Hierarchical answering method, electronic equipment and computer readable storage medium

Abstract

The invention discloses a hierarchical answering method, electronic equipment and a computer readable storage medium, and relates to the technical field of machine question answering, and the method comprises the steps: querying related document contents matched with an original question text from a document knowledge base under the condition that the original question text is received; based on the related document content, the original question text and the hierarchical answer cue word template, constructing hierarchical answer cue words; inputting the graded answer prompt words into a preset question and answer model to obtain an answer result, and displaying the answer result in a question and answer dialogue interface; wherein the hierarchical answer prompt word is used for indicating the preset question and answer model to perform hierarchical answer on the original question text according to the matching degree between the related document content and the original question text, the level of the answer result is in positive correlation with the matching degree, and the confidence of the answer result is in positive correlation with the level of the answer result; the answer results of different levels have different answer content forms. According to the method, the answer result credibility of the intelligent question-answering system can be clarified.

Description

Technical Field

[0001] This application relates to the field of machine question-answering technology, and in particular to graded answering methods, electronic devices, and computer-readable storage media. Background Technology

[0002] Over time, organizations such as businesses, schools, and medical institutions have accumulated a large number of unstructured documents in their daily operations, including rules and regulations, operation manuals, training videos, teaching videos, and product introduction PPTs (PowerPoint presentations). This information is stored in a scattered manner and in various formats, leading to inefficient knowledge acquisition. At the same time, the need for organizational members to efficiently acquire accurate knowledge is growing.

[0003] Against this backdrop, knowledge-based intelligent question-answering systems have become important tools for improving information retrieval efficiency, reducing communication costs, and ensuring business compliance. The accuracy, credibility, and traceability of their answers directly affect user experience and decision-making quality.

[0004] However, existing intelligent question-answering systems generally suffer from unclear credibility of their answers in practical applications, making it difficult for users to determine the authenticity and reliability of the responses. Summary of the Invention

[0005] The main purpose of this application is to provide a graded answering method, electronic device, and computer-readable storage medium, which aims to solve the technical problem of unclear credibility of the answer results of intelligent question answering systems.

[0006] To achieve the above objectives, this application provides a tiered response method, the method comprising: Upon receiving the original question text input through the question-and-answer dialogue interface, the system queries a preset document knowledge base for relevant document content that matches the original question text. Based on the relevant document content, the original question text, and the preset hierarchical answer prompt word template, a hierarchical answer prompt word is constructed; The graded answer prompts are input into a preset question-and-answer model to obtain the answer to the original question text, and the answer is displayed in the question-and-answer dialogue interface. The graded response prompts are used to instruct the preset question-and-answer model to provide graded responses to the original question text based on the degree of matching between the relevant document content and the original question text. The grade of the response is positively correlated with the degree of matching, and the confidence level of the response is positively correlated with the grade of the response. Different grades of response results have different response content formats.

[0007] Furthermore, to achieve the above objectives, this application also provides a graded response device, the device comprising: The retrieval module is used to query relevant document content that matches the original question text from a preset document knowledge base when the original question text is received through the question-and-answer dialogue interface. The prompting module is used to construct hierarchical answer prompts based on the relevant document content, the original question text, and a preset hierarchical answer prompt template; The answer module is used to input the hierarchical answer prompts into a preset question-and-answer model, obtain the answer result of the original question text, and display the answer result in the question-and-answer dialogue interface; The graded response prompts are used to instruct the preset question-and-answer model to provide graded responses to the original question text based on the degree of matching between the relevant document content and the original question text. The grade of the response is positively correlated with the degree of matching, and the confidence level of the response is positively correlated with the grade of the response. Different grades of response results have different response content formats.

[0008] In addition, to achieve the above objectives, this application also provides an electronic device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the graded response method as described above.

[0009] In addition, to achieve the above objectives, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the graded response method described above.

[0010] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the graded response method described above.

[0011] This embodiment of the application, upon receiving the original question text input through the question-and-answer dialogue interface, first queries relevant document content matching it from a preset document knowledge base. Then, based on the relevant document content, the original question text, and a preset hierarchical answer prompt word template, it constructs hierarchical answer prompt words. Finally, it inputs the hierarchical answer prompt words into a preset question-and-answer model to obtain the model-generated answer result, which is then displayed on the document dialogue interface. Since the hierarchical answer prompt words can instruct the preset question-and-answer model to generate answer results with different answer content formats according to the degree of matching between the relevant document content and the original question text, this embodiment of the application, without changing the underlying model architecture, guides the model to output answer results with differentiated answer content formats based on the availability of reliable knowledge content (i.e., relevant document content) and the degree of matching between the reliable knowledge content and the user's question. This allows the answer results seen by the user in the question-and-answer dialogue interface to implicitly reflect the sufficiency of its generation basis, thereby improving the subjective judgment ability of the authenticity and reliability of the answer results without relying on external explanations. This alleviates, to some extent, the technical problem of unclear answer credibility caused by the opaque source of answers in existing intelligent question-and-answer systems. Attached Figure Description

[0012] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating the first embodiment of the graded response method of this application; Figure 2 This is a flowchart illustrating the second embodiment of the graded response method of this application. Figure 3 This is a flowchart illustrating the third embodiment of the graded response method in this application. Figure 4 This is a schematic diagram of the module structure of the hierarchical response device according to an embodiment of this application; Figure 5 This is a schematic diagram of the hardware operating environment of the electronic device involved in the hierarchical response method in this application embodiment.

[0015] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0017] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0018] Currently, intelligent question-answering systems generally suffer from unclear credibility in practical applications, making it difficult for users to determine the authenticity and reliability of the answers.

[0019] To address the aforementioned issues, the main solution of this application embodiment is as follows: Upon receiving the original question text input through a question-and-answer dialogue interface, relevant document content matching the original question text is queried from a preset document knowledge base; based on the relevant document content, the original question text, and a preset graded answer prompt word template, graded answer prompt words are constructed; the graded answer prompt words are input into a preset question-and-answer model to obtain the answer result for the original question text, and the answer result is displayed in the question-and-answer dialogue interface; wherein, the graded answer prompt words are used to instruct the preset question-and-answer model to provide a graded answer to the original question text according to the degree of matching between the relevant document content and the original question text, and the level of the answer result is positively correlated with the degree of matching, the confidence level of the answer result is positively correlated with the level of the answer result, and different levels of answer results have different answer content formats.

[0020] This embodiment of the application, upon receiving the original question text input through the question-and-answer dialogue interface, first queries relevant document content matching it from a preset document knowledge base. Then, based on the relevant document content, the original question text, and a preset hierarchical answer prompt word template, it constructs hierarchical answer prompt words. Finally, it inputs the hierarchical answer prompt words into a preset question-and-answer model to obtain the model-generated answer result, which is then displayed on the document dialogue interface. Since the hierarchical answer prompt words can instruct the preset question-and-answer model to generate answer results with different answer content formats according to the degree of matching between the relevant document content and the original question text, this embodiment of the application, without changing the underlying model architecture, guides the model to output answer results with differentiated answer content formats based on the availability of reliable knowledge content (i.e., relevant document content) and the degree of matching between the reliable knowledge content and the user's question. This allows the answer results seen by the user in the question-and-answer dialogue interface to implicitly reflect the sufficiency of its generation basis, thereby improving the subjective judgment ability of the authenticity and reliability of the answer results without relying on external explanations. This alleviates, to some extent, the technical problem of unclear answer credibility caused by the opaque source of answers in existing intelligent question-and-answer systems.

[0021] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0022] This application proposes a hierarchical response method according to a first embodiment.

[0023] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the hierarchical response method of this application.

[0024] In this embodiment, the graded response method may include steps S100~S300: Step S100: Upon receiving the original question text input through the question-and-answer dialogue interface, query the relevant document content that matches the original question text from the preset document knowledge base; It should be noted that, in this embodiment, the question-and-answer dialogue interface refers to the front-end display interface through which users interact with the intelligent question-and-answer system (which can be implemented through an intelligent agent). This interface can be a web-based chat window, a dialog box in a mobile application, or a chatbot conversation area within an enterprise collaboration platform. This interface supports text input, message sending, answer rendering, and media content presentation, and is the core channel for users to initiate questions and receive system responses.

[0025] The original question text refers to the question content in natural language form that the user inputs into the intelligent question-answering system through the question-answering interaction interface. It is the input starting point of the entire intelligent question-answering process.

[0026] A document knowledge base refers to a collection of knowledge pre-built based on various unstructured documents accumulated by an organization in its daily operations, supporting efficient retrieval functions based on semantics or keywords.

[0027] It is important to note that in constructing this document knowledge base based on various unstructured documents (typically including unstructured documents in different formats such as video, audio, PPT, image, and text), a unified conversion strategy is adopted for unstructured documents in non-text format: First, the unstructured documents in non-text format are converted into text format using appropriate conversion tools or algorithms. For example, for unstructured documents in video or audio format, speech recognition technology can be used to convert their content into text, thus obtaining their unstructured text format; for unstructured text in PPT or PDF (Portable Document Format) format, optical character recognition technology and a specific parser can be used to extract the text information, thus obtaining their unstructured text format. After the above processing, all unstructured documents that originally existed in non-text format are converted into unstructured text format documents that are easy to process and analyze, and stored in the document knowledge base in a unified format, forming a knowledge foundation that can be used for subsequent retrieval and analysis.

[0028] By uniformly converting multimodal unstructured documents such as videos, audios, PPTs, PDFs, and images into standardized text formats before constructing the document knowledge base, this embodiment achieves deep integration and efficient management of heterogeneous data sources. This processing method not only eliminates semantic barriers and technical obstacles between different document formats but also provides a highly consistent data foundation for subsequent knowledge retrieval and intelligent question answering. Since all document content exists in text format, the system can directly use a unified natural language processing workflow for index building, semantic vectorization, and similarity calculation, avoiding parsing failures or information loss caused by format differences, and significantly improving the integrity and stability of the document knowledge base construction. At the same time, the above-mentioned text processing method greatly enhances the system's compatibility and scalability. Whether the newly added content is the text transcription of training videos, the recognition results of meeting recordings, or the PPT content of product manuals, it can be seamlessly integrated into the document knowledge base through the same conversion path, without the need to develop separate adaptation modules for each document format.

[0029] More importantly, the unified text format allows semantic-based vector retrieval, keyword matching, and contextual analysis to function effectively. This ensures that when faced with user inquiries, the system can accurately extract relevant document content from cross-media and cross-type source materials, thereby guaranteeing the comprehensiveness and authority of the answers. Therefore, this document knowledge base not only possesses excellent maintainability and scalability but also provides a reliable source of facts for judging the "sufficiency of evidence" in the tiered answering method, fundamentally supporting the subsequent technical goal of generating differentiated answers based on the strength of knowledge evidence.

[0030] It should also be noted that, in this embodiment, the relevant document content includes at least document fragments extracted from a document knowledge base using at least one retrieval technique that match the original question text. A document fragment refers to a specific part or subset extracted from a complete document. This relevant document content serves as the factual basis for generating a credible answer, and the degree of matching between it and the original question text directly affects the accuracy and authority of the answer.

[0031] In this embodiment, after receiving the original question text input through the question-and-answer dialogue interface, the system immediately initiates the retrieval process. Using semantic vector retrieval, keyword matching, or a hybrid retrieval strategy, it locates and extracts the most relevant document content from the document knowledge base, which serves as the basic input for generating tiered answers.

[0032] This embodiment achieves knowledge support for user questions by querying relevant document content matching the original question text from a pre-set document knowledge base. Its technical advantages are: providing traceable factual evidence for subsequent answer generation, avoiding the risk of large-scale model "illusions" or "fabricated answers," and is a prerequisite for achieving credible question-and-answer. Simultaneously, this step ensures that the system's output answers are highly correlated with the organization's actual knowledge, improving the accuracy and practicality of the question-and-answer results, and laying a solid data foundation for the next step of tiered response based on matching degree.

[0033] Step S200: Based on the relevant document content, the original question text, and the preset hierarchical answer prompt word template, construct hierarchical answer prompt words; It should be noted that the tiered answer prompt template refers to a standardized prompt framework pre-designed and configured within the system. Embedded within it are instruction logic and format constraints that guide the large model (i.e., the pre-defined question-answering model) to generate different levels of answer results based on the sufficiency of knowledge (i.e., relevant document content). This tiered answer prompt template exists in the form of natural language plus structured markup, explicitly instructing the large model to adopt different answer strategies based on the varying degrees of matching between the knowledge base and the user's question. For example, "If there is a direct match, please cite the original text and indicate the source; if only indirect relevant information exists, please answer based on contextual reasoning; if there is no relevant information, please truthfully inform us and provide suggestions based on your own knowledge."

[0034] Tiered answer prompts refer to the final input prompts generated by dynamically integrating the original question text, the relevant document content obtained from the query, and the tiered answer prompt template. These tiered answer prompts not only contain the user's question and contextual information, but more importantly, they encapsulate the control instruction to "generate an answer result of the corresponding level based on the degree of matching between the relevant document content and the original question text," thereby guiding the preset question-answering model to reflect differentiated expression when generating answers.

[0035] In this embodiment, the system injects the original question text and relevant document content into a tiered answer prompt template. After formatted splicing and semantic integration, a complete prompt with tiered guidance capabilities is generated. For example, when the relevant document content contains a clause that perfectly matches the question, the prompt will guide the model to generate an answer in the form of "According to Article X of the 'XXX System'..."; when only some relevant information exists, the system guides the model to use expressions such as "Based on the analysis of existing data, it is possible..."; if there is no relevant content, the prompt will prompt the model to explicitly state "No relevant information found".

[0036] This embodiment achieves fine-grained control over the output behavior of a large language model by constructing prompt words containing hierarchical control logic. Its technical advantage lies in the fact that it achieves differentiated expression of answer formats solely through prompt word engineering without modifying the model's underlying parameters or architecture. This is a lightweight, flexible, and rapidly iterative technical solution. The core function of this step is to transform the implicit states of "whether there is knowledge basis" and "the matching strength of the basis" into explicit instructions that the model can understand. This provides a decision-making basis for subsequently generating answer results with confidence differentiation, effectively improving the controllability and transparency of the question-answering system.

[0037] Step S300: Input the graded answer prompts into the preset question-and-answer model to obtain the answer result of the original question text, and display the answer result in the question-and-answer dialogue interface; The graded response prompts are used to instruct the preset question-and-answer model to provide graded responses to the original question text based on the degree of matching between the relevant document content and the original question text. The grade of the response is positively correlated with the degree of matching, and the confidence level of the response is positively correlated with the grade of the response. Different grades of response results have different response content formats.

[0038] It should be noted that the preset question-answering model refers to a large model (large language model or multimodal large model) deployed in the system. It is trained on a general corpus and can accept external prompt words as input, enabling it to complete natural language understanding and generation tasks. This preset question-answering model can be an open-source model or a commercial model, which has powerful language generation capabilities. However, in this embodiment, it is guided to generate answer results with structural differences by receiving tiered answer prompt words.

[0039] The content format of an answer refers to the comprehensive characteristics of the answer in terms of language expression, sentence structure, information completeness, and citation style. For example: A high-level response should begin with "According to Article 3.2 of the Employee Handbook..." and directly quote the original paragraph. Intermediate level answer: Use phrases like "based on certain information, we infer..." to demonstrate logical reasoning. A low-level response: Clearly state "No relevant information was found in the current knowledge base" to avoid misleading users.

[0040] It should also be noted that, in this embodiment, tiered response refers to the process by which the question-answering model generates response results of different levels, confidence levels, and content formats based on the response strategy instructions in the tiered response prompts and the degree of matching between the relevant document content and the original question text. This process realizes a hierarchical response mechanism from "based on evidence" to "supplementary speculation" and then to "explanation without evidence".

[0041] In this embodiment, the system inputs the constructed hierarchical answer prompts into a preset question-and-answer model, obtains the generated answer results, and renders them onto the question-and-answer dialogue interface corresponding to the user's question for the user to view. Since the answer results at different levels have significant differences in language style, format, and information presentation (i.e., different answer content formats), the user can intuitively perceive the reliability of the answer result without additional explanation.

[0042] This embodiment completes a closed-loop output from "knowledge retrieval" to "credible expression" by inputting tiered answer prompts into a preset question-and-answer model and displaying differentiated answer results. Its technical effect is that it significantly improves users' ability to judge the authenticity and reliability of answers, alleviating the "uncertainty of credibility" problem caused by the lack of transparency in the source of answers in existing intelligent question-and-answer systems. Especially in high-risk scenarios involving interpretation of regulations and confirmation of processes, users can quickly assess the authority and applicability of answers based on their format, thereby enhancing their trust in the system and their willingness to use it. This step, as the final execution link of the entire method, realizes the visual presentation of technical value and is key to improving the quality of organizational knowledge services and the human-computer interaction experience.

[0043] This embodiment, upon receiving the original question text input through the question-and-answer dialogue interface, first queries a pre-set document knowledge base for relevant document content that matches the original question text. Then, based on the relevant document content, the original question text, and a pre-set hierarchical answer prompt template, it constructs hierarchical answer prompts. Finally, it inputs these hierarchical answer prompts into a pre-set question-and-answer model to obtain the model-generated answer and displays it on the document dialogue interface. Because these hierarchical answer prompts instruct the pre-set question-and-answer model to generate different answer content formats based on the degree of matching between the relevant document content and the original question text, this embodiment, without changing the underlying model architecture, guides the model to output differentiated answer content formats based on the availability of reliable knowledge content (i.e., relevant document content) and the degree of matching between the reliable knowledge content and the user's question. This allows the answer results seen by the user in the question-and-answer dialogue interface to implicitly reflect the sufficiency of its generation basis, thereby improving the subjective judgment ability regarding the authenticity and reliability of the answer results without relying on external explanations. This alleviates, to some extent, the technical problem of unclear answer credibility caused by the opaque source of answers in existing intelligent question-and-answer systems.

[0044] In one feasible implementation, the hierarchical answer prompt template carries multiple levels of answer strategies, and different levels of answer strategies correspond to instruct the preset question-and-answer model to generate answer results in different answer content formats; The tiered answer prompts carry a target answer strategy, which is the answer strategy among the multiple levels of answer strategies that corresponds to the degree of matching, and the level of the target answer strategy is positively correlated with the degree of matching.

[0045] It should be noted that, in this embodiment, the answer strategy refers to a set of rule-based instructions preset in the hierarchical answer prompt template, used to explicitly instruct the preset question-answering model to generate answers under specific knowledge conditions. Each level of the answer strategy corresponds to a level of answer result and specifies the answer content format that the answer result at that level should possess, such as language style, dependence on information sources, citation methods, and rigor of expression. For example, a high-level answer strategy requires the model to strictly cite the original text based on the document fragment and indicate the source; a medium-level answer strategy allows the model to perform logical reasoning and induction based on relevant document content; a low-level answer strategy instructs the model to provide general suggestions and explicitly declare knowledge gaps based on its own training knowledge when there is no direct or indirect matching content (i.e., the matching degree is too low, including the absence of relevant document content matching the original question text, i.e., the relevant document content is empty). Multiple levels of answer strategies together constitute a hierarchical, progressively decreasing response system, ensuring that the system can output matching answer behavior according to the sufficiency of knowledge support.

[0046] "Target answer strategy" refers to the optimal answer strategy dynamically selected by the system from multiple levels of answer strategies during the question-and-answer process, based on the degree of matching between the content of the relevant documents retrieved and the original question text. This matching degree can be evaluated using quantitative indicators such as semantic similarity score, keyword coverage, and vector space distance, and mapped to a preset strategy level range. For example, when the matching degree is higher than the first threshold, the first answer strategy (high-confidence citation) is selected; when the matching degree is between the second and first thresholds, the second answer strategy (medium-confidence inference) is selected; and when the matching degree is lower than the second threshold, the third answer strategy (low-confidence explanatory) is selected, where the first threshold is greater than the second threshold. This target answer strategy is embedded in the final constructed hierarchical answer prompts, serving as the core instruction unit controlling the model's output behavior.

[0047] Furthermore, this matching degree can also be assessed by inputting relevant document content and the original question text into a large model, which then evaluates the degree of matching between the two. It's easy to understand that in this process, pre-prepared criteria for judging the matching degree can be input into the large model, meaning the three elements together form a set of prompts specifically designed to guide the large model in evaluating the matching degree.

[0048] It is worth mentioning that this implementation method can determine the target answer strategy before constructing the hierarchical answer prompt words, so that the hierarchical answer prompt words only carry the target answer strategy, so as to avoid other levels of answer strategies interfering with the model's reasoning. Alternatively, all levels of answer strategies, as well as the criteria for judging the degree of matching or the already determined degree of matching, can be put into the hierarchical answer prompt words, and the preset question-and-answer model can judge on its own to determine the target answer strategy from multiple levels of answer strategies before performing hierarchical answering.

[0049] This implementation method achieves refined and structured guidance of the output behavior of the preset question-answering model by pre-setting multiple levels of answer strategies in the hierarchical answer prompt template and dynamically selecting the target answer strategy based on the matching degree. Its technical effect is that it transforms the originally ambiguous "do I know?" question into a quantifiable decision-making process of "how much do I know?" and "how do I express that knowledge?", enabling the system to adaptively adjust its answering method according to the strength of the knowledge basis (i.e., the degree of matching), thereby constructing an intelligent question-answering mechanism with a hierarchical sense and credibility gradient. This mechanism not only improves the transparency and interpretability of the answer results but also effectively avoids the "illusion" risk caused by large models forcing answers when lacking evidence, enhancing user trust in the system's output. More importantly, by positively correlating the matching degree with the answer strategy, the system can externally control the confidence of the answer without retraining the model, possessing good engineering practicality and deployment flexibility, and providing organizational-level knowledge service systems with auditable, traceable, and hierarchical intelligent response capabilities.

[0050] In one feasible implementation, the relevant document content includes document fragments of the target document, wherein the target document is a document in the document knowledge base that matches the original question text; The multiple levels of response strategies include a first response strategy, a second response strategy, and a third response strategy, wherein the first response strategy is at a higher level than the second response strategy, and the second response strategy is at a higher level than the third response strategy. When the target answering strategy is the first answering strategy, the answering result includes the first answering text generated by the preset question-and-answer model directly referencing the document fragment; When the target answering strategy is the second answering strategy, the answering result includes the second answer text obtained by the preset question-answering model based on the document fragment reasoning; When the target answering strategy is the third answering strategy, the answering result includes the third answer text obtained by the preset question-answering model based on the model's own knowledge reasoning.

[0051] It should be noted that, in this embodiment, the first answer strategy refers to the highest-level answer strategy among multiple levels. It instructs the preset question-answering model to strictly generate an answer based on the original text of the document fragment when it confirms that a document fragment highly matches the original question text exists within the relevant document content. This first answer strategy requires the model to output content through direct citation and prioritizes authoritative and traceable expressions such as "According to Article X of the 'Regulations X'..." or "Document stipulates..." to ensure that the answer content is completely consistent with the organization's internal knowledge. This strategy is suitable for scenarios with sufficient knowledge base and high semantic matching.

[0052] The second response strategy refers to a response strategy that falls between the highest and lowest levels among multiple response strategies. It instructs the pre-defined question-answering model to generate an answer based on semantic understanding, logical reasoning, or inductive summarization of the retrieved document fragments when there is a partial match or indirect relationship between the relevant document content and the original question text (such as contextual relevance or similar but not directly corresponding themes). This second response strategy allows the model to appropriately paraphrase or interpret the original content in the document fragments, typically using expressions such as "combining existing data analysis..." to demonstrate the derivation process. It is suitable for scenarios where knowledge exists but cannot be directly used to answer the user's question, requiring further interpretation and derivation to provide an answer.

[0053] The third-level response strategy is the lowest level among multiple response strategies. It instructs the pre-defined question-answering model to infer and generate a response based on general knowledge learned during its training when no relevant document content matching the original question text is found in the document knowledge base, or when the degree of matching between the retrieved relevant document content and the original question text is too low. This third-level response strategy does not rely on internal organizational knowledge sources, and the output content is not directly traceable. It typically uses phrases such as "No relevant information was found," "Generally speaking…," or "Based on common practices…" to clearly indicate to the user that the response is based on the model's own knowledge, aiming to avoid silent responses and reduce the risk of misleading questions. It is suitable for scenarios involving knowledge gaps or questions beyond the scope of the question.

[0054] It should also be noted that, in this embodiment, the first response text refers to the response content generated by the preset question-and-answer model under the guidance of the first response strategy, strictly based on the original text of the document fragments in the relevant document content. Its core feature is direct quotation. This type of response text usually begins with sentences such as "According to Article X of the 'XXX System'..." or "The document clearly states..." and completely retains the language expression of the original document, only making necessary semantic connection processing to ensure the accuracy and authority of information transmission.

[0055] The second response text refers to the response content generated by the pre-defined question-and-answer model after semantic understanding and logical reasoning based on the original content of the document fragment, guided by the second response strategy. It does not directly copy the original text, but presents information through induction, explanation, or contextualized paraphrasing. It is suitable for scenarios where the document content is indirectly related to the user's question but requires further interpretation.

[0056] The third-response text refers to the response generated by the pre-defined question-answering model when it determines that there is no valid matching content in the document knowledge base under the guidance of the third-response strategy. The model then relies on the knowledge it has learned during training to reason and generate the response. Although such responses have some reference value, they lack direct support from the organization's internal knowledge. They are usually expressed in the form of "generally speaking...", "usually recommended...", or "no relevant information was found at present, but according to general practice..." to indicate the difference in the source of their basis.

[0057] This implementation method establishes a three-tiered response mechanism of "document citation—knowledge reasoning—model self-awareness," achieving refined layering and traceability control of the answer generation mode. Its technical effects are as follows: Firstly, when there is highly relevant knowledge, the system prioritizes outputting the first answer text using original text citation, maximizing the accuracy and compliance of the answer, suitable for high-risk decision-making scenarios such as policy inquiries and process confirmations. Secondly, when the knowledge basis is weak but related information exists, the second answer text achieves a semantic bridge from the original data to the user's question, improving the system's adaptability to ambiguous questions. Thirdly, when there is a complete lack of internal knowledge support, the third answer text avoids the interaction interruption caused by "no answer" and implicitly hints to the user about the externality and speculative nature of the answer through differences in expression, preventing misleading. This three-tiered strategy not only constructs a clear credibility gradient, allowing users to judge the authority of the answer from its form, but also enhances the transparency and controllability of the entire question-and-answer system through the binding relationship between the strategy and the content source. More importantly, this mechanism does not require fine-tuning or retraining of the preset question-answering model. It can dynamically adjust the output behavior simply by guiding the strategy in the prompt words. It has good scalability and engineering application value, and provides a hierarchical response solution for organizational-level intelligent question-answering systems that takes into account accuracy, practicality and security.

[0058] In one feasible implementation, the relevant document content also includes the document identification information and document source information of the target document; If the target answer strategy is not the third answer strategy, the answer result also includes the document identification information and the document source information.

[0059] It should be noted that, in this embodiment, document identification information refers to metadata information used to uniquely identify the target document. This typically includes the document's name, number, version number, release date, or internal code, such as "Employee Handbook_V2.1" or "Procurement Process Specification - 2024 Edition." This document identification information helps users accurately locate the original knowledge carrier, avoiding misunderstandings caused by documents with the same name, multiple versions, or updated content.

[0060] Document source information refers to information used to indicate the organizational unit or knowledge attribution path to which the target document belongs, such as department name (e.g., "Human Resources Department"), knowledge category directory (e.g., "System Management / Personnel Management"), document archiving path, or system source identifier (e.g., "From Enterprise Knowledge Base - Training Materials Module"). This document source information reflects the authority and scope of application of the target document, helping users determine the applicable context and responsible party for its content.

[0061] This implementation significantly enhances the traceability and credibility of question-and-answer output by incorporating document identification information and document source information into the answer result in scenarios where a third-party answer strategy is not employed. The technical effect is that when the system generates an answer based on internal organizational knowledge (i.e., using a first or second answer strategy), the added document identification and source information constitute a complete knowledge reference chain. This allows users not only to know the answer content but also to clearly identify which document the answer originates from, which department issued it, and which knowledge category it belongs to, thus achieving a transparent response that is "verifiable and has a clear source." This mechanism effectively increases users' subjective trust in the authenticity of the answer, and has significant practical value, especially in key business scenarios involving policy implementation, compliance review, and responsibility definition. Simultaneously, by setting the conditional logic of "only displaying when the target answer strategy is not a third-party answer strategy," the system avoids forcibly associating false sources when the model relies on its own knowledge to answer, preventing users from being misled into believing that the answer comes from an internal document, further ensuring the honesty of information expression and the standardization of system behavior. Therefore, this implementation method achieves automated annotation and credibility-level presentation of knowledge citations without increasing the cognitive burden on users, providing strong support for building a highly transparent and auditable enterprise-level intelligent question-answering system.

[0062] It is easy to understand that, in addition to document identification information and document source information, the relevant document content can further include fragment identification information used to uniquely identify document fragments, so that, in the case of the third answer strategy in the answer strategy section, the fragment identification information of the document fragments can be included in the answer results.

[0063] In one feasible implementation, the relevant document content also includes a media segment link corresponding to the document segment, wherein the media segment link points to the media segment from which the document segment originates; If the target response strategy is not the third response strategy, the response result also includes the media clip link.

[0064] Those skilled in the art will recognize that media refers to any form of content carrier used to carry and transmit information, including but not limited to text, images, audio, video, presentations (such as PowerPoint), web pages, spreadsheets, PDF documents, etc.

[0065] A media segment refers to a logical subunit or time interval extracted from a complete media resource, corresponding to specific semantic content. For example, in a training video, the time interval from 4 minutes 20 seconds to 5 minutes 10 seconds explaining the "annual leave application process" constitutes a media segment; in a product description PowerPoint presentation, the slide on page 6 about "core parameters" can be considered a media segment; and in a digital policy document, the content of the second section of chapter 3 can also be divided into a text-based media segment. A media segment is the smallest referable unit in a media resource, typically possessing independent information expression functions, and can be located and indexed through metadata or structural tags.

[0066] A media clip link is a digital reference identifier used to locate and access the aforementioned media clip. It typically manifests as a Uniform Resource Identifier (URI), a hyperlink address, or an access path with anchor points / timestamps. This media clip link not only points to the storage location of the original media resource but also precisely marks the start and end times, page numbers, chapter numbers, or data areas of the target content within that media resource. Users can directly jump to the specific location of the relevant content by clicking, enabling fast and accurate content navigation.

[0067] In this embodiment, there is a semantic mapping and source-stream correspondence between text-based document fragments in the document knowledge base and the media fragments linked by media fragment links. Specifically, during the construction of the document knowledge base, the system has performed content extraction and structured parsing on various original media resources (such as unstructured documents in different document formats like videos, audio, PPTs, PDFs, and images), transforming their non-text media fragments into text-based document fragments that can be stored in the document knowledge base, and establishing an association index between the two (e.g., a media fragment link pointing to the media fragment). For example, the text "The 2024 budget approval process will adopt an online countersigning mechanism" generated from a meeting recording after speech recognition is stored in the knowledge base as a text-based document fragment, and the system binds it to a media fragment link pointing to the time interval of the statement in the original audio. Therefore, this document fragment is not generated independently, but serves as a semantic summary and retrieval entry point of the original media content, maintaining content consistency and spatiotemporal traceability with its source media fragment.

[0068] This implementation method achieves a verifiable jump from structured text answers to original knowledge carriers by introducing media clip links into the answer results. Its technical effect is that when the system generates non-minimum-level answer results based on internal organizational knowledge (i.e., the target answer strategy is the first or second answer strategy), users can not only obtain the first or second answer text generated by the model from document clips, but also directly view the actual context of the document clip in the original media resource by clicking the media clip link. This includes auxiliary information such as explanatory context, charts, tone, or document layout, thereby enhancing the depth of understanding and trust in the answer. Especially in scenarios with high credibility requirements such as policy interpretation, process confirmation, and training review, this mechanism significantly improves the transparency and knowledge traceability capabilities of the question-and-answer system.

[0069] Meanwhile, by limiting the inclusion of the media clip link to "only when the target answer strategy is not the third answer strategy," it ensures that the link is only displayed when the answer is supported by genuine knowledge, avoiding the introduction of false sources when the model generates speculative answers based on its own knowledge, and effectively maintaining the authenticity of the system output and the accuracy of user perception. Therefore, this implementation method, based on more general terminology, constructs a complete knowledge service closed loop from "question → retrieval → generation → source tracing," further enhancing the practicality, reliability, and auditability of the intelligent question-answering system in complex organizational environments.

[0070] In one feasible implementation, after the step of displaying the answer result in the question-and-answer dialogue interface in step S300 above, the graded answering method may further include step S400: Step S400: If the answer result includes the media segment link, generate a media presentation component in the question-and-answer dialogue interface based on the media segment link, and display the media segment through the media presentation component.

[0071] It should be noted that, in this embodiment, the media presentation component refers to a visualization function module deployed within the question-and-answer dialogue interface. It is used to parse the content type and location information pointed to by the media segment link, and directly render and play the corresponding media segment in an embedded manner without exiting the current dialogue environment. This media presentation component can dynamically adapt its presentation format according to the media type pointed to by the link. For example, when the link points to a video resource, a video window with playback controls is generated; when the link points to an audio resource, an audio waveform and a play button are generated; when the link points to a PPT or PDF document, a lightweight document preview box with page turning is generated. The media presentation component supports automatic positioning to the start time point or target page number of the media segment and can be configured for various interactive modes such as automatic playback, manual triggering, or displaying only thumbnails, ensuring that users can conveniently and intuitively access the original knowledge content.

[0072] When displaying media clips through this media presentation component, it is not limited to static display, but also supports operation functions such as play, pause, progress adjustment, page switching, and volume control. This allows users to directly experience the original media information in the current question-and-answer context without leaving the current dialogue interface to search for or open external applications, thereby achieving a seamless knowledge acquisition experience of "ask and see, see and understand instantly".

[0073] It's worth noting that the aforementioned media presentation components can be rendered and displayed in the question-and-answer dialogue interface using front-end markup languages ​​or data formats such as HTML (HyperText Markup Language), JSON (JavaScript Object Notation), or Markdown. Furthermore, a large model (including a pre-defined question-and-answer model) can generate the front-end code corresponding to the media presentation component based on media clip links, and return it to the client. The client then implements this front-end code in the question-and-answer dialogue interface to generate the media presentation component and display the media clip. Alternatively, the client can generate the media presentation component itself based on media clip links, or a general media presentation component can be pre-integrated into the client; simply embedding the media clip link will display the media clip pointed to by that link.

[0074] This implementation achieves a deep integration of knowledge response from "text summary" to "original context restoration" by dynamically generating media presentation components and directly displaying media clips within the question-and-answer dialogue interface. Its technical effects are: significantly improving users' efficiency in understanding and trusting the answer content. When a user receives an answer quoted from a training video or meeting transcript, they can not only see the text summary but also immediately view the actual expression of the content in the original media, including non-textual information such as tone of voice, auxiliary charts, and demonstration logic, thus gaining a more comprehensive grasp of the knowledge background and intent. Especially in scenarios that rely on contextual awareness, such as complex process explanations, interpretations of regulations and clauses, and technical operation guidance, this mechanism effectively compensates for the information gaps in pure text answers, enhancing the practicality and immersive experience of the intelligent question-and-answer system. Simultaneously, by integrating the display of media clips within the dialogue interface, it avoids operational interruptions and cognitive fragmentation caused by frequent jumps to external systems, improving the coherence of human-computer interaction and the smoothness of the user experience. More importantly, this component is only triggered to generate when the answer contains genuine knowledge evidence (i.e., the target answer strategy is the first or second answer strategy), ensuring the accuracy and credibility of its use cases and preventing the accidental display of irrelevant media content in the model's speculative answers. Therefore, this implementation constructs a multi-dimensional question-and-answer mode that is "readable, searchable, and viewable," further enhancing the transparency, traceability, and user-perceived value of the enterprise knowledge service system, and providing higher-level interactive capabilities for organizational-level intelligent question-and-answer platforms.

[0075] Based on the first embodiment described above, a hierarchical response method according to the second embodiment of this application is proposed.

[0076] In the second embodiment of this application, the same or similar content as in the above embodiments can be referred to the above description, and will not be repeated hereafter.

[0077] Please refer to Figure 2 , Figure 2 This is a flowchart illustrating the second embodiment of the graded response method in this application.

[0078] In this embodiment, before the step of querying relevant document content matching the original question text from a preset document knowledge base, the hierarchical answering method may further include steps S500~S600: Step S500: Upon receiving the original question text input through the question-and-answer dialogue interface, determine whether the original question text is a document knowledge query type. It should be noted that, in this embodiment, document knowledge query questions refer to questions that explicitly express the intention to search, confirm, or cite specific documents or policy clauses within the organization. Document knowledge query questions typically have the following linguistic characteristics: they contain normative terms such as "according to," "based on," "in accordance with," "refer to," "is there a regulation," "how is it stipulated," and "which article"; or they directly mention specific document names, policy numbers, process steps, etc., such as "What are the regulations regarding annual leave in the employee handbook?" "Which approval process is required for project initiation?" "In which document is the contract approval authority?" etc. The core demand of this type of question is to obtain authoritative, traceable, and structured internal knowledge information, rather than open discussions or general knowledge consultations.

[0079] In this embodiment, the process of determining whether the original question text belongs to the document knowledge query category is as follows: The system uses natural language understanding technology to perform semantic analysis and intent recognition on the original question text to determine whether it belongs to the aforementioned document knowledge query category. This determination process can be implemented through rule matching, keyword recognition, intent classification models (such as text classifiers built based on pre-trained language models like BERT), or a hybrid method. For example, the system can pre-build a keyword library and sentence templates for document query intents, and combine them with contextual semantic analysis to calculate the matching degree between the question text and the document query intent; or it can directly call a lightweight classification model to map the question text to intent tags such as "document knowledge query" or "non-document knowledge query". This determination result serves as a preliminary decision-making basis for whether to initiate the knowledge base retrieval process.

[0080] Step S600: If it is determined whether the original question text is a document knowledge query type, relevant document content matching the original question text is queried from the preset document knowledge base.

[0081] It should be noted that in this embodiment, the execution of step S600 is conditional upon the judgment result of step S500. That is, the retrieval operation of the document knowledge base is triggered only when the system determines that the original question text belongs to the document knowledge query category; if it is determined to be a non-document knowledge query category (such as casual conversation, general knowledge questions, emotional expression, etc.), the knowledge base retrieval process can be skipped, and the answer can be directly generated by the preset question-answering model based on general knowledge, or the user can be guided to clarify the intention of asking the question.

[0082] This embodiment achieves intelligent routing and resource optimization scheduling of the question-and-answer process by introducing a document query intent recognition and judgment mechanism before document knowledge base retrieval. Its technical effects are twofold: First, it effectively avoids performing high-overhead document knowledge base retrieval operations indiscriminately on all questions, significantly reducing system resource consumption and response latency, especially when facing a large number of non-document knowledge query interactions (such as greetings, casual conversations, and general questions), thus improving overall service efficiency and response speed. Second, by accurately identifying the user's document knowledge query intent, it ensures that knowledge base retrieval is activated only in necessary scenarios, thereby improving the targeting and effectiveness of retrieval behavior and preventing noise content introduced by irrelevant searches from interfering with subsequent answer generation. More importantly, this mechanism enhances the system's semantic understanding and interactive intelligence, enabling the system to distinguish between "fact-confirming" and "open-ended discussion" questions and adopt differentiated processing strategies accordingly, providing refined process control capabilities for building a highly efficient and accurate enterprise-level intelligent question-and-answer system. Therefore, while ensuring the integrity of the hierarchical response mechanism in the first embodiment, this embodiment further introduces intent-driven conditional retrieval logic, realizing a leap from "passive response" to "active understanding," and providing more adaptable and scalable technical architecture support for intelligent knowledge services in complex organizational environments.

[0083] Based on the second embodiment described above, a graded response method according to the third embodiment of this application is proposed.

[0084] In the third embodiment of this application, the same or similar content as the above embodiments can be referred to the above description, and will not be repeated hereafter.

[0085] Please refer to Figure 3 , Figure 3 This is a flowchart illustrating the third embodiment of the graded response method in this application.

[0086] In this embodiment, the step S100 above, which involves querying relevant document content from a preset document knowledge base that matches the original question text, may include steps S110 to S120: Step S110: Obtain the data organization pattern information of the preset document knowledge base, and determine the retrieval element information corresponding to the original question text based on the data organization pattern information; It should be noted that, in this embodiment, data organization pattern information refers to a set of metadata used to describe the structured organizational logic within the document knowledge base, reflecting the design features of knowledge entries in terms of storage, indexing, and association. This data organization pattern information typically includes, but is not limited to: document classification systems (such as regulations, processes, and training), document attribute fields (such as document name, number, version number, issuing department, effective date, and keyword tags), document hierarchical relationships (such as chapter-clause, main file-attachment), and document index structure. Data organization pattern information is the fundamental support for the searchability of the document knowledge base, determining how the system efficiently locates and extracts relevant content.

[0087] Generally speaking, the system can obtain the schema information of the document knowledge base by calling external tools to perform database structure query operations, thereby obtaining the data organization mode information of the document knowledge base.

[0088] Retrieval element information refers to a set of key parameters or condition combinations extracted based on the semantic content of the original question text and the requirements of the question-and-answer task, combined with the data organization pattern (i.e., database schema) of the document knowledge base. This retrieval element information can include not only traditional keywords (such as "annual leave," "reimbursement," and "approval"), but also structured query dimensions, such as: target document type (such as "institutional document"), department (such as "human resources department"), time range (such as "effective in 2024"), document level (such as "Chapter 3, Section 2"), or media source type (such as "training video"). By analyzing the data organization pattern of the document knowledge base, the system can identify which fields support exact matching, which support semantic vector retrieval, and which can be used for filtering or sorting, thus intelligently transforming natural language questions into multi-dimensional retrieval elements adapted to the knowledge base architecture.

[0089] Step S120: Construct a knowledge base query instruction based on the retrieval element information, and use the knowledge base query instruction to query relevant document content that matches the original question text from the document knowledge base.

[0090] It should be noted that, in this embodiment, the knowledge base query command refers to an executable query command generated based on the retrieval element information and conforming to the document knowledge base access interface specification. This knowledge base query command can manifest in various technical forms depending on the underlying retrieval mechanism: If you use a keyword + filter method for retrieval, the query command may be an SQL or SQL-like statement; If vectorized semantic retrieval is used, the query instruction includes a high-dimensional numerical array after encoding the question text into vectors, and calls the approximate nearest neighbor search interface of the vector database. If a hybrid retrieval strategy is adopted, the query command can encapsulate multi-stage operation logic, first narrowing the candidate set through structured fields, and then performing semantic similarity calculation on the result subset.

[0091] It is worth mentioning that, in this embodiment, preferably, the document knowledge base is a graph database. In this case, the knowledge base query statement can be a graph database query statement, such as Cypher. At the same time, it can also be combined with Fulltext search for supplementary retrieval, thereby deduplicating the results obtained by the two retrieval methods.

[0092] In this embodiment, the system executes the knowledge base query command to efficiently locate and return the most relevant document fragments and their associated metadata (such as document identification information, source information, media fragment links, etc.) in the document knowledge base, which serve as the factual basis for generating subsequent graded answers.

[0093] This embodiment achieves intelligent reconstruction of the knowledge base query process by introducing a data organization pattern-driven retrieval element extraction mechanism. Its technical advantages lie in changing the traditional coarse-grained processing mode of "question → direct retrieval" in question-answering systems, and instead adopting a step-by-step optimization path of "understanding the knowledge base structure → mapping question semantics → constructing precise queries," significantly improving the accuracy and adaptability of retrieval. Especially when facing complex questions or scenarios with multiple constraints, this method can fully utilize the structured metadata advantages of the document knowledge base, avoiding false positives or false negatives caused by relying solely on full-text fuzzy matching. Furthermore, since the generation of retrieval element information depends on the understanding of the data organization pattern, the system has stronger adaptability—when the knowledge base structure is adjusted or new fields are added, only the data organization pattern information needs to be updated synchronously to automatically adapt to the new retrieval logic, without retraining the model or modifying the core code, thus possessing good maintainability and scalability.

[0094] In one feasible implementation, the step of determining the retrieval element information corresponding to the original question text based on the data organization pattern information in step S110 above can be: standardizing the original question text based on the data organization pattern information to obtain the standardized question text corresponding to the original question text, and extracting the retrieval element information from the standardized question text.

[0095] Specifically, in this embodiment, normalization refers to the process of converting the unstructured, colloquial, or diverse original question text input by the user into a standardized question expression that conforms to the semantic structure and field system of the knowledge base, based on the data organization pattern information of the document knowledge base. This process is essentially a structure-aligned semantic reconstruction, aiming to eliminate the ambiguity and expression differences of natural language, ensuring that the question expression is consistent with metadata such as field naming, classification systems, and terminology standards in the knowledge base, thereby improving the accuracy of subsequent information extraction and retrieval matching.

[0096] For example, if the document knowledge base uses "travel reimbursement standards" as the formal term, and a user asks, "How much accommodation can I be reimbursed for on a business trip?", the system, after obtaining the data organization pattern information, can identify structural features in the knowledge base such as "Expense Type: Travel", "Item Category: Reimbursement Standards", and "Field Naming: Accommodation Expense Limit". It can then standardize the original question text to: "Please check the limits for accommodation expenses in the travel reimbursement standards." Similarly, if the knowledge base organizes documents by "Department + Year + Document Type" (e.g., "Human Resources Department_2024_Employee Handbook"), and a user asks, "What does the onboarding process document issued by HR this year say?", the system can standardize it based on this organizational pattern to: "Please find the relevant employee onboarding process documents issued by the Human Resources Department in 2024."

[0097] Furthermore, "extracting retrieval element information from the standardized question text" refers to, based on the standardization of the question text, using techniques such as Named Entity Recognition (NER), keyword extraction, dependency parsing, or rule template matching to identify and structure-output key elements that can be used to construct query instructions. These elements typically include: Document attribute classes: Document name, version number, year, publishing department; Content themes: keywords, hashtags, business areas; Logical constraints include: time range, applicable objects, and numerical conditions.

[0098] It is easy to understand that the above-mentioned normalization process can be completed by a large model. By simply setting appropriate prompt words and inputting the original question text and data organization pattern information into the large model, the normalized question text output by the large model can be obtained, and the large model can further extract the retrieval element information from the normalized question text.

[0099] It is worth mentioning that, after standardizing the original question text to obtain standardized question text, the above step S200, which constructs graded answer prompts based on the relevant document content, the original question text, and the preset graded answer prompt template, can be described as: constructing graded answer prompts based on the relevant document content, the standardized question text, and the preset graded answer prompt template. Correspondingly, the degree of matching between the relevant document content and the original question text can be considered as the degree of matching between the relevant document content and the standardized question text.

[0100] This implementation method introduces a mechanism of "information standardization based on data organization patterns → element extraction," achieving an efficient transformation from "semantically ambiguous input" to "structurally precise query." Its technical advantages are twofold: First, the standardization process acts as a "semantic bridge" between the user's language and the system's knowledge structure, effectively alleviating retrieval failures caused by inconsistent terminology, arbitrary expressions, or abbreviation habits, significantly improving the system's fault tolerance and user experience. Second, by reconstructing the question into a standardized form aligned with the knowledge base architecture, the system can more accurately identify which fields can be used for filtering and which can be used for semantic expansion, thereby generating more targeted retrieval elements and avoiding the high noise or low recall problems caused by relying solely on keyword matching in traditional methods.

[0101] In addition to the above-mentioned method of directly standardizing the original question text through the large model, the large model can also generate guiding questions or clarifying questions to guide users to gradually clarify their query intent in an interactive manner, and gradually build complete search element information in multiple rounds of dialogue.

[0102] Specifically, in this implementation, after the system receives the original question text, the large model (which can be a preset question-and-answer model) can determine whether the original question text has sufficient structured information to support accurate retrieval based on its understanding of the data organization pattern information of the document knowledge base. If the original question is found to be vague, lacking key attributes (such as not specifying the year, department, or document type), having non-standard terminology, or being ambiguous, the system will not immediately perform a retrieval operation. Instead, the large model will automatically generate one or more clarifying questions to proactively ask the user follow-up questions in order to obtain necessary supplementary information.

[0103] Through this interactive clarification, the system can dynamically collect missing search dimensions (such as time, department, business type, etc.) during the dialogue process, and gradually build a complete set of search element information.

[0104] This implementation upgrades the traditional "single-match" model to an "interactive understanding + progressive retrieval" mechanism, effectively addressing the challenge of users struggling to ask precise questions due to a lack of understanding of the organization's knowledge structure. Especially in enterprise environments, ordinary employees are often unfamiliar with the formal naming, classification logic, or version systems of policy documents, resulting in incomplete or poorly expressed questions. Through proactive guidance from a large-scale model, the system not only lowers the knowledge threshold for users but also enhances the intelligence and inclusiveness of the question-and-answer interaction.

[0105] In one feasible implementation, the document knowledge base adopts a three-level data organization model of department level, document level and fragment level, wherein the department level is higher than the document level, and the document level is higher than the fragment level; The knowledge base query instructions include multi-level query conditions targeting the department level, the document level, and the fragment level; The retrieval element information includes department entity information, document entity information, and fragment entity information. The department entity information is used to construct query conditions for the department level, the document entity information is used to construct query conditions for the document level, and the fragment entity information is used to construct query conditions for the fragment level.

[0106] It should be noted that, in this embodiment, the three-level data organization model refers to a technical architecture that organizes all knowledge content in the document knowledge base in a hierarchical and structured manner according to organizational management logic and information granularity. Specifically: The departmental level is the highest level, used to identify the responsible party for knowledge content. It typically corresponds to functional or business departments within a company, such as "Human Resources," "Finance," "R&D," or "Compliance Office." This level serves as the first dimension for knowledge classification, supporting the allocation of knowledge resources and the management of access permissions based on organizational units.

[0107] The document hierarchy is an intermediate level, located below the department hierarchy. It represents a specific knowledge carrier published or managed by a particular department, i.e., a single document entity, such as the "Employee Handbook," "Travel Expense Management System," or "New Employee Onboarding Training PPT." Each document has a unique metadata identifier (i.e., document identification information, such as name and number) and belongs to a specific department.

[0108] The fragment level is the lowest level, located below the document level. It represents semantically independent content units extracted from a complete document, i.e., document fragments, such as "Chapter 3, Article 5: Annual Leave Calculation Method" or "Page 8: Business Trip Accommodation Standards." This level is the basic unit for knowledge retrieval and citation, possessing the characteristics of being locatable, annotable, and associative to media sources.

[0109] This three-tiered data organization model constitutes a top-down, progressively refined knowledge tree structure, realizing a systematic organization from "organizational affiliation" to "knowledge carrier" and then to "specific content." This not only facilitates the classification, storage, and access control of knowledge but also provides clear path navigation for precise retrieval.

[0110] In this embodiment, the retrieved element information is divided into three dimensions of entity information under this three-level structure: Departmental entity information: refers to the name or abbreviation of the organizational unit that matches the departmental level and is identified from the original question text or the standard question text. It is used to construct filtering conditions for the departmental level and narrow the search scope to the knowledge domain managed by a specific department.

[0111] Document entity information: refers to the identified document name, number, version number, document type, or the content topic covered by the document, which is used to further locate specific documents under the target department and form precise matching conditions at the document level.

[0112] Fragment entity information: refers to key keywords, clause names, process steps or numerical conditions that are semantically related to the user's question, such as "number of annual leave days", "reimbursement limit", "approval process", "probation period regulations", etc., which are used to retrieve the most relevant document fragments and constitute the semantic matching basis at the fragment level.

[0113] Based on the above three types of retrieval element information, the "knowledge base query command" constructed by the system adopts a multi-level query mechanism, which includes composite query conditions for the three levels.

[0114] This implementation method achieves structured and hierarchical access to the document knowledge base by introducing a three-level data organization model and a multi-level retrieval element mapping mechanism. Its technical effects are: significantly improving the accuracy and efficiency of knowledge retrieval in complex organizational environments; avoiding interference from irrelevant cross-department documents through departmental-level filtering; ensuring that queries focus on the correct knowledge carriers through document-level positioning; and accurately extracting specific clauses or regulations required by users through fragment-level semantic matching. This multi-level collaborative retrieval strategy effectively reduces noise from large-scale semantic searches and improves the recall and precision of relevant document content.

[0115] Meanwhile, this three-tiered data organization model enhances the system's maintainability and access control capabilities. Different departments can independently manage their subordinate documents, supporting differentiated updates and version control; fine-grained annotation at the fragment level facilitates subsequent citation, tracing, and multimedia association. Furthermore, this model provides users with a clear knowledge navigation path, allowing even non-professionals to gradually understand the knowledge system through the logic of "department → document → clause."

[0116] More importantly, this implementation deeply couples the natural language problem parsing process with the internal structure of the knowledge base, so that document knowledge base retrieval no longer relies on single keyword or vector matching, but is based on an intelligent reasoning process based on an understanding of organizational structure and knowledge logic, which significantly improves the professionalism, credibility and practicality of intelligent question answering systems in enterprise-level application scenarios.

[0117] It's easy to understand that in practical applications, different organizational units within the same organization exist in a nested hierarchical structure. For example, a certain department belongs to a certain group, a certain team belongs to a certain department, and a certain project team belongs to a certain team. This nesting relationship constitutes a tree-like or network-like organizational structure topology, reflecting the actual structure of the enterprise in terms of management, functions, and division of responsibilities. Therefore, in this implementation, the departmental hierarchy does not simply refer to a set of departments at the same level, but rather a recursive hierarchical system that supports multi-level organizational affiliation, capable of expressing the affiliation and inheritance relationships between multiple levels of organizational units, from group, region, department, team to project team.

[0118] Because user queries in practical applications are often colloquial, incomplete, or use non-standard terminology, the system may not be able to accurately extract departmental or document entity information. To improve the system's robustness and usability, this implementation introduces a default fallback mechanism based on data organization pattern information to address scenarios where key search elements are missing. Specifically: When the department entity information cannot be determined from the user input, the system can use the highest-level organizational unit in the department hierarchy (such as "Group Headquarters," "Company Headquarters," or "Headquarters") as the default search scope based on the data organization model information. This highest level usually represents the organization as a whole, and its published policies, standards, and processes have global applicability, making it a reasonable starting point for searching most common questions. For example, when a user asks "How is annual leave calculated?" without specifying a department, the system can automatically set "Group Headquarters" as the default department entity and prioritize searching company-wide documents such as the "Employee Handbook" or "Attendance Management System," thereby avoiding search failures due to missing information.

[0119] When document entity information cannot be determined from user input, the system can leave the document entity information blank and construct a knowledge base query instruction based solely on the identified department and content entity information. In this case, the retrieval process will perform semantic matching and / or keyword matching based on content entity information across all documents under the specified department (or the default highest department) to find the document fragments most relevant to the user's question. For example, for the question "What materials are needed for expense reimbursement?", if the specific document name is not explicitly mentioned, the system can perform keyword or vector semantic retrieval on all financial documents (such as "Expense Reimbursement Process" and "Travel Management Regulations") within the scope of "Finance Department" or "Group Headquarters" to extract fragments containing content such as "reimbursement materials," "required documents," and "attachment requirements" as relevant document content.

[0120] Through the above mechanism, the system realizes intelligent downgrade retrieval in the case of partial information loss, which not only ensures the continuity of the retrieval process, but also preserves the semantic relevance judgment ability to the maximum extent.

[0121] Furthermore, this implementation follows a "scope-first, condition-optimized" retrieval strategy to address common issues in document knowledge queries such as missing results, matching bias, and expression differences, ensuring that retrieval success rate and result relevance are maximized under different semantic coverage conditions. Specifically: I. Scope-First Logic: When executing knowledge base query commands, the system prioritizes scope-first logic. This means it first constructs a department-limited query based on extracted department entity information, narrowing the search scope to the most likely relevant organizational units to improve retrieval efficiency and result accuracy. This strategy, by reducing the search space and avoiding cross-departmental noise interference, is particularly suitable for scenarios where there are many policy documents with the same name or similar themes but different scopes of application within an organization. However, if the query returns no results (i.e., no relevant documents were found in the target department), or if the results are insufficient to support the basis for generating the first or second answer strategy (e.g., only vague mentions, no specific clauses), the system will automatically trigger a scope expansion mechanism: removing the department filtering conditions and regenerating and executing a knowledge base query command applicable to all departments.

[0122] II. Condition Optimization Mechanism: When the initial query (whether department-limited or full-scope) returns no results or the relevance score is below a preset threshold, the system further activates the condition optimization mechanism to semantically enhance and reconstruct the expression of the original search elements, thereby improving the matching success rate. This mechanism includes the following core strategies: Synonym replacement: Based on a pre-built enterprise terminology ontology or a general / domain-specific word vector model, the original keywords are replaced with semantically similar expressions. For example, "attendance" can be replaced with "attendance," "clocking in," or "sign-in"; "reimbursement" can be replaced with "expense application" or "expense reimbursement"; "probation period" can be replaced with "internship period" or "traineeship stage." The replaced keywords are used to reconstruct the query conditions at the fragment level, expanding the semantic matching boundaries.

[0123] Keyword expansion: While retaining the core semantics, introduce hypernyms, related phrases, or common collocations to form richer query expressions. For example, "attendance time" can be expanded to "working hours", "off work hours", "flexible working hours", or "standard working hours"; "training process" can be expanded to "registration method", "course arrangement", or "graduation assessment".

[0124] This implementation achieves precise positioning through "range priority" and semantic completion through "condition optimization," with the two working together to form a closed-loop, adaptive knowledge retrieval enhancement system. Its technical advantages are: without relying on manual intervention, the system can proactively address real-world challenges such as missing information, terminology differences, and organizational complexity, significantly improving the success rate of retrieving relevant document content and the quality of semantic matching. This provides a more stable and reliable knowledge input foundation for tiered answering methods, further enhancing the practicality, robustness, and user satisfaction of enterprise-level intelligent question-answering systems in real-world business scenarios.

[0125] It should be noted that the above embodiments / implementations are only used to assist in understanding this application and do not constitute a limitation on the hierarchical response method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0126] In addition, please refer to Figure 4 , Figure 4 This is a schematic diagram of the module structure of the hierarchical response device according to an embodiment of this application.

[0127] This application also provides a graded response device, which includes: The retrieval module 10 is used to query relevant document content that matches the original question text from a preset document knowledge base when it receives the original question text input through the question-and-answer dialogue interface. The prompting module 20 is used to construct hierarchical answer prompts based on the relevant document content, the original question text, and the preset hierarchical answer prompt template; The answer module 30 is used to input the hierarchical answer prompts into a preset question-and-answer model to obtain the answer result of the original question text, and to display the answer result in the question-and-answer dialogue interface; The graded response prompts are used to instruct the preset question-and-answer model to provide graded responses to the original question text based on the degree of matching between the relevant document content and the original question text. The grade of the response is positively correlated with the degree of matching, and the confidence level of the response is positively correlated with the grade of the response. Different grades of response results have different response content formats.

[0128] The graded response device provided in this application, employing the graded response method described in the above embodiments, can solve the technical problem of unclear reliability of response results in intelligent question-answering systems. Compared with the prior art, the beneficial effects of the graded response device provided in this application are the same as those of the graded response method provided in the above embodiments, and other technical features in the graded response device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0129] In addition, please refer to Figure 5 , Figure 5 This is a schematic diagram of the hardware operating environment of the electronic device involved in the hierarchical response method in this application embodiment.

[0130] This application also provides an electronic device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the hierarchical response method in the above embodiments.

[0131] The following is for reference. Figure 5 It shows a structural schematic diagram of an electronic device suitable for implementing the embodiments of this application. The electronic device may include, but is not limited to, terminal devices such as mobile phones, laptops, tablets, and desktop computers, or any electronic device capable of performing the above functions. Figure 5 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0132] like Figure 5As shown, the electronic device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the electronic device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. The communication device 1009 allows the electronic device to communicate wirelessly or wiredly with other devices to exchange data. Although the diagrams show electronic devices with various systems, it should be understood that it is not required to implement or have all of the systems shown. More or fewer systems may be implemented alternatively.

[0133] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0134] The electronic device provided in this application, employing the hierarchical answering method described in the above embodiments, can solve the technical problem of unclear reliability of the answer results in intelligent question-answering systems. Compared with the prior art, the beneficial effects of the electronic device provided in this application are the same as those of the hierarchical answering method provided in the above embodiments, and other technical features of the electronic device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0135] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0136] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of protection of the above claims.

[0137] In addition, this application also provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to perform the steps of the graded response method in the above embodiments.

[0138] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory, read-only memory (erasable programmable read-only memory), optical fibers, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0139] The aforementioned computer-readable storage medium may be an electronic device or contained within an electronic device; or it may exist independently and not assembled into or contained within an electronic device.

[0140] The aforementioned computer-readable storage medium carries one or more programs that, when executed by an electronic device, cause the electronic device to: upon receiving an original question text input through a question-and-answer dialogue interface, query relevant document content matching the original question text from a preset document knowledge base; construct graded answer prompts based on the relevant document content, the original question text, and a preset graded answer prompt template; input the graded answer prompts into a preset question-and-answer model to obtain an answer result for the original question text, and display the answer result in the question-and-answer dialogue interface; wherein the graded answer prompts are used to instruct the preset question-and-answer model to provide a graded answer to the original question text based on the degree of matching between the relevant document content and the original question text, and the level of the answer result is positively correlated with the degree of matching, the confidence level of the answer result is positively correlated with the level of the answer result, and different levels of answer results have different answer content formats.

[0141] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0142] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0143] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0144] The computer-readable storage medium provided in this application stores computer-readable program instructions (i.e., a computer program) for performing the steps of the above-described graded response method, which can solve the technical problem of unclear reliability of the response results in intelligent question-answering systems. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as the beneficial effects of the graded response method provided in the above embodiments, and will not be repeated here.

[0145] Furthermore, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the graded response method in the above embodiments.

[0146] The computer program product provided in this application can solve the technical problem of unclear credibility of the answer results in intelligent question-answering systems. Compared with the prior art, the beneficial effects of the computer program product provided in the embodiments of this application are the same as the beneficial effects of the hierarchical answering method provided in the above embodiments, and will not be repeated here.

[0147] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A hierarchical answering method, characterized by, The method comprises: In the case of receiving original question text input through a question and answer dialogue interface, querying relevant document content matching the original question text from a preset document knowledge base; Based on the relevant document content, the original question text and a preset hierarchical answer prompt word template, a hierarchical answer prompt word is constructed; The hierarchical answer prompt word is input into a preset question and answer model to obtain an answer result of the original question text, and the answer result is displayed in the question and answer dialogue interface; Wherein, the hierarchical answer prompt word is used to instruct the preset question and answer model to perform hierarchical answering on the original question text according to the matching degree between the relevant document content and the original question text, and the level of the answer result is positively correlated with the matching degree, the confidence of the answer result is positively correlated with the level of the answer result, and different levels of answer results have different answer content forms.

2. The hierarchical answering method of claim 1, wherein, The hierarchical answer prompt word template carries a plurality of levels of answer strategies, and different levels of answer strategies correspond to instructing the preset question and answer model to generate answer results with different answer content forms; The target answer strategy in the hierarchical answer prompt word is the answer strategy corresponding to the matching degree in the plurality of levels of answer strategies, and the level of the target answer strategy is positively correlated with the matching degree.

3. The hierarchical answering method of claim 2, wherein, Before the step of querying relevant document content matching the original question text from a preset document knowledge base, the method comprises: In the case of receiving original question text input through a question and answer dialogue interface, determining whether the original question text is a document knowledge query type; In the case of determining whether the original question text is a document knowledge query type, querying relevant document content matching the original question text from a preset document knowledge base.

4. The hierarchical answering method of claim 3, wherein, The step of querying relevant document content matching the original question text from a preset document knowledge base comprises: Obtain data organization mode information of a preset document knowledge base, and based on the data organization mode information, determine retrieval element information corresponding to the original question text; Based on the retrieval element information, construct a knowledge base query instruction, and query relevant document content matching the original question text from the document knowledge base through the knowledge base query instruction.

5. The hierarchical answering method of claim 4, wherein, The document knowledge base adopts a three-level data organization mode of department level, document level and segment level, wherein the department level is higher than the document level, and the document level is higher than the segment level; The knowledge base query instruction includes multi-level query conditions for the department level, the document level and the segment level; The retrieval element information includes department entity information, document entity information and segment entity information, wherein the department entity information is used to construct query conditions for the department level, the document entity information is used to construct query conditions for the document level, and the segment entity information is used to construct query conditions for the segment level.

6. The hierarchical answering method according to any one of claims 2 to 5, wherein, The related document content includes a document segment of a target document, where the target document is a document in the document knowledge base that matches the original question text; The multiple levels of answering strategies include a first answering strategy, a second answering strategy, and a third answering strategy, where the first answering strategy has a higher level than the second answering strategy, and the second answering strategy has a higher level than the third answering strategy; In a case where the target answering strategy is the first answering strategy, the answering result includes first answering text generated by the preset question and answer model directly referencing the document segment; In a case where the target answering strategy is the second answering strategy, the answering result includes second answering text inferred by the preset question and answer model based on the document segment; In a case where the target answering strategy is the third answering strategy, the answering result includes third answering text inferred by the preset question and answer model based on model self-knowledge.

7. The hierarchical answering method of claim 6, wherein, The related document content further includes document identification information and document source information of the target document; In a case where the target answering strategy is not the third answering strategy, the answering result further includes the document identification information and the document source information.

8. The hierarchical answering method of claim 6, wherein, The related document content further includes a media segment link corresponding to the document segment, where the media segment link points to a media segment from which the document segment is derived; In a case where the target answering strategy is not the third answering strategy, the answering result further includes the media segment link.

9. The hierarchical answering method of claim 8, wherein, After the step of displaying the answering result in the question and answer dialogue interface, the method further includes: In a case where the answering result includes the media segment link, generating a media presentation component in the question and answer dialogue interface based on the media segment link, and displaying the media segment through the media presentation component.

10. An electronic device, comprising: The electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the computer program, when executed by the processor, implements the steps of the hierarchical answering method according to any one of claims 1 to 9.

11. A computer readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program, when executed by a processor, implements the steps of the hierarchical answering method according to any one of claims 1 to 9.

Images