A specific-scene-oriented knowledge base construction and intelligent retrieval decision method

Abstract

This invention relates to the fields of big data processing and artificial intelligence, and discloses a knowledge base construction and intelligent retrieval decision-making method for specific scenarios. Addressing the problems of shallow semantic understanding, discrete retrieval results, and lack of dynamic perception in existing technologies, this invention first utilizes hybrid coding to extract features from multimodal heterogeneous data and constructs structured argumentation units based on the Thurmin model; secondly, it establishes a multi-dimensional semantic association index containing dynamic attributes; furthermore, it introduces an external data stream monitoring mechanism to automatically trigger updates to the state bits of logical nodes based on real-time features; finally, it uses a dynamic weight optimization algorithm based on node state bits for chain-like reasoning and decision generation. This invention achieves a deep integration of logic and perception, effectively addressing real-time risks in fields such as healthcare and finance, and significantly improving the accuracy and reliability of intelligent decision-making.

Description

Technical Field

[0001] This invention relates to the fields of big data processing and artificial intelligence technology, and in particular to a knowledge base construction and intelligent retrieval decision-making method for specific scenarios. Background Technology

[0002] Existing large language model knowledge base technologies (such as RAG) mainly rely on text slicing and vector retrieval. While they can solve general question-answering problems to some extent, they still have significant limitations in professional fields such as medicine and finance. On the one hand, traditional knowledge bases are mostly based on shallow semantic similarity matching, making it difficult to deeply analyze the complex argumentation logic (such as claim-evidence chains) in professional documents. Furthermore, they lack effective structured fusion methods for multimodal heterogeneous data containing a large number of numerical tables, trend charts, and medical images. As a result, the search results are often discrete document fragments, which cannot form decision recommendations with rigorous logical support.

[0003] On the other hand, existing knowledge bases, once constructed, are often static and lack the ability to perceive real-time dynamic changes in the physical or digital world. In scenarios with extremely high timeliness requirements, such as emergency treatment and high-frequency trading, the external environment (such as patient vital signs and market fluctuations) is constantly changing. Static knowledge entries, if unable to be dynamically corrected based on real-time data streams, are prone to providing outdated or even erroneous decision-making basis. Lacking effective risk mitigation and dynamic adaptation mechanisms, they are unable to meet the intelligent application needs of high-risk scenarios. Summary of the Invention

[0004] This invention provides a knowledge base construction and intelligent retrieval decision-making method for specific scenarios. Through deep logic reconstruction and real-time dynamic perception mechanisms, it solves the problems of shallow semantic understanding in professional fields and static lag in knowledge bases, and significantly improves the accuracy, timeliness and risk response capabilities of intelligent decision-making in complex environments.

[0005] This invention provides a knowledge base construction and intelligent retrieval decision-making method for specific scenarios, including:

[0006] S1. Obtain multimodal heterogeneous data in specific scenarios and perform feature extraction using hybrid coding. Define a logical element schema containing multiple nodes based on the Thurmin argumentation model, and construct structured argumentation units with the logical element schema as constraints. Among them, nodes include claims, guarantees, evidence, support, limitations, and rebuttals.

[0007] S2. Store the structured argumentation unit in a hybrid database and store the multimodal heterogeneous data in association. Establish a high-dimensional semantic vector index for the claim node and a full-text sparse index for the multimodal heterogeneous data. Configure dynamic attribute fields in the limiting node and the rebuttal node. The dynamic attribute fields include status bits, trigger thresholds and data flow mapping interfaces.

[0008] S3. Monitor the external data stream associated with a specific scenario in real time through the data stream mapping interface, calculate the real-time data characteristics, and when the real-time data characteristics meet the trigger threshold, locate the corresponding argument unit associated with the external data stream in the hybrid database, and flip the status bits of the limiting node or the rebuttal node in the argument unit in real time to obtain the updated node status bits.

[0009] S4. In response to the user's search request, retrieve the candidate argument unit set from the hybrid database using a hybrid search strategy, obtain the updated node status bit, calculate the decision score of each candidate argument unit using a dynamic weight optimization algorithm, sort and chain-reason the candidate argument units according to the decision score, and generate an intelligent decision scheme containing risk warnings; wherein, the decision score is jointly determined by the static logic strength and the dynamic state coefficient determined based on the node status bit.

[0010] Furthermore, S1 specifically includes:

[0011] S101. Obtain the original data stream of the multimodal heterogeneous data, and preprocess the original data stream into unstructured text stream, semi-structured table stream and unstructured image stream.

[0012] S102. For different types of data after splitting, feature extraction is performed to generate text semantic vectors, table structured encoding, and image multimodal feature vectors. For semi-structured table streams, a large table model or linearization processing method is used to convert the tables into text descriptions or high-dimensional vectors that retain numerical relationships and logical structures. For unstructured image streams, feature vectors are generated using a multimodal embedding model. If the image stream contains macroscopic statistical charts, a text summary of the chart is first generated using a multimodal model, and then the text summary and image features are jointly embedded.

[0013] S103. Define the data structure of the logical element Schema, including six nodes: claim, guarantee, evidence, support, limitation, and rebuttal. Reserve mapping fields in the Schema for associating with the original data. The mapping fields include research design ID and argumentation ID.

[0014] S104. Input the extracted text semantic vector into the large language model, and identify the core arguments and deductive logic through prompting engineering, and map them as claim nodes and guarantee nodes respectively.

[0015] S105. The extracted table structured coding and image multimodal feature vectors are used as factual evidence and directly mapped to evidence nodes, and externally cited authoritative standards are associated and mapped to supporting nodes.

[0016] S106. Identify the constraints and potential limitations in the original data, map them to limiting nodes and rebuttal nodes respectively, and initialize the dynamic attribute interface of the rebuttal nodes; wherein, the initialization configuration specifically includes: if the original data does not explicitly mention the rebuttal content, then initialize an empty rebuttal node container in the logical element schema; configure a dynamic monitoring interface for the rebuttal node container, the interface including status bit parameters, trigger threshold parameters and data flow source identifier, and set the initial status bit to an inactive state;

[0017] S107. Each generated node is encapsulated as a structured argument unit, assigned a unique identifier, and a bidirectional index relationship is established between the structured argument unit and the original data stream.

[0018] Furthermore, in S104 and S105, the mapping is specifically performed as follows:

[0019] When constructing claim nodes and guarantee nodes, a large language model is used to filter out pure theoretical descriptions that are not empirical, and only the derivation logic based on empirical data is retained.

[0020] When constructing supporting nodes, reference tags in the text stream are parsed and linked to industry regulations, medical guidelines, or journal evaluation indicators in external knowledge bases, serving as external theoretical support to enhance the credibility of the nodes.

[0021] Furthermore, S2 specifically includes:

[0022] S201. Initialize the hybrid database storage architecture and build a hybrid storage environment that includes a core storage engine and a raw data storage area. Among them, a relational database that supports vector extension is used as the core storage engine to store structured argumentation units and vector data; a distributed object storage system is used as the raw data storage area to store unstructured raw multimodal data.

[0023] S202. The structured argumentation unit is serialized and stored in the core storage engine, and the original multimodal data is stored in the original data storage area. A traceability association between the two is established through a foreign key. The specific method for establishing the traceability association is as follows: a foreign key is established in the core storage engine, and the research design ID in the argumentation unit is pointed to the access path of the corresponding original data in the object storage system.

[0024] S203. Extract the claim node text from the stored argument unit, generate semantic vectors, and use the HNSW algorithm or IVF algorithm to construct an approximate nearest neighbor index for the semantic vectors of the claim nodes to support intent-based fuzzy matching.

[0025] S204. Extract text content from the stored raw multimodal data, and construct a full-text sparse index for the text content of the raw multimodal data using an inverted index or the BM25 algorithm to support accurate retrieval for entity names or numerical ranges.

[0026] S205. Traverse the stored argument units, locate the limiting nodes and rebuttal nodes, inject dynamic attribute fields into them, and complete the construction from static storage to dynamic logical network.

[0027] Further, in S205, the dynamic attribute field includes:

[0028] Status bit: A boolean field used to identify the effective status of the node's logical conditions. The initial value is set to inactive.

[0029] Trigger threshold: A JSON object field used to store the specific conditional logic for triggering the state bit to flip.

[0030] Data Stream Mapping Interface: A string field used to store a unique identifier for the external data stream bound to this node;

[0031] Timeliness tag: The timestamp field is used to record the time of the last change of the status bit in order to determine the validity of the current status.

[0032] Furthermore, S3 specifically includes:

[0033] S301. Establish a persistent listening channel for external data streams based on the data stream mapping interface, and maintain an active connection pool;

[0034] S302. Set a time sliding window for the external data stream that is connected, clean and aggregate the data within the window, and calculate the real-time data features; wherein, the calculation method of the real-time data features is as follows: remove outliers within the time sliding window, calculate the moving average, volatility or rate of change of the data within the window, and standardize the calculation results into a data format consistent with the definition of the trigger threshold.

[0035] S303. When the calculated real-time data features arrive, perform a reverse index query using the data stream mapping interface as the key to locate the corresponding argument unit in the hybrid database associated with the external data stream, and perform a state update operation.

[0036] S304. Read the trigger threshold field of the limited node or rebuttal node in the corresponding argumentation unit, and logically compare the real-time data features with the trigger threshold; wherein, the comparison method is: parsing the JSON format condition logic in the trigger threshold field, performing a comparison for a single indicator and a multi-condition composite comparison based on AND and OR logic;

[0037] S305. If the trigger threshold is met, perform an atomic update operation to flip the state bit of the node. When the real-time data feature meets the trigger condition, flip the state bit from inactive to active. If the real-time data feature falls back to the safe range, restore the state bit to inactive according to the preset hysteresis strategy.

[0038] S306. Update the timeliness tag of the node to the current time, write the state change event to the log, and publish the logical state change event through the message queue.

[0039] Furthermore, S4 specifically includes:

[0040] S401. Receive the user's natural language retrieval request, generate a high-dimensional query vector using an embedding model, and extract entity nouns and numerical constraints using natural language processing tools to generate sparse query conditions.

[0041] S402. Based on a hybrid database architecture, the semantic recall path and the precise recall path are executed in parallel, and the results are merged using a fusion algorithm to generate an initial set of candidate argument units; wherein, the semantic recall path uses a high-dimensional query vector to perform an approximate nearest neighbor search in the claim node index; the precise recall path uses sparse query conditions to perform full-text retrieval in the original multimodal data index; the fusion algorithm uses a reciprocal sorting fusion algorithm to deduplicate and merge the results from the two paths;

[0042] S403. Traverse the set of candidate argument units and read the current state bit and update time of the limiting node and the rebuttal node of each argument unit.

[0043] S404. Using a dynamic weight optimization algorithm, combined with the retrieval matching degree and dynamic state coefficient, calculate the final decision score for each candidate argument unit.

[0044] S405. Sort the candidate argumentation units according to the final decision score, set a safety threshold, remove argumentation units whose final decision score is lower than the safety threshold, and select the top K argumentation units after sorting as the final reference.

[0045] S406. Input the selected argumentation units into the large language model, construct prompts including background, evidence and real-time status, instruct the large language model to perform chain reasoning, and generate intelligent decision-making solutions, which include three structured parts: risk prompts based on activation nodes, core suggestions based on high-scoring claims, and evidence support based on original multimodal data.

[0046] Furthermore, in S404, the final decision score The calculation formula is:

[0047]

[0048] in, Cosine similarity score for vector retrieval; The matching score for sparse retrieval; These are preset weight hyperparameters; This is a dynamic state coefficient, and its value is determined by the node state bits obtained by S403:

[0049] If the rebuttal node's status bit is active, then Set it as a penalty factor to reduce the weight;

[0050] If the node status bit is restricted to an active state, then Set it as an enhancement factor to increase the weight;

[0051] If all status bits are inactive, then Set as the baseline value.

[0052] The present invention also provides a knowledge base construction and intelligent retrieval decision-making device for specific scenarios. Based on the knowledge base construction and intelligent retrieval decision-making method for specific scenarios described above, the device includes:

[0053] The acquisition module is used to acquire multimodal heterogeneous data in specific scenarios and perform feature extraction using hybrid coding. Based on the Toulmin argumentation model, it defines a logical element schema containing multiple nodes, and constructs structured argumentation units with the logical element schema as constraints. Among them, nodes include claims, guarantees, evidence, support, limitations, and rebuttals.

[0054] The storage module is used to store the structured argumentation units in a hybrid database and to store the multimodal heterogeneous data in association. It establishes a high-dimensional semantic vector index for the claim nodes, establishes a full-text sparse index for the multimodal heterogeneous data, and configures dynamic attribute fields in the limiting nodes and rebuttal nodes. The dynamic attribute fields include status bits, trigger thresholds, and data flow mapping interfaces.

[0055] The dynamic update module is used to monitor external data streams associated with specific scenarios in real time through the data stream mapping interface, calculate real-time data features, and when the real-time data features meet the trigger threshold, locate the corresponding argument unit associated with the external data stream in the hybrid database, and flip the status bits of the limiting node or rebuttal node in the argument unit in real time to obtain the updated node status bits.

[0056] The generation module is used to respond to the user's search request, recall a set of candidate argument units from the hybrid database using a hybrid search strategy, obtain the updated node status bits, calculate the decision score of each candidate argument unit using a dynamic weight optimization algorithm, sort and chain-reason the candidate argument units according to the decision scores, and generate an intelligent decision scheme containing risk warnings; wherein, the decision score is jointly determined by the static logic strength and the dynamic state coefficient determined based on the node status bits.

[0057] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0058] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0059] The beneficial effects of this invention are as follows:

[0060] This invention achieves deep logical reconstruction and real-time value discovery of heterogeneous data by integrating the Toulmin argumentation model, multimodal hybrid coding, and dynamic perception technology. First, it utilizes multimodal structured extraction technology to transform non-textual evidence such as tables and images into reasonable logical nodes, addressing the problem of shallow semantic understanding of professional data. Second, it innovatively establishes a logical node state update mechanism driven by external data flow, enabling the validity of argumentation units to be adjusted based on real-time environmental changes (such as sudden changes in disease progression or market fluctuations), overcoming the static lag of traditional knowledge bases. Finally, the decision-making scheme generated through a dynamic weight optimization algorithm possesses both strong evidence support and real-time risk warnings, significantly improving the accuracy, reliability, and security of intelligent decision-making in specific scenarios. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of a method flow according to an embodiment of the present invention.

[0062] Figure 2 This is a schematic diagram of the device structure according to an embodiment of the present invention.

[0063] Figure 3 This is a schematic diagram of the internal structure of a computer device according to an embodiment of the present invention.

[0064] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0065] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0066] The specific scenarios described in this invention include, but are not limited to, fields that require logical reasoning based on empirical evidence, such as medical diagnosis, financial risk control, industrial operation and maintenance, and legal consulting.

[0067] like Figure 1 As shown, this invention provides a knowledge base construction and intelligent retrieval decision-making method for specific scenarios, which specifically includes the following steps:

[0068] S1. Obtain multimodal heterogeneous data in specific scenarios and use hybrid coding for feature extraction. Based on the Turmin argumentation model, define a logical element schema containing multiple nodes, and construct structured argumentation units with the logical element schema as constraints. Among them, nodes include claims, guarantees, evidence, support, limitations, and rebuttals.

[0069] Step S1 employs a scheme that integrates the Thurmin model and multimodal coding technology to transform unstructured multimodal heterogeneous data into a structured graph that is machine-understandable and logically reasonable; specifically, it includes the following sub-steps:

[0070] S101. Preprocessing of multimodal data splitting

[0071] First, acquire the original document (such as a PDF file, scanned copy, or data stream) for a specific scenario (e.g., medical record analysis, financial research report review). Use document parsing tools to read the binary stream of the original document, and divide the data stream into three processing channels based on content characteristics: unstructured text stream, semi-structured table stream, and unstructured image stream.

[0072] For unstructured text streams: perform cleaning, segmentation and slicing to remove irrelevant characters and noise.

[0073] For semi-structured table streams: identify table boundaries, extract table headers and cell data, perform structured restoration for complex tables (such as those spanning multiple rows and columns), and maintain the correspondence between rows and columns.

[0074] For unstructured image streams: extract illustrations, charts (such as candlestick charts, medical images) or vector graphics (such as CAD drawings) from the document and record their contextual location information in the original document.

[0075] S102, Multimodal Feature Extraction Based on Hybrid Coding

[0076] Differential coding techniques are used for feature extraction for different types of data after S101 splitting:

[0077] ①Text feature extraction: Use a pre-trained language model (such as BERT) to vectorize and embed text slices to generate text semantic vectors.

[0078] ② Table feature extraction: Using the Table-LLM or linearization method, the table is converted into a text description or high-dimensional vector code containing structural information to preserve the numerical relationships and logical structure in the table.

[0079] ③ Image Feature Extraction: Image data is processed using a multimodal embedding model (such as CLIP). For regular images, multimodal feature vectors are generated; for macroscopic images containing rich information (such as statistical charts), a text summary is first generated using a large multimodal model, and then the summary and image features are jointly embedded.

[0080] S103, Define Logical Element Schema Constraints

[0081] Based on the Toulmin argumentation model, a standardized JSON Schema (data schema) is defined as a structural constraint for information extraction. This schema contains six core fields:

[0082] 1) Claim: A conclusive viewpoint or recommendation;

[0083] 2) Evidence: Factual data supporting the claim;

[0084] 3) Warrant: The logical rule connecting evidence and claims;

[0085] 4) Supporting: The theoretical basis or legal regulations that support the guarantee;

[0086] 5) Qualifier: The scope or conditions under which the claim is valid;

[0087] 5) Rebuttal: This weakens the exceptions or risks to the claim. Additionally, the schema reserves the fields `study_design_ids` (study design ID) and `toulmin_argument_ids` (argument ID) to establish a many-to-many mapping between argument units and the original data.

[0088] S104. Logical Mapping of Textual Semantic Features (Constructing Claims and Guarantees)

[0089] The text content extracted in S102 and the schema defined in S103 are input into the large language model. Through prompt engineering, the model is instructed to identify core arguments from the text and map them as claim nodes; and to identify logical descriptions explaining the derivation process in the text and map them as guarantee nodes. In this process, the model filters out purely theoretical descriptions that are not empirical to ensure the empirical nature of the extracted content.

[0090] S105. Entity Association Based on Multimodal Features (Constructing Evidence and Support)

[0091] The table structure features and image feature vectors extracted in S102 are used as factual evidence and directly mapped to evidence nodes in the Toulmin model. For example, the extracted net profit table from financial statements or CT image features are directly attached to the Evidence field, enabling logical reasoning to be supported by multimodal data.

[0092] External authoritative standards (such as industry regulations, medical guidelines, or journal authority indicators) referenced in text or databases are associated and mapped as supporting nodes to enhance the credibility of the Warrant.

[0093] S106. Initialization configuration of dynamic nodes (construction constraints and rebuttals)

[0094] According to the S103 Schema, identify explicitly mentioned limitations (such as those applicable only to adults) or potential limitations (such as insufficient sample size) from the raw data, and map them as limiting nodes and rebuttal nodes respectively. In particular, for argument units that do not explicitly mention rebuttal content, initialize an empty rebuttal node container and set dynamic attribute interfaces (including status bits, trigger thresholds, etc.) for it, in preparation for dynamic activation based on external data streams (such as real-time sensor data or market conditions) in the subsequent S3 step.

[0095] S107. Encapsulation and ID Mapping of Structured Argumentation Units

[0096] The generated nodes are assembled into a standard JSON object to form a complete structured argument unit. A unique identifier (ID) is assigned to each argument unit, and a bidirectional index relationship is established between the unit and the original multimodal data fragments in S101 (such as original PDF page numbers and original image IDs) through an ID array to ensure that the original source of evidence can be traced during subsequent retrieval.

[0097] S2. Store the structured argumentation unit in a hybrid database and store the multimodal heterogeneous data in association. Establish a high-dimensional semantic vector index for the claim node and a full-text sparse index for the multimodal heterogeneous data. Configure dynamic attribute fields in the limiting node and the rebuttal node. The dynamic attribute fields include status bits, trigger thresholds and data flow mapping interfaces.

[0098] Step S2 employs a hybrid database architecture and dynamic attribute injection technology to construct a multi-dimensional knowledge base that supports both deep logical retrieval and real-time state awareness; specifically, it includes the following sub-steps:

[0099] S201. Initialize the hybrid database storage architecture

[0100] A hybrid storage architecture based on "relational database + vector database + document database" is constructed. In this embodiment, a relational database that supports vector extension is preferably used as the core storage engine to store structured argumentation units and vector data; a distributed object storage system is used as the raw data storage area to store unstructured raw multimodal data (such as original PDF documents and high-definition medical image files).

[0101] S202, The associated storage of structured argumentation units and raw data

[0102] The structured argument units generated in step S1 are serialized in JSON format and stored in the argument knowledge table of the core storage engine. Simultaneously, the raw multimodal data obtained in step S1 is stored in the object storage system, and its access path (URL or URI) is obtained. A foreign key relationship is established in the argument knowledge table, pointing the study_design_ids field in the argument unit to the storage path of the raw data, thereby establishing a persistent traceability link between logical nodes (such as evidence nodes) and physical data (such as the original chart).

[0103] S203. Construct a high-dimensional semantic vector index for claim nodes.

[0104] Extract the text of claim nodes from all stored argument units and use the embedding model (consistent with the model used in S102) to generate fixed-dimensional semantic vectors. In the core storage engine, use the HNSW (Hierarchical Navigable Small World) algorithm or the IVF (Inverted File) algorithm to construct an approximate nearest neighbor (ANN) index for the semantic vectors of the claim nodes. This index aims to support intent-based fuzzy matching in subsequent searches, addressing the issue of inconsistencies between user query expressions and technical terms.

[0105] S204. Construct a full-text sparse index of the original data.

[0106] For the raw multimodal data stored in the object storage system, the text content (including text after OCR recognition) is extracted, and a full-text sparse index is constructed using inverted indexing or the BM25 algorithm. This sparse index is used to supplement the shortcomings of vector indexes in precise matching, especially for precise retrieval of specific entity names (such as drug chemical names, stock codes) and numerical ranges (such as growth rate > 5%), ensuring the accuracy of the knowledge base in finding detailed data.

[0107] S205, Dynamic Attribute Injection of Limitation and Rebuttal Nodes

[0108] Traverse the stored argument units, locate the qualifying and rebuttal nodes, and inject the following dynamic attribute fields using the database's schema extension function or NoSQL field features:

[0109] ① Status Bit: Boolean type, initial value is 0 (inactive / valid). When this bit flips to 1, it indicates that the logical condition of this node is in effect (e.g., the rebuttal is effective, causing the claim to be invalid).

[0110] ② Trigger Threshold: A JSON object that stores the specific conditions for triggering the state to flip. For example, {"metric":"heart_rate","operator":">","value":120}.

[0111] ③ Data Stream Mapping Interface (Stream_ID): String type, storing a unique identifier of the external data stream bound to this node (such as the Kafka Topic name or API Endpoint).

[0112] ④Timestamp: Records the timestamp of the last change to the status bit, used to determine whether the current status has expired.

[0113] The above steps complete the construction from static data storage to dynamic logical network, providing the necessary data foundation and index support for real-time perception in step S3 and dynamic weight calculation in step S4.

[0114] S3. Monitor the external data stream associated with a specific scenario in real time through the data stream mapping interface, calculate the real-time data characteristics, and when the real-time data characteristics meet the trigger threshold, locate the corresponding argument unit associated with the external data stream in the hybrid database, and flip the status bits of the limiting node or the rebuttal node in the argument unit in real time to obtain the updated node status bits.

[0115] Step S3 employs an event-driven architecture and atomic state update technology to map real-time changes in the physical or digital world into dynamic constraints in the logical graph, achieving real-time awareness and logical self-adaptation of the knowledge base; specifically, it includes the following sub-steps:

[0116] S301. Establish a persistent listening channel for external data streams.

[0117] Start the multi-threaded listening service based on the data stream mapping interface (Stream_ID) configured in step S205.

[0118] In financial scenarios, the exchange's market data API can be connected via WebSocket or FIX protocol to subscribe to real-time trading data streams for specific stocks or indices.

[0119] In medical scenarios, the device connects to the Internet of Medical Technology (IoMT) gateway via MQTT or HL7 protocol and subscribes to the vital signs data stream (such as heart rate, blood oxygen, and blood pressure) transmitted from the monitor.

[0120] For industrial scenarios, connect to SCADA systems or PLC interfaces to acquire sensor readings. Maintain an active connection pool to ensure low latency and high availability of data transmission.

[0121] S302, Sliding window calculation of real-time data features

[0122] For high-frequency data streams, a time sliding window (such as the most recent 10 seconds or the most recent 5 data points) is set for preprocessing and feature calculation to eliminate instantaneous noise interference.

[0123] ① Data cleaning: Remove outliers (null values ​​or physically impossible values) caused by network jitter or sensor malfunction.

[0124] ② Feature aggregation: Calculate statistical features within the sliding window, such as moving average (MA), volatility, or rate of change.

[0125] ③ Format alignment: Standardize the calculated real-time features into a data format consistent with the trigger threshold definition in the database (e.g., unified units, unified precision).

[0126] S303, Indexing and Transaction Locking of Related Argument Units

[0127] When the real-time features calculated by S302 arrive, a reverse index query is performed in the argument knowledge table of the hybrid database using the Stream_ID of the data stream as the key to quickly locate all structured argument units that have subscribed to the data stream. To ensure data consistency in a high-concurrency environment, a database row-level lock or optimistic locking mechanism is used to lock the target argument unit to be updated, preventing read-write conflicts during state determination.

[0128] S304, Logical determination of trigger threshold

[0129] Read the trigger threshold field of the limiting or rebuttal node in the locked argument unit. Parse the JSON-formatted conditional logic and compare the real-time data features calculated by S302 with the threshold.

[0130] Example of comparison logic: If the threshold is defined as {"metric":"price_drop","operator":"≥","value":0.05} (the drop exceeds 5%), and the real-time calculated drop feature is 0.06, then it is determined that the trigger condition is met.

[0131] Multiple conditional compounding: Supports compound condition judgments using "AND" and "OR" logic, such as simultaneously satisfying "blood pressure > 160" and "heart rate > 100".

[0132] S305, Atomic Flipping of Logic Node State Bits

[0133] Once the triggering condition is met, an atomic update operation is performed to modify the status bit (Status_Bit) of the node.

[0134] Activation operation: Flip the status bit from 0 (default / invalid) to 1 (activated / valid). For example, when a refutation node is activated, it means that the solution described by the argument unit has a very high risk and has been logically refuted.

[0135] Recovery Operation: If the real-time data characteristics fall back to a safe range (e.g., the drop recovers to within 1%), the status bit can be restored from 1 to 0 according to the preset hysteresis strategy, and the alarm can be lifted.

[0136] S306, Timeliness Tag Updates and Status Broadcasts

[0137] While completing the state bit flip, update the timeliness label of the node to the current time.

[0138] Persistent Log: This state change record is written to the operation log for subsequent auditing and backtracking.

[0139] Message broadcasting: Logical state change events are published via message queues (such as Kafka) to notify downstream S4 decision modules or other subscription systems that the logical attributes of the knowledge point have undergone a substantial change. This mechanism ensures that intelligent decision-making is always based on the latest logical state, realizing a leap from static knowledge to dynamic wisdom.

[0140] S4. In response to the user's search request, retrieve the candidate argument unit set from the hybrid database using a hybrid search strategy, obtain the updated node status bit, calculate the decision score of each candidate argument unit using a dynamic weight optimization algorithm, sort and chain-reason the candidate argument units according to the decision score, and generate an intelligent decision scheme containing risk warnings; wherein, the decision score is jointly determined by the static logic strength and the dynamic state coefficient determined based on the node status bit.

[0141] Step S4 employs a hybrid retrieval strategy and a dynamic weighted scoring algorithm, upgrading the retrieval process from simple text relevance matching to a multi-dimensional decision-making process incorporating logic and risk perception. This ensures that the final generated solution not only meets semantic requirements but also mitigates potential risks in a real-time environment. Specifically, it includes the following sub-steps:

[0142] S401. Intent parsing and vectorization of retrieval requests

[0143] Receive natural language search requests input by the user (e.g., emergency treatment plan for sudden chest pain in patients with hypertension or investment strategy for the consumer electronics sector today). Using an embedding model similar to that in step S102, convert the user's search request into a high-dimensional query vector. Simultaneously, use natural language processing tools (such as Jieba segmentation or spaCy) to extract keywords from the search request, identify entity nouns (e.g., hypertension, chest pain) and numerical constraints, and generate sparse query conditions.

[0144] S402, Hybrid Recall (Dual-path parallel)

[0145] Based on a hybrid database architecture, two-way recall operations are executed in parallel to generate a set of candidate argument units:

[0146] ① Semantic Recall Path: Using the query vector generated by S401, perform an approximate nearest neighbor (ANN) search in the claim node vector index of the core storage engine to recall the Top-N argument units that are semantically similar in logical claims.

[0147] ② Precise Recall Path: Using the sparse query conditions generated by S401, Boolean queries or BM25 searches are performed in the full-text sparse index of the original data storage area to recall the Top-N argument units that precisely match the original evidence (such as table values ​​or specific terms).

[0148] Results fusion: The Reciprocal Rank Fusion (RRF) algorithm is used to deduplicate and merge the two recall results to generate an initial set of candidate argument units.

[0149] S403, Synchronization acquisition of real-time status bits

[0150] Iterate through the set of candidate argument units generated by S402. For each argument unit in the set, retrieve the current attribute values ​​of its limiting node and rebuttal node through a database query operation, specifically including:

[0151] Status Bit: Get the latest status (0 or 1) after real-time flipping based on the external data stream in step S3.

[0152] Time stamp: Checks whether the update time of the status bit is within the allowed time window (such as the most recent 1 minute). If it times out, it is marked as unknown status.

[0153] S404. Decision Score Calculation Based on Dynamic State Coefficients

[0154] The final decision score for each candidate argument unit is calculated using a dynamic weight optimization algorithm. The calculation formula is as follows:

[0155]

[0156] in, Cosine similarity score for vector retrieval; The matching score for sparse retrieval (BM25 Score). For the preset weight hyperparameters (such as...) =0.7, =0.3);

[0157] This is a dynamic state coefficient, and its value is determined by the node state bit obtained by S403: if the rebuttal node's state bit is 1 (active state, indicating high risk), then... Setting a penalty factor (such as 0.1 or a negative value) significantly reduces the recommendation weight of this scheme; if the node state bit is limited to 1 (active state, indicating that specific constraints are met), then... Set it as an enhancement factor (e.g., 1.2) to improve the matching degree of the scheme; if all status bits are 0 (default status), then... Set it to 1.

[0158] S405, Risk Perception Ranking and Cutoff

[0159] Calculated according to S404 The candidate argument units are sorted in descending order.

[0160] Safety cut-off: Set a safety threshold to directly remove argument units with scores below the threshold, especially high-risk units whose scores drop sharply due to the activation of rebuttal nodes, to ensure that they do not enter the final generation stage, thereby realizing a circuit breaker mechanism at the knowledge base level.

[0161] Top-K cutoff: Select the top K (e.g., Top-5) argument units after sorting as the final reference.

[0162] S406, Intelligent Solution Generation Based on Chain Reasoning

[0163] The Top-K argument units selected by S405 are input into the large language model to construct a structured Prompt containing background information, evidence chain and real-time state.

[0164] Prompt construction: Explicitly mark the activation criteria in the Prompt (e.g., Note: Current patient heart rate > 100, the following scheme is for reference only).

[0165] Chained Reasoning (CoT): Instructs the large language model to reason according to the logical chain of "analyzing claims → citing multimodal evidence → combining real-time risk status → generating final recommendations".

[0166] Output Generation: The final output is a structured intelligent decision-making solution, which consists of three parts: ① Risk Warning: Based on the activated rebuttal / limitation nodes, the taboos or warnings in the current environment are displayed first; ② Core Recommendation: Action guidelines generated based on high-scoring claims; ③ Evidence Support: List the original multimodal data sources that support the recommendation (such as table screenshots, reference links).

[0167] like Figure 2 As shown, the present invention also provides a knowledge base construction and intelligent retrieval decision-making device for specific scenarios. Based on the knowledge base construction and intelligent retrieval decision-making method for specific scenarios described above, the device includes:

[0168] Module 1 is used to acquire multimodal heterogeneous data in specific scenarios and perform feature extraction using hybrid coding. Based on the Toulmin argumentation model, a logical element schema containing multiple nodes is defined, and structured argumentation units are constructed using the logical element schema as constraints. Among them, nodes include claims, guarantees, evidence, support, limitations, and rebuttals.

[0169] Storage module 2 is used to store the structured argumentation unit in a hybrid database and to store the multimodal heterogeneous data in association. It establishes a high-dimensional semantic vector index for the claim node, establishes a full-text sparse index for the multimodal heterogeneous data, and configures dynamic attribute fields in the limiting node and the rebuttal node. The dynamic attribute fields include status bits, trigger thresholds, and data flow mapping interfaces.

[0170] The dynamic update module 3 is used to monitor the external data stream associated with a specific scenario in real time through the data stream mapping interface, calculate the real-time data characteristics, and when the real-time data characteristics meet the trigger threshold, locate the corresponding argument unit associated with the external data stream in the hybrid database, and flip the status bits of the limiting node or the rebuttal node in the argument unit in real time to obtain the updated node status bits.

[0171] The generation module 4 is used to respond to the user's search request, recall the set of candidate argument units from the hybrid database using a hybrid search strategy, obtain the updated node status bits, calculate the decision score of each candidate argument unit using a dynamic weight optimization algorithm, sort and chain-reason the candidate argument units according to the decision scores, and generate an intelligent decision scheme containing risk warnings; wherein, the decision score is jointly determined by the static logic strength and the dynamic state coefficient determined based on the node status bits.

[0172] Each of the above modules is used to perform the corresponding steps in the knowledge base construction and intelligent retrieval decision-making method for specific scenarios. The specific implementation methods are as described in the above method embodiments, and will not be repeated here.

[0173] like Figure 3 As shown, the present invention also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores all data required for the process of building a knowledge base and implementing intelligent retrieval decision-making methods for specific scenarios. The network interface allows communication with external terminals via a network connection. The computer program is executed by the processor to implement the knowledge base construction and intelligent retrieval decision-making methods for specific scenarios.

[0174] Those skilled in the art will understand that Figure 3The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment on which the present application is applied.

[0175] An embodiment of this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-described knowledge base construction and intelligent retrieval decision-making methods for a specific scenario.

[0176] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by hardware related to computer program instructions. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the above method embodiments. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM), such as dynamic RAM (used as main storage) or static RAM (commonly used as cache memory). By way of illustration and not limitation, RAM has various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), and Rambus DRAM (RDRAM).

[0177] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0178] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A knowledge base construction and intelligent retrieval decision-making method for specific scenarios, characterized in that, include: S1. Obtain multimodal heterogeneous data in specific scenarios and perform feature extraction using hybrid coding. Define a logical element schema containing multiple nodes based on the Thurmin argumentation model, and construct structured argumentation units with the logical element schema as constraints. Among them, nodes include claims, guarantees, evidence, support, limitations, and rebuttals. S2. Store the structured argumentation unit in a hybrid database and store the multimodal heterogeneous data in association. Establish a high-dimensional semantic vector index for the claim node and a full-text sparse index for the multimodal heterogeneous data. Configure dynamic attribute fields in the limiting node and the rebuttal node. The dynamic attribute fields include status bits, trigger thresholds and data flow mapping interfaces. S3. Monitor the external data stream associated with a specific scenario in real time through the data stream mapping interface, calculate the real-time data characteristics, and when the real-time data characteristics meet the trigger threshold, locate the corresponding argument unit associated with the external data stream in the hybrid database, and flip the status bits of the limiting node or the rebuttal node in the argument unit in real time to obtain the updated node status bits. S4. In response to the user's search request, retrieve the candidate argument unit set from the hybrid database using a hybrid search strategy, obtain the updated node status bit, calculate the decision score of each candidate argument unit using a dynamic weight optimization algorithm, sort and chain-reason the candidate argument units according to the decision score, and generate an intelligent decision scheme containing risk warnings; wherein, the decision score is jointly determined by the static logic strength and the dynamic state coefficient determined based on the node status bit.

2. The knowledge base construction and intelligent retrieval decision-making method for specific scenarios according to claim 1, characterized in that, S1 specifically includes: S101. Obtain the original data stream of the multimodal heterogeneous data, and preprocess the original data stream into unstructured text stream, semi-structured table stream and unstructured image stream. S102. For different types of data after splitting, feature extraction is performed to generate text semantic vectors, table structured encoding, and image multimodal feature vectors. For semi-structured table streams, a large table model or linearization processing method is used to convert the tables into text descriptions or high-dimensional vectors that retain numerical relationships and logical structures. For unstructured image streams, feature vectors are generated using a multimodal embedding model. If the image stream contains macroscopic statistical charts, a text summary of the chart is first generated using a multimodal model, and then the text summary and image features are jointly embedded. S103. Define the data structure of the logical element Schema, including six nodes: claim, guarantee, evidence, support, limitation, and rebuttal. Reserve mapping fields in the Schema for associating with the original data. The mapping fields include research design ID and argumentation ID. S104. Input the extracted text semantic vector into the large language model, and identify the core arguments and deductive logic through prompting engineering, and map them as claim nodes and guarantee nodes respectively. S105. The extracted table structured coding and image multimodal feature vectors are used as factual evidence and directly mapped to evidence nodes, and externally cited authoritative standards are associated and mapped to supporting nodes. S106. Identify the constraints and potential limitations in the original data, map them to limiting nodes and rebuttal nodes respectively, and initialize the dynamic attribute interface of the rebuttal nodes; wherein, the initialization configuration specifically includes: if the original data does not explicitly mention the limiting or rebuttal content, initialize an empty limiting node container or a rebuttal node container respectively, and set a dynamic monitoring interface for it. The interface includes a status bit parameter, a trigger threshold parameter and a data flow source identifier, and set the initial status bit to an inactive state; S107. Each generated node is encapsulated as a structured argument unit, assigned a unique identifier, and a bidirectional index relationship is established between the structured argument unit and the original data stream.

3. The knowledge base construction and intelligent retrieval decision-making method for specific scenarios according to claim 2, characterized in that, In S104 and S105, the mapping method is as follows: When constructing claim nodes and guarantee nodes, a large language model is used to filter out pure theoretical descriptions that are not empirical, and only the derivation logic based on empirical data is retained. When constructing supporting nodes, reference tags in the text stream are parsed and linked to industry regulations, medical guidelines, or journal evaluation indicators in external knowledge bases, serving as external theoretical support to enhance the credibility of the nodes.

4. The knowledge base construction and intelligent retrieval decision-making method for specific scenarios according to claim 1, characterized in that, S2 specifically includes: S201. Initialize the hybrid database storage architecture and build a hybrid storage environment that includes a core storage engine and a raw data storage area. Among them, a relational database that supports vector extension is used as the core storage engine to store structured argumentation units and vector data; a distributed object storage system is used as the raw data storage area to store unstructured raw multimodal data. S202. The structured argumentation unit is serialized and stored in the core storage engine, and the original multimodal data is stored in the original data storage area. A traceability association between the two is established through a foreign key. The specific method for establishing the traceability association is as follows: a foreign key is established in the core storage engine, and the research design ID in the argumentation unit is pointed to the access path of the corresponding original data in the object storage system. S203. Extract the claim node text from the stored argument unit, generate semantic vectors, and use the HNSW algorithm or IVF algorithm to construct an approximate nearest neighbor index for the semantic vectors of the claim nodes to support intent-based fuzzy matching. S204. Extract text content from the stored raw multimodal data, and construct a full-text sparse index for the text content of the raw multimodal data using an inverted index or the BM25 algorithm to support accurate retrieval for entity names or numerical ranges. S205. Traverse the stored argument units, locate the limiting nodes and rebuttal nodes, inject dynamic attribute fields into them, and complete the construction from static storage to dynamic logical network.

5. The knowledge base construction and intelligent retrieval decision-making method for specific scenarios according to claim 4, characterized in that, In step S205, the dynamic attribute field includes: Status bit: A boolean field used to identify the effective status of the node's logical conditions. The initial value is set to inactive. Trigger threshold: A JSON object field used to store the specific conditional logic for triggering the state bit to flip. Data Stream Mapping Interface: A string field used to store a unique identifier for the external data stream bound to this node; Timeliness tag: The timestamp field is used to record the time of the last change of the status bit in order to determine the validity of the current status.

6. The knowledge base construction and intelligent retrieval decision-making method for specific scenarios according to claim 1, characterized in that, S3 specifically includes: S301. Establish a persistent listening channel for external data streams based on the data stream mapping interface, and maintain an active connection pool; S302. Set a time sliding window for the external data stream that is connected, clean and aggregate the data within the window, and calculate the real-time data features; wherein, the calculation method of the real-time data features is as follows: remove outliers within the time sliding window, calculate the moving average, volatility or rate of change of the data within the window, and standardize the calculation results into a data format consistent with the definition of the trigger threshold. S303. When the calculated real-time data features arrive, perform a reverse index query using the data stream mapping interface as the key to locate the corresponding argument unit in the hybrid database associated with the external data stream, and perform a state update operation. S304. Read the trigger threshold field of the limited node or rebuttal node in the corresponding argumentation unit, and logically compare the real-time data features with the trigger threshold; wherein, the comparison method is: parsing the JSON format condition logic in the trigger threshold field, performing a comparison for a single indicator and a multi-condition composite comparison based on AND and OR logic; S305. If the trigger threshold is met, perform an atomic update operation to flip the state bit of the node. When the real-time data feature meets the trigger condition, flip the state bit from inactive to active. If the real-time data feature falls back to the safe range, restore the state bit to inactive according to the preset hysteresis strategy. S306. Update the timeliness tag of the node to the current time, write the state change event to the log, and publish the logical state change event through the message queue.

7. The knowledge base construction and intelligent retrieval decision-making method for specific scenarios according to claim 1, characterized in that, S4 specifically includes: S401. Receive the user's natural language retrieval request, generate a high-dimensional query vector using an embedding model, and extract entity nouns and numerical constraints using natural language processing tools to generate sparse query conditions. S402. Based on a hybrid database architecture, the semantic recall path and the precise recall path are executed in parallel, and the results are merged using a fusion algorithm to generate an initial set of candidate argument units; wherein, the semantic recall path uses a high-dimensional query vector to perform an approximate nearest neighbor search in the claim node index; the precise recall path uses sparse query conditions to perform full-text retrieval in the original multimodal data index; the fusion algorithm uses a reciprocal sorting fusion algorithm to deduplicate and merge the results from the two paths; S403. Traverse the set of candidate argument units and read the current state bit and update time of the limiting node and the rebuttal node of each argument unit. S404. Using a dynamic weight optimization algorithm, combined with the retrieval matching degree and dynamic state coefficient, calculate the final decision score for each candidate argument unit. S405. Sort the candidate argumentation units according to the final decision score, set a safety threshold, remove argumentation units whose final decision score is lower than the safety threshold, and select the top K argumentation units after sorting as the final reference. S406. Input the selected argumentation units into the large language model, construct prompts including background, evidence and real-time status, instruct the large language model to perform chain reasoning, and generate intelligent decision-making solutions, which include three structured parts: risk prompts based on activation nodes, core suggestions based on high-scoring claims, and evidence support based on original multimodal data.

8. The knowledge base construction and intelligent retrieval decision-making method for specific scenarios according to claim 7, characterized in that, In S404, the final decision score The calculation formula is: in, Cosine similarity score for vector retrieval; The matching score for sparse retrieval; These are preset weight hyperparameters; This is a dynamic state coefficient, and its value is determined by the node state bits obtained by S403: If the rebuttal node's status bit is active, then Set it as a penalty factor to reduce the weight; If the node status bit is restricted to an active state, then Set it as an enhancement factor to increase the weight; If all status bits are inactive, then Set as the baseline value.

9. A knowledge base construction and intelligent retrieval decision-making device for a specific scenario, based on the knowledge base construction and intelligent retrieval decision-making method for a specific scenario as described in any one of claims 1-8, characterized in that, The device includes: The acquisition module is used to acquire multimodal heterogeneous data in specific scenarios and perform feature extraction using hybrid coding. Based on the Toulmin argumentation model, it defines a logical element schema containing multiple nodes, and constructs structured argumentation units with the logical element schema as constraints. Among them, nodes include claims, guarantees, evidence, support, limitations, and rebuttals. The storage module is used to store the structured argumentation units in a hybrid database and to store the multimodal heterogeneous data in association. It establishes a high-dimensional semantic vector index for the claim nodes, establishes a full-text sparse index for the multimodal heterogeneous data, and configures dynamic attribute fields in the limiting nodes and rebuttal nodes. The dynamic attribute fields include status bits, trigger thresholds, and data flow mapping interfaces. The dynamic update module is used to monitor external data streams associated with specific scenarios in real time through the data stream mapping interface, calculate real-time data features, and when the real-time data features meet the trigger threshold, locate the corresponding argument unit associated with the external data stream in the hybrid database, and flip the status bits of the limiting node or rebuttal node in the argument unit in real time to obtain the updated node status bits. The generation module is used to respond to the user's search request, recall a set of candidate argument units from the hybrid database using a hybrid search strategy, obtain the updated node status bits, calculate the decision score of each candidate argument unit using a dynamic weight optimization algorithm, sort and chain-reason the candidate argument units according to the decision scores, and generate an intelligent decision scheme containing risk warnings; wherein, the decision score is jointly determined by the static logic strength and the dynamic state coefficient determined based on the node status bits.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images