Intelligent collaborative question-answering system and method for enterprise-level schedule planning and knowledge base

Abstract

The present invention relates to the technical field of computer systems for natural language processing or semantic processing, and discloses an intelligent collaborative question-answering system and method for enterprise-level schedule planning and knowledge base. Among them, the system includes modules such as a vectorization processing module and a RAG knowledge base module. The system converts the user's natural language input and enterprise knowledge data into semantic vectors respectively, and retrieves relevant enterprise knowledge fragments from the vector database based on semantic similarity. The large language model core module parses the user's intention, generates a preliminary answer, and identifies whether it contains a schedule operation request. If there is a schedule operation request, the system disassembles the request and distributes it to the corresponding agent through the multi-agent collaboration module to obtain the task processing result. Finally, the preliminary answer and the task processing result are semantically consistent through the context fusion module, and the large language model generates a comprehensive reply and returns it to the user end, realizing a unified intelligent response of enterprise knowledge and schedule services.

Description

Technical Field

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

[0002] In existing technologies, enterprises often use calendar management systems and knowledge base management systems to handle time scheduling and information query tasks, respectively. Some systems incorporate natural language processing or conversational interfaces to enhance the user interaction experience. Meanwhile, the combination of vectorized search and large language models has seen initial application in question-answering systems, enabling semantically based knowledge retrieval and intelligent responses. However, these systems are often focused on a single function and lack unified coordination capabilities.

[0003] The following problems are common in existing technologies: First, the schedule management and knowledge retrieval systems are independent of each other, and users need to switch between multiple platforms, which affects usage efficiency; second, the intelligent question-answering system finds it difficult to combine heterogeneous knowledge within the enterprise with dynamic task requests, and is unable to perform structured parsing and task scheduling on natural language input containing operation-type requests; third, there is a lack of a semantic fusion mechanism between knowledge response results and operation execution results, which leads to the fragmentation of output information, affecting overall consistency and response accuracy.

[0004] In order 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 intelligence level of information acquisition and operation instruction processing.

[0006] This application provides an intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base, including:

[0007] 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;

[0008] A RAG knowledge base module is configured to retrieve, based on the first semantic vector, an enterprise knowledge segment corresponding to a second semantic vector semantically related to the first semantic vector in a vector database;

[0009] A large language model core module is used to receive the user's natural language input and the enterprise knowledge fragment, perform semantic understanding and task analysis, generate a preliminary answer containing knowledge response content, and identify whether the user's natural language input contains a schedule operation request;

[0010] A multi-agent collaboration module is configured to, upon recognizing that the user's natural language input contains a schedule operation request, decompose the schedule operation request into one or more subtasks based on the results of task analysis, and dispatch the subtasks to corresponding functional agents for execution, including at least calling a schedule management agent to access enterprise schedule data and generate task processing results;

[0011] A context fusion module, configured to receive the preliminary answer and the task processing result, and contextually fuse the two based on a semantic structure consistency strategy;

[0012] Among them, the large language model core module is also used to receive the results of context fusion, generate a comprehensive reply, and output it to the user end to complete the question and answer response.

[0013] The beneficial effects of this application mainly include: (1) achieving a unified response for enterprise knowledge retrieval and task operations. By using natural language input for both knowledge recall and task parsing, the system can simultaneously process query and instruction requests in one interactive process, thereby improving the intelligence level of enterprise information services. (2) improving the accuracy and contextual coherence of question and answer content. By using the context fusion module to align the semantic structure of the preliminary answer with the task processing result, the problem of the separation between knowledge response and actual operation results in traditional systems is avoided, and the overall consistency and practicality of the results are enhanced. (3) constructing a semantic-driven multi-agent dynamic collaboration mechanism. The system can automatically decompose tasks based on language understanding results and intelligently allocate them to functional components, adapting to the complex and changing needs of enterprise-level users, and has good scalability and task adaptation capabilities. (4) significantly improving user efficiency and experience. Users can obtain knowledge answers and operation execution results at the same time through only one natural language input, eliminating the steps of switching between multiple platforms and repeated operations, and optimizing the workflow of enterprise employees. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a schematic diagram of an intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base provided in the first embodiment of the present application.

[0015] Figure 2 This is a flowchart of an intelligent collaborative question-answering method for enterprise-level schedule planning and knowledge base provided in the second embodiment of this application. DETAILED DESCRIPTION

[0016] The following description sets forth many specific details to facilitate a thorough understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the scope of the present application. Therefore, the present application is not limited to the specific implementations disclosed below.

[0017] The first embodiment of this application provides an intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base. Figure 1 , which is a schematic diagram of the first embodiment of this application. Figure 1 The first embodiment of the present application provides a detailed description of intelligent collaborative question-answering for enterprise-level schedule planning and knowledge base.

[0018] The intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base includes a vectorized 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.

[0019] The vectorization processing module 101 is 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.

[0020] Vectorization processing module 101 is used to implement semantic representation conversion in an intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge bases. Its core function is to preprocess and embed code input natural language text and enterprise knowledge data, generating semantically comparable vector representations to support subsequent semantic retrieval and intelligent understanding. This module consists of two main processing paths, corresponding to the vectorization tasks of user natural language input and pre-stored enterprise knowledge data.

[0021] After the user interaction layer receives the natural language input, the vectorization processing module 101 first performs language normalization operations on the input, including pre-processing processes such as word segmentation, removal of stop words, and unification of word forms, and sends it to the selected semantic embedding model. The model can be an open source or self-developed pre-trained language model, such as BERT, Sentence-BERT, bge-large-zh and other semantic embedding models based on the Transformer architecture, which are used to generate a first semantic vector corresponding to the user input. The first semantic vector is a high-dimensional dense vector, usually with a dimension between 256 and 1024, and can capture the contextual semantics and keyword information in the user input. Its data structure supports rapid comparison and semantic matching with the enterprise knowledge semantic vectors in the vector database.

[0022] Regarding the processing path of enterprise knowledge data, 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 and Word documents) during the system deployment phase or data update node. The multi-source data will be converted into text-based corpus through a unified information extraction interface, and will also be sent to the same type of semantic embedding model for encoding through the preprocessing process. The generated second semantic vector is used to construct and maintain the vector database for the RAG knowledge base module 102. The vector entries will be bound to the unique identifier and metadata information of the original text fragment (such as document source, creation time, project number, etc.) to ensure that knowledge location and tracking can be achieved in the subsequent recall process.

[0023] Furthermore, the vectorization processing module 101 must support incremental updates and batch vector refresh mechanisms during the construction of the second semantic vector. Specifically, the module should include a data change detection and task scheduling mechanism to trigger the re-vectorization of some or all documents based on knowledge base update events. Upon completion, the updated second semantic vector is written to the vector database via an interface. This process must maintain vector consistency with the RAG knowledge base module 102 to avoid recall mismatches.

[0024] To ensure comparability between the two types of vectors, the embedding model parameters used by vectorization processing module 101 should be consistent, or the two models should be jointly trained using a cross-domain training strategy that shares a semantic space, thereby ensuring that the first semantic vector and the second semantic vector are comparable in distance within the same semantic representation space. In certain embodiments, the system can also fine-tune the model for different data sources to improve semantic fit, but the vector output still needs to undergo a uniform transformation to maintain vector space compatibility.

[0025] Finally, to support integration with the large language model core module 103, the vectorization processing module 101 must provide a standardized API interface. The output first and second semantic vectors should use consistent vector format and floating-point precision, and, when necessary, attach position information, sentence boundaries, or paragraph context to support subsequent combined processing of multimodal inputs. The module must also include an anomaly detection and fallback mechanism, outputting a recognizable empty vector marker or triggering error logging in the event of embedding failure or missing semantic information to ensure system robustness.

[0026] In practical applications, user natural language input and enterprise knowledge data are the two core objects of semantic computing in intelligent collaborative question-answering systems. To more clearly illustrate their specific meaning and role, we can use a typical enterprise scenario as an example.

[0027] Suppose an enterprise has deployed the intelligent collaborative question-and-answer system described in this invention. A marketing manager initiates the following voice or text request through the client: "Please help me check the minutes of last month's promotion meeting for the X series products and schedule a feedback meeting with the design department this Friday afternoon." This sentence is the user's natural language input, expressed in a natural, unstructured manner, and contains two parallel semantic intents: a knowledge query, namely, to query a certain type of historical meeting records; and an operational request, namely, to schedule a meeting with a specific department. The vectorization processing module receives this input and converts it into a first semantic vector using an embedding model, enabling the system to understand its semantic content and perform subsequent knowledge retrieval and task identification.

[0028] In contrast, enterprise knowledge data refers to the various information resources accumulated by an enterprise during its operations. This data typically exists in structured, semi-structured, or unstructured forms, and includes content such as meeting minutes, project reports, personnel schedules, policy documents, and business logs. For example, in the above scenario, the enterprise knowledge data in the system might include a Word document titled "Minutes of the X Series March 2024 Promotion Meeting," which details the meeting time, participants, discussion topics, and decision outcomes. It might also include a meeting record in the database, numbered PM-0322, with the meeting topic marked "X Series Quarterly Promotion Report Meeting," the date set as March 15, 2024, and the summary stating, "The discussion focused on social platform advertising conversion rates and offline channel effectiveness feedback."

[0029] During system deployment or data synchronization, this enterprise knowledge data is parsed by the vectorization processing module and uniformly fed into the semantic embedding model. This generates a second semantic vector, which is then stored in the vector database. This way, when a user initiates a natural language request, the system can retrieve the knowledge fragment corresponding to the second semantic vector most relevant to the user's first semantic vector from the vector database, such as the aforementioned meeting minutes. This provides high-quality context for the subsequent generation of knowledge responses by the core module of the large language model.

[0030] Through this mechanism, users' natural language input is semantically connected to enterprise knowledge data, enabling the system to not only understand user needs but also quickly locate relevant information from within the vast enterprise knowledge, achieving unified coordination between knowledge responses and task operations. This bidirectional mapping capability based on semantic vectors is the foundation for efficient enterprise-level question-answering and scheduling.

[0031] In summary, vectorization processing module 101 is not only a precursor to semantic computing, but its output, the first semantic vector, directly determines the retrieval accuracy of RAG knowledge base module 102. The second semantic vector, on the other hand, forms the semantic representation foundation of the entire knowledge system and is key to implementing semantically driven intelligent question-answering and operation scheduling. The module's architectural design should ensure that it meets the real-time and security requirements of enterprise-level scenarios in terms of processing speed, vector consistency, model scalability, and interface compatibility.

[0032] Furthermore, the vectorized processing module is further configured to:

[0033] After receiving the user's natural language input, calling an embedding model based on the Transformer architecture to convert the user's natural language input into a first semantic vector, and appending an input timestamp and a context position index to the first semantic vector to preserve the user's language input timing information;

[0034] Before converting the pre-stored enterprise knowledge data into a second semantic vector, a corresponding semantic modeling strategy is selected according to the format type of the enterprise knowledge data, wherein a rule embedding method is invoked for structured data, and an embedding model identical to the first semantic vector is invoked for a natural language document to obtain a preliminary second semantic vector;

[0035] performing vector normalization processing on the first semantic vector and the preliminary second semantic vector respectively, and calculating a vector consistency index between the two based on cosine similarity in the same semantic space;

[0036] When the vector consistency index is less than a preset semantic similarity threshold, a vector tuning operation is performed on the preliminary second semantic vector to enhance its cross-source semantic alignment capability with the first semantic vector. The tuned second semantic vector is used for the RAG knowledge base module to retrieve the enterprise knowledge fragment corresponding to the second semantic vector that is semantically related to the first semantic vector in the vector database.

[0037] In the intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base, the vectorization processing module not only has basic natural language semantic representation conversion capabilities, but also further undertakes key functions such as consistent processing of semantic representation formats between different types of data, cross-modal semantic fusion, and quality control. Especially in enterprise-level application scenarios, faced with 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 foundation for subsequent modules, significantly improving the retrieval accuracy of the RAG knowledge base module.

[0038] During system startup or operation, when a user issues a natural language request through the terminal, the vectorization processing module immediately receives the natural language input. This input may be a single-sentence instruction, such as "Schedule a meeting with Manager Wang," or a complex request, such as "Find the sales data report for the second quarter of last year and remind me to submit the meeting materials tomorrow afternoon," or a paragraph-style information description. After receiving this type of input, the system encodes it by calling a class of embedding models based on the Transformer architecture (such as ChatGLM, Baichuan, BERT, RoBERTa, or a customized lightweight language model). The encoding process is not a simple word-level or sentence-level mapping, but rather captures semantic elements such as contextual dependencies, structural relationships, subject-verb collocation, temporal logic, and verb intention through a multi-layer attention mechanism, thereby generating a high-dimensional dense representation, which is the first semantic vector.

[0039] To make the generated first semantic vector temporally interpretable, the system also injects two additional information dimensions into the semantic representation. The first is the input timestamp of the current request, which records the exact time when the semantic request was issued. This timestamp is usually expressed in UNIX time format or ISO standard format and is mapped into a vector channel or additional feature vector by the embedding model through the time embedding mechanism. The second is the context position index, which is mainly enabled in multi-round dialogue scenarios and is used to indicate the relative position of the input in the entire dialogue, such as "3rd round", "Reply to the previous question", "Follow up the first arrangement", etc. This helps the system understand the continuity of the context, which is particularly critical for expressions such as "make another appointment" and "postpone yesterday's meeting".

[0040] In contrast to user input processing, enterprise knowledge data comes from a vastly diverse set of sources. The system may contain structured business records in databases (such as meeting databases, task lists, and customer management systems), semi-structured logs (such as JSON-formatted OA approval processes, report interface outputs, and 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, and PowerPoint presentations). To ensure that different data types can be represented and compared in a unified semantic space, the vectorization processing module dynamically selects appropriate modeling strategies based on the data format.

[0041] Specifically, for structured data, the system adopts a rule-based embedding method, which is to first perform field segmentation and content extraction on the data through field templates, and then assign specific semantic labels or embedding rules based on 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 and 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 an embedding model that is exactly the same 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.

[0042] 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 the vector consistency index, that is, to judge the semantic closeness between the 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 moduli. 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.

[0043] If the similarity result falls below a preset semantic similarity threshold (e.g., 0.6 or 0.7), the system determines that the current preliminary second semantic vector does not adequately match the first semantic vector in the semantic space. This inconsistency can arise from a variety of reasons, such as inconsistent embedding model training data styles, deviations in temporal context or task orientation between the document data and the user request, or failure of the embedding rules to capture certain semantic details. To correct this inconsistency, the system triggers a vector tuning operation.

[0044] 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 proportion of fields such as "time" and "place" that correspond to high-frequency keywords in user requests; 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 capabilities; 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.

[0045] After the tuning process is complete, the generated new vector is normalized again and replaces the original preliminary second semantic vector as the official second semantic vector for use by the RAG knowledge base module. Within the vector database, the RAG knowledge base module uses efficient indexing structures (such as HNSW, IVF, and PQ) to search for enterprise knowledge fragments semantically related to the first semantic vector, ensuring that the question-answering output is semantically accurate, temporally relevant, and consistent with the business context.

[0046] In summary, the vectorized processing module implements a multi-dimensional, deeply integrated, and tunable semantic vector construction and alignment mechanism from user natural language input to multi-source enterprise knowledge data. It plays the core role of a bridge between input understanding and knowledge retrieval in the entire intelligent question-answering system. Through temporal information modeling, format adaptation strategy, semantic consistency detection and tuning mechanism, it ensures that the system has the ability to handle complex requests and heterogeneous knowledge in actual enterprise applications.

[0047] Furthermore, the vector processing module performs the vector tuning operation, including the following steps:

[0048] Extracting the generation time of the enterprise knowledge data corresponding to the preliminary second semantic vector, and embedding the generation time into the time dimension channel of the preliminary second semantic vector to construct a temporal second semantic vector with a time series label;

[0049] Calculate the implicit temporal vector representation based on the temporal expression content involved in the context of the first semantic vector, and perform temporal alignment evaluation between the vector and the temporal second semantic vector;

[0050] If the evaluation result indicates that there is a significant inconsistency in temporality between the two, the temporal weight parameter of the preliminary second semantic vector is adjusted, and attention enhancement is performed in combination with the context position index to improve the temporal matching degree;

[0051] The adjusted temporal second semantic vector is used as the tuned second semantic vector.

[0052] In the intelligent collaborative question-and-answer system for enterprise-level calendar planning and knowledge bases described in this invention, when the vectorization processing module detects that the semantic consistency index between the first semantic vector and the preliminary second semantic vector falls below a preset threshold, a vector tuning operation is triggered. This tuning operation specifically targets the time dimension, with the core goal of improving the temporal semantic matching accuracy between knowledge fragments and user input. This is particularly useful for handling user questions that involve time expressions, such as "first quarter of this year," "last month," or "most recent regular meeting."

[0053] 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 the knowledge fragment from the metadata. If this information is not explicitly stored, the system can also extract time expressions appearing in the text through content analysis, such as "March 15, 2023", "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 consistency in subsequent processing. After the time information is extracted, the system encodes the timestamp into a dense time vector and injects it into the original preliminary second semantic vector as a time dimension channel.

[0054] Injection can be achieved through vector concatenation, position channel expansion, or feature fusion strategies. The most direct approach is concatenation, which appends a time vector channel of equal or fixed length to the original semantic vector, creating a temporal second semantic vector with temporal characteristics. For example, if the original semantic vector is 768-dimensional, the time channel can be set to 64 dimensions, resulting in a final vector of 832 dimensions. Alternatively, a parallel embedding approach can be used, generating a context-independent time vector through a separate time embedding model. This vector is then weighted using attention weights and fused into the main vector channel.

[0055] Secondly, the system needs to perform temporal semantic recognition and modeling on the user input content represented by the first semantic vector. The specific method is to use a temporal expression recognizer (such as a temporal NER system based on rules, statistics, or a Transformer model) 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," and "fuzzy time"). The extracted time expression will be linked to the current time of the system to generate one or more implicit temporal vectors. For example, "last month" will be parsed as "a time period one natural month before the current date" and embedded into a vector representing the user's temporal intention, forming a vector pair of equal dimensions with the temporal second semantic vector.

[0056] The system then compares the semantic distance between the temporal second semantic vector and the implicit temporal vector to assess whether they align at the temporal intent level. This comparison can be performed using standard cosine similarity calculations, or by introducing distance functions based on temporal logic, such as determining the overlap between two time periods or the offset between the start and end points. If the evaluation indicates a significant inconsistency, the system performs a tuning operation to adjust the temporal semantic structure of the preliminary second semantic vector to better align with the user's temporal intent.

[0057] The tuning process mainly includes two directions of optimization: the first is to adjust the weight parameters of the time dimension channel. In the original temporal second semantic vector, time information and content semantics are usually mixed. By adjusting the scalar amplification coefficient of the time channel, the contribution of this channel to the overall vector can be enhanced or reduced, thereby controlling the dominance of the time attribute in subsequent matching. The second 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 focuses more on time-related paragraphs or sentences in the document. The context position index can be based on the document structure hierarchy, such as title, chapter number, paragraph number, or based on the relative position identification of the sentence in the document, combined with the temporal entity distribution in the document, to further locate the semantic concentration area.

[0058] Finally, the optimized temporal second semantic vector is the optimized second semantic vector. This vector aligns content semantics with user temporal intent, providing a stronger representation of temporal relevance. This vector serves as an index for knowledge fragment retrieval in the RAG knowledge base module, retrieving enterprise knowledge fragments corresponding to the second semantic vector that is semantically related to the first semantic vector from the vector database.

[0059] The entire optimization process starts with the raw vector and goes through a series of steps, including time extraction, temporal modeling, semantic alignment, time-dimension weighting, and attention redistribution, forming a complete data flow loop. This ensures precise alignment of user intent with the temporal context of knowledge during semantic retrieval. This mechanism offers significant advantages in handling directive queries containing temporal expressions, time-sensitive multi-round conversations, and scenarios where time-sensitive knowledge retrieval is required within enterprises.

[0060] The RAG knowledge base module 102 is configured to retrieve, based on the first semantic vector, enterprise knowledge segments corresponding to second semantic vectors semantically related to the first semantic vector in a vector database.

[0061] The RAG knowledge base module 102 is used to implement semantic vector-based enterprise knowledge retrieval in an intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base. Its core function is to efficiently search for a second semantic vector semantically related to the first semantic vector generated by the user's natural language input in the vector database, and obtain the corresponding enterprise knowledge fragment based on this, so as to support the subsequent semantic understanding and response generation completed by the large language model core module 103.

[0062] During implementation, the RAG knowledge base module 102 receives a 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. The dimension is usually 256, 512 or higher, and its semantic structure remains 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. The vector database has pre-stored a large number of second semantic vectors generated by enterprise knowledge data and established a mapping relationship with the original knowledge fragments. In order to improve retrieval efficiency, this module usually uses a database engine that supports fast approximate nearest neighbor (ANN) search, such as Milvus, and improves query performance while ensuring recall accuracy by constructing acceleration structures such as IVF (inverted file index) and HNSW (graph structure index).

[0063] During the retrieval process, the RAG knowledge base module 102 selects the most relevant second semantic vectors based on a preset similarity threshold. It typically returns the enterprise knowledge fragments corresponding to the top-K (e.g., the top 3-5) vectors and extracts information such as their original text content, data source, and related tags. These knowledge fragments may include internal enterprise knowledge content such as project summaries, meeting minutes, regulations, product specifications, and operating manuals. The module also incorporates a confidence scoring mechanism, assigning a semantic match confidence level to each recalled fragment for subsequent context reconstruction and multi-fragment fusion.

[0064] In some embodiments, the RAG knowledge base module 102 also includes content pre-processing and post-processing capabilities. For example, it performs operations such as syntactic segmentation, keyword highlighting, and context completion before outputting knowledge fragments, making the fragments more suitable as input context for the larger language model. When processing semantically ambiguous or ambiguous input, the module can incorporate multi-vector extended retrieval, which introduces multiple copies of the semantically enhanced first semantic vector and independently performs retrieval tasks on each copy. Finally, the deduplicated results are merged to enhance recall coverage and robustness.

[0065] The enterprise knowledge fragments output by the RAG knowledge base module 102 must retain their relevance to the original enterprise knowledge data and have direct semantic relevance to the first semantic vector. Its output is not only the text content itself, but also can include associated metadata structures, such as the department to which the knowledge belongs, update time, document type, etc., to assist the large language model core module 103 in completing context understanding and accurate response.

[0066] It should be noted that the RAG knowledge base module 102 does not directly participate in the response generation process. Instead, it provides a credible, controllable, and traceable semantic context basis for the large language model core module 103 by providing enterprise knowledge fragments that are closely related to the semantics of user input, ensuring that the subsequently generated answer content is based on internal enterprise knowledge rather than relying entirely on the pre-trained parameters of the language model, thereby effectively reducing the risk of hallucinations and improving accuracy and professionalism in enterprise-level scenarios.

[0067] Therefore, the RAG knowledge base module 102 plays the role of a semantic retrieval bridge in the entire system and is the key link between user intentions and corporate knowledge. Its technical implementation relies on the synergy of semantic embedding consistency, vector indexing mechanism, similarity measurement model and fragment scheduling strategy to ensure that the entire question-answering process has stable, efficient and scalable knowledge support capabilities.

[0068] Enterprise knowledge fragments are text units with independent semantic expression, extracted from various knowledge carriers within an enterprise. They are typically the smallest usable information units that constitute an enterprise knowledge system. These fragments can originate from structured data (such as project names and responsible person records in a database), semi-structured data (such as JSON-formatted business logs, OA approval process information), or unstructured data (such as meeting minutes, training documents, product specifications, operating manuals, internal notices, etc.). For example, the description of "Product A in the Northeast region achieved 25% year-on-year growth in Q3 2023" in a PDF quarterly sales report can be segmented and extracted as an enterprise knowledge fragment for subsequent semantic matching and question-and-answer support. Similarly, the stipulation in an employee handbook that "new employees must complete Level 3 safety training within 30 days of joining" can serve as an independent enterprise knowledge fragment, providing precise information for users asking, "When do I need to complete training?" Enterprise knowledge fragments are typically mapped to their source documents to ensure traceability and contextual supplementation during use.

[0069] The large language model core module 103 is configured to receive the user's natural language input and the enterprise knowledge fragment, 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;

[0070] The large language model core module 103 is also used to receive the results of context fusion, generate a comprehensive reply, and output it to the user end to complete the question-answer response.

[0071] The large language model core module 103 is the core intelligent processing unit in the intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base. It is mainly responsible for deep semantic understanding and task logic analysis of users' natural language input and enterprise knowledge fragments, thereby realizing knowledge response generation and operation intention recognition, and coordinating subsequent modules in the system to jointly complete the final question-answering response when necessary.

[0072] During the system operation, the large language model core module 103 first receives natural language input from the user and the enterprise knowledge fragments retrieved by the RAG knowledge base module 102. The user's natural language input is generally an open-ended question, an imperative expression, or a mixed request, which may involve various scenarios such as enterprise knowledge consultation, historical event query, task scheduling, schedule change, etc. The enterprise knowledge fragments are semantically related knowledge support texts, which usually have the characteristics of high information density, clear semantics, and complete context. The large language model core module 103 combines the two as input to construct context corpus to drive the language model for intelligent reasoning and answer generation.

[0073] In practice, the large language model core module 103 can utilize current mainstream natural language generation model architectures, such as open-source Transformer-based models (e.g., ChatGLM, Baichuan, MOSS, DeepSeek, etc.) or customized, fine-tuned enterprise models. This module possesses the ability to model long-text semantics. Using a self-attention mechanism, it captures keywords, contextual relationships, and intended expressions in user input, aligning them with the specialized terminology and factual information in the knowledge fragments to generate preliminary answers containing knowledge-based responses. These preliminary answers are not simply surface-level text outputs, but rather clearly structured, semantically relevant question-answering results that accurately address the user's query needs.

[0074] During the process of generating preliminary answers, the large language model core module 103 also executes the task parsing process in parallel. Task parsing includes, but is not limited to, multiple subtasks such as intent recognition, entity extraction, operation type determination, parameter extraction, and conditional constraint determination. This process is based on semantic analysis of natural language input to determine whether the user's natural language input contains a schedule operation request. If so, the module constructs a logical representation structure to identify the task type (e.g., query, create, update), target entity (e.g., meeting, reminder, collaboration object), timing requirements (e.g., "next Monday morning"), and other constraints.

[0075] Once it determines that a user request includes a calendar-related operation, the large language model core module 103 no longer directly outputs a final response. Instead, it acts as a coordinator, passing the task analysis results to the multi-agent collaboration module 104, which then completes the specific task execution. At the same time, the large language model core module 103 retains the preliminary answer content so that it can participate in the generation of the final response after obtaining the operation results.

[0076] After the multi-agent collaboration module 104 completes the task and returns the schedule processing results, 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 results, and is unified and integrated through a semantic structure consistency strategy, which has a high degree of semantic coherence and information integrity. The large language model core module 103 performs pragmatic judgment and language optimization on this fusion result, ultimately generating a comprehensive response that meets the user's expectations and outputting this comprehensive response to the user end, completing the entire question-answer response process.

[0077] It should be pointed out that in order 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 prompt template management, model parameter control, output stability calibration mechanism, and abnormal response guarantee mechanism to improve the practicality and engineering deployment capabilities of the system.

[0078] To sum up, the large language model core module 103 not only undertakes the responsibilities of semantic understanding, knowledge retrieval and language generation, but also realizes an intelligent collaborative question-and-answer process that integrates knowledge and operations through 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.

[0079] The multi-agent collaboration module 104 is used to decompose the schedule operation request into one or more subtasks based on the result of task analysis when it is recognized that the user's natural language input contains a schedule operation request, and schedule them to the corresponding functional agent for execution, at least including calling the schedule management agent to access the enterprise schedule data and generate task processing results.

[0080] The multi-agent collaboration module 104 is a key functional module for processing schedule operation requests in the intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base. Its main function is to further decompose the request into one or more subtasks based on the structured results of task analysis, on the premise that the large language model core module 103 recognizes that the user's natural language input contains schedule operation requests, and schedules each subtask to the corresponding functional agent for execution in an orderly manner, so as to realize the system's automatic response to user requests and operation implementation.

[0081] This module is usually built based on an extensible agent framework, and each functional agent is defined as an independently running task execution unit with clear functional boundaries and data access rights. Specifically, after the large language model core module 103 completes the semantic understanding and task parsing of the user's natural language input, if it is determined that the input belongs to or contains a schedule operation request, such as "help me arrange a meeting with Manager Zhang at 3 pm tomorrow", the task semantics will be structured into a set of parsing parameters including operation type (such as "create a meeting"), time parameters (such as "tomorrow at 3 pm"), object entities (such as "Manager Zhang"), location or conference room, etc. After receiving the parsing parameters, the multi-agent collaboration module 104 will decompose the request into one or more logical subtasks according to the preset task scheduling rules or dynamic task mapping mechanism, such as "find Manager Zhang's free time", "book a conference room", "send a meeting notice", etc.

[0082] After the task is decomposed, the multi-agent collaboration module 104 assigns the subtask to the preset functional agent according to the type and execution requirements of the subtask. The functional agent can be a schedule management agent with database access capabilities, or an API proxy agent connected to a third-party office system (such as Exchange, DingTalk, Feishu, enterprise WeChat calendar module, etc.). The system interacts with the agent through RESTful interfaces, message queues, or local function calls. After receiving the task instruction, the schedule management agent will access the company's internal schedule database or calendar service, extract the time schedule information of the relevant personnel, determine whether the target time period is available, and create, modify or cancel the meeting according to the meeting arrangement rules. At the same time, it generates a structured task processing result, which usually includes the operation result status (success or failure), the created event ID, the meeting link, the schedule summary text, and other content.

[0083] The multi-agent collaboration module 104 must also include a task status monitoring and callback mechanism. This mechanism allows it to promptly collect the execution results of each subtask after each functional agent completes its task, encapsulate and aggregate them in a standard format, and then integrate and process them with the context fusion module 105 and the large language model core module 103. Furthermore, to improve execution stability and robustness, the multi-agent collaboration module 104 could further incorporate error handling mechanisms and task retry logic. For example, if a conference room reservation fails, it could automatically try alternative times or conference rooms and report the results back to the system.

[0084] In order to support the traceability and auditability of the task process, the multi-agent collaboration module 104 also needs to log the key steps in each task parsing and scheduling process, including task decomposition parameters, distribution path, execution status, exception information and user feedback, etc., to provide a data basis for system operation and maintenance and agent performance evaluation.

[0085] To sum up, the multi-agent collaboration module 104 not only assumes the responsibility of converting language understanding results into executable operations, but also builds an automated control center for enterprise schedule task processing in a modular and programmable manner, ensuring that the system can accurately, efficiently and securely call enterprise resources after receiving complex natural language requests, complete structured task execution and feedback results, and lay a data foundation for the generation of subsequent comprehensive responses.

[0086] Furthermore, the multi-agent collaboration module is also used to:

[0087] Based on the schedule operation request in the user's natural language input identified by the large language model core module, a task graph template corresponding to the semantic structure of the schedule operation request is matched from a preset semantic-driven task decomposition mapping table, and a task graph including multiple task nodes and their dependencies is generated according to the task graph template;

[0088] Perform semantic parsing on each task node in the task graph, extract the target operation type, target entity constraint, and context time parameters corresponding to each subtask, and generate a task description object containing instruction parameters;

[0089] Input the task description object into the task scheduling engine, schedule each task node in order according to the dependency relationship in the task graph, and map each task node to a corresponding functional agent, so as to call at least a schedule management agent to execute the task operation corresponding to the task node;

[0090] The task execution results returned by the functional agent are collected, and all task execution results are structured and aggregated according to their logical structure in the task graph, and a task processing result set is output for the context fusion module to perform fusion processing.

[0091] In the intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base described in the present invention, the multi-agent collaboration module not only undertakes the function of converting schedule operation requests into executable subtasks, but also further realizes the process of mapping from natural language intent structure to task graph structure, and through semantic parsing and task scheduling mechanism, each task node in the task graph is efficiently delivered to the corresponding functional agent for processing, and finally completes the structured aggregation of task processing results, providing complete and consistent operation execution information support for the subsequent context fusion module.

[0092] The technical solution provided in this embodiment relies primarily on the semantic understanding of user natural language input by the large language model core module. When the system detects that the natural language input contains a scheduling request, such as "Schedule a project review meeting with Manager Wang and the Finance Department next Tuesday afternoon," the multi-agent collaboration module receives the analysis results from the large language model core module, which typically include the action intent type (e.g., create a meeting), the participants (Manager Wang, the Finance Department), the time condition (next Tuesday afternoon), and other contextual prompts.

[0093] In order to convert this natural language request into a structured task representation that can be executed within the system, a semantically driven task decomposition mapping table is preset in the multi-agent collaboration module. This mapping table uses semantic structure as the matching key to map different types of natural language input structures to corresponding task graph templates. For example, the "arrangement + object + time" structure will be matched to a standard task graph template with four nodes: "Check participant schedule availability → Query meeting room → Create meeting → Send notification". During the semantic matching process, the system can use BERT-based classifiers or rule tree parsing to identify the sentence structure, ensuring a high structural similarity between the input and the task graph.

[0094] After matching the task graph template, the system generates an initial task graph, where each task node represents an independent, executable subtask and includes dependencies between tasks. Next, the system performs semantic parsing on each task node in the task graph, extracting the corresponding target operation type (e.g., conference room query, schedule confirmation, notification push), target entity constraints (e.g., involved personnel, conference room resources), and contextual time parameters (e.g., time window, specific time point, duration), and encapsulating these parsed results into task description objects. Each task description object is a structured data structure that contains all the parameters required for task execution, ensuring high interpretability and executable capabilities.

[0095] After the task description object is constructed, the system will hand it over to the task scheduling engine for scheduling. The task scheduling engine determines the order and concurrency of task execution based on the node dependencies defined in the task graph. For example, "querying personnel schedules" must be executed before "creating meetings", while "sending meeting notifications" can be executed in parallel with other nodes. The scheduling engine also calls the corresponding functional agent according to the task type, which currently includes at least the schedule management agent, and may also include the conference room resource management agent, personnel status synchronization agent, notification distribution agent, etc. Each functional agent acts 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, conference resource platform, etc.), and performs the specific operations corresponding to the task node.

[0096] After all task nodes are completed, each functional agent returns a standardized task execution result. This result typically includes the operation status (success, failure, conflict), the operation object identifier (such as the meeting ID or conference room number), time confirmation information (such as the finalized time period), and prompt information. The multi-agent collaboration module aggregates these execution results and performs structured aggregation based on the original task graph logical structure. This generates a complete, ordered, and contextually unified set of task processing results without disrupting task dependencies. This set not only preserves the execution results of each subtask but also represents the execution process and operation logic in a nested structure or critical path format, ensuring semantic perception and content integration by the subsequent context fusion module.

[0097] This series of processing flows is logically highly closed-loop: from the structured parsing of natural language, the generation and node distribution of task graphs, to the operation execution and result aggregation of functional agents, the output of each step is directly used by the next step, and each processing unit has a stable data dependency relationship. The technical path described in this embodiment is significantly superior to traditional static rule-based systems. It not only has semantic self-adaptation capabilities, but can also be expanded to support more task graph templates and agent types. It is suitable for various complex collaborative operation requests in enterprises and is an indispensable technical support core for the intelligent collaborative question-answering capabilities of the present invention.

[0098] Furthermore, the multi-agent collaboration module includes the following steps when performing structured aggregation:

[0099] Before the task scheduling engine schedules the task node to the functional agent, a global context state object is constructed, wherein the global context state object includes a time window, an entity reference mapping, and user role information extracted from the user natural language input;

[0100] In the process of concurrently scheduling multiple task nodes, the global context state object is injected into the operating environment of each functional agent, so that each functional agent can interpret parameters and execute tasks based on consistent context state constraints when executing tasks;

[0101] After receiving the task execution result of the functional agent, calculate the task execution confidence score based on the semantic fit between each functional agent's historical execution record and the current task description object, and append the confidence score to the corresponding task execution result;

[0102] When constructing the task processing result set, candidate results are weighted screened and prioritized according to the confidence score in each task execution result to generate a structured task processing result set with the highest overall semantic consistency and context fit.

[0103] In the intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge bases described in the present invention, the multi-agent collaboration module, in the process of performing structured aggregation, does not simply splice or sequentially arrange the task execution results returned by the functional agents. Instead, it completes a task result set construction process with high semantic consistency and global context adaptability based on the dynamic cross-evaluation between the dialogue context, user intention structure, agent behavior history, and task execution status. The core of the technical solution provided by this embodiment is to establish a result integration mechanism of "context state drive + confidence weighted sorting" to minimize semantic conflicts, state drift, and deviation from user expectations between asynchronous agent execution results, and significantly improve the structural clarity and logical unity of the final answer.

[0104] During the implementation 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 will build a global context state object based on the user natural language input generated by the large language model core module. This object is a state container that can be shared by multiple subtasks, mainly including time window information extracted from user input (such as "next Wednesday afternoon" or "the last working day of this month"), entity reference mapping relationship (such as "he" refers to "Manager Wang", "that meeting" refers to the "quarterly review meeting" mentioned above) and user role information (such as whether the initiator is a project manager, department head or ordinary employee). These information constitute restrictive constraints in actual semantic reasoning and are the prerequisite for ensuring the consistency of task execution.

[0105] Once the global context state object is established, the system will inject it into the operating 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 the context state object, so that it can perform parameter interpretation, query scheduling, command execution, etc. based on a unified semantic background. For example, when scheduling 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 avoidance judgments in advance. 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.

[0106] After all functional agents complete the corresponding operations of 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 of each execution result based on the degree of match between the historical behavior records of each functional agent (such as the execution accuracy of past tasks, average response time, positive correlation with user feedback, etc.) and the semantic structure in the current task description object. This score is not a hard-coded setting, but is generated based on dynamic factors such as content semantic fit, context fit, and model prediction distribution. A multi-factor fusion algorithm can be used, such as a weighted classifier or a clustered cosine similarity scoring model. Each task execution result will be accompanied by this confidence score as an additional attribute to guide subsequent aggregation optimization.

[0107] After all results and scores are prepared, the system will enter the critical stage of structured aggregation, which is to build the final set of task processing results. This process will maintain the overall structural logic based on the original dependency order in the task graph, and introduce confidence scores as the sorting basis for candidate results at the same level, performing weighted screening and priority sorting operations. Specifically, if multiple candidate results provide different execution feedback for the same subtask, such as multiple time options or conference room conflicts in a meeting schedule, the system will retain the entry with the highest confidence score, or perform fusion between semantically similar candidates (for example, returning two alternative time periods for user confirmation). During this process, the system will also recheck 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 reordered in a timely manner to ensure that the entire task processing result set has logical integrity and causal consistency in structure.

[0108] The resulting set of task processing results not only semantically reflects the user's original input intent, but also maintains clear information, minimal conflict, and contextual consistency at the execution level. This allows for seamless integration of knowledge responses and operational feedback in subsequent context fusion modules. This fusion process is far superior to the "serial execution + result concatenation" mechanism used in traditional static process systems and is particularly well-suited for complex operational requests in enterprise-level Q&A involving multi-role collaboration, conditional execution, and time-sensitive scheduling.

[0109] Furthermore, the multi-agent collaboration module uses the following scoring function when calculating the task execution confidence score of each task execution result:

[0110]

[0111] in, Indicates the The task execution confidence score corresponding to the task execution result is used to measure the comprehensive credibility of the task execution result relative to other candidate results in the multi-agent collaborative scheduling process, and the value range is a positive real number. Corresponding to the first task returned by the functional agent mapped by multiple task nodes generated by task graph decomposition during a certain user request processing. A valid execution result, ,in The number of all subtasks executed under this request.

[0112] Indicates the The semantic structure matching score between a task execution result and its corresponding task description object is derived from the graph embedding cross entropy inverse score calculated after structural alignment of the semantic graph constructed by the task description object (including subject-verb-object structure, restrictive conditions, target object, etc.) with the semantic graph implied by the execution result, which measures the degree of semantic overlap between the two at the structural level. The unit is , the larger the value, the more consistent the semantic structure.

[0113] Indicates the The context adaptation index of a functional agent under the currently loaded global context state object is specifically defined as the agent’s context adaptation index in the past The proportion of successful completion of context-sensitive tasks when similar context conditions (such as time window, user role, reference mapping) are loaded for the first time, expressed as a percentage (%), which is used to measure the agent's familiarity with the enterprise context and the consistency of parameter interpretation.

[0114] Indicates the The path depth of a task node in the current task graph structure, that is, the shortest hierarchical distance of the task node relative to the root node, is used to represent the information hierarchical position of the task in the overall semantic reasoning chain. The unit is layer, and its value is a natural number.

[0115] It represents the arithmetic mean of the path depths of all task nodes in the current task graph, and is used as a structural center reference value of the task graph to evaluate whether a task node is a structural deviation node.

[0116] :Indicates the The deviation of a task in the path structure. If the value is large, it means that the task is a non-backbone node, which affects the scoring weight.

[0117] Indicates the The semantic main line consistency coefficient between the task execution result and the main line intention proposed by the user's natural language input is The dimensionless score in the interval is derived from the weighted cosine similarity calculation between the semantic vector generated by the task execution result and the intent vector generated after parsing the user input. The closer it is to 1, the more closely the task responds to the user's true intention.

[0118] Represents 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 and is determined by experience during the system tuning process.

[0119] Indicates the The historical behavior confidence weight of a functional agent is composed of two parts: one is the average of the user feedback scores received by the agent in the past period (such as the last 100 tasks), and the other is the execution stability confidence factor of the current task type. The two are multiplied and normalized to This parameter is used to express the system's confidence in the long-term performance of each agent during result fusion. The execution stability confidence factor is determined by the fluctuation of the agent's historical success rate for the task type. The smaller the fluctuation, the closer the factor is to 1, indicating greater stability.

[0120] This scoring function integrates multiple technical factors such as semantic structure consistency, context adaptability, structural path dominance, and intent response accuracy, and flexibly differentiates the scoring results through nonlinear function transformation and exponential adjustment control, ensuring that the final confidence level not only conforms to the objective structural constraints of logical reasoning, but also reflects the consistency of language semantics and user intent responsiveness. It is suitable for dynamic intelligent question-answering systems in multi-role and multi-task scenarios in enterprises.

[0121] In the actual implementation process, when a user issues a complex 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 last year", the system will simultaneously trigger two processes: knowledge-based query and schedule operation. The core module of the large language model first parses the user's semantic mainline keywords as "arrange a meeting", "next Wednesday afternoon", "Manager Zhang" and other operational intentions, and at the same time identifies that the query part involves knowledge targets such as "last year" and "meeting minutes". Subsequently, in the multi-agent collaboration module, the task is broken down into multiple task nodes, including "querying the minutes of last year's review meeting", "checking whether Manager Zhang and the marketing department are free next Wednesday afternoon", "assigning a meeting room", "creating a meeting object", "generating a meeting notice" and other sub-tasks, which are respectively executed by different functional agents.

[0122] Suppose now that To calculate the confidence score of a task execution result (e.g. "allocate a meeting room") , the system will extract and calculate each parameter item by item according to the aforementioned scoring function.

[0123] First, the system constructs a semantic graph of the task, marking nodes such as the verb "allocation", the object "conference room", the condition "next Wednesday afternoon", and the participants "marketing department, manager Zhang". It then aligns the structure with the semantic graph of the information returned in the task execution result (such as "conference room A-302, 14:00-16:00"), uses graph embedding to perform similarity evaluation, and finally obtains the graph cross entropy inverse score. , for example .

[0124] Secondly, the system counted the proportion of the agent that successfully completed the task in the past 100 executions of the "conference room allocation" task when loading a specific time window and entity reference context, and found that there were 92 successes, and obtained the context adaptation index .

[0125] Then, calculate the path depth of the task node in the task graph . Let the root node be "Parsing User Input" and the path be "Schedule Meeting Check availability Assign meeting room", then the node is at the third layer; at the same time, the average path depth of all task nodes is , get the structure offset value .

[0126] The system also generates an intent vector based on the user's original input through an intent extraction mechanism. The intent vector is centered around "arranging a meeting" and its constraints, with the semantic core being the time-participant pairing. A semantic vector is also generated in the task execution results, representing "allocating spatial resources to specific time and personnel." The weighted cosine similarity calculation is performed on these two vectors to obtain the mainline consistency score. .

[0127] For the main line deviation index factor , which is set to 5 in the system default configuration. This value is set by the system during the debugging phase based on actual business needs and error tolerance.

[0128] Finally, the confidence weight of the agent's historical behavior The confidence factor of the current "conference room scheduling" task type (0.95) is multiplied by the past task satisfaction score (average 0.92) to get .

[0129] Substituting the above into the scoring formula:

[0130]

[0131] Therefore, the execution result of the task "Assign meeting room" will be assigned a confidence score This score will be input into the structured aggregation stage together with the scores of other subtasks to guide the system to select which subtask results should be retained, merged, or presented to the user as alternatives.

[0132] The context fusion module 105 is 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.

[0133] The context fusion module 105 is a key functional component used to integrate language understanding results and task execution results in the intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base. Its main function is to receive two types of information inputs, namely preliminary answers and task processing results, and generate context fusion results with coherent content, complete information and unified logic through a fusion process based on a semantic structure consistency strategy, thereby providing accurate and consistent semantic support for the final comprehensive reply.

[0134] The premise for the operation of this module is that the system has generated a preliminary answer by the large language model core module 103 during operation. The preliminary answer is based on the semantic understanding and content generation of the user's natural language input and the context composed of the enterprise knowledge fragments, and usually includes the answer content for knowledge-based queries. At the same time, if the large language model core module 103 recognizes that the user's natural language input contains a schedule operation request during task analysis, the multi-agent collaboration module 104 will complete the task decomposition and scheduling, and the corresponding functional agent will perform the specific operation to generate a task processing result. The task processing result may include confirmation information of the successful creation of the meeting, feedback on time conflicts, description of the task execution status, etc., usually in the form of structured or semi-structured data.

[0135] The main task of the context fusion module 105 is to effectively fuse the above two types of texts and data contents with different sources and semantic structures, so that they can reach consistency in terms of language style, logical order, semantic connection, etc., thereby improving the overall readability, completeness and user experience of the final answer. 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 the subject units, semantic centers, time directions, entity references and other elements of the two through feature extraction, and compare and align them based on the semantic structure consistency strategy. This consistency strategy can be implemented in the form of rule-driven, semantic graph matching or small model-assisted evaluation. Its core goal is to determine whether there are conflicts, duplications, omissions or logical jumps between the two inputs at the semantic level, and adjust them through language processing methods such as reordering, replacement, expansion and connection.

[0136] In specific implementations, the context fusion module 105 can perform operations such as automatic summarization, semantic insertion, and paraphrase reconstruction by designing prompt templates or calling a small-scale language model. This allows the fused text to not only cover the user's knowledge query but also clearly provide feedback on the task's execution status, forming a semantic response with a bidirectional closed-loop feature. For example, if a user asks, "Please help me schedule a meeting with Manager Wang tomorrow afternoon and tell me the sales volume of the products he was responsible for last year," the initial answer might be, "In 2023, the sales volume of the products that Manager Wang was responsible for was XX billion yuan." The task processing result might be, "Conference Room A has been reserved for you at 3:00 PM tomorrow. The meeting with Manager Wang has been successfully scheduled." The context fusion module 105 integrates these two parts through cohesive expression into "In 2023, the sales volume of the products that Manager Wang was responsible for was XX billion yuan. At the same time, Conference Room A has been reserved for you at 3:00 PM tomorrow. The meeting schedule has been successfully completed." This completes the fusion output with consistent semantics and complete information.

[0137] Furthermore, the context fusion module 105 must also possess error detection and noise reduction capabilities. This ensures that when there are inconsistencies, omissions, or conflicts between the preliminary answer and the task processing result, it can output prompt information or invoke a pre-set fallback template for structural reorganization, preventing erroneous information from being directly transmitted to the user. Furthermore, this module must preserve the traceability of the original semantic units, allowing the large language model core module 103 to perform semantic corrections or content supplementation based on user feedback when generating a comprehensive response.

[0138] In summary, the context fusion module 105 plays a key role in connecting the upper and lower parts of the system of the present invention. It not only connects the static response based on knowledge generation and the dynamic response based on task execution, but also realizes the deep integration of natural language and system behavior results through the semantic structure consistency strategy, ensuring that the comprehensive response output to the user is accurate and credible, naturally expressed, and in line with the context, thereby significantly improving the overall intelligent interaction capability and enterprise-level application adaptability of the system.

[0139] Furthermore, the context fusion module is further configured to:

[0140] Performing semantic structure extraction operations on the preliminary answer and the task processing result respectively to obtain their respective syntactic structures, coreference chains, temporal relationship labels, and context position vectors, and generating semantic structure representation objects;

[0141] Based on the matching relationship between the semantic structure representation objects, calling a semantic structure consistency scoring function, wherein the semantic structure consistency scoring function calculates a structural consistency score between the two based on syntactic subtree alignment, nested structure overlap and attention weight cross-mapping;

[0142] When the structural consistency score is lower than a preset fusion threshold, a conflict mitigation mechanism is triggered, which includes rewriting the conflicting content in the preliminary answer or the task processing result based on contextual semantic priority to make it logically connected, with clear priorities and cause-effect relationships;

[0143] Based on the rewritten semantic structure, a semantic dominant order determination is performed. According to the intention direction and information order extracted from the user's natural language input, it is decided whether the preliminary answer or the task processing result is used as the main semantic content. On the basis of maintaining the original syntactic structure, a fused text structure is generated as input for the core module of the large language model to generate a comprehensive reply.

[0144] In the intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base described in the present invention, the context fusion module is not only used to simply combine the knowledge response content and the task processing results, but also further introduces deep fusion mechanisms such as semantic structure modeling, structural consistency judgment, conflicting content rewriting and semantic dominance sorting to achieve the unity of semantic coordination, structural clarity and context consistency in natural language generation.

[0145] During system operation, when a user submits a complex natural language input containing a knowledge query and an operation request, the large language model core module first generates the preliminary answer to the user's knowledge-based question. Simultaneously, after processing by the multi-agent collaboration module, the system obtains the task processing result to respond to execution-related instruction requests. These two types of content originate from different system sub-processes and have different language styles, structural layouts, and semantic focuses. Therefore, they cannot be simply spliced ​​together, otherwise it will easily lead to chaotic response structure, logical contradictions, and even semantic fragmentation. To address this issue, the context fusion module performs semantic structure extraction operations on these two types of content separately.

[0146] This semantic structure extraction operation utilizes a hybrid technology approach based on 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 connections between each entity, action, and time node. Simultaneously, the reference chain recognition module identifies the specific entities pointed to by pronouns such as "he," "it," and "that," reconstructing the entity chain within the text. The temporal expression recognition module extracts the relationship between temporal phrases and events, such as the sequence between "last Monday's meeting" and "the topic raised by Manager Zhang." Furthermore, the sentence's contextual position vector is extracted from the original text, identifying its logical paragraph, subtask path position, and information role (background description, execution results, advisory content, etc.). All of this information together constitutes a semantic structure representation object for subsequent consistency comparison.

[0147] After completing the above-mentioned structure extraction, the context fusion module will evaluate the matching relationship between the two semantic structure representation objects and call the semantic structure consistency scoring function to score them. This scoring function is no longer a traditional string similarity matching, but an indicator system built on a multi-layer semantic structure. It mainly includes three parts: the first is the degree of alignment of syntactic subtrees, that is, the similarity score of the semantic trunk and its modified structure in the two texts at the dependency tree level; the second is the nested structure overlap, such as comparing the hierarchical structures of conditional clauses, causal relationships, and adverbial phrases in the two to see whether there is nesting inconsistency or master-slave relationship conflict; the third is based on the cross-mapping of the attention weights of the Transformer model to evaluate the relative semantic attention overlap of the two inputs in the shared attention space. These indicators together constitute the structural consistency score, which is used to determine whether the preliminary answer and the task processing result can be directly integrated in the current context.

[0148] If the structural consistency score falls below the system's set fusion threshold, for example due to conflicting subject statements, chronological errors, or unclear references, the system triggers a conflict mitigation mechanism. This conflict mitigation mechanism is a key process in fusion quality control, focusing on strategically rewriting semantic expressions without altering the factual content. The system first determines the primary and secondary relationship between the two pieces of content based on contextual semantic priorities. This determination is typically based on the user's input intent structure or sentence focus. For example, if the user input "I want to know the summary report" comes before "arrange a meeting," the system will use the preliminary answer as the primary semantic content, preserving its original language. It will then perform semantic paraphrasing on conflicting parts of the task processing result, for example, rewriting "The meeting has been scheduled" to "The topic will be further discussed in the scheduled meeting," preserving the information while avoiding structural conflicts. The entire rewriting process relies on a trained text reconstruction model and contextual reinforcement mechanism to ensure that the output is both faithful to the original meaning and natural in language.

[0149] After the rewriting is completed, the system needs to determine the structural order of the final fusion output, that is, to perform semantic dominance order judgment. The principle based on this step is the order of intent between the questions and tasks raised in the user's natural language input, while taking into account the logical fluency of the language presentation after fusion. If the user focuses on knowledge questions first, the system will place the preliminary answer in the front section and connect the subsequent task execution information with a transitional statement; on the contrary, if the user is dominated by the operation request, the task processing result will be placed in the main position, and additional knowledge supplements will be added with an insertion structure or post-instruction. At this time, the system will try its best to keep the original syntactic structure intact, avoid generating redundant or unnatural repetitive sentences, and output the fused text structure, which is complete, well-organized, and semantically consistent.

[0150] The fused text structure will be used as the final prompt input and then further processed by the core module of the large language model. Combining language generation capabilities with the contextual prompt mechanism, it will generate a comprehensive response with high semantic consistency, natural tone, and clear structure. This ensures that the output result accurately restores the system processing flow in terms of content and conforms to the cognitive logic of human users in terms of expression. It is one of the final forms of the system output response.

[0151] Furthermore, the process of generating the fused text structure in the context fusion module includes the following steps:

[0152] Extracting semantic mainline keywords, target behavior categories, and context triggering conditions involved in the user's natural language input, and injecting them into the fusion control module as fusion guidance vectors;

[0153] selecting, based on a matching result between the semantic type of the fused guidance vector and the content label of the fused text structure, a language generation template that best matches the structure from a predefined prompt template library, wherein the prompt templates include a sequential template, an embedded description template, a segmented combination template, and a semantic interleaving template;

[0154] Binding the fused text structure to the selected language generation template by position mapping, filling each semantic paragraph into the designated position according to the structure slot defined in the template, and adjusting the generated intonation, subject-verb order, and conjunction configuration according to the context content type;

[0155] The filled template text structure is used as the fused text structure and input into the core module of the large language model, driving it to generate a comprehensive response with high semantic consistency, natural tone and clear structure and output it to the user end.

[0156] In the intelligent collaborative question-and-answer system for enterprise-level schedule planning and knowledge bases described in the present invention, in order 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 controls the semantic consistency and mitigates conflicts of the content at the structural level, but also completes the standardized expression of the fused text structure through a template-driven generation mechanism, thereby ensuring that the final output content has both natural language and meets the actual requirements of semantic dominance and structural clarity. The key link in the fused text structure generation process provided by this embodiment is the technical essence of the process, which is to unify the complex structure and heterogeneous source semantic units into prompt inputs that conform to the language generation specification by combining semantic guidance with template binding, and drive the large language model core module to generate a final comprehensive answer with strong usability and complete expression.

[0157] In this embodiment, the generation process of the fused text structure begins with 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, extracting semantic mainline keywords (such as verb-centered words such as "arrange", "query", and "summary"), target behavior categories (such as "create meeting", "check report", "remind execution" and other operation types) and context trigger conditions (such as time, subject and object restriction information such as "next Monday", "about new products", and "with Manager Wang"). These elements are encoded into 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 language input by the user, thereby providing clear structural guidance.

[0158] The fusion control module then semantically matches the fusion guidance vector with the content labels in the fused text structure output by the context fusion module. These content labels are typically structural markers attached to different semantic segments during the previous fusion process, such as "knowledge response segment," "task execution result segment," "notes segment," and "secondary background segment." This annotation mechanism helps the system identify the structural functions of each semantic segment in the fused content, ensuring consistency between the main semantics and the structural presentation during matching. After matching semantic types with content labels, the system selects a language generation template from a predefined prompt template library that best matches the current content structure. This template library includes various types of template structures, including sequential templates for scenarios with clear priorities, such as "knowledge first, then execution"; embedded explanation templates for inserting one type of content into another as additional explanation; segmented combination templates for expressing parallel structures; and semantic interleaving templates for complex contexts with closely interwoven content, such as time-driven parallel task descriptions.

[0159] After determining the appropriate language generation template, the system will map and bind each semantic paragraph in the fused text structure to the structural slots defined in the template. For example, if the template has preset slots such as "<Main Answer Paragraph>", "<Background Description Paragraph>", and "<Task Result Prompt Paragraph>", the system will fill the corresponding paragraphs into the corresponding slot positions based on the content labels of each paragraph in the fused text structure. During this process, the system will also automatically fine-tune the language based on the contextual content type, including adjusting the tone intensity of the generated sentences, the order of the subject-predicate structure, the sentence style (declarative, prompting, summarizing), and the use of logical conjunctions, such as the appropriate use of "therefore", "next", "at the same time", "please note" and so on to enhance the naturalness of semantic transitions and the logical coherence of the language.

[0160] After the slots are filled, the result is the filled template text structure. This structure is an intermediate expression between the fused semantic content and the generated language form. It combines structural clarity, complete semantic coverage, and natural language. It is ready to be directly processed by the core module of the large language model as prompt input. The system sends it to the core module of the large language model. After receiving the filled template text structure, the large language model will fine-tune and optimize the coherence of each paragraph based on its language modeling capabilities, further improving the naturalness of the language expression and user readability, and ultimately generating comprehensive response content with high semantic consistency, natural tone, and clear structure. The response content will be returned directly to the user end as the system's final response to the user's original natural language input.

[0161] The entire process, from semantic guidance vector generation, template matching, content binding, to prompt input construction, strictly respects and controls the original semantic structure. This avoids language issues such as redundancy, skipping, and repetition common in traditional spliced ​​responses, significantly improving the system's language organization capabilities and output quality in complex semantic fusion scenarios. This mechanism is not only applicable to mixed question scenarios, but also provides a good framework for expanding more structural templates and adapting to language styles in multiple fields.

[0162] In the above embodiment, an intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base is provided. Correspondingly, this application also provides an intelligent collaborative question-answering method for enterprise-level schedule planning and knowledge base. Figure 2 , which is a flowchart of an embodiment of an intelligent collaborative question-answering method for enterprise-level schedule planning and knowledge bases in this application. Since this embodiment, the second embodiment, is substantially similar to the first embodiment, its description is relatively simple. For relevant details, please refer to the partial description of the first embodiment. The method embodiment described below is merely illustrative.

[0163] The second embodiment of the present application provides an intelligent collaborative question-answering method for enterprise-level schedule planning and knowledge base, including:

[0164] Step S201: receiving a natural language input from a user, and converting the natural language input from the user into a first semantic vector through a vectorization processing step, while converting pre-stored enterprise knowledge data into a second semantic vector;

[0165] Step S202: Based on the first semantic vector, searching a vector database for an enterprise knowledge fragment corresponding to a second semantic vector semantically related to the first semantic vector;

[0166] Step S203: Inputting the user's natural language input and the enterprise knowledge fragment into the large language model, performing semantic understanding and task analysis, generating a preliminary answer including knowledge response content, and identifying whether the user's natural language input contains a schedule operation request;

[0167] Step S204: If the recognition result indicates that the user's natural language input includes a schedule operation request, the schedule operation request is decomposed into one or more subtasks based on the result of the task parsing, and the subtasks are dispatched to corresponding functional agents for execution, which at least includes calling a schedule management agent to access enterprise schedule data and generate task processing results.

[0168] Step S205: receiving the preliminary answer and the task processing result, and performing context fusion on the preliminary answer and the task processing result based on a semantic structure consistency strategy;

[0169] Step S206: The context-fused content is input into the large language model again to generate a comprehensive response, which is then output to the user end to complete the question-answer response.

[0170] Although the present application is disclosed as above with the preferred embodiments, it is not intended to limit the present application. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. An intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base, characterized by: include: 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 is configured to retrieve, based on the first semantic vector, an enterprise knowledge segment corresponding to a second semantic vector semantically related to the first semantic vector in a vector database; A large language model core module is used to receive the user's natural language input and the enterprise knowledge fragment, perform semantic understanding and task analysis, generate a preliminary answer containing knowledge response content, and identify whether the user's natural language input contains a schedule operation request; A multi-agent collaboration module is configured to, upon recognizing that the user's natural language input contains a schedule operation request, decompose the schedule operation request into one or more subtasks based on the results of task analysis, and dispatch the subtasks to corresponding functional agents for execution, including at least calling a schedule management agent to access enterprise schedule data and generate task processing results; A context fusion module, configured to receive the preliminary answer and the task processing result, and contextually fuse the two based on a semantic structure consistency strategy; 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 end to complete the question-answer response; The context fusion module is further configured to: Performing semantic structure extraction operations on the preliminary answer and the task processing result respectively to obtain their respective syntactic structures, coreference chains, temporal relationship labels, and context position vectors, and generating semantic structure representation objects; Based on the matching relationship between the semantic structure representation objects, calling a semantic structure consistency scoring function, wherein the semantic structure consistency scoring function calculates a structural consistency score between the two based on syntactic subtree alignment, nested structure overlap and attention weight cross-mapping; When the structural consistency score is lower than a preset fusion threshold, a conflict mitigation mechanism is triggered, which includes rewriting the conflicting content in the preliminary answer or the task processing result based on contextual semantic priority to make it logically connected, with clear priorities and cause-effect relationships; Based on the rewritten semantic structure, a semantic dominant order determination is performed. According to the intention direction and information order extracted from the user's natural language input, it is decided whether the preliminary answer or the task processing result is used as the main semantic content. On the basis of maintaining the original syntactic structure, a fused text structure is generated as input for the core module of the large language model to generate a comprehensive reply.

2. The intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base according to claim 1 is characterized in that: The vectorized processing module is further configured to: After receiving the user's natural language input, calling an embedding model based on the Transformer architecture to convert the user's natural language input into a first semantic vector, and appending an input timestamp and a context position index to the first semantic vector to preserve the user's language input timing information; Before converting the pre-stored enterprise knowledge data into a second semantic vector, a corresponding semantic modeling strategy is selected according to the format type of the enterprise knowledge data, wherein a rule embedding method is invoked for structured data, and an embedding model identical to the first semantic vector is invoked for a natural language document to obtain a preliminary second semantic vector; performing vector normalization processing on the first semantic vector and the preliminary second semantic vector respectively, and calculating 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, a vector tuning operation is performed on the preliminary second semantic vector to enhance its cross-source semantic alignment capability with the first semantic vector. The tuned second semantic vector is used for the RAG knowledge base module to retrieve the enterprise knowledge fragment corresponding to the second semantic vector that is semantically related to the first semantic vector in the vector database.

3. The intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base according to claim 2 is characterized in that: The vector processing module performs the vector tuning operation, including the following steps: Extracting the generation time of the enterprise knowledge data corresponding to the preliminary second semantic vector, and embedding the generation time into the time dimension channel of the preliminary second semantic vector to construct a temporal second semantic vector with a time series label; Calculate the implicit temporal vector representation based on the temporal expression content involved in the context of the first semantic vector, and perform temporal alignment evaluation between the vector and the temporal second semantic vector; If the evaluation result indicates that there is a significant inconsistency in temporality between the two, the temporal weight parameter of the preliminary second semantic vector is adjusted, and attention enhancement is performed in combination with the context position index to improve the temporal matching degree; The adjusted temporal second semantic vector is used as the tuned second semantic vector.

4. The intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base according to claim 1 is characterized in that: The multi-agent collaboration module is also used to: Based on the schedule operation request in the user's natural language input identified by the large language model core module, a task graph template corresponding to the semantic structure of the schedule operation request is matched from a preset semantic-driven task decomposition mapping table, and a task graph including multiple task nodes and their dependencies is generated 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 parameters 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 in order according to the dependency relationship in the task graph, and map each task node to a corresponding functional agent, so as to call at least a schedule management agent to execute the task operation corresponding to the task node; The task execution results returned by the functional agent are collected, and all task execution results are structured and aggregated according to their logical structure in the task graph, and a task processing result set is output for the context fusion module to perform fusion processing.

5. The intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base according to claim 4 is characterized in that: The multi-agent collaboration module performs structured aggregation in the process of performing structured aggregation, including the following steps: Before the task scheduling engine schedules the task node to the functional agent, a global context state object is constructed, wherein the global context state object includes a time window, an entity reference mapping, and user role information extracted from the user natural language input; In the process of concurrently scheduling multiple task nodes, the global context state object is injected into the operating environment of each functional agent, so that each functional agent can interpret parameters and execute tasks based on consistent context state constraints when executing tasks; After receiving the task execution result of the functional agent, calculate the task execution confidence score based on the semantic fit between each functional agent's historical execution record and the current task description object, and append the confidence score to the corresponding task execution result; When constructing the task processing result set, candidate results are weighted screened and prioritized according to the confidence score in each task execution result to generate a structured task processing result set with the highest overall semantic consistency and context fit.

6. The intelligent collaborative question-answering system for enterprise-level schedule planning and knowledge base according to claim 1 is characterized in that: The process of generating the fused text structure in the context fusion module includes the following steps: Extracting semantic mainline keywords, target behavior categories, and context triggering conditions involved in the user's natural language input, and injecting them into the fusion control module as fusion guidance vectors; selecting, based on a matching result between the semantic type of the fused guidance vector and the content label of the fused text structure, a language generation template that best matches the structure from a predefined prompt template library, wherein the prompt templates include a sequential template, an embedded description template, a segmented combination template, and a semantic interleaving template; Binding the fused text structure to the selected language generation template by position mapping, filling each semantic paragraph into the designated position according to the structure slot defined in the template, and adjusting the generated intonation, subject-verb order, and conjunction configuration according to the context content type; The filled template text structure is used as the fused text structure and input into the core module of the large language model, driving it to generate a comprehensive response with high semantic consistency, natural tone and clear structure and output it to the user end.

7. An intelligent collaborative question-answering method for enterprise-level schedule planning and knowledge base, characterized by: include: Receiving a natural language input from a user, and converting the natural language input from the user into a first semantic vector through a vectorization processing step, while converting pre-stored enterprise knowledge data into a second semantic vector; Based on the first semantic vector, searching a vector database for an enterprise knowledge fragment corresponding to a second semantic vector semantically related to the first semantic vector; Input the user's natural language input and the enterprise knowledge fragment into the large language model, perform semantic understanding and task analysis, generate a preliminary answer containing knowledge response content, and identify whether the user's natural language input contains a schedule operation request; If the recognition result indicates that the user's natural language input includes a schedule operation request, the schedule operation request is decomposed into one or more subtasks according to the result of the task parsing, and the subtasks are dispatched to corresponding functional agents for execution, which at least includes calling a schedule management agent to access enterprise schedule data and generate task processing results; receiving the preliminary answer and the task processing result, and performing context fusion on the preliminary answer and the task processing result based on a semantic structure consistency strategy; Performing semantic structure extraction operations on the preliminary answer and the task processing result respectively to obtain their respective syntactic structures, coreference chains, temporal relationship labels, and context position vectors, and generating semantic structure representation objects; Based on the matching relationship between the semantic structure representation objects, calling a semantic structure consistency scoring function, wherein the semantic structure consistency scoring function calculates a structural consistency score between the two based on syntactic subtree alignment, nested structure overlap and attention weight cross-mapping; When the structural consistency score is lower than a preset fusion threshold, a conflict mitigation mechanism is triggered, which includes rewriting the conflicting content in the preliminary answer or the task processing result based on contextual semantic priority to make it logically connected, with clear priorities and cause-effect relationships; Based on the rewritten semantic structure, a semantic dominance order determination is performed. According to the intent direction and information order extracted from the user's natural language input, a decision is made as to whether the preliminary answer or the task processing result should be used as the main semantic content. A fused text structure is generated while maintaining the original syntactic structure, and used as input for the core module of the large language model to generate a comprehensive response. The context-fused content is input into the large language model again to generate a comprehensive response, which is then output to the user end to complete the question-answering response.

Images