An adaptive rag-based low-altitude regulation knowledge base question-answering system

Abstract

The application relates to the technical field of intelligent query of low-altitude flight regulations, and discloses a low-altitude regulation knowledge base question-answering system based on adaptive RAG, which comprises the following modules: a specification document preprocessing module, which is used for data cleaning of low-altitude multi-source heterogeneous regulation documents, extraction of multi-dimensional information, and establishment of a multi-dimensional label system; an adaptive RAG knowledge base construction module, which is used for structured storage of a clause unit and a label system structure and construction of a double-mode knowledge base; an adaptive routing and differentiated retrieval module, which is used for complexity discrimination of user queries and adaptive selection of different retrieval paths; a context integration and generation module, which is used for splicing of retrieval results, source information and user queries after sorting of the retrieval results, and generation of an interpretable answer based on reference materials; and a feedback optimization module, which is used for closed-loop optimization through periodical incremental fine-tuning of the model. The application can deeply understand the space-time constraint characteristics of the low-altitude field, dynamically resolve multi-source regulation conflicts, and realize efficient and accurate intelligent question-answering.

Description

Technical Field

[0001] This invention relates to the field of intelligent query technology for low-altitude flight regulations, specifically to a question-and-answer system for a low-altitude regulations knowledge base based on adaptive RAG. Background Technology

[0002] With the rapid development of the low-altitude economy, the development and utilization of airspace below 1,000 meters has entered a phase of explosive growth. The implementation and operation of applications such as drone logistics delivery, electric vertical takeoff and landing (eVTOL) manned aircraft, and low-altitude sightseeing tourism have placed unprecedentedly high demands on the refined management of airspace resources, the safety assurance of flight activities, and regulatory compliance reviews. Against this backdrop, the knowledge system in the low-altitude field exhibits highly complex characteristics. Its core content encompasses a large number of heterogeneous documents, including Civil Aviation Administration regulations, local regulations, CCAR technical standards, and airspace delineation maps. The sheer volume of text, frequent updates, dense technical terminology, and complex cross-document relationships present significant challenges for practitioners in obtaining accurate and complete regulatory information.

[0003] Faced with the increasingly complex regulatory inquiry needs, traditional methods of knowledge acquisition mainly rely on manual searching and professional interpretation, where regulatory experts or flight management personnel consult paper documents or electronic databases to find answers. This model is not only inefficient and unable to meet the need for immediate compliance assessments, but it is also highly susceptible to overlooking implicit constraints hidden in multiple documents due to human oversight, leading to misunderstandings of regulations and potentially causing serious flight safety hazards.

[0004] To improve the efficiency and accuracy of knowledge acquisition, retrieval-enhanced generation technology has opened up new avenues for intelligent question-answering systems. This technology, by introducing external knowledge bases, provides large language models with real-time, traceable factual evidence, effectively suppressing illusions during model generation and improving the professionalism and reliability of answers to a certain extent. Therefore, it has been widely applied in various fields such as finance, healthcare, and general document question answering. However, directly applying the retrieval-enhanced generation technology, which has achieved success in general fields, to the highly specialized vertical field of low-altitude regulations results in significant "incompatibility," failing to meet the extreme requirements for operational security in this field. Specifically: (1) Lack of semantic awareness regarding three-dimensional space and timeliness: When performing information retrieval, the core mechanism of the general-purpose RAG system relies on the semantic similarity calculation between texts. This two-dimensional text matching method is completely unable to understand the three-dimensional spatial attributes and dynamic timeliness requirements unique to the low-altitude domain. For example, when the regulation text contains the restriction of "120 meters altitude", the general model cannot automatically distinguish whether this value represents the "true altitude" relative to the ground or the "altitude" based on a certain reference plane. This lack of semantic awareness is very likely to cause aircraft to mistakenly enter the no-fly zone. At the same time, the general-purpose vector retrieval model cannot effectively handle the timeliness priority relationship between temporary notices to air traffic and long-term effective static regulations. It may provide users with expired airspace opening information as valid evidence, thereby directly causing safety accidents.

[0005] (2) Knowledge fragmentation leads to errors in the applicability of regulations: General RAG systems commonly use fixed character lengths or text segmentation by natural paragraphs during the knowledge base construction phase. This approach leads to serious knowledge fragmentation problems when dealing with logically rigorous legal provisions. A complete legal restriction clause often has its applicable premises, specific constraints, and exceptions distributed across multiple paragraphs, or requires reference to the "preliminary clauses" in other chapters for accurate understanding. Mechanical segmentation forcibly separates the scope of application information, such as "applies only to light drones," from its corresponding specific restrictions, such as "prohibited from flying in specific areas," resulting in the retrieved information being taken out of context. This leads to serious errors in the applicability of regulations in the answers generated based on this fragmented information.

[0006] (3) Difficulty in handling conflicts of multi-source heterogeneous data: The management of low-altitude airspace involves multiple levels of effectiveness, such as national laws, departmental regulations and local regulations. When there are subtle differences or even conflicts between higher-level laws and lower-level laws, or management methods in different places on the same issue, the general RAG system often recalls and presents contradictory clauses to users at the same time due to the lack of awareness and resolution mechanism of the legal effectiveness levels, leaving users at a loss and unable to obtain clear and compliant operation guidance.

[0007] Therefore, developing a dedicated RAG question-and-answer system that can deeply understand the spatiotemporal constraints of the low-altitude domain and dynamically resolve multi-source regulatory conflicts has become an urgent need to ensure the safe and efficient operation of low-altitude economic activities. Summary of the Invention

[0008] The present invention aims to provide a question-and-answer system based on an adaptive RAG (Regulatory Information Group) knowledge base for low-altitude regulations. This system can deeply understand the spatiotemporal constraints of the low-altitude domain and dynamically resolve multi-source regulatory conflicts, thereby achieving efficient and accurate intelligent question-and-answer.

[0009] The basic solution provided by this invention is: a question-and-answer system for a low-altitude regulatory knowledge base based on adaptive RAG, comprising: The standard document preprocessing module is used to clean the data of multi-source heterogeneous regulatory documents in the low-altitude field, and based on the principle of legal atomicity, it segments the documents into multiple semantically complete clause units with clauses as the smallest semantic units. At the same time, it extracts the three-dimensional spatial constraint information, dynamic timeliness information and regulatory validity level information from each clause unit through named entity recognition technology, and establishes a multi-dimensional tag system for each clause unit, including spatial tags, time tags and validity tags. The adaptive RAG knowledge base construction module is used to structure and store the preprocessed clause units and their multi-dimensional tag system to build a dual-mode knowledge base that includes vector indexes and knowledge graphs. The adaptive routing and differential retrieval module is used to determine the complexity of the natural language query input by the user and adaptively select a preset retrieval path based on the determination result. The retrieval path includes: a direct generation path for common sense questions, a single-round vector retrieval path for simple professional questions, and a multi-round sub-query and graph reasoning retrieval path for complex questions. The context integration and generation module is used to sort the search results by relevance and combine them with the source information and user query to form a structured context, which is then input into the big language generation model to generate an interpretable answer based on the cited materials. The feedback optimization module is used to collect explicit or implicit user feedback on the generated results, convert it into training samples, and periodically perform incremental fine-tuning on the models on which each module of the system depends, thereby achieving closed-loop optimization.

[0010] The working principle and advantages of this invention are as follows: This invention presents a question-and-answer system for a low-altitude regulatory knowledge base based on adaptive RAG (Regulatory Information Group). This system possesses a deep understanding of the spatiotemporal constraints in the low-altitude domain and can dynamically resolve multi-source regulatory conflicts, achieving efficient and accurate intelligent question answering. The key points are: This solution fundamentally overcomes the inherent shortcomings of general RAG technology in handling spatiotemporally sensitive, structurally complex, and conflicting regulatory knowledge. It proposes an adaptive routing RAG architecture suitable for the low-altitude domain, combining the "legal atomicity" principle in the regulatory domain with the explicit modeling of "spatiotemporal attributes" in the low-altitude domain, and achieving closed-loop optimization of the entire process from knowledge preprocessing, storage and retrieval to generation and feedback at the system level.

[0011] The preprocessing module for standard documents employs clause-level semantic segmentation and assigns multi-dimensional labels such as space, time, and level of effectiveness to each clause unit. This ensures the logical integrity of the legal provisions, enabling the system to accurately understand the essential difference between "true height" and "altitude," and to perceive the timeliness priority of temporary notices and static regulations. This avoids compliance misjudgments caused by knowledge fragmentation and lays a solid data foundation for subsequent high-precision retrieval.

[0012] Building upon this foundation, the adaptive RAG knowledge base construction module innovatively establishes a dual-mode knowledge base integrating vector indexes and knowledge graphs. This enhances the ability to distinguish between terms that are similar in form but different in meaning, while also enabling the system to handle cross-document associations and conflicts in regulatory validity. Furthermore, the adaptive routing and differentiated retrieval module adaptively selects direct generation, single-round retrieval, or multi-round sub-queries and graph reasoning paths based on the complexity of user questions, balancing efficiency and depth. The subsequent generation module mandates answers based on source information to ensure interpretability; the feedback optimization module collects user signals and periodically incrementally fine-tunes the model, forming a continuous evolutionary closed loop. This significantly improves the accuracy, adaptability, and reliability of low-altitude regulatory question answering, meeting the requirements of policy consultation and regulatory auditing, and adapting to the evolving knowledge needs of the low-altitude economy. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the system structure of a question-and-answer system for a low-altitude regulatory knowledge base based on adaptive RAG according to the present invention. Figure 2 This is a schematic diagram illustrating the retrieval path division of an embodiment of the low-altitude regulatory knowledge base question-and-answer system based on adaptive RAG of the present invention. Detailed Implementation

[0014] The following detailed explanation illustrates the specific implementation methods: Example 1 The basic implementation examples are as follows: Figure 1 As shown: A question-and-answer system for a low-altitude regulatory knowledge base based on adaptive RAG, including a regulatory document preprocessing module, an adaptive RAG knowledge base construction module, an adaptive routing and differential retrieval module, a context integration and generation module, and a feedback optimization module.

[0015] The preprocessing module for the standard documents aims to transform multi-source, heterogeneous regulatory documents in the low-altitude domain into structured, semantically complete clause units rich in spatiotemporal tags, laying a data foundation for the subsequent construction of the knowledge base.

[0016] The specification document preprocessing module is used to perform the following operations: S1 performs data cleaning on multi-source heterogeneous regulatory documents in the low-altitude airspace.

[0017] In this embodiment, the multi-source heterogeneous regulatory documents in the low-altitude domain include, but are not limited to: CCAR series regulations (PDF format) issued by the Civil Aviation Administration of China, local drone management regulations (Word / PDF), airspace delineation maps (GeoJSON format), NOTAMs (semi-structured text), industry standards, and enterprise operation manuals.

[0018] Specifically, the data cleaning involves employing customized cleaning strategies for data of different formats. For PDF documents, the MinerU document parsing engine (which is based on layout analysis algorithms and OCR text recognition technology) is used to automatically remove headers, footers, page numbers, charts and other distractions, and extract the plain text content.

[0019] For GeoJSON spatial domain maps, parse their geometric coordinate set (such as the latitude and longitude of polygon vertices) and attribute fields (such as spatial domain type, height range, and effective time), and convert them into structured text descriptions.

[0020] For NOTAM announcements, key fields are extracted using regular expressions: announcement number, effective start time (e.g., "FROM 2410151200 TILL 2410152359"), affected airspace (e.g., "WITHIN 5NM OF 395300N / 1162400W"), and restriction content (e.g., "NO FLY ZONE ACTIVE"), and these are standardized into a unified timestamp and spatial coordinate representation.

[0021] S2, based on the principle of legal atomicity, divides clauses into multiple semantically complete clause units by using clauses as the smallest semantic unit.

[0022] Specifically, to ensure the logical integrity of the regulatory provisions, the common RAG segmentation method based on fixed token length is abandoned, and instead a clause-level semantic segmentation strategy is adopted. The specific steps include: Using Article as the smallest semantic unit, regular expressions are used to match the article numbering patterns in Chinese regulations, such as "Article [0123456789100]+" and "Article [0-9]+", to locate the starting position of each article; Combining semantic boundary detection algorithms with natural language processing models, the start and end positions of clauses are identified. In this embodiment, an NLP semantic boundary detection model (fine-tuned on legal and regulatory corpora) is introduced to identify the semantic end points of clause texts, avoiding truncating complete clauses. For long clauses containing multiple clauses (such as "Clause 1", "Clause 2") or multiple items (such as "(I)", "(II)"), sub-clauses are segmented only under the premise of semantic independence, and their association index with the parent clause is retained. For example, each item is treated as an independent sub-unit, but the index ID of its parent clause is forcibly retained to form a tree structure, ensuring that "scope of application" and "specific provisions" are not separated, and solving the problem of compliance misjudgment caused by knowledge fragmentation.

[0023] After the above processing, the length of each segmented text unit is controlled between 300 and 500 characters (this range is based on statistics of typical regulations such as CCAR-91 and CCAR-92: the average clause length is 387 characters and the standard deviation is 112 characters. Therefore, selecting this range can ensure semantic integrity while taking into account retrieval efficiency).

[0024] Meanwhile, named entity recognition (NER) technology is used to extract three-dimensional spatial constraint information, dynamic timeliness information and legal effect level information from each clause unit, and a multi-dimensional tag system containing spatial tags, time tags and effect tags is established for each clause unit.

[0025] The spatial labels include one or more of the following: true altitude, altitude, corrected sea level pressure altitude, airspace type, airspace sector identifier, no-fly zone identifier, and geographic coordinate range; and are mapped to a unified spatial coordinate or altitude reference.

[0026] The time stamps include one or more of the following: effective date, repeal date, specific time period of day, and effective time window of temporary notices of navigation.

[0027] The validity labels include: information on the national, ministerial, or local level of the regulations or the civil aviation standard scenario level to which they pertain.

[0028] All tags correspond one-to-one with the clause units, and the final output is a structured index file in JSON format for use by subsequent modules.

[0029] The adaptive RAG knowledge base construction module is used to structure and store the preprocessed clause units and their multi-dimensional tag system, and construct a dual-mode knowledge base containing vector indexes and knowledge graphs (i.e., construct a "regulation-spatiotemporal" dual-layer knowledge graph based on Neo4j).

[0030] The vector index is constructed based on a low-altitude domain-specific embedding model, which is obtained in the following way: Based on a general vector model ("bge-large-zh-v1.5"), LoRA fine-tuning was performed using positive and negative sample pairs specific to the low-altitude domain. The rank was set to 8, the scaling factor to 32, and low-rank adaptation was applied only to the attention layer weights of the base model, while other parameters were frozen. A contrastive loss function (InfoNCE) was used during training. Evaluation on a manually annotated low-altitude terminology test set showed that the fine-tuned model achieved approximately 18% higher accuracy in low-altitude terminology semantic similarity assessment compared to the base model, significantly enhancing the recall accuracy of vector retrieval.

[0031] The fine-tuned model was applied to all clause units, generating 768-dimensional vector representations, which were then stored in the Milvus distributed vector database. To support efficient retrieval, HNSW indexing combined with PQ compression was employed to achieve high-performance vector recall.

[0032] Among the positive and negative sample pairs, the positive sample pairs include legal clauses and their rewritten questions or term explanations; specifically including: ① rewriting legal clauses as questions (e.g., rewriting "civil unmanned aerial vehicles need to be registered with real names" as "do drones need to be registered with real names?"); ② pair of term abbreviations and full explanations (e.g., "true altitude - actual height, i.e., vertical distance relative to the ground").

[0033] Negative sample pairs include text pairs that are semantically conflicting or similar in form but different in meaning in a low-altitude context; for example, “controlled airspace” and “regulated airspace”, “restricted flight zone” and “no-fly zone”, as well as text pairs containing ambiguous altitude descriptions.

[0034] In this embodiment, the data sources for positive and negative sample pairs cover the "Interim Regulations on the Flight Management of Civil Unmanned Aerial Vehicles", CCAR-91 / 92, local low-altitude airspace management regulations, and historical NOTAMs (notices to air traffic) for the past three years.

[0035] The knowledge graph uses legal provisions, airspace entities, aircraft types, and operational scenarios as nodes, and is associated with various edge types, including hierarchical relationships, referencing relationships, and spatiotemporal constraints.

[0036] The detailed definitions of each node are as follows: Regulatory nodes: Attributes include regulatory name, issuing agency, document number, and level of legal force; Clause node: Attributes include clause number, text content, effective date, and repeal date; Airspace entity node: Attributes include airspace name, type (such as no-fly zone, restricted zone), geometric extent (GeoJSON), altitude reference, etc. Aircraft type node: Attributes include model, category (micro / light / small / medium / large), maximum takeoff weight, etc.; Running scene node: Attributes include scene name (such as "logistics delivery", "aerial photography", "agricultural plant protection"), typical height range, etc.

[0037] The detailed definition of the hierarchical relationship is as follows: from regulations to chapters and clauses, a tree structure is established, such as "Regulations → Chapter 2 → Article X → Section X".

[0038] The detailed definition of the reference relationship is as follows: Through regular expression matching and dependency parsing, the expression in the clause that explicitly points to other clauses or regulations (such as "in accordance with Article 23 of the Law on Penalties for Administration of Public Security") is identified, and reference edges are automatically created.

[0039] The detailed definitions of spatiotemporal constraints are as follows: For example, “RESTRICTS_HEIGHT” connects the airspace node and the clause node, indicating the height restriction of the airspace; “VALID_IN” connects the clause node and the airspace node, indicating that the clause applies to a specific airspace; “APPLIES_TO” connects the clause node and the aircraft type or scenario node.

[0040] The construction of the knowledge graph includes a citation chain maintenance mechanism: when a legal provision is detected to reference other provisions or regulations, a citation relationship edge is automatically established between the citing provision node and the target provision node. This ensures that when retrieving a citation provision, the underlying provision being referenced is recalled by traversing the graph, enabling multi-hop reasoning within the legal context. For example, if a local regulation references a national regulation, and the national regulation in turn references another law, the system will sequentially associate the local provision with the referenced national provision and the legal provision, forming a citation chain.

[0041] Furthermore, newly added documents (such as new regulatory documents or NOTAM releases) are processed by the normative document preprocessing module. Their clause units and multi-dimensional tagging system are automatically embedded into the vector index, and relationships (such as citation relationships and spatial domain associations) are established with existing nodes in the knowledge graph. This eliminates the need to rebuild the entire knowledge base, adapting to the dynamic evolution of low-altitude regulations. The update frequency can be set according to the data release cycle, for example, checking for NOTAM updates daily and synchronizing the regulatory database weekly.

[0042] Through the above design, on the one hand, this module uses LoRA fine-tuning of the embedding model with a low-altitude domain-specific corpus, significantly enhancing its ability to distinguish between similar-sounding but semantically different terms such as "controlled airspace" and "managed airspace," thus improving the accuracy of vector recall. On the other hand, this module constructs a two-layer "regulatory-spatial" knowledge graph based on a graph database, abstracting core elements such as regulatory clauses, airspace entities, aircraft types, and operational scenarios into nodes, and associating them through hierarchical relationships, reference relationships, and spatiotemporal constraints. In particular, it establishes explicit reference chains for "referral clauses" commonly found in regulations. This structured storage method enables the system not only to perform semantic similarity-based retrieval but also to perform multi-hop reasoning along the reference edges in the graph, automatically recalling referenced preceding clauses or associated airspace constraints. This gives it the ability to handle complex cross-document associations and regulatory validity conflicts, a key breakthrough that general RAGs cannot achieve.

[0043] The adaptive routing and differentiated retrieval module is used to determine the complexity of the natural language query input by the user, and adaptively select a preset retrieval path based on the determination result to achieve accurate and efficient recall.

[0044] The complexity determination is implemented through a lightweight large language model fine-tuned by a single instruction. This lightweight large language model is fine-tuned based on a low-level domain question-and-answer dataset labeled with query complexity tags (e.g., a dataset containing 3000 typical low-level user queries, independently labeled with complexity tags by three domain experts). It takes the user's natural language query as input and outputs a routing parameter (Route), which indicates the retrieval path to which the current query should belong. .

[0045] like Figure 2 As shown, the retrieval path includes: Route A – a direct generation path for common-sense questions; the answers to these questions do not depend on specific regulations, such as "What is true height?".

[0046] The processing strategy for this path is to skip the retrieval, directly optimize the user query using a prompt (e.g., add "please explain in plain language"), and submit it to the context integration and generation module.

[0047] Route B – A single-round vector search path for simple, specialized questions; such questions can be answered directly in the knowledge base through a single vector search, such as “What is the flight restriction altitude in area C of region A?”.

[0048] The processing strategy for this path is as follows: perform a single-round vector retrieval. The query is converted into a vector using a fine-tuned embedding model, and an approximate nearest neighbor search is performed in Milvus to recall the Top-20 text blocks; then, the MMR algorithm is used for rearrangement, balancing relevance and diversity, and the Top-3 are selected as the final results.

[0049] Route C – a multi-round sub-query and graph reasoning retrieval path for complex problems; such problems need to be decomposed into multiple sub-problems and integrated with graph reasoning, such as “compliance of xx flying car operation in location A and location B”.

[0050] The processing strategy for this path includes the following sub-steps: A subquery generation model (such as the LLM_subquery problem splitting model) can be used to break down the original complex query into multiple subqueries. For example, the query "I have an X-type flying car. Is it legal to fly it 120 meters away from point A to take pictures of the night view?" can be broken down into: ① "What type of aircraft is the X-type flying car?", ② "What are the airspace attributes and altitude restrictions for point A?", and ③ "What are the regulations for nighttime flights in this area?".

[0051] For each subquery, entity links are created to locate the corresponding node in the knowledge graph; that is, key entities in the subquery (such as "X type" or "A location") are mapped to graph nodes.

[0052] Parallel vector retrieval is performed, and graph expansion retrieval is carried out on the retrieved core clause nodes. Multi-hop traversal is performed along the reference relationship edges or spatiotemporal constraint relationship edges to recall related clauses, as well as related spatial domain or timeliness information.

[0053] All recall results are merged and input into the reranking model, which outputs the top-relevant clauses. For example, all vector retrieval results and graph expansion results are merged and deduplicated to form a candidate pool (usually 30-50 items), which is then input into the bge-reranker-large cross-encoder for reranking. The relevance score of each candidate block is output, and the Top-5 are selected as the final retrieval results.

[0054] The context integration and generation module is used to sort the search results by relevance and combine them with the source information and user query to form a structured context, which is then input into the large language generation model (using LLM_generate and fine-tuned by instructions on the low-altitude question-and-answer corpus) to generate an interpretable answer based on the cited materials (such as attaching a list of cited sources in a structured format for easy auditing).

[0055] The construction of the structured context includes the following sub-steps: Each retrieved clause is labeled with the name of the source regulation, clause number, and page number.

[0056] The clause units are arranged in descending order of relevance score and concatenated to form a context; the user query is appended at the end to form a complete generated input.

[0057] When inputting into the large language generation model, the system uses pre-set prompt templates to force the model to strictly adhere to the cited materials in its answers, and prompts consulting agencies when information is uncertain. For example, the prompt templates may stipulate, "Please answer the questions strictly according to the above-mentioned cited materials, and do not add any information that is not cited," "If the cited materials do not explicitly mention xx information, it must be stated that 'please consult the local civil aviation administration department to confirm the specific xx information,'" and "When there are conflicts of regulations, they shall be handled according to the principle of priority of legal effect: local regulations take precedence over national regulations (if local regulations are stricter), and the basis shall be explained."

[0058] Preferably, this module also performs rule validation on the generated answers, such as checking whether they contain explicit annotations for "true altitude / altitude"; if not, it automatically adds a hint. If the answer involves numerical altitude, it compares it with the spatial altitude limits in the knowledge base to ensure consistency. If a potential contradiction is found, the system returns a warning and requires the user to verify.

[0059] The feedback optimization module is used to collect explicit or implicit user feedback on the generated results, convert it into training samples, and periodically perform incremental fine-tuning on the models on which each module of the system depends, thereby achieving closed-loop optimization.

[0060] The closed-loop optimization specifically includes the following operations: The system records the original question, generated answer, search path, and model confidence level when providing user feedback, generating structured feedback data. In this embodiment, a "Report Error" button is added at the end of each answer, allowing users to submit correction text. Simultaneously, the system implicitly monitors user behavior, such as copying answers, asking the question again, and excessively short page dwell time, as signals of dissatisfaction. Each interaction record is saved as structured JSON, including: the original query, generated answer, search path, confidence level of each search block, user feedback text (if any), and behavioral characteristics.

[0061] Through similarity analysis and confidence assessment, the feedback data is classified into route correction samples, retrieval negative samples, or adversarial sample generation. In this embodiment, when a user submits a correction, the system treats the original query and the correction text as a pair to generate route correction samples (used to optimize the routing model), retrieval negative samples (marking erroneous blocks in the original retrieval results as negative examples), and adversarial sample generation (the original answer and the correction text are used as positive and negative generation pairs).

[0062] For implicit feedback, a confidence assessment model (in this embodiment, a binary classifier based on BERT is selected and trained on historical feedback data) is used to determine whether it is a valid negative sample. Samples with a confidence score higher than 0.8 are added to the training pool, those with a confidence score lower than 0.3 are discarded, and the intermediate part is pushed to the manual annotation platform.

[0063] Each week, newly added samples are manually reviewed and then included in the training set. The lightweight large language model, reordering model, and large language generation model are incrementally fine-tuned using the newly added samples regularly to continuously optimize system performance.

[0064] To illustrate the system workflow intuitively, we will use a specific low-altitude flight compliance query as an example.

[0065] The user inputs a query: "I have a DJI Mavic 3 and plan to fly it to a height of 120 meters in Talent Park, Nanshan District, Shenzhen to take night photos. Is this legal?" The system processing is as follows: The adaptive routing and differential retrieval module identifies that the query contains "specific model (Mavic 3)", "specific geographical location (Nanshan Talent Park)", "altitude value (120 meters)" and "time attribute (night view / night)", and determines it to be a high-complexity compound query, outputs Route=C, and enters the complex path; The subquery generation model is broken down into three subqueries: ① "What type of drone is the Mavic 3?"; ② "What are the airspace attributes and altitude restrictions of Talent Park in Nanshan District, Shenzhen?"; ③ "What are the regulations for nighttime flights in this area?"; The entity link maps "Mavic 3" to the "Light Drone" node in the graph, and "Talent Park" to the "Shenzhen Bay Restricted Area" node; Vector retrieval retrieved the clauses regarding height restrictions in the "Interim Measures for the Management of Civil Micro-Light Unmanned Aerial Vehicles in Shenzhen". Graph expansion along the "RESTRICTS_HEIGHT" edge revealed that the area has a height restriction of 100 meters (instead of the general 120 meters). At the same time, the "CITES" reference edge linked to the penalty clauses of the "Public Security Administration Punishment Law", and the NOTAM on that day confirmed that there was no temporary no-fly zone. The rearranged Top-5 contains the key information mentioned above; The generated model outputs a detailed compliance analysis based on the cited materials, clearly stating: "1. Aircraft type compliance: The Mavic 3 is a light drone, which meets the access requirements for this area; 2. Altitude risk: Although the national standard limits the altitude to 120 meters, some areas in the Shenzhen Bay area have a height limit of 100 meters, so it is recommended to reduce the altitude; 3. Time restrictions: Warning lights must be turned on when flying in this area at night; 4. Legal consequences: Violations may result in penalties under the Public Security Administration Punishment Law," and the source of each regulation is indicated.

[0066] This case demonstrates that the system can accurately understand spatiotemporal constraints, resolve regulatory conflicts, and provide highly interpretable compliance guidance.

[0067] This embodiment provides a low-altitude regulatory knowledge base question-and-answer system based on adaptive RAG, which effectively solves the problems of lack of spatial timeliness perception, knowledge fragmentation, and difficulty in resolving regulatory conflicts in the prior art. It can deeply understand the spatiotemporal constraint characteristics of the low-altitude domain and dynamically resolve multi-source regulatory conflicts, achieving efficient and accurate intelligent question-and-answer.

[0068] Example 2 A question-and-answer system for a low-altitude regulatory knowledge base based on adaptive RAG (Rapid Air Regulator) is proposed. Building upon the first embodiment, this system further introduces a real-time airspace situation data access module and a multimodal graphic and textual joint reasoning module. This enables the system to not only query static regulations but also dynamically deduce spatiotemporal constraints by combining information such as real-time weather conditions, dynamic airspace occupancy status, radar detection data, and charts in airspace maps and the original PDF of NOTAMs (Notices to Airmen). This allows the system to answer real-time and dynamic questions such as "Can I fly now?" and "Is there a risk of temporary no-fly zone in the next two hours?"

[0069] The real-time airspace situation data access module is used for real-time situation awareness and data access. Specifically, this module is used to perform the following operations: Access real-time weather APIs (such as wind speed, visibility, and precipitation), air traffic control radar data (real-time aircraft position and density), dynamic airspace occupancy status (such as temporary mission occupancy), and real-time NOTAM push streams; and convert unstructured real-time data (such as weather radar images and NOTAM text streams) into structured spatiotemporal events. For example, identify "strong convective areas" in weather radar images as "temporary no-fly risk areas" and assign them a time window (such as being valid for the next 30 minutes).

[0070] Real-time events are dynamically inserted into the knowledge graph as temporary nodes and a "real-time constraint" relationship is established with static regulatory nodes. For example, "strong convective areas" are dynamically associated with "no-fly clauses" and an expiration timestamp is set, which will be automatically removed after the timeout.

[0071] The multimodal image-text joint reasoning module incorporates a large multimodal model (such as Qwen-VL or GPT-4V) for image-text joint understanding. For airspace demarcation maps in regulatory documents (such as JPG aeronautical charts in PDFs) and schematic diagrams in NOTAMs, this model performs image-text joint understanding. For example, it identifies "red areas as no-fly zones and blue areas as altitude-restricted zones" from an airspace map and extracts their latitude, longitude, and altitude ranges.

[0072] Furthermore, the spatial information obtained from the analysis is aligned and verified with the descriptions in the clause text. For example, if the clause states "the no-fly zone is..." Figure 1 If the value is "shown", the system will automatically associate it. Figure 1 The analysis results ensure the accuracy of spatial constraints.

[0073] When a user queries a location-related question (such as "Can I fly over Shenzhen Bay Park?"), this system not only retrieves the text terms but also relevant airspace maps and images. It then performs spatial calculations using the geometric information in the images and the user's location to determine whether the user has fallen into a no-fly zone.

[0074] The adaptive routing and differentiated retrieval module also performs real-time constraint solving during the retrieval process: The user query (e.g., "I am currently at Nanshan Talent Park, with a wind speed of 5 m / s. Can I fly a Mavic from 3 to 120 meters?") is transformed into a problem with spatiotemporal constraints. The system performs logical deduction based on the following conditions: Static regulations: aircraft type classification, altitude restrictions, airspace types; Real-time situation: Whether the current wind speed exceeds the wind resistance capability of this aircraft model; whether the real-time radar density in this area triggers the "congested airspace" restriction; Dynamic events: Are there any temporary no-fly notices in effect? ​​Are there any weather warnings? Image and text analysis results: Whether the user's GPS coordinates fall within the red area of ​​the airspace map.

[0075] When static regulations conflict with real-time situation (e.g., regulations allow flight, but real-time radar density is too high), the system outputs a risk warning based on the preset "safety priority" rule (e.g., "it is recommended to suspend flight when there is a real-time situation conflict") and provides the basis (e.g., "radar monitoring shows that the density of aircraft in the current area exceeds the safety threshold, it is recommended to wait for 10 minutes").

[0076] The context integration and generation module combines the generative model with the inference results to output a dynamic response supported by real-time data, such as: "Flying is currently permitted, but please note: ① Wind speed of 5 m / s is within the safe range of Mavic 3; ② The real-time radar density in this area is low, with no risk of congestion; ③ There are no temporary no-fly notices in the next 30 minutes. It is recommended to maintain visual line of sight during flight."

[0077] This embodiment provides a low-altitude regulatory knowledge base question-and-answer system based on adaptive RAG, which, compared to Embodiment 1, can provide additional real-time decision support for flight compliance, resulting in better low-altitude flight safety assurance.

[0078] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A question-and-answer system for a low-altitude regulatory knowledge base based on adaptive RAG, characterized in that, include: The standard document preprocessing module is used to clean the data of multi-source heterogeneous regulatory documents in the low-altitude field, and based on the principle of legal atomicity, it segments the documents into multiple semantically complete clause units with clauses as the smallest semantic units. At the same time, it extracts the three-dimensional spatial constraint information, dynamic timeliness information and regulatory validity level information from each clause unit through named entity recognition technology, and establishes a multi-dimensional tag system for each clause unit, including spatial tags, time tags and validity tags. The adaptive RAG knowledge base construction module is used to structure and store the preprocessed clause units and their multi-dimensional tag system to build a dual-mode knowledge base that includes vector indexes and knowledge graphs. The adaptive routing and differential retrieval module is used to determine the complexity of the natural language query input by the user and adaptively select the preset retrieval path based on the determination result; The retrieval paths include: a direct generation path for common sense questions, a single-round vector retrieval path for simple professional questions, and a multi-round sub-query and graph reasoning retrieval path for complex questions; The context integration and generation module is used to sort the search results by relevance and combine them with the source information and user query to form a structured context, which is then input into the big language generation model to generate an interpretable answer based on the cited materials. The feedback optimization module is used to collect explicit or implicit user feedback on the generated results, convert it into training samples, and periodically perform incremental fine-tuning on the models on which each module of the system depends, thereby achieving closed-loop optimization.

2. The low-altitude regulatory knowledge base question-answering system based on adaptive RAG as described in claim 1, characterized in that, In the aforementioned preprocessing module for specification documents, the segmentation based on clauses as the smallest semantic unit specifically includes: Regular expressions are used to match clause numbers, and semantic boundary detection algorithms from natural language processing models are combined to identify the start and end positions of clauses. For long clauses containing multiple clauses or items, sub-clauses are segmented under the premise of semantic independence, and their association index with the parent clause is preserved.

3. The question-and-answer system for a low-altitude regulatory knowledge base based on adaptive RAG as described in claim 1, characterized in that, In the adaptive RAG knowledge base construction module, the vector index is constructed based on a low-altitude domain-specific embedding model, which is obtained through the following method: Based on a general vector model, LoRA fine-tuning is performed using positive and negative sample pairs specific to the low-altitude domain. Among the positive and negative sample pairs, positive sample pairs include legal clauses and their rewritten questions or terminological explanations, while negative sample pairs include text pairs that are semantically conflicting or similar in form but different in meaning in the low-altitude context.

4. The low-altitude regulatory knowledge base question-answering system based on adaptive RAG as described in claim 1, characterized in that, The knowledge graph uses legal provisions, airspace entities, aircraft types, and operational scenarios as nodes, and is associated with various edge types, including hierarchical relationships, referencing relationships, and spatiotemporal constraints. The construction of the knowledge graph includes a reference chain maintenance mechanism: when a reference to other clauses or regulations is detected in a legal clause, a reference relationship edge is automatically established between the reference clause node and the target clause node, ensuring that the referenced basic clause is recalled by traversing the graph when retrieving the referenced clause, thus realizing multi-hop reasoning of the legal context.

5. A question-and-answer system for a low-altitude regulatory knowledge base based on adaptive RAG as described in claim 1, characterized in that, In the adaptive routing and differential retrieval module, the complexity determination is achieved through a lightweight large language model finely tuned by a single instruction; The lightweight large language model is fine-tuned based on a low-level domain question-answering dataset labeled with query complexity, and outputs routing parameters, which are used to indicate the retrieval path to which the current query should belong.

6. A question-and-answer system for a low-altitude regulatory knowledge base based on adaptive RAG as described in claim 5, characterized in that, The multi-round sub-query and graph reasoning retrieval path for complex problems includes the following sub-steps: A model that generates a subquery breaks down a complex original query into multiple subqueries. For each subquery, an entity link is created, locating the corresponding node in the knowledge graph; Parallel vector retrieval is performed, and graph expansion retrieval is carried out on the retrieved core clause nodes. Multi-hop traversal is performed along the reference relationship edges or spatiotemporal constraint relationship edges to recall related clauses. All recall results are merged and input into the re-ranking model, which outputs the most relevant clause units.

7. A question-and-answer system for a low-altitude regulatory knowledge base based on adaptive RAG as described in claim 6, characterized in that, In the context integration and generation module, the construction of the structured context includes the following sub-steps: Each retrieved clause is labeled with the name of the source regulation, clause number, and page number; Arrange the clause units in descending order of relevance score and then combine them to form a context. When inputting into the large language generation model, the system uses preset prompt templates to force the model to answer strictly according to the cited materials, and prompts the consulting agency when the information is uncertain.

8. A question-and-answer system for a low-altitude regulatory knowledge base based on adaptive RAG as described in claim 7, characterized in that, In the feedback optimization module, the closed-loop optimization includes: Record the original questions, generated answers, retrieval paths, and model confidence scores of user feedback to generate structured feedback data; Through similarity analysis and confidence assessment, the feedback data is classified into routing error correction samples, retrieved negative samples, or generated adversarial samples. The lightweight large language model, reordering model, and large language generation model are incrementally fine-tuned periodically using new samples to continuously optimize system performance.

9. A question-and-answer system for a low-altitude regulatory knowledge base based on adaptive RAG as described in claim 1, characterized in that, The spatial labels include one or more of the following: true altitude, altitude, corrected sea-level pressure altitude, airspace type, airspace sector identifier, no-fly zone identifier, and geographic coordinate range; The time stamps include one or more of the following: effective date, repeal date, specific time period of day, and effective time window of temporary notices of navigation; The validity labels include: information on the national, ministerial, or local level of the regulations or the civil aviation standard scenario level to which they pertain.

10. A question-and-answer system for a low-altitude regulatory knowledge base based on adaptive RAG as described in claim 1, characterized in that, After the new document is processed by the standard document preprocessing module, its clause units and multi-dimensional tag system are automatically embedded into the vector index and a relationship with existing nodes is established in the knowledge graph.

Images