefings, etc.) in the system. To ensure that different data types can be expressed in a unified semantic space and have comparability, the vectorization processing module needs to dynamically select appropriate modeling strategies according to the types of data formats.

[0031] Specifically, for structured data, the system uses a rule embedding method, that is, the data is first segmented and the content is extracted through field templates, and then specific semantic labels or embedding rules are assigned according to the meaning of each field (such as "meeting title", "time", "person in charge", "status"). Common methods include pre-trained vocabulary matching, numerical regularization, field splicing vectorization, etc. After each field value is converted into a low-dimensional or medium-dimensional vector, the module splices these fragments in a fixed order or fuses them through a multi-layer perceptron to form a preliminary second semantic vector. For natural language document content, the system uses the same embedding model as the first semantic vector to ensure that the generated results are isomorphic in the semantic space and avoid embedding offsets caused by model differences.

[0032] After completing the above vector generation, the system will perform vector normalization operations on the first semantic vector and the preliminary second semantic vector respectively. The typical method is L2 norm normalization. The purpose of this step is to standardize each semantic vector to unit length, so as to facilitate the use of cosine similarity for directional semantic matching and remove the interference of semantic strength (modulus) on similarity calculation. After normalization, the system will perform the calculation of vector consistency index, that is, judge the semantic proximity between two vectors in the semantic space. The formula for calculating cosine similarity is the dot product of the first vector and the second vector divided by the product of their respective modulus lengths. Since it has been normalized, the calculation result is the cosine value of the angle between the two unit vectors, and the value range is [-1, 1]. The closer to 1, the more similar.

[0033] When the similarity result is lower than the system preset semantic similarity threshold (for example, 0.6 or 0.7), the system will determine that the current preliminary second semantic vector does not match the first semantic vector in the semantic space. This inconsistency may be due to multiple reasons, such as inconsistent embedding model training data style, document data and user requests deviate in time context or task orientation, or embedding rules fail to capture certain semantic details. In order to fix this inconsistency, the system will trigger a vector tuning operation.

[0034] Tuning operations generally include but are not limited to the following strategies: First, re-weight the embedding fields of structured data, such as increasing the semantic weight of fields corresponding to high-frequency keywords in user requests, such as "time" and "place"; second, re-embedding more representative paragraphs of natural language documents, such as prioritizing the extraction of parts containing time expressions or imperative sentences; third, introducing knowledge context completion fragments, that is, extracting additional sentences or paragraphs semantically related to the target content from the same document to form input with stronger semantic coverage; fourth, using the attention mechanism to adjust the embedding input weights so that the embedding model pays more attention to the alignment path between the problem entity and the key content in the document.

[0035] After the optimization operation is completed, the generated new vector will be normalized again and replace the original preliminary second semantic vector as the official second semantic vector for the RAG knowledge base module to call. In the vector database, the RAG knowledge base module can use this second semantic vector as a retrieval query through an efficient indexing structure (such as HNSW, IVF, PQ, etc.) to find enterprise knowledge fragments semantically related to the first semantic vector, thus ensuring that the Q&A output has semantic accuracy, temporal relevance, and business context consistency.

[0036] In summary, the vectorization processing module implements a multi-dimensional, deeply integrated, and tunable semantic vector construction and alignment mechanism from user natural language input to enterprise knowledge multi-source data, playing a core bridging role between input understanding and knowledge invocation in the entire intelligent Q&A system. Through temporal information modeling, format adaptation strategies, semantic consistency detection, and tuning mechanisms, it ensures that the system has the ability to handle complex requests and heterogeneous knowledge in actual enterprise applications.

[0037] Furthermore, when the vectorization processing module performs the vector optimization operation, it includes the following steps: Extract the generation time of the enterprise knowledge data corresponding to the preliminary second semantic vector and embed this generation time into the time dimension channel of the preliminary second semantic vector to construct a temporal second semantic vector with temporal tags. Based on the time expression content involved in the context of the first semantic vector, calculate its implicit temporal vector representation and conduct a temporal alignment evaluation with the temporal second semantic vector. If the evaluation result shows a significant temporal inconsistency between the two, adjust the time weight parameter of the preliminary second semantic vector and enhance the attention by combining the context position index to improve its temporal matching degree. Use the adjusted temporal second semantic vector as the optimized second semantic vector.

[0038] In the intelligent collaborative Q&A system for enterprise-level schedule planning and knowledge base described in the present invention, when the vectorization processing module finds that the semantic consistency index between the first semantic vector and the preliminary second semantic vector is lower than the preset threshold, it triggers the vector optimization operation. This optimization operation specifically focuses on the time dimension, and its core purpose is to improve the semantic matching accuracy of knowledge fragments and user input in terms of tense, especially when dealing with questions with time expressions raised by users, such as "the first quarter of this year", "last month", or "the most recent regular meeting", etc.

[0039] First, the system needs to extract time information from the enterprise knowledge data corresponding to the preliminary second semantic vector. This step usually includes obtaining the document generation time, database write time, or log record time of this knowledge fragment from the metadata. If this information is not explicitly stored, the system can also extract the time expressions appearing in the text through content analysis, such as "March 15, 2023", "the second quarter of 2024", etc. Regardless of the form of the information source, the system must standardize it into a unified timestamp representation, such as the ISO 8601 standard format, to ensure the consistency of subsequent processing. After the time information is extracted, the system encodes this timestamp into a dense time vector and injects it into the original preliminary second semantic vector as a time dimension channel.

[0040] The injection method can adopt vector concatenation, position channel expansion, or feature fusion strategies. Among them, the concatenation form is the most direct, that is, appending an equal-length or fixed-length time vector channel after the original semantic vector to make it a temporal second semantic vector with temporal characteristics. For example, if the original semantic vector is 768-dimensional and the time channel can be set to 64-dimensional, the final vector is 832-dimensional; the parallel embedding method can also be used, generating a context-independent time vector through another time embedding model and fusing it into the main vector channel after weighting by attention weights.

[0041] Secondly, the system needs to perform time semantic recognition and modeling on the user input content represented by the first semantic vector. The specific method is to use a time expression recognizer (such as a time NER system based on rules, statistics, or Transformer models) to extract time-related entities or phrases from the original natural language and standardize them into structured time ranges, time points, or time types (such as "relative time", "absolute time", "fuzzy time"). The extracted time expressions will be linked with the system's current time to generate one or more implicit temporal vectors. For example, "last month" will be parsed as "the time period one natural month before the current date" and embedded into a vector representing the user's temporal intention, forming a vector pair with the same dimension as the temporal second semantic vector.

[0042] Then, the system evaluates whether the temporal second semantic vector and the implicit temporal vector are consistent at the time intention level by comparing the semantic distance between them. This comparison can adopt the standard cosine similarity calculation method or introduce a distance function based on time logic, such as judging the overlap rate of two time periods, the offset size of the start and end points, etc. If the evaluation result shows a significant inconsistency between them, the system will perform a tuning operation to adjust the time semantic structure of the preliminary second semantic vector to make it more in line with the user's time intention.

[0043] The optimization process mainly includes optimizations in two directions: One is to adjust the weight parameters of the time dimension channel. In the original second semantic vector of the tense, time information and content semantics are usually mixedly represented. By adjusting the scalar amplification coefficient of the time channel, the contribution degree of this channel to the overall vector can be enhanced or reduced, so as to control the dominance of the time attribute in subsequent matching. The other is to adjust the position encoding in the original input of the embedding model in combination with the context position index, so that the attention mechanism can focus more on the time-related paragraphs or sentences in the document. The context position index can be based on the document structure hierarchy, such as headings, chapter numbers, paragraph numbers, or can be based on the relative position identification of sentences in the document. Combining with the distribution of time entities in the document, the semantic concentration area can be further located.

[0044] Finally, the optimized second semantic vector of the tense is the optimized second semantic vector. This vector has both the ability to express content semantics and align with the user's time intention, and has a stronger time correlation expression effect. This vector will be used as the index basis for knowledge fragment retrieval in the RAG knowledge base module, and is used to recall the enterprise knowledge fragments corresponding to the second semantic vectors semantically related to the first semantic vector from the vector database.

[0045] The entire optimization process starts from the original vector and goes through continuous processing links such as time extraction, tense modeling, semantic alignment, time dimension weight adjustment, and attention redistribution to form a complete data flow closed loop, ensuring that the user's intention and the knowledge time background can be accurately aligned in the semantic retrieval stage. This mechanism has significant advantages in processing instruction-based queries with time expressions, time-sensitive multi-round conversations, and scenarios where enterprises need to retrieve knowledge according to timeliness internally.

[0046] The RAG knowledge base module 102 is used to retrieve the enterprise knowledge fragments corresponding to the second semantic vectors semantically related to it in the vector database based on the first semantic vector.

[0047] The RAG knowledge base module 102 is used to implement semantic vector-based enterprise knowledge retrieval in the intelligent collaborative Q&A system for enterprise-level schedule planning and knowledge base. Its core function is to efficiently find the second semantic vector semantically related to it in the vector database according to the first semantic vector generated from the user's natural language input, and obtain the corresponding enterprise knowledge fragments accordingly to support the subsequent semantic understanding and response generation completed by the large language model core module 103.

[0048] During the implementation process, the RAG knowledge base module 102 receives the first semantic vector from the vectorization processing module 101. The first semantic vector is a dense vector representation obtained by converting the user's natural language input through a semantic embedding model, usually with dimensions of 256, 512, or higher, and its semantic structure is in the same vector space as the second semantic vector. The RAG knowledge base module 102 first performs efficient matching in the constructed vector database by calling a vector similarity retrieval algorithm, such as methods based on inner product, cosine similarity, or Euclidean distance. A large number of second semantic vectors generated from enterprise knowledge data are pre-stored in the vector database and mapped to the original knowledge fragments. To improve the retrieval efficiency, this module usually uses a database engine that supports fast approximate nearest neighbor (ANN) search, such as Milvus, and improves the query performance by constructing acceleration structures such as IVF (inverted file index) and HNSW (graph structure index) while ensuring the recall accuracy.

[0049] During the retrieval process, the RAG knowledge base module 102 can screen a number of the most relevant second semantic vectors according to a preset similarity threshold, usually returning the enterprise knowledge fragments corresponding to the Top-K (for example, the first 3 to 5) vectors, and extracting information such as their original text content, data source, and relevant tags. These knowledge fragments can include internal enterprise knowledge content such as project summaries, meeting minutes, system specifications, product descriptions, and operation manuals. The module can also attach a confidence scoring mechanism to assign a semantic matching confidence to each recalled fragment for subsequent context reconstruction and multi-fragment fusion.

[0050] In some implementation manners, the RAG knowledge base module 102 also includes content preprocessing and postprocessing capabilities. For example, operations such as syntactic segmentation, keyword highlighting, and context supplementation are performed before the knowledge fragment is output to make the fragment more suitable as the input context of the large language model. When processing semantically ambiguous or ambiguous inputs, the module can combine multi-vector extended retrieval, that is, introduce multiple copies of the first semantic vector after semantic enhancement, and independently perform retrieval tasks on each copy, and finally merge and deduplicate the results to enhance the coverage and robustness of the recall.

[0051] The enterprise knowledge fragments output by the RAG knowledge base module 102 must retain the relevance to the original enterprise knowledge data and have a direct semantic relevance to the first semantic vector. Its output is not only the text content itself but can also include associated metadata structures, such as the department to which the knowledge belongs, the update time, the document type, etc., to assist the large language model core module 103 in completing context understanding and accurate responses.

[0052] It should be noted that the RAG knowledge base module 102 does not directly participate in the answer generation process. Instead, by providing enterprise knowledge fragments that are closely related semantically to the user input, it provides a reliable, controllable, and traceable semantic context basis for the large language model core module 103, ensuring that the subsequent generated answer content is based on the enterprise's internal knowledge rather than relying solely on the pre-trained parameters of the language model, thereby effectively reducing the hallucination risk and enhancing the accuracy and professionalism in enterprise-level scenarios.

[0053] Therefore, the RAG knowledge base module 102 plays the role of a semantic retrieval bridge in the entire system and is a key connection link between the user's intention and enterprise knowledge. Its technical implementation relies on the coordinated action of semantic embedding consistency, vector indexing mechanism, similarity measurement model, and fragment scheduling strategy to ensure that the entire question-and-answer process has stable, efficient, and scalable knowledge support capabilities.

[0054] Enterprise knowledge fragments refer to text units extracted from various internal enterprise knowledge carriers and having the ability of independent semantic expression, usually the smallest available information units constituting the enterprise knowledge system. These fragments can be sourced from structured data (such as project name and responsible person records in a database), semi-structured data (such as business logs in JSON format, OA approval process information), or unstructured data (such as meeting minutes, training documents, product manuals, operation manuals, internal notices, etc.). For example, the description "Product A in the Northeast region had a 25% year-on-year increase in Q3 2023" in a quarterly sales report in PDF format can be segmented and extracted as an enterprise knowledge fragment for subsequent semantic matching and question-and-answer support. Another example is the regulation in the employee handbook "New employees must complete the three-level safety training within 30 days of joining the company", which can also be used as an enterprise knowledge fragment with independent meaning to provide an accurate basis for the user's inquiry "When do I need to complete the training". Enterprise knowledge fragments usually maintain a mapping relationship with their source documents to achieve traceability and context supplementation during use.

[0055] The large language model core module 103 is used to receive the natural language input of the user and the enterprise knowledge fragments, perform semantic understanding and task parsing, generate a preliminary answer containing knowledge response content, and identify whether the natural language input of the user contains a schedule operation request; Among them, the large language model core module 103 is also used to receive the result of context fusion, generate a comprehensive reply, and output it to the user side to complete the question-and-answer response.

[0056] The core module 103 of the large language model is the core intelligent processing unit in the intelligent collaborative Q&A system for enterprise-level schedule planning and knowledge base. It is mainly responsible for deeply semantic understanding and task logic parsing of the user's natural language input and enterprise knowledge fragments, and then realizing knowledge response generation and operation intention recognition, and coordinating subsequent modules in the system to jointly complete the final Q&A response when necessary.

[0057] During the operation of the system, the core module 103 of the large language model first receives the natural language input from the user and the enterprise knowledge fragments retrieved by the RAG knowledge base module 102. The natural language input of the user is generally an open-ended question, an imperative expression, or a mixed request, and may involve various scenarios such as enterprise knowledge consultation, historical event query, task scheduling, schedule change, etc. The enterprise knowledge fragments are knowledge support texts with semantic relevance, usually characterized by high information density, clear semantics, and complete context. The core module 103 of the large language model takes the two together as input to construct a context corpus for driving the language model to perform intelligent reasoning and answer generation.

[0058] In implementation, the core module 103 of the large language model can adopt the current mainstream natural language generation model architecture, such as open-source models based on the Transformer structure (such as ChatGLM, Baichuan, MOSS, DeepSeek, etc.) or enterprise-customized fine-tuning models. This module has the ability of long-text semantic modeling, and can capture keywords, context relationships, and intention expressions in the user input through the self-attention mechanism, and form semantic alignment with professional terms and factual information in the knowledge fragments, so as to generate a preliminary answer containing knowledge response content. The preliminary answer is not just a surface text output, but a problem-solving result with clear structure and semantic fit, which can accurately respond to the query requirements mentioned by the user.

[0059] During the process of generating the preliminary answer, the core module 103 of the large language model also executes the task parsing process in parallel. Task parsing includes but is not limited to multiple subtasks such as intention recognition, entity extraction, operation type determination, parameter extraction, and condition restriction judgment. This process is based on the semantic analysis of the natural language input to judge whether the natural language input of the user contains a schedule operation request. If so, the module will construct a logical representation structure to identify the task type (such as query, create, update), target entity (such as meeting, reminder, collaboration object), timing requirements (such as "next Monday morning"), and other constraint conditions.

[0060] Once it is determined that the user request contains a schedule operation request, the large language model core module 103 will no longer directly output the final response. Instead, as a coordinator, it will pass the task parsing result to the multi-agent collaboration module 104, which will complete the specific task execution. At the same time, the large language model core module 103 will retain the preliminary answer content to participate in the generation of the final response after obtaining the operation result.

[0061] After the multi-agent collaboration module 104 completes the task execution and returns the schedule processing result, the large language model core module 103 also needs to receive the fusion result generated by the context fusion module 105. This fusion result is based on the preliminary answer and the task processing result, and is unified and integrated through a semantic structure consistency strategy, and already has high semantic coherence and information integrity. The large language model core module 103 performs pragmatic judgment and language optimization on this fusion result, finally generates a comprehensive response that meets the user's expectations, and outputs this comprehensive response to the user side to complete the entire question-and-answer response process.

[0062] It should be noted that to ensure the stability and accuracy of the large language model core module 103 when processing complex, multi-round, and multi-modal enterprise queries, the module should be equipped with a Prompt template management, model parameter control, output stability calibration mechanism, and an exception response fallback mechanism to improve the practicality and engineering deployment ability of the system.

[0063] In summary, the large language model core module 103 not only undertakes the responsibilities of semantic understanding, knowledge invocation, and language generation, but also realizes an intelligent collaborative question-and-answer process that integrates knowledge and operations through the linkage with the task execution module and the context fusion module. It is the core technical support for the intelligence, practicality, and enterprise adaptability of this system.

[0064] The multi-agent collaboration module 104 is used to disassemble the schedule operation request into one or more subtasks according to the result of task parsing and schedule them to the corresponding functional agents for execution when it is recognized that the natural language input of the user contains a schedule operation request, and at least includes invoking the schedule management agent to access enterprise schedule data and generate task processing results.

[0065] The multi-agent collaboration module 104 is a key functional module for processing schedule operation requests in an intelligent collaborative question-and-answer system for enterprise-level schedule planning and knowledge base. Its main role is to further disassemble the request into one or more subtasks according to the structured result of task parsing on the premise that the large language model core module 103 recognizes that the natural language input of the user contains a schedule operation request, and orderly schedule each subtask to the corresponding functional agent for execution to achieve the automatic response and operation implementation of the system to the user request.

[0066] This module is typically built on an extensible agent framework. Each functional agent is defined as an independent task execution unit with clear functional boundaries and data access permissions. Specifically, after the large language model core module 103 completes semantic understanding and task parsing of the user's natural language input, if it determines that the input belongs to or contains a scheduling operation request, such as "Help me schedule a meeting with Manager Zhang at 3 pm tomorrow", it will structure the task semantics into a set of parsing parameter sets including operation type (such as "Create meeting"), time parameter (such as "3 pm tomorrow"), object entity (such as "Manager Zhang"), location or meeting room, etc. After receiving this parsing parameter, the multi-agent collaboration module 104 disassembles the request into one or more logical subtasks, such as "Find Manager Zhang's available time", "Reserve a meeting room", "Send a meeting notice", etc., according to the preset task scheduling rules or dynamic task mapping mechanism.

[0067] After the task disassembling is completed, the multi-agent collaboration module 104 distributes it to the preset functional agents according to the type and execution requirements of the subtasks. The functional agents can be a schedule management agent with database access capabilities, or an API proxy agent connecting to third-party office systems (such as Exchange, DingTalk, Feishu, enterprise WeChat calendar module, etc.). The system realizes the interaction with the agents through methods such as RESTful interfaces, message queues, or local function calls. After receiving the task instruction, the schedule management agent will access the enterprise internal schedule database or calendar service, extract the time arrangement information of relevant personnel from it, judge whether the target time period is available, and perform operations such as creating, modifying, or canceling the meeting according to the meeting arrangement rules. At the same time, it generates a structured task processing result, which usually contains operation result status (success or failure), created event ID, meeting link, schedule summary text, etc.

[0068] The multi-agent collaboration module 104 also needs to have a task status monitoring and callback mechanism, which can collect the execution results of each subtask in a timely manner after each functional agent completes the task execution, and encapsulate and summarize them in a standard format for subsequent integration processing by the context fusion module 105 and the large language model core module 103. In addition, to improve the stability and robustness of execution, the multi-agent collaboration module 104 can further introduce an error handling mechanism and task retry logic. For example, after the meeting room reservation fails, it will automatically try alternative times or meeting rooms and feedback the results to the system.

[0069] To support the traceability and auditability of the task process, the multi-agent collaboration module 104 also needs to record the key steps in each task parsing and scheduling process, including task disassembling parameters, distribution paths, execution status, exception information, and user feedback, etc., providing a data basis for system operation and maintenance and agent performance evaluation.

[0070] In summary, the multi-agent collaboration module 104 not only undertakes the responsibility of converting the language understanding result into an executable operation, but also constructs an automated control center for enterprise schedule task processing in a modular and orchestratable manner, ensuring that the system can accurately, efficiently, and securely call enterprise resources after receiving a complex natural language request, complete the execution of structured tasks and provide feedback, laying a data foundation for the generation of subsequent comprehensive responses.

[0071] Furthermore, the multi-agent collaboration module is also used for: Based on the schedule operation requests in the user's natural language input recognized by the large language model core module, match a task graph template corresponding to the semantic structure of the schedule operation requests from a preset semantic-driven task decomposition mapping table, and generate a task graph including multiple task nodes and their dependencies according to the task graph template; Semantically parse each task node in the task graph, extract the target operation type, target entity constraint, and context time parameter corresponding to each subtask, and generate a task description object containing instruction parameters; Input the task description object into the task scheduling engine, schedule each task node orderly according to the dependencies in the task graph, and map each task node to the corresponding functional agent, so as to call at least the schedule management agent to execute the task operations corresponding to the task nodes; Collect the task execution results returned by the functional agents, and structurally aggregate all task execution results according to their logical structure in the task graph, and output a set of task processing results for the context fusion module to perform fusion processing.

[0072] In the intelligent collaborative Q&A system for enterprise-level schedule planning and knowledge base of the present invention, the multi-agent collaboration module not only undertakes the function of converting schedule operation requests into executable subtasks, but further realizes the process of mapping from the natural language intention structure to the task graph structure, and through the semantic parsing and task scheduling mechanism, efficiently delivers each task node in the task graph to the corresponding functional agent for processing, and finally completes the structural aggregation of task processing results, providing complete and consistent operation execution information support for the subsequent context fusion module.

[0073] The technical solution provided in this embodiment first relies on the semantic understanding function of the core module of the large language model for the user's natural language input. When the system detects that the natural language input contains a schedule operation request, such as "Arrange a project review meeting with Manager Wang and the finance department next Tuesday afternoon", the multi-agent collaboration module will receive the parsing result of the core module of the large language model, which usually includes the operation intention type (such as creating a meeting), participating objects (Manager Wang, finance department), time conditions (next Tuesday afternoon), and other context cues.

[0074] To convert this natural language request into a structured task representation executable within the system, a set of semantic-driven task decomposition mapping tables are preset in the multi-agent collaboration module. This mapping table uses semantic structures as matching keys to map different types of natural language input structures to corresponding task graph templates. For example, the "Arrange + object + time" structure will match a standard task graph template with four nodes: "Check the schedule availability of participants → Query the meeting room → Create the meeting → Send the notification". The system can use a BERT-based classifier or rule tree parsing during the semantic matching process to discriminate the sentence structure, ensuring a high structural similarity between the input and the task graph.

[0075] After matching the task graph template, the system generates an initial task graph, where each task node represents an independent executable subtask, and also includes the sequential execution dependencies between tasks. Next, the system performs semantic parsing operations on each task node in this task graph, extracts the corresponding target operation type (such as meeting room query, schedule confirmation, notification push), target entity constraints (such as the personnel and meeting room resources involved), and context time parameters (such as time window, specific time point, duration, etc.), and encapsulates these parsing results into task description objects. Each task description object is a structured data structure that contains all the parameters required for task execution and has good interpretability and executability.

[0076] After the task description objects are constructed, the system hands them over to the task scheduling engine for scheduling. The task scheduling engine determines the sequential and concurrent relationships of task execution based on the node dependencies defined in the task graph. For example, "Query the schedule of personnel" must be executed before "Create the meeting", while "Send the meeting notification" can be executed in parallel with other nodes. The scheduling engine also calls the corresponding functional agents according to the task type, currently including at least the schedule management agent, and may also include the meeting room resource management agent, personnel status synchronization agent, notification distribution agent, etc. Each functional agent, as an independent adjustable module, receives the task description object through a unified API interface, accesses the corresponding data resources (such as the enterprise calendar system, meeting resource platform, etc.), and executes the specific operations corresponding to the task nodes.

[0077] After all task nodes are executed, each functional intelligent agent will return the task execution results in a standardized format, which usually includes the operation status (success, failure, conflict), the operation object identifier (such as meeting ID, conference room number), the time confirmation information (such as the finally determined time period), and the prompt information, etc. The multi-agent collaboration module summarizes these execution results and performs structured aggregation processing according to the initial task graph logical structure, that is, without destroying the task dependency relationship, generating a complete, ordered, and context-unified set of task processing results. This set not only retains the result information of each subtask execution, but also represents the execution process and operation logic in a nested structure or critical path manner, ensuring that the subsequent context fusion module can perform semantic perception and content integration on it.

[0078] This series of processing flows is highly closed-loop logically: from the structured parsing of natural language, the generation and node distribution of the task graph, to the operation execution and result summary of the functional intelligent agent, the output of each step is directly used by the next step, and there is a stable data dependency relationship between each processing unit. The technical path described in this embodiment is significantly superior to the traditional static rule-based system, not only having semantic self-adaptability, but also being extensible to support more task graph templates and intelligent agent types, applicable to various complex collaborative operation requests in enterprises, and is an indispensable core technical support for constructing the intelligent collaborative question-answering ability of the present invention.

[0079] Furthermore, the process of the multi-agent collaboration module performing structured aggregation includes the following steps: Before the task scheduling engine schedules task nodes to the functional intelligent agent, construct a global context state object, which includes the time window, entity reference mapping, and user role information extracted from the user's natural language input; During the concurrent scheduling of multiple task nodes, inject the global context state object into the running environment of each functional intelligent agent, so that each functional intelligent agent can perform parameter interpretation and task execution based on consistent context state constraints when executing tasks; After receiving the task execution results of the functional intelligent agent, calculate the task execution confidence score based on the semantic fit degree between the historical execution record of each functional intelligent agent and the current task description object, and attach the confidence score to the corresponding task execution result; When constructing the set of task processing results, perform weighted screening and priority sorting on the candidate results according to the confidence scores in each task execution result, so as to generate a structured set of task processing results with the highest overall semantic consistency and context fit degree.

[0080] In the intelligent collaborative Q&A system for enterprise-level schedule planning and knowledge base described in the present invention, during the process of performing structured aggregation in the multi-agent collaboration module, it is not simply a matter of splicing or arranging the task execution results returned by the functional agents in sequence. Instead, based on the dynamic cross-evaluation among the dialogue context, user intention structure, agent behavior history, and task execution status, a process of constructing a task result set with high semantic consistency and global context adaptability is completed. The technical solution provided in this embodiment focuses on establishing a result integration mechanism of "context state-driven + confidence weighted sorting" to minimize semantic conflicts, state drift, and deviation from user expectations among the execution results of asynchronous agents, and significantly improve the structural clarity and logical unity of the final reply.

[0081] During the implementation process of the present invention, structured aggregation is not a post-processing module but a state coordination mechanism throughout the entire multi-task scheduling and execution life cycle. First, before the task scheduling engine is ready to distribute the task nodes in the task graph to each functional agent, the multi-agent collaboration module constructs a global context state object based on the user's natural language input generated by the large language model core module. This object is a state container that can be shared by multiple sub-tasks, mainly including time window information extracted from the user input (such as "next Wednesday afternoon" or "the last working day of this month"), entity reference mapping relationships (such as "he" refers to "Manager Wang", "that meeting" refers to the "quarterly review meeting" mentioned earlier), and user role information (such as the initiator being a project manager, department head, or ordinary employee). These information constitute restrictive constraints in actual semantic reasoning and are the premise for ensuring the consistency of task execution.

[0082] Once the global context state object is established, the system injects it into the running environment of each functional agent. When the task graph nodes are distributed, each agent not only receives the corresponding task description object but also automatically loads this context state object, enabling it to execute based on a unified semantic background during processes such as parameter interpretation, query scheduling, and command execution. For example, when arranging a meeting, if the user role is a department manager, the system can automatically apply a higher-level meeting permission template; if there are holidays in the time window, the schedule management agent can make an early avoidance judgment. This mechanism ensures that when multiple task nodes are scheduled concurrently, each functional agent can still maintain the consistency of task parameter interpretation and the logical unity of behavior output in different context scenarios.

[0083] After all functional agents complete the operations corresponding to the task nodes and return the task execution results, the multi-agent collaboration module will collect these results and perform in-depth semantic analysis. First, the system will calculate the task execution confidence score for each execution result based on the matching degree between the historical behavior records of each functional agent (such as indicators like the execution accuracy rate of past tasks, average response time, and positive correlation with user feedback) and the semantic structure in the current task description object. This score is not set by hard coding, but is generated by combining dynamic factors such as content semantic fit, context match, and model prediction distribution. A multi-factor fusion algorithm can be used, such as a weighted classifier or a clustering cosine similarity scoring model. Each task execution result will be attached with this confidence score as an additional attribute to guide subsequent aggregation optimization.

[0084] After all results and scores are prepared, the system will enter the key stage of structured aggregation, that is, constructing the final task processing result set. This process will maintain the overall structural logic according to the original dependency order in the task graph, and introduce the confidence score as the sorting basis among candidate results at the same level to perform weighted screening and priority sorting operations. Specifically, if multiple candidate results provide different execution feedbacks for the same subtask, such as multiple time options for meeting arrangements or meeting room conflicts, the system will retain the entry with the highest confidence score, or perform fusion among candidates with similar semantics (for example, return two alternative time periods for the user to confirm). During this process, the system will also re-check the dependency consistency between each node to ensure that the downstream task path changes caused by the selection of high-confidence results can be reorganized and re-sorted in a timely manner to ensure that the entire task processing result set has logical integrity and causal consistency in structure.

[0085] The finally output task processing result set not only reflects the intention of the user's original input semantically, but also maintains the structural characteristics of clear information, minimum conflicts, and consistent context at the execution result level, enabling the subsequent context fusion module to seamlessly connect when integrating knowledge answers and operation feedback. This fusion processing effect is far better than the "serial execution + result splicing" mechanism in traditional static process systems, and is particularly suitable for complex operation requests involving multi-role collaboration, condition-triggered execution, and time-sensitive scheduling in enterprise-level question answering.

[0086] Furthermore, when calculating the task execution confidence score for each task execution result, the multi-agent collaboration module adopts the following scoring function: Among them, represents the The task execution confidence score corresponding to a task execution result, which is used to measure the comprehensive credibility of the task execution result relative to other candidate results during the multi-agent collaborative scheduling process, and the value range is positive real numbers. The subscript corresponds to the th valid execution result returned by the functional agents mapped by multiple task nodes generated by task graph decomposition during a certain user request processing, where is the number of all subtasks executed under this request.

[0087] represents the semantic structure matching score between the th task execution result and its corresponding task description object. Its value comes from the reverse score of the graph embedding cross-entropy calculated after structurally aligning the semantic graph constructed by the task description object (including subject-predicate-object structure, limiting conditions, target objects, etc.) with the semantic graph implied by this execution result, measuring the semantic overlap degree between the two at the structural level. The unit is , and the larger the value, the more consistent the semantic structure.

[0088] represents the th context adaptation index of the functional agent under the currently loaded global context state object, which is specifically defined as the proportion of successfully completing context-sensitive tasks by this agent when loading similar context conditions (such as time window, user role, reference mapping) in the past times. The unit is percentage (%), which is used to measure the familiarity of this agent with the enterprise context and the consistency of parameter interpretation.

[0089] represents the th path depth of the task node in the current task graph structure, that is, the shortest hierarchical distance of this task node relative to the root node, which is used to represent the information hierarchical position of this task in the overall semantic reasoning chain. The unit is layer, and its value is a natural number.

[0090] represents the arithmetic mean of the path depths of all task nodes in the current task graph, which is used as the structural center reference value of the task graph to evaluate whether a certain task node is a structurally deviated node.

[0091] :represents the th offset amplitude of the task in the path structure. If this value is large, it means that this task is a non-main trunk node and affects the scoring weight.

[0092] represents the The semantic main line consistency coefficient between the task execution result and the main intention mentioned in the user's natural language input, whose value is A dimensionless score within the range, which is derived from the weighted cosine similarity calculation between the semantic vector generated by the task execution result and the intention vector generated after parsing the user input. The closer it is to 1, the closer the task responds to the user's true intention.

[0093] Indicates the semantic main line deviation exponential factor, which is used to amplify the penalty effect on confidence when the main line consistency is insufficient. It is an adjustable positive real constant, determined by experience during the system tuning process.

[0094] Indicates the Historical behavior confidence weight of the th functional agent, whose value consists of two parts: one is the average value of the user feedback scores received by the agent within a certain past period (such as the last 100 tasks), and the other is the execution stability confidence factor of the current task type. After multiplying the two, it is normalized to the

[0095] This scoring function fuses multiple technical elements such as semantic structure consistency, context adaptation ability, structure path dominance, and intention response accuracy, and flexibly differentiates the scoring results through non-linear function transformation and exponential adjustment control, ensuring that the final confidence not only conforms to the objective structural constraints of logical reasoning, but also reflects the consistency of language semantics and the user intention response degree, and is applicable to the dynamic intelligent question-answering system in the enterprise multi-role and multi-task scenario.

[0096] In the actual implementation process, when the user sends a composite natural language request through the terminal, such as "Please arrange a product review meeting with the Marketing Department and Manager Zhang next Wednesday afternoon and tell me the minutes of the meeting in the same period last year", the system will trigger two processes: the knowledge-based query and the schedule operation at the same time. The large language model core module first parses the main semantic line keywords of the user as operation intentions such as "arrange a meeting", "next Wednesday afternoon", "Manager Zhang", etc., and at the same time identifies the knowledge targets involved in the query part such as "the same period last year", "meeting minutes", etc. Subsequently, in the multi-agent collaboration module, the task is decomposed into multiple task nodes, including "query the minutes of the review meeting last year", "check whether Manager Zhang and the Marketing Department are available next Wednesday afternoon", "allocate a meeting room", "create a meeting object", "generate a meeting notice", etc., and are respectively handed over to different functional agents for execution.

[0097] Suppose now for the The execution result of one task (e.g., "allocate a meeting room") needs to calculate the confidence score of task execution. The system will extract and calculate each parameter item by item according to the aforementioned scoring function.

[0098] First, the system constructs a semantic graph of this task, marking nodes such as the verb "allocate", the object "meeting room", the condition "next Wednesday afternoon", and the participants "Marketing Department, Manager Zhang", and then performs structural alignment with the semantic graph of the information returned in the task execution result (such as "Meeting Room A - 302, 14:00 - 16:00"), and uses the graph embedding method for similarity evaluation to finally obtain the reverse graph cross-entropy score. , for example, it is .

[0099] Secondly, the system counts the proportion of successful task completions of this agent in the past 100 executions of the "meeting room allocation" task under the condition of loading a specific time window and entity reference context, and finds that there are 92 successful times to obtain the context adaptation index. .

[0100] Then, calculate the path depth of this task node in the task graph. . Assume that the root node is "parse user input", and the path is "arrange a meeting check availability allocate a meeting room", then this node is at the 3rd layer; at the same time, count the average path depth of all task nodes as to obtain the structure deviation value. .

[0101] The system also generates an intent vector for the user's original input through the intent extraction mechanism, mainly based on "arrange a meeting" and its constraints, and the semantic center is the time - participant pairing. In the task execution result, a semantic vector is also generated, indicating "allocate space resources for specific time and personnel". Calculate the weighted cosine similarity of these two vectors to obtain the main line consistency score. .

[0102] Regarding the main line deviation index factor , it is set to 5 in the system default configuration. This value is set by the system according to actual business requirements and error tolerance during the debugging phase.

[0103] Finally, the confidence weight of the agent's historical behavior is obtained by multiplying the past task satisfaction score (with an average of 0.92) by the confidence benchmark factor (0.95) of the current "meeting room scheduling" task type, that is .

[0104] Substitute the above items into the scoring formula: As a result, the execution result of the task "allocate meeting room" will be given a confidence score . This score will be jointly input into the structured aggregation stage together with the scores of other subtasks, guiding the system to select which subtask results should be preferentially retained, fused, or presented to the user as alternative options.

[0105] The context fusion module 105 is used to receive the preliminary answer and the task processing result, and perform context fusion on the two based on the semantic structure consistency strategy.

[0106] The context fusion module 105 is a key functional component in the intelligent collaborative Q&A system for enterprise-level schedule planning and knowledge base, which is used to integrate the results of language understanding and task execution. Its main role is to receive the input of two types of information, namely the preliminary answer and the task processing result, and through the fusion process based on the semantic structure consistency strategy, generate a context fusion result with coherent content, complete information, and unified logic, so as to provide accurate and consistent semantic support for the final comprehensive answer.

[0107] The premise for the module to work is that during the operation of the system, the large language model core module 103 has generated a preliminary answer. This preliminary answer performs semantic understanding and content generation based on the context composed of the user's natural language input and the enterprise knowledge fragments, and usually contains the answer content for knowledge-based queries. At the same time, if the large language model core module 103 identifies that the user's natural language input contains a schedule operation request during task parsing, it will complete the task decomposition and scheduling through the multi-agent collaboration module 104, and the corresponding functional agent will execute the specific operation, thus generating a task processing result. This task processing result may include confirmation information for successful meeting creation, feedback on time conflicts, task execution status descriptions, etc., and is usually in the form of structured or semi-structured data.

[0108] The main task of the context fusion module 105 is to effectively fuse the above two types of texts and data content with different sources and semantic structures, so that they are consistent in terms of language style, logical order, semantic connection, etc., thereby improving the overall readability, integrity and user experience of the final response. In implementation, the context fusion module 105 will first perform content annotation and semantic partitioning on the preliminary answer and the task processing result, identify elements such as the theme unit, semantic center, time reference, entity reference, etc. of the two through feature extraction, and perform comparison and alignment based on the semantic structure consistency strategy. This consistency strategy can be implemented by means of rule-driven, semantic graph matching or small model-assisted evaluation. Its core goal is to judge whether there are problems such as conflicts, repetitions, omissions or logical jumps between the two inputs at the semantic level, and make adjustments through language processing means such as reordering, replacement, extension, connection, etc.

[0109] In specific implementation, the context fusion module 105 can perform operations such as automatic summarization, semantic insertion, paraphrasing and reconstruction by designing a Prompt template or calling a small-scale language model, so that the fused text not only covers the knowledge query raised by the user, but also clearly feedbacks the execution status of the task, forming a semantic response with the characteristics of a two-way closed loop. For example, when the user asks "Please help me arrange a meeting with Manager Wang tomorrow afternoon and tell me the sales volume of the products he was in charge of last year", the preliminary answer may be "The sales volume of the products in charge of Manager Wang in 2023 was XX billion yuan", and the task processing result may be "Conference room A has been reserved for you at 3:00 pm tomorrow, and the meeting arrangement with Manager Wang is successful". The context fusion module 105 integrates these two parts through cohesive expressions as "The sales volume of the products in charge of Manager Wang in 2023 was XX billion yuan. At the same time, conference room A has been reserved for you at 3:00 pm tomorrow, and the meeting arrangement has been successfully completed", so as to complete the fusion output with consistent semantics and complete information.

[0110] In addition, the context fusion module 105 also needs to have the ability of error detection and noise reduction to ensure that when there are inconsistencies, omissions or conflicts between the preliminary answer and the task processing result, it can output prompt information or call a preset fallback template for structural reorganization to prevent incorrect information from being directly transmitted to the user side. At the same time, this module also needs to retain the traceability of the original semantic unit, which is convenient for the core module 103 of the large language model to perform semantic correction or content supplementation according to the user feedback when generating a comprehensive response.

[0111] In summary, the context fusion module 105 plays a crucial role in the system of the present invention. It not only connects the static responses generated based on knowledge and the dynamic responses based on task execution, but also realizes the deep fusion of natural language and the system behavior results through the semantic structure consistency strategy, ensuring that the comprehensive reply output to the user is both accurate and credible, and natural and context - compliant in expression, thus significantly enhancing the overall intelligent interaction ability of the system and the adaptability of enterprise - level applications.

[0112] Furthermore, the context fusion module is further configured to: Perform semantic structure extraction operations on the preliminary answer and the task processing result respectively, obtain their syntactic structures, anaphoric chains, time - relation labels, and context position vectors, and generate semantic structure representation objects; Based on the matching relationship between the semantic structure representation objects, call a semantic structure consistency scoring function, and the semantic structure consistency scoring function calculates the structural consistency score between the two based on syntactic sub - tree alignment, nested structure overlap degree, and attention weight cross - mapping; When the structural consistency score is lower than a preset fusion threshold, trigger a conflict resolution mechanism, and the conflict resolution mechanism includes rewriting the conflicting content in the preliminary answer or the task processing result based on the context semantic priority, making its logical connection, primary and secondary clear, and cause - effect distinct; Based on the rewritten semantic structure, perform semantic dominance order determination. According to the intention direction and information order extracted from the user's natural language input, decide whether to use the preliminary answer or the task processing result as the main semantic content, and generate a fused text structure while maintaining the original syntactic structure, which is used as the input for the large - language model core module to generate a comprehensive reply.

[0113] In the intelligent collaborative Q&A system for enterprise - level schedule planning and knowledge base of the present invention, the context fusion module is not only used to simply combine the knowledge response content and the task processing result, but further introduces deep fusion mechanisms such as semantic structure modeling, structural consistency judgment, conflict content rewriting, and semantic dominance ranking, so as to achieve the unity of semantic coordination, structural clarity, and context consistency in natural language generation.

[0114] During system operation, after the user enters a composite natural language input containing knowledge query and operation requests, the large language model core module first generates the preliminary answer to respond to the user's knowledge-based questions. At the same time, through the processing of the multi-agent collaboration module, the system obtains the task processing result to respond to the relevant instruction requests. These two types of content come from different system sub-processes, with different language styles, structural layouts, and semantic focuses. Therefore, they cannot be simply spliced together, otherwise it is easy to cause chaotic reply structures, logical contradictions, and even semantic fragmentation. To solve this problem, the context fusion module will perform semantic structure extraction operations on these two pieces of content respectively.

[0115] This semantic structure extraction operation adopts a technical path based on a hybrid of deep language modeling and symbolic parsing. It can extract a complete syntactic structure tree from the text (for example, through a dependency parser or the intermediate attention matrix in a Transformer), and annotate the logical relationships between each entity, action, and time node. At the same time, through the anaphora chain recognition module, it identifies the specific entity pointed to by pronouns such as "he", "it", "that", etc., and reconstructs the internal entity chain of the text; through the time expression recognition module, it extracts the relationship between time phrases and events, such as the sequence between "the meeting last Monday" and "the topics proposed by Manager Zhang"; in addition, it also extracts the context position vector of the sentence in the original text, identifying its logical paragraph, sub-task path position, and information role (background description, execution result, suggestive content, etc.). All these pieces of information together constitute a semantic structure representation object for subsequent consistency comparison.

[0116] After the above-mentioned structure extraction is completed, the context fusion module will evaluate the matching relationship between the two semantic structure representation objects and call a semantic structure consistency scoring function to score them. This scoring function is no longer a traditional string similarity match, but an index system built on multiple levels of semantic structures, mainly including three parts: one is the alignment degree of syntactic sub-trees, that is, the similarity score of the semantic backbones and their modifying structures in the two texts at the dependency tree level; the second is the overlap degree of nested structures, for example, comparing whether there are inconsistencies in the hierarchical structures of conditional clauses, causal relationships, and adverbial phrases in the two, or conflicts in the master-slave relationships; the third is the cross-mapping based on the attention weights of the Transformer model, evaluating the relative semantic attention overlap degree of the two inputs in the shared attention space. These indicators together constitute the structure consistency score, which is used to determine whether the preliminary answer and the task processing result can be directly fused in the current context.

[0117] If the structural consistency score is lower than the fusion threshold set by the system, for example, due to contradictory subject statements, incorrect time sequence logic, or unclear references in the two pieces of content, the system will trigger the conflict resolution mechanism. The conflict resolution mechanism is a key process in fusion quality control, and its core lies in strategically rewriting the semantic expression without changing the factual content. The system will first determine the primary and secondary relationships in the two pieces of content based on the semantic priority in the context, which is usually carried out according to the intended structure or focus words in the user input. For example, if "want to know the summary report" in the user input comes before "arrange a meeting", the system takes the preliminary answer as the main semantic content and retains its original language form, while paraphrasing the conflicting part in the task processing result. For example, it rewrites "The meeting has been arranged" as "This issue will be further discussed in the arranged meeting" to retain the information but avoid structural confrontation. The entire rewriting process relies on the trained text reconstruction model and context enhancement mechanism to ensure that the output is both faithful to the original meaning and natural in language.

[0118] After the rewriting is completed, the system needs to determine the structural order of the final fusion output, that is, perform the semantic dominance order determination. This step is based on the principle of the order of the questions and tasks in the user's natural language input, and at the same time considers the logical fluency of the language presentation after fusion. If the user first focuses on knowledge questions, the system places the preliminary answer at the front and connects the subsequent task execution information in a connecting language form; on the contrary, if the user is dominated by operation requests, the task processing result is placed in the main position, and additional knowledge supplements are added in an inserted structure or postscript. At this time, the system will try to keep the original syntactic structure intact, avoid generating redundant or unnatural repeated sentences, and output the fused text structure, which is complete, well-organized, and semantically consistent.

[0119] The fused text structure will be used as the final Prompt input to the core module of the large language model for further processing. Combining the language generation ability and context prompt mechanism, it generates a comprehensive response with high semantic consistency, natural tone, and clear structure, ensuring that the output result accurately restores the system processing process in content and conforms to the cognitive logic of human users in expression. It is one of the final forms of the system output response.

[0120] Furthermore, the generation process of the fused text structure in the context fusion module includes the following steps: Extract the semantic main line keywords, target behavior categories, and context trigger conditions involved in the user's natural language input, and inject them into the fusion control module as fusion guiding vectors; Select the language generation template that best matches the structure from a predefined Prompt template library according to the matching result between the semantic type of the fusion guidance vector and the content tags of the fused text structure, where the Prompt templates include sequential templates, embedded description templates, segmented combination templates, and semantic interleaving templates; Perform position mapping binding on the fused text structure and the selected language generation template, fill each semantic paragraph into the specified position according to the structure slots defined in the template, and at the same time adjust the generated intonation, subject-verb order, and conjunction configuration according to the context content type; Use the filled template text structure as the fused text structure, input and send it to the core module of the large language model, drive it to generate a comprehensive reply with high semantic consistency, natural tone, and clear structure, and output it to the user side.

[0121] In the intelligent collaborative Q&A system for enterprise-level schedule planning and knowledge base described in the present invention, to achieve high-quality integration between the knowledge-based response content corresponding to the user's natural language input and the task processing results, the context fusion module not only performs semantic consistency control and conflict mitigation on the content at the structural level, but further completes the standardized expression of the fused text structure through a template-driven generation mechanism, thereby ensuring that the final output content not only has language naturalness, but also meets the actual requirements of semantic dominance and structural clarity. The key link in the generation process of the fused text structure provided in this embodiment, the technical essence of which lies in unifying and organizing the structurally complex and heterogeneous-source semantic units into a Prompt input that conforms to the language generation specification through the combination of semantic guidance and template binding, driving the core module of the large language model to generate a highly usable and complete final comprehensive reply.

[0122] In this embodiment, the generation process of the fused text structure starts from the extraction of semantic elements in the fusion control stage. The system first performs structural analysis on the key expressions in the user's natural language input, and extracts the semantic main line keywords (such as verb central words like "arrange", "query", "summarize", etc.), target behavior categories (such as operation types like "create a meeting", "consult a report", "remind to execute", etc.), and context trigger conditions (such as time, theme, and object restriction information like "next Monday", "about the new product", "with Manager Wang"). These elements are encoded as high-dimensional semantic vectors, collectively referred to as fusion guidance vectors, which are used to guide the subsequent template selection process. This process ensures that the system can identify the main content and operation focus of the information expression from the user's input language, thereby providing clear structural guidance.

[0123] Subsequently, the fusion control module semantically matches the fusion guidance vector with the content tags in the fused text structure output by the context fusion module. These content tags are usually structural markers attached to different semantic paragraphs in the previous fusion process, such as "knowledge response segment", "task execution result segment", "precautions segment", "secondary background segment", etc. This annotation mechanism helps the system identify the structural functions of each semantic segment in the fused content, and then ensures that the main line semantics is consistent with the structural presentation method during the matching. After the matching of semantic types and content tags is completed, the system selects a language generation template that best suits the current content structure from a predefined Prompt template library. This template library can include various types of template structures. Among them, the sequential template is used for scenarios where the content has clear priorities, such as "knowledge first and then execution"; the embedded description template is suitable for inserting one type of content into another as additional explanation; the segmented combination template can be used for the expression requirements of parallel structures; the semantic interleaving template is suitable for complex contexts where the content is closely intertwined, such as the description of time-driven parallel tasks, etc.

[0124] After determining the appropriate language generation template, the system maps and binds each semantic paragraph in the fused text structure to the structure slots defined in the template one by one. For example, if the template presets slots such as "<main answer paragraph>", "<background description paragraph>", "<task result prompt paragraph>", etc., the system will fill the corresponding paragraphs into the corresponding slot positions according to the content tags of each paragraph in the fused text structure. During this process, the system will also automatically perform language fine-tuning based on the context content type, including adjusting the tone intensity of the generated sentences, the order of the subject-predicate structure, the sentence style (declarative, prompt, summary), and the use of logical connectives, such as reasonably using "therefore", "next", "meanwhile", "please note", etc. to enhance the naturalness of semantic transition and the logical coherence of the language.

[0125] After completing the slot filling, the generated result is the filled template text structure. This structure is an intermediate expression form between the fused semantic content and the generated language form, with both structural clarity, semantic coverage integrity, and language naturalness, and is already qualified to be directly used as the Prompt input for processing by the large language model core module. The system sends it to the large language model core module. After receiving this filled template text structure, the large language model will fine-tune and coherently optimize each paragraph according to its language modeling ability, further improving the naturalness and user readability at the language expression level, and finally generating a comprehensive response content with high semantic consistency, natural tone, and clear structure. This response content will be directly returned to the user side as the system's final response to the user's original natural language input.

[0126] Throughout the process, from the generation of semantic guidance vectors, template matching, content binding to prompt input construction, strict respect and control for the original semantic structure are maintained, avoiding common language problems such as redundancy, jumpiness, and repetition in traditional concatenated responses, and greatly improving the system's language organization ability and output expression quality in complex semantic fusion scenarios. This mechanism is not only applicable to hybrid problem scenarios but also provides a good framework for expanding more structural templates and adapting to multi-domain language styles.

[0127] In the above embodiments, an intelligent collaborative Q&A system for enterprise-level schedule planning and knowledge base is provided. Correspondingly, the present application also provides an intelligent collaborative Q&A method for enterprise-level schedule planning and knowledge base. Please refer to Figure 2 , which is a flowchart of an embodiment of the intelligent collaborative Q&A method for enterprise-level schedule planning and knowledge base of the present application. Since this embodiment, that is, the second embodiment, is basically similar to the first embodiment, the description is relatively simple, and the relevant parts can refer to the partial description of the first embodiment. The method embodiments described below are merely illustrative.

[0128] An intelligent collaborative Q&A method for enterprise-level schedule planning and knowledge base provided by the second embodiment of the present application includes: Step S201: Receive the natural language input of the user, and through a vectorization processing step, convert the natural language input of the user into a first semantic vector, and at the same time convert the pre-stored enterprise knowledge data into a second semantic vector; Step S202: Based on the first semantic vector, retrieve in the vector database the enterprise knowledge fragments corresponding to the second semantic vectors that are semantically related to the first semantic vector; Step S203: Input the natural language input of the user and the enterprise knowledge fragments into a large language model, perform semantic understanding and task parsing, generate a preliminary answer containing knowledge response content, and identify whether the natural language input of the user contains a schedule operation request; Step S204: In the case where the recognition result indicates that the natural language input of the user contains a schedule operation request, according to the result of the task parsing, disassemble the schedule operation request into one or more subtasks, and schedule them to the corresponding functional intelligent agents for execution, at least including calling the schedule management intelligent agent to access enterprise schedule data and generate a task processing result; Step S205: Receive the preliminary answer and the task processing result, and perform context fusion on the preliminary answer and the task processing result based on the semantic structure consistency strategy; Step S206: Input the context-fused content into the large language model again, generate a comprehensive response, and output it to the user side to complete the Q&A response.

[0129] Although the present application is disclosed above with preferred embodiments, it is not intended to limit the present application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the protection scope of the present application shall be subject to the scope defined by the claims of the present application.

Claims

1. An intelligent collaborative Q&A system for enterprise-level schedule planning and knowledge base, characterized in that Comprising: A vectorization processing module, configured to receive a user's natural language input, convert the user's natural language input into a first semantic vector, and convert pre-stored enterprise knowledge data into a second semantic vector; A RAG knowledge base module, configured to retrieve, in a vector database, enterprise knowledge fragments corresponding to the second semantic vector semantically related to the first semantic vector based on the first semantic vector; A large language model core module, configured to receive the user's natural language input and the enterprise knowledge fragments, perform semantic understanding and task parsing, generate a preliminary answer including knowledge response content, and identify whether the user's natural language input contains a schedule operation request; A multi-agent collaboration module, configured to, when it is identified that the user's natural language input contains a schedule operation request, disassemble the schedule operation request into one or more subtasks according to the result of task parsing, and schedule them to corresponding functional agents for execution, at least including invoking a schedule management agent to access enterprise schedule data and generate a task processing result; A context fusion module, configured to receive the preliminary answer and the task processing result, and perform context fusion on the two based on a semantic structure consistency strategy; Wherein, the large language model core module is further configured to receive the result of context fusion, generate a comprehensive response, and output it to the user terminal to complete the question and answer response.

2. The intelligent collaborative Q&A system for enterprise-level schedule planning and knowledge base according to claim 1, characterized in that, The vectorization processing module is further configured to: After receiving the user's natural language input, invoke an embedding model based on the Transformer architecture, convert the user's natural language input into a first semantic vector, and append an input timestamp and a context position index to the first semantic vector to retain the user language input timing information; Before converting the pre-stored enterprise knowledge data into a second semantic vector, select a corresponding semantic modeling strategy according to the format type of the enterprise knowledge data, wherein for structured data, call a rule embedding method, and for natural language documents, call the same embedding model as the first semantic vector to obtain a preliminary second semantic vector; Perform vector normalization processing on the first semantic vector and the preliminary second semantic vector respectively, and calculate a vector consistency index between the two based on cosine similarity in the same semantic space; When the vector consistency index is less than a preset semantic similarity threshold, perform a vector tuning operation on the preliminary second semantic vector to enhance its cross-source semantic alignment ability with the first semantic vector, and the tuned second semantic vector is used for the RAG knowledge base module to retrieve, in the vector database, enterprise knowledge fragments corresponding to the second semantic vector semantically related to the first semantic vector.

3. The intelligent collaborative Q&A system for enterprise-level schedule planning and knowledge base according to claim 2, wherein When the vectorization processing module performs the vector tuning operation, it includes the following steps: Extract the generation time of the enterprise knowledge data corresponding to the preliminary second semantic vector, and embed the generation time into the time dimension channel of the preliminary second semantic vector to construct a temporal second semantic vector with a timing label; Calculate the implicit tense vector representation based on the time expression content involved in the context of the first semantic vector, and perform tense alignment evaluation with the second semantic vector of tense; If the evaluation result shows a significant inconsistency in tense between the two, adjust the time weight parameter of the preliminary second semantic vector, and perform attention enhancement in combination with the context position index to improve its time matching degree; Use the adjusted second semantic vector of tense as the optimized second semantic vector.

4. The intelligent collaborative Q&A system for enterprise-level schedule planning and knowledge base according to claim 1, characterized in that The multi-agent collaboration module is also used for: Based on the schedule operation request in the user natural language input recognized by the large language model core module, match the task graph template corresponding to the semantic structure of the schedule operation request from the preset semantic-driven task decomposition mapping table, and generate a task graph including multiple task nodes and their dependencies according to the task graph template; Perform semantic parsing on each task node in the task graph, extract the target operation type, target entity constraint, and context time parameter corresponding to each subtask, and generate a task description object including instruction parameters; Input the task description object into the task scheduling engine, schedule each task node in an orderly manner according to the dependencies in the task graph, and map each task node to the corresponding functional agent, so as to call at least the schedule management agent to execute the task operation corresponding to the task node; Collect the task execution results returned by the functional agent, and perform structured aggregation on all task execution results according to their logical structure in the task graph, and output a set of task processing results for the context fusion module to perform fusion processing.

5. The intelligent collaborative Q&A system for enterprise-level schedule planning and knowledge base according to claim 4, wherein The process of the multi-agent collaboration module performing structured aggregation includes the following steps: Before the task scheduling engine schedules the task node to the functional agent, construct a global context state object, which includes the time window, entity reference mapping, and user role information extracted from the user natural language input; During the process of concurrently scheduling multiple task nodes, inject the global context state object into the running environment of each functional agent, so that each functional agent can perform parameter interpretation and task execution based on consistent context state constraints when executing tasks; After receiving the task execution results of the functional agent, calculate the task execution confidence score based on the semantic fit degree between the historical execution record of each functional agent and the current task description object, and attach the confidence score to the corresponding task execution result; When constructing the set of task processing results, perform weighted screening and priority sorting on the candidate results according to the confidence scores in each task execution result to generate a structured set of task processing results with the highest overall semantic consistency and context fit degree.

6. The intelligent collaborative Q&A system for enterprise-level schedule planning and knowledge base according to claim 1, wherein The context fusion module is also used for: Perform semantic structure extraction operations on the preliminary answer and the task processing results respectively, obtain their syntactic structures, reference chains, time relation labels, and context position vectors, and generate semantic structure representation objects; Based on the matching relationship between objects represented by the semantic structure, a semantic structure consistency scoring function is called. The semantic structure consistency scoring function calculates the structural consistency score between the two based on syntactic subtree alignment, nested structure overlap degree, and attention weight cross-mapping; When the structural consistency score is lower than the preset fusion threshold, a conflict resolution mechanism is triggered. The conflict resolution mechanism includes rewriting the conflicting content in the preliminary answer or the task processing result based on the context semantic priority to make its logical connection, primary and secondary clear, and cause and effect distinct; Based on the rewritten semantic structure, perform semantic dominant order determination. According to the intention direction and information order extracted from the user's natural language input, decide whether to use the preliminary answer as the main semantic content or the task processing result as the main semantic content, and generate a fused text structure while maintaining the original syntactic structure as the input for the large language model core module to generate a comprehensive response.

7. The intelligent collaborative Q&A system for enterprise-level schedule planning and knowledge base according to claim 6, wherein The generation process of the fused text structure in the context fusion module includes the following steps: Extract the semantic mainline keywords, target behavior categories, and context trigger conditions involved in the user's natural language input, and inject them into the fusion control module as fusion guidance vectors; According to the matching result between the semantic type of the fusion guidance vector and the content label of the fused text structure, select the language generation template that best matches the structure from the predefined Prompt template library. The Prompt template includes sequential templates, embedded description templates, segmented combination templates, and semantic interleaving templates; Perform position mapping binding between the fused text structure and the selected language generation template, fill each semantic paragraph into the specified position according to the structure slots defined in the template, and adjust the generated intonation, subject-verb order, and conjunction configuration according to the context content type; Use the filled template text structure as the fused text structure, input and send it to the large language model core module to drive it to generate a comprehensive response with high semantic consistency, natural tone, and clear structure and output it to the user side.

8. An intelligent collaborative Q&A method for enterprise-level schedule planning and knowledge base, characterized in that, Including: Receive the user's natural language input, and through the vectorization processing step, convert the user's natural language input into a first semantic vector, and at the same time convert the pre-stored enterprise knowledge data into a second semantic vector; Based on the first semantic vector, retrieve the enterprise knowledge fragments corresponding to the second semantic vector semantically related to the first semantic vector in the vector database; Input the user's natural language input and the enterprise knowledge fragments into the large language model, perform semantic understanding and task parsing, generate a preliminary answer containing knowledge response content, and identify whether the user's natural language input contains a schedule operation request; In the case where the recognition result indicates that the user's natural language input contains a schedule operation request, according to the result of the task parsing, disassemble the schedule operation request into one or more subtasks and schedule them to the corresponding functional intelligent agents for execution, at least including calling the schedule management intelligent agent to access enterprise schedule data and generate a task processing result; Receive the preliminary answer and the task processing result, and perform context fusion on the preliminary answer and the task processing result based on the semantic structure consistency strategy; Input the content after context fusion into the large language model again to generate a comprehensive reply, and output it to the user side to complete the Q&A response.

Images