A large-model-based intelligent decision-making system and method

Abstract

This invention relates to the field of artificial intelligence decision-making technology, specifically to an intelligent decision-making system and method based on a large model. The system includes: receiving raw decision request streams from heterogeneous business sources; performing semantic normalization and knowledge enhancement processing on these streams to generate a structured decision query containing a core problem statement, key fact triples, and a knowledge index; invoking a large language model inference engine to generate a preliminary decision reasoning chain based on the structured query; using a predefined decision logic constraint graph to perform consistency verification and completion of the reasoning chain, generating a verified complete decision logic tree; and performing pattern matching in an executable atomic operation library based on the logic tree to map and assemble an executable decision action plan. This invention improves the understanding accuracy, reasoning reliability, and decision executability of large models in complex decision-making scenarios through structured query input and external logic verification.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence decision-making technology, and in particular to an intelligent decision-making system and method based on a large model. Background Technology

[0002] In the field of AI-assisted decision-making, existing technologies primarily rely on large language models to directly process users' raw requests. These requests are typically input in natural language, and the model performs end-to-end analysis and response based on its training data and internal parameters. This approach is efficient for simple problems, but its processing is highly dependent on the original quality of the input text and lacks standardized information preprocessing.

[0003] The current approach suffers from limitations in its input and processing mechanisms. Requests from different business channels exhibit semantic differences, and the accompanying contextual data is often unstructured text, containing significant redundancy and noise. This makes it difficult for the model to accurately identify core decision-making intentions and key facts, resulting in a weak foundation for reasoning. Furthermore, the model's autonomously generated reasoning process is entirely internalized, lacking real-time interactive verification with external domain rules and logical structures. Consequently, its output may deviate from actual business constraints or generate logical contradictions, making it difficult to guarantee the credibility and operability of the decisions. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent decision-making system and method based on a large model.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent decision-making method based on a large model, comprising:

[0006] Receive raw decision request streams from heterogeneous service sources, the raw decision request streams containing request identifiers, problem description text, associated context data, and expected output format indications;

[0007] Semantic normalization and knowledge enhancement processing are performed on the original decision request stream to generate a structured decision query. The structured decision query includes a semantically aligned core question statement, key fact triples extracted from the associated context data, and an index of pre-set knowledge entries related to the question domain.

[0008] The large language model inference engine is invoked, and the structured decision query is used as input to drive the large language model inference engine to generate a multi-step inference chain, thereby obtaining a preliminary decision inference chain.

[0009] Based on a predefined decision logic constraint graph, consistency verification and completion operations are performed on the preliminary decision reasoning chain to generate a verified complete decision logic tree.

[0010] Based on the complete decision logic tree, pattern matching is performed in the preset executable atomic operation library to map and assemble an executable decision action plan.

[0011] As a further aspect of the present invention, the step of performing semantic normalization and knowledge enhancement processing on the original decision request stream to generate a structured decision query specifically includes:

[0012] The problem description text in the original decision request stream is parsed, and word segmentation, stop word removal, and stemming are performed to generate a standardized word sequence.

[0013] Named entity recognition and relation extraction are performed on the standardized lexical sequence to extract entities, attributes and relationships between entities from the problem description text and associated context data, forming the key fact triples;

[0014] By utilizing the core terms in the standardized word sequence, the domain knowledge graph is queried to obtain the concepts, attributes, and rules that are semantically associated with the core terms, thus forming the pre-set knowledge entry index;

[0015] The standardized word sequence, the key fact triples, and the pre-defined knowledge item index are filled and combined according to a predefined structured query template to construct the data structure of the structured decision query.

[0016] Add metadata to the structured decision query, the metadata including at least the request identifier, processing timestamp, and version identifier used in the semantic normalization process.

[0017] As a further aspect of the present invention, the step of invoking the large language model inference engine, taking the structured decision query as input, and driving the large language model inference engine to execute multi-step inference chain generation to obtain a preliminary decision inference chain, specifically involves:

[0018] The core question statements, key fact triples, and pre-set knowledge item indexes in the structured decision query are constructed into a large language model prompt text according to a specific prompt template;

[0019] The prompt text of the large language model is input into the large language model inference engine, which has been fine-tuned by instructions.

[0020] The large language model reasoning engine, based on the prompt text of the large language model, uses chain-like thinking generation technology to gradually output reasoning steps, intermediate conclusions, and plans for subsequent steps, forming a preliminary reasoning chain composed of multiple reasoning step nodes.

[0021] For each reasoning step node in the preliminary reasoning chain, its reasoning basis is extracted, and the reasoning basis points to the key fact triple or a specific item in the pre-set knowledge item index;

[0022] The generation order, content, and corresponding reasoning basis of all reasoning step nodes in the preliminary reasoning chain are recorded and encapsulated together into the data structure of the preliminary decision reasoning chain.

[0023] As a further aspect of the present invention, the step of performing consistency verification and completion operations on the preliminary decision reasoning chain based on a predefined decision logic constraint graph to generate a verified complete decision logic tree specifically includes:

[0024] Load a predefined decision logic constraint graph corresponding to the problem domain in the structured decision query. The decision logic constraint graph represents the decision elements, logical rules and constraints in the domain in a graph structure.

[0025] Traverse each reasoning step node in the preliminary decision reasoning chain and match the decision elements involved in the reasoning step node with the nodes in the decision logic constraint graph;

[0026] For a successfully matched decision element, check whether the logical rules applied by the reasoning step node are consistent with the rules associated with the decision element in the decision logic constraint graph.

[0027] If an inconsistency is found, the content of the inference step node is replaced or modified with consistent logical rules according to the decision logic constraint graph.

[0028] If it is found that the preliminary decision reasoning chain is missing a decision element or reasoning branch that must be considered in the decision logic constraint graph, then a new reasoning step node is generated based on the decision logic constraint graph and inserted into the appropriate position of the preliminary decision reasoning chain.

[0029] All the reasoning steps that have been verified, corrected and completed are reorganized into a tree structure according to their logical dependencies. The root node of the tree structure is the decision problem, the leaf nodes are the decision conclusions, and the intermediate nodes are the reasoning steps, thus forming the complete decision logic tree.

[0030] As a further aspect of the present invention, the step of performing pattern matching, mapping, and assembling an executable decision action plan based on the complete decision logic tree in a preset executable atomic operation library specifically involves:

[0031] The complete decision logic tree is analyzed to extract the decision intent and target state implied in all leaf nodes and key intermediate nodes;

[0032] Traverse the executable atomic operation library, which stores a variety of predefined basic operation units with explicit inputs, outputs and execution interfaces;

[0033] Each decision intention or target state is semantically matched with the basic operation units in the executable atomic operation library, and one or more basic operation units that can realize its function are found as candidates for each intention or state.

[0034] Based on the logical order and dependencies between nodes in the complete decision logic tree, the execution timing and data flow between the selected basic operation units are determined.

[0035] All selected basic operation units are logically connected according to the determined execution sequence and data flow direction, and specific input parameters for each basic operation unit are configured. The input parameters are derived from the context information of the corresponding node in the complete decision logic tree.

[0036] The connected and configured sequence of basic operation units is encapsulated into the executable decision action plan, which includes a series of ordered operation instructions and the data transfer relationships between the operation instructions.

[0037] As a further aspect of the present invention, it also includes the step of simulating and risk-assessing the executable decision-making action plan:

[0038] Obtain a snapshot of the current system state, which contains the current values ​​of all environmental variables and resource states related to the decision action plan;

[0039] In an isolated simulation environment, the system state snapshot is loaded as the initial state, and each operation instruction in the decision action plan is executed sequentially.

[0040] It records in real time the output results of each operation command during the simulation execution, the modifications to the state of the simulation environment, and the simulation events triggered.

[0041] The final system state achieved by the simulation is compared with the preset expected target state in the complete decision logic tree, and the target achievement deviation is calculated.

[0042] Analyze the data recorded during the simulation execution to identify the sequence of operation instructions that leads to execution failure, resource conflicts, or violation of constraints, and mark them as potential risk points.

[0043] As a further aspect of the present invention, it also includes a step of optimizing and adjusting the executable decision-making action plan based on simulation results:

[0044] Based on the calculated deviation in target achievement, locate the decision logic tree node or operational instruction in the decision action plan that caused the main deviation.

[0045] For the located node or instruction, backtrack and query its upstream dependent nodes and reasoning basis in the complete decision logic tree;

[0046] Based on the analysis results of the potential risk points, reassess the sufficiency of the relevant reasoning or the applicability of the relevant operational instructions;

[0047] Within the framework of the decision logic constraint graph, local re-reasoning is performed on the decision logic tree nodes that cause deviations, or the corresponding operation instructions in the decision action plan are replaced, parameters are adjusted, or sorting is optimized.

[0048] Based on the re-inference results or the optimized sequence of operation instructions, an updated decision action plan fragment is generated and used to replace the corresponding part in the original decision action plan, thus forming an optimized decision action plan.

[0049] As a further aspect of the present invention, it also includes the step of executing the optimized decision-making action plan and monitoring its actual effects:

[0050] The optimized decision action plan is decomposed into discrete execution tasks, and the execution tasks are assigned to the corresponding execution agents or execution engines;

[0051] The execution of each task is triggered sequentially, and before each task begins, its input conditions and preset premises are verified to match the current actual system state.

[0052] During task execution, real-time data is collected on execution logs, output results, and the impact of the execution process on the actual system state.

[0053] After all tasks are completed, the actual execution output is summarized and compared with the preset expected target state in the complete decision logic tree;

[0054] The data collected during the actual execution process, the comparative analysis results of the final output and the expected results, along with the original decision request identifier, are stored in the historical decision case library.

[0055] As a further aspect of the present invention, it also includes a step of iteratively learning the decision-making process based on historical execution data:

[0056] Decision cases are periodically extracted from the historical decision case library. The decision cases include the original decision request, the generated decision logic tree, the decision action plan, the actual execution results, and the effect evaluation data.

[0057] The quality of the decision logic tree and decision action plan is evaluated by using the actual execution results and effect evaluation data in the decision case as feedback signals.

[0058] For decision-making cases where the results did not meet expectations, analyze the reasoning flaws, knowledge gaps, or constraint violations in the decision-making logic tree, and analyze the operational inappropriateness or environmental mismatch in the decision-making action plan.

[0059] Using the analysis results, targeted additions, corrections, or optimizations are made to the fine-tuning data of the large language model inference engine, the decision logic constraint graph, or the executable atomic operation library;

[0060] The optimized model, graph, and operation library are applied to the new decision request processing flow, forming a closed loop from decision-making to execution and then to feedback optimization.

[0061] As a further aspect of the present invention, the present invention also includes an intelligent decision-making system based on a large model, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the intelligent decision-making method based on a large model as described above.

[0062] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0063] By performing semantic normalization and knowledge augmentation on the original decision request stream, unstructured natural language requests and context are transformed into structured queries containing aligned core question statements, key fact triples, and pre-defined knowledge indexes. This process essentially builds a high-quality, information-enhanced, standardized input interface for subsequent large-scale model inference. Semantic alignment of the core question statements eliminates ambiguity in requests from different sources, providing the model with a unified semantic understanding benchmark. Automatic extraction of key fact triples from the context filters out noise, transforming scattered data into structured facts that can be precisely processed by the machine. The associated pre-defined knowledge index injects real-time, controllable domain knowledge into the model. This structured input method improves the accuracy of large models' understanding of the essence of complex decision problems, reduces illusions and misjudgments caused by ambiguous input or missing information, and lays a solid foundation for generating reliable inference chains.

[0064] A predefined decision logic constraint graph is introduced as an external verification and completion mechanism to perform consistency checks and logical completion on the initial reasoning chains generated by the large language model. This graph encapsulates a network of rules, constraints, and logical relationships specific to the domain. The initial reasoning chain is compared with the graph to check whether its reasoning steps conflict with established rules and whether there are any breakpoints in logical jumps. When inconsistencies or omissions are found, the system can automatically correct or fill in the graph, thereby regularizing the model's endogenous, potentially unstable reasoning paths into a complete decision logic tree that conforms to the logic of the external domain. This process combines the creative reasoning capabilities of the large model with the constraints of the external deterministic knowledge system, ensuring the rigor, consistency, and interpretability of the final decision logic. The verified decision logic tree has a clear structure, facilitating pattern matching with downstream executable atomic operations, making the transformation from decision reasoning to actual action plans direct, efficient, and reliable. Attached Figure Description

[0065] Figure 1 This is a flowchart of the intelligent decision-making method based on a large model as described in this invention;

[0066] Figure 2 A flowchart for generating structured decision queries;

[0067] Figure 3 A flowchart for generating the initial decision-making reasoning chain;

[0068] Figure 4 To simulate and simulate the entire process of system status monitoring trend chart;

[0069] Figure 5 This is a dynamic stacked area diagram of the data collected during the decision-making and execution process. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0071] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0072] See Figure 1 The system receives raw decision request streams from heterogeneous business sources. Each request in this stream contains a request identifier, a problem description text, associated context data, and an indication of the expected output format. The system first performs semantic normalization and knowledge enhancement processing on the raw decision request streams. This process transforms the unstructured input into a structured decision query containing a semantically aligned core problem statement, key fact triples extracted from the associated context data, and an index of pre-defined knowledge entries related to the problem domain. Subsequently, the system calls the large language model inference engine, using the generated structured decision query as input to drive the engine to perform multi-step inference chain generation, producing a preliminary decision inference chain containing multiple inference step nodes and their basis. Then, based on a predefined decision logic constraint graph, the system performs consistency verification and logical completion operations on this preliminary chain, generating a verified, structured, tree-like complete decision logic tree. Finally, based on this complete decision logic tree, the system performs pattern matching and semantic mapping in a pre-defined executable atomic operation library, assembling the decision logic nodes into an executable decision action plan with a clear execution sequence and data flow direction.

[0073] See Figure 2, in an embodiment of the present invention, an intelligent decision-making method based on a large model involves performing semantic normalization and knowledge enhancement processing on an original decision request stream to generate a structured decision query. The original decision request stream comes from heterogeneous business sources such as enterprise cloud service platforms, server monitoring systems, or user operation and maintenance work order systems. The original decision request stream contains a request identifier, a problem description text, associated context data, and an expected output format indication. The problem description text is a natural language statement such as "The CPU usage rate of cloud service ECS-08 has been continuously above 90% for 2 hours, accompanied by a timeout in the business interface response". The associated context data includes structured database records such as the deployment configuration of the server, business access logs, and historical operation and maintenance records. The expected output format indication specifies that the output is in JSON format. Parse the problem description text in the original decision request stream for word segmentation, stop word removal, and stemming processing to generate a standardized token sequence. The word segmentation process divides the problem description text into word units. The stop word removal operation filters out common non-content words such as "de" (的), "le" (了). The stemming converts words to their base forms, for example, "accompanied by" (伴有) is reduced to "accompany" (伴). The standardized token sequence is represented as sequence S = [w1, w2,..., wn], where wi represents the i-th standardized token, and n is the length of the standardized token sequence. The standardized token sequence generated from the example text is ["cloud server", "ECS-08", "CPU", "usage rate", "90%", "continuous", "2", "hours", "accompany", "interface", "response timeout"]. A data comparison is made between the original problem description text and the standardized token sequence. The original text contains punctuation and redundant words, while the standardized token sequence removes stop words and stems the words.

[0074] In some embodiments, named entity recognition and relation extraction are performed on the standardized lexical sequence. Entities, attributes, and relationships between entities are extracted from the problem description text and associated context data to form key fact triples. Named entity recognition identifies entity types in the standardized lexical sequence, such as server, metric, time, and fault phenomenon. Relationship extraction analyzes relationships between entities, such as "utilization rate reached", "continuous", and "accompanying". Key fact triples extracted from the example text include (ECS-08, CPU utilization rate, above 90%), (ECS-08, abnormal duration, 2 hours), and (ECS-08, accompanying fault, interface response timeout). Key fact triples are supplemented from the associated context data, such as (ECS-08, deployment business, e-commerce transaction interface) and (ECS-08, configuration, 4 cores 8G). The set of key fact triples is represented as a structured list. Optionally, the core terms in the standardized lexical sequence are queried from the domain knowledge graph to obtain semantically related concepts, attributes, and rules, forming a pre-defined knowledge entry index. The core terms are selected from the standardized lexical sequence based on word frequency or semantic importance, such as "CPU usage is too high." The IT operations and maintenance domain knowledge graph is queried to obtain related concepts such as "process usage is too high" and "resource bottleneck," attributes such as "high CPU usage can easily lead to interface timeouts," and rules such as "CPU usage exceeding 80% for 1 hour requires immediate intervention and investigation." The pre-defined knowledge entry index includes concept references, attribute lists, and rule identifiers. The core term "CPU usage is too high" is compared with the knowledge graph entries to demonstrate the mapping from terms to knowledge entries.

[0075] In practice, standardized lexical sequences, key fact triples, and pre-defined knowledge item indexes are populated and combined according to a predefined structured query template to construct a structured decision query data structure. The structured query template defines fields including "problem statement," "fact set," and "knowledge index." Standardized lexical sequences are populated into the "problem statement" field, key fact triples into the "fact set" field, and pre-defined knowledge item indexes into the "knowledge index" field. The structured decision query object contains text data, factual data, and knowledge data. The original decision request stream and the structured decision query form a data comparison; the original stream is unstructured input, while the structured decision query is in a standardized format. This can be understood as attaching metadata to the structured decision query. The metadata includes at least a request identifier, a processing timestamp, and a version identifier used during semantic normalization. This metadata is attached to the structured decision query object as metadata fields. The request identifier tracks the original request, the processing timestamp records the processing time, and the version identifier indicates the processing algorithm version. This metadata enhances the traceability of the structured decision query.

[0076] See Figure 3In one embodiment of the present invention, a big language-based intelligent decision-making method involves calling a big language model inference engine and taking a structured decision query as input to obtain a preliminary decision reasoning chain. The structured decision query includes a semantically aligned core question statement such as "evaluating the cause and handling solution of high CPU usage anomaly in cloud server ECS-08", key fact triples such as (ECS-08, CPU utilization, above 90%), (ECS-08, deployment business, e-commerce transaction interface) and a pre-set knowledge entry index such as knowledge entry K001 pointing to "common causes and troubleshooting process of high CPU usage". The core question statement, key fact triples, and pre-defined knowledge item indexes from the structured decision query are constructed into a large language model prompt text according to a specific prompt template. This template includes role settings, background information injection, task instructions, and output format requirements. During construction, the core question statement serves as the main body of the task instructions, key fact triples are embedded with background information in list form, and pre-defined knowledge item indexes are inserted as reference markers. The discrete data of the structured decision query is compared with the coherent narrative of the large language model prompt text. The structured decision query is field-based data, while the large language model prompt text is a sequence of natural language instructions that conforms to the understanding habits of the large language model. This can be formally represented as:

[0077]

[0078] in: It is a text that expresses the core issue. It is a set of natural language descriptions of key fact triples. It is the knowledge summary text corresponding to the pre-defined knowledge entry index. It is a text of instructions requiring chain reasoning, with symbols. This indicates the sequential concatenation operation of text segments.

[0079] In some embodiments, the large language model prompt text is input to the large language model inference engine after instruction fine-tuning. The large language model inference engine is a neural network model pre-trained on a large-scale corpus and fine-tuned for multi-step inference tasks. The input process passes the large language model prompt text sequence to the input layer of the large language model inference engine through the application programming interface. Optionally, the large language model inference engine, based on the large language model's prompt text, employs chain-like thinking generation technology to progressively output inference steps, intermediate conclusions, and plans for subsequent steps, forming a preliminary inference chain composed of multiple inference step nodes. This chain-like thinking generation technology guides the model to explicitly generate the intermediate inference process. For the example scenario, the text output by the large language model inference engine might include: "Step 1: Identify the core anomaly: ECS-08 CPU usage above 90% for 2 hours, accompanied by e-commerce transaction interface response timeout. Step 2: Combining server configuration (4 cores 8GB) and deployed business, determine that resource load exceeds the threshold or process abnormal usage. Step 3: Based on knowledge K001, high CPU usage prioritizes investigating three causes: abnormal processes, sudden increases in business access volume, and unreasonable service configuration..." This output text is parsed into discrete inference step nodes. The preliminary inference chain contrasts with the direct answer output; the preliminary inference chain demonstrates the inference process, while the direct answer only provides the conclusion. In practice, for each inference step node in the initial inference chain, its reasoning basis is extracted. The reasoning basis points to a key fact triple or a specific item in the pre-set knowledge entry index. The extraction process identifies entities or knowledge identifiers mentioned in the description of the inference step node through text matching or dependency relationship analysis. For the inference step node "combining server configuration (4 cores 8G) and deployment business", the extracted reasoning basis is a reference to the key fact triple (ECS-08, configuration, 4 cores 8G) and (ECS-08, deployment business, e-commerce transaction interface). For the inference step node "based on knowledge K001", the extracted reasoning basis is a reference to the identifier K001 in the pre-set knowledge entry index. The content of the inference step node and the extracted reasoning basis form a data association.

[0080] It is understandable that the generation order, content, and corresponding reasoning basis of all reasoning step nodes in the initial reasoning chain are recorded and encapsulated together into a data structure for the initial decision reasoning chain. During the recording process, each reasoning step node is assigned a unique sequence number, its content is stored as a text string, and its reasoning basis is stored as a list of pointers to data items in the structured decision query. The data structure is serialized using JSON or Protocol Buffers format. The natural language text of the initial reasoning chain contrasts with the structured encapsulation of the initial decision reasoning chain. The natural language text is for human reading, while the structured encapsulation is for programmatic processing and subsequent verification.

[0081] In one embodiment of the present invention, consistency verification and completion operations are performed on the preliminary decision reasoning chain based on a predefined decision logic constraint graph. The preliminary decision reasoning chain includes multiple serial reasoning step nodes, such as "Step 1: Identify core anomalies" and "Step 2: Identify core anomalies". The decision logic constraint graph represents the decision elements, logical rules, and constraints within the domain in a graph structure. The predefined decision logic constraint graph corresponding to the problem domain is loaded and used in structured decision queries. For the IT operations and maintenance domain, the decision logic constraint graph includes graph nodes such as "server anomaly indicators", "server configuration", "deployment services", "troubleshooting rules", and "necessity of handling", as well as logical edges connecting these nodes, such as "priority investigation required", "strong reminder", and "related configuration". The graph is stored in the form of a file or database. The loading process reads the graph data into memory to form a queryable graph data structure. Traverse each reasoning step node in the initial decision reasoning chain, and match the decision elements involved in the reasoning step node with the nodes in the decision logic constraint graph. The matching operation is based on text similarity or a predefined element-node mapping table. For the reasoning step node "judged as resource load exceeding threshold or abnormal process usage", the extracted decision element "CPU high usage fault type judgment" is matched with the node "fault type assessment" in the decision logic constraint graph.

[0082] In some embodiments, for a successfully matched decision element, it is checked whether the logical rules applied by the inference step node are consistent with the rules associated with the decision element in the decision logic constraint graph. The consistency check is achieved by comparing the implicit or explicit logical judgments in the inference step node with the rule objects linked to the graph node. The node "fault type assessment" in the decision logic constraint graph is associated with the rule "must be combined with abnormal indicators, server configuration, and business access volume for comprehensive judgment". If the inference step node only mentions indicators and configurations but ignores business access volume, it is determined to be inconsistent. If an inconsistency is found, the content of the inference step node is replaced or corrected with consistent logical rules according to the decision logic constraint graph. The correction operation uses text generation technology to add missing elements to the node description, and the example node is corrected to "judged as resource load exceeding the threshold or abnormal process usage: combined with CPU utilization of over 90%, server 4-core 8G configuration, and e-commerce transaction interface business access volume for comprehensive assessment". The original node of the preliminary decision inference chain and the corrected node content form a data comparison. Optionally, if it is found that the preliminary decision-making reasoning chain is missing a decision element or reasoning branch that must be considered in the decision logic constraint graph, a new reasoning step node is generated based on the decision logic constraint graph and inserted into the appropriate position of the preliminary decision-making reasoning chain. The decision logic constraint graph stipulates that investigating "high CPU usage" must include the branch "investigating server host resource preemption". Since the preliminary chain is missing this branch, a new node "Generate investigation branch: It is necessary to assess the host resource usage status, because there is a possibility of host resource preemption on cloud servers" is generated based on the graph and inserted after the fault type judgment node.

[0083] In practical implementation, all the verified, corrected, and completed reasoning step nodes are reorganized into a tree structure according to their logical dependencies. The root node of the tree structure is the decision problem "Investigate high CPU usage of ECS-08," the leaf nodes are decision conclusions such as "It is recommended to first investigate abnormal processes, then check business access volume, and simultaneously check the host machine status," and the intermediate nodes are reasoning steps such as fault type judgment and investigation branch division. The reorganization process analyzes the conditional, causal, or temporal relationships between nodes to establish parent-child node connections. The linear sequence of the preliminary decision reasoning chain is compared with the hierarchical structure of the complete decision logic tree. The linear sequence represents a single path, while the hierarchical structure represents multi-branch reasoning. (Complete decision logic tree) This can be formally represented as:

[0084]

[0085] in: It is a set of nodes, containing the root node, intermediate nodes, and leaf nodes. It is a set of directed edges connecting nodes, representing the logical flow direction. It is a collection of content and information stored on each node.

[0086] Based on the complete decision logic tree, pattern matching is performed in the pre-defined executable atomic operation library to generate a decision action plan. The leaf nodes of the complete decision logic tree contain the decision conclusion "It is recommended to first check the abnormal process, then check the business access volume, and simultaneously check the host machine status." The intermediate nodes implicitly contain the decision intent "To obtain the server operation data required for troubleshooting." The complete decision logic tree is parsed to extract the decision intent and target state implied in all leaf nodes and key intermediate nodes. The operation traversal tree nodes are extracted and the node text is parsed using rules. For the leaf node "It is recommended to first check the abnormal process, then check the business access volume, and simultaneously check the host machine status," the extracted target states are "Query the ECS-08 process usage," "Statistics on e-commerce transaction interface access volume," and "Check the ECS-08 host machine resource usage status." The executable atomic operation library is traversed. The executable atomic operation library stores a variety of predefined basic operation units with clear input, output, and execution interfaces. A basic operation unit is, for example, "Execute server process query (input: server I..."). The operations library consists of three parts: "Query Instructions", "Statistics on Business Interface Access (Input: Interface ID, Statistical Time Range)" and "Query Cloud Server Host Status (Input: Server ID, Adapted to IT Operation and Maintenance Scenarios)". The library exists as code modules or service interfaces. Each decision intent or target state is semantically matched with the basic operation units in the executable atomic operation library. One or more basic operation units capable of implementing the function are found as candidates for each intent or state. Semantic matching is achieved by comparing the similarity between the text description and the operation unit's metadata. For example, the target state "Query ECS-08 Process Usage" matches the basic operation unit "Execute Server Process Query".

[0087] It is understandable that, based on the logical order and dependencies between nodes in the complete decision logic tree, the execution sequence and data flow of the selected basic operation units are determined. Dependency analysis shows that "statistical analysis of e-commerce transaction interface access volume" and "checking the resource usage status of ECS-08 host machine" can be executed in parallel. The operation "querying ECS-08 system load status" corresponding to the intermediate node "needs to confirm the basic running status of the server" should be executed before troubleshooting abnormal processes. All selected basic operation units are logically connected according to the determined execution sequence and data flow, and the specific input parameters of each basic operation unit are configured. The input parameters come from the context information of the corresponding node in the complete decision logic tree. For "execute server process query", the input parameters are configured as ECS-08 server ID and top-p query command, all derived from key fact triples and node content. The connection operation forms a directed acyclic graph to represent the dependencies between operations. The independent operation units before logical connection and the ordered operation graph after connection form a data comparison. The connected and configured sequence of basic operation units is encapsulated into an executable decision action plan. The decision action plan contains a series of ordered operation instructions and the data transfer relationships between the operation instructions. The encapsulation format is JSON or YAML script, which clearly lists the executor, parameters and subsequent operation triggering conditions for each operation.

[0088] In one embodiment of the present invention, an executable decision action plan is simulated and risk-assessed. The executable decision action plan includes a series of ordered operation instructions, such as "querying the ECS-08 system load status" and "executing server process queries." The decision action plan is encapsulated as a JSON script. A snapshot of the current system state is obtained. The system state snapshot contains the current values ​​of all environment variables and resource states related to the decision action plan. Environment variables include, for example, the service status of cloud platform operation and maintenance tools and the configuration of remote connection permissions for servers. Resource states include, for example, the status of the remote query interface of ECS-08 (available / unavailable). The system state snapshot is obtained by querying the real-time interfaces or databases of various IT business systems and is stored in the form of key-value pairs or objects. The real-time data of the system state snapshot is compared with the assumed state on which the decision action plan is based. In the isolated simulation environment, a system state snapshot is loaded as the initial state, and each operation instruction in the decision action plan is executed sequentially. The isolated simulation environment is a sandbox computing environment isolated from the real production system. The loading process maps the key-value pairs of the system state snapshot to the memory variables of the simulation environment. The logic of each operation instruction in the decision action plan is executed sequentially by the simulation engine, which parses and interprets it. No real server operations or resource consumption are generated during the simulation execution.

[0089] In some embodiments, the output results of each operation instruction, modifications to the state of the simulation environment, and simulated events triggered during the simulation execution are recorded in real time. Recorded output results include, for example, "ECS-08 system load query successful, current load average is 8.2"; recorded state modifications include, for example, "ECS-08 process query interface status changed from 'idle' to 'querying'"; recorded simulated events include, for example, "triggering 'remote query timeout' alarm". All records are timestamped and associated with the operation instruction ID. The final system state achieved by the simulation execution is compared with the preset expected target state in the complete decision logic tree to calculate the target achievement deviation. The expected target state originates from the leaf nodes of the complete decision logic tree, such as "Completed preliminary investigation and data collection for ECS-08 CPU high utilization fault", and is quantified into a series of sub-targets such as "Server process utilization data obtained", "Business interface access volume counted", and "Host status data obtained". Finally, the comparison between the simulated system state and the expected target state calculates the achievement degree of each sub-target and synthesizes the overall deviation value. The calculation can be expressed as:

[0090]

[0091] in: It is the number of sub-targets. It is the first The weight of each sub-objective, It is the first time after the simulation execution The state values ​​of each sub-target It is the first The expected state value of each sub-objective. It is a state matching function that outputs 1 when two state values ​​match, and 0 otherwise. The analysis process analyzes the data recorded during the simulation execution, identifies the sequences of operation instructions that lead to execution failure, resource conflicts, or violations of constraints, and marks them as potential risk points. The analysis process scans the recorded logs to find events where the operation output is "failure," resource state modification conflicts occur, or the operation triggers a business rule alarm. The identified risk instruction sequences and their contexts are extracted and labeled. See Table 1.

[0092] Table 1: Simulation Execution Records and Risk Identification Table

[0093] Operation Instruction ID Operation Instruction Description Simulation output results Simulation environment state modification Trigger simulated event Risk Marking OP001 Query ECS-08 system load status Success: Load average is 8.2 ECS-08 System Load Data Update none none OP002 Execute ECS-08 process query Success: Returns the top 10 processes by usage data. The ECS-08 process query interface is set to "Querying". none none OP003 Statistics on e-commerce transaction interface visits Failure: Log analysis tool service not started. none "Service unavailable" event yes OP004 Query ECS-08 host status Success: Returned host resource data Conflicts with the default rule "verify host machine permissions first" "Rule violation" alert yes

[0094] Based on the simulation results, the executable decision-making action plan is optimized and adjusted. According to the calculated deviation from the target achievement, the decision logic tree nodes or operational instructions in the decision-making action plan that cause the main deviation are located. This location is achieved through deviation contribution analysis, identifying the impact on the overall deviation. The sub-objective that contributes the most is identified, and the decision logic tree leaf node responsible for this sub-objective and the operation instructions that implement it are traced back. For the identified node or instruction, its upstream dependent nodes and reasoning basis in the complete decision logic tree are traversed back. For the operation instruction "statistics e-commerce transaction interface access volume", the tree is traced back to the leaf node "obtain business access volume data" in the decision logic tree, and further back to its parent node "investigate the cause of the sudden increase in business access volume". The reasoning basis supporting this node is found to be the key fact triple (ECS-08, deployment business, e-commerce transaction interface) and knowledge item K001. Based on the analysis results of potential risk points, the sufficiency of the relevant reasoning basis or the applicability of the relevant operation instructions are reassessed. The reassessment process checks whether the failure of the operation instruction "statistics e-commerce transaction interface access volume" is due to resource constraints (log analysis tool service not started), and the reasoning basis does not include tool service availability information. At the same time, it is checked whether the constraint rules violated by the operation instruction "query ECS-08 host status" were fully considered in the original decision logic.

[0095] In practical implementation, within the framework of the decision logic constraint graph, local re-reasoning is performed on decision logic tree nodes that cause deviations, or the corresponding operation instructions in the decision action plan are replaced, parameters are adjusted, or the order is optimized. Local re-reasoning targets the node "investigating the cause of a sudden increase in business access volume." Considering resource constraints within the decision logic constraint graph framework, alternative inspection solutions may be deduced, such as "directly querying e-commerce transaction interface access volume through the cloud platform monitoring panel." The operation instruction is replaced by changing the original instruction "statistically count e-commerce transaction interface access volume" to "query interface access volume from the cloud platform monitoring panel." Parameter adjustment may modify the target interface or time range of the query, and order optimization may advance the operation of verifying the service status of the tool. It can be understood that based on the re-reasoning results or the optimized operation instruction sequence, an updated decision action plan fragment is generated and used to replace the corresponding part in the original decision action plan, forming an optimized decision action plan. When generating the updated fragment, it is ensured that the new operation instructions are compatible with the data interfaces of the preceding and following instructions. The replacement operation finds the position of the original instruction sequence in the data structure of the decision action plan and overwrites it. The optimized decision action plan needs to be re-verified and tested in a simulation environment.

[0096] See Figure 4This is a system status monitoring trend chart simulating the entire process, primarily used for dynamic analysis of cloud platform resources and tool services in IT operations and maintenance decision-making scenarios. This chart fully recreates the dynamic changes in the status of cloud platform resources, operations and maintenance tool services, and server interfaces during IT operations and maintenance decision-making simulations. It can be used to pinpoint key time points of tool service unavailability and resource access conflicts, verify the resource and tool compatibility of the decision-making process, and provide data support for optimizing decision-making plans. Through linkage with preset alarm thresholds, it can automatically trigger warnings such as service unavailability and interface access timeouts during the simulation, helping decision-makers identify potential risks in advance. By transforming the abstract decision-making process into an intuitive dynamic trend of resource and service changes, it allows non-technical personnel such as IT operations and maintenance personnel and decision engineers to quickly understand the resource flow and tool call logic in the decision-making process, reducing cross-role communication costs.

[0097] In one embodiment of the present invention, an optimized decision-making action plan is executed and its actual effect is monitored. The optimized decision-making action plan includes an updated sequence of operation instructions, such as "querying interface access volume from the cloud platform monitoring panel." The decision-making action plan is simulated, deduced, and optimized. The optimized decision-making action plan is decomposed into discrete execution tasks and assigned to corresponding execution agents or execution engines. The decomposition process is based on the atomicity of the operation instructions and the execution subject. The operation instruction "querying interface access volume from the cloud platform monitoring panel" is decomposed into sub-tasks such as "calling the cloud platform monitoring API," "parseing access volume data," and "generating data statistical reports." The assignment process routes tasks according to the capability registry of the execution agent. For example, API call tasks are assigned to the company's cloud platform interface call engine, and operation and maintenance tasks are assigned to the server operation and maintenance execution agent. The unified plan before and after task assignment is compared with the discrete tasks. The execution of each task is triggered sequentially. Before each task begins, its input conditions and preset premises are verified to match the current actual system state. The triggering operation is initiated by the workflow engine based on task dependencies. The verification operation checks whether the task's input parameters, such as the server ID, are valid and whether the required tools and services, such as the monitoring API, are in place. If the verification fails, execution is paused and an alarm is issued. The actual system state is a real-time snapshot of the company's production environment and may differ from the simulated environment state.

[0098] In some embodiments, during task execution, execution logs, output results, and data on the impact of the execution process on the actual system state are collected in real time. Execution logs are collected by recording task start time, end time, and execution agent identifier through embedded points. Output results are collected, such as interface access statistics and the top 5 processes by occupancy. Status impact data is collected, such as the server query interface changing from "idle" to "in use." The collected data stream is synchronized with the task execution stream. After all tasks are completed, the actual execution output is summarized and compared with the preset expected target state in the complete decision logic tree. The summary operation aggregates the final output results of all sub-tasks to form a comprehensive report. The expected target state is derived from the leaf node definition of the decision logic tree, such as "Complete ECS-08 CPU high occupancy fault troubleshooting data collection and preliminary analysis." The comparison analysis calculates the matching degree between the actual achieved state and the expected state of each sub-target, and analyzes the reasons for deviations. It is understandable that the data collected during the actual execution process, the final output, and the expected comparative analysis results, along with the original decision request identifier, are stored in the historical decision case library. The stored procedure serializes and stores the execution log, output results, system status snapshot, comparative analysis report, and decision request identifier into the database to form a traceable historical decision case record. The structured storage of decision cases is compared with the transient data stream of the execution process.

[0099] The decision-making process is iteratively learned based on historical execution data. Decision cases are periodically extracted from the historical decision case library. Each case includes the original decision request, the generated decision logic tree, the decision action plan, the actual execution results, and the effect evaluation data. Extraction is scheduled periodically, such as weekly, and the extracted cases must cover various scenarios, including both successes and failures to meet expectations. The actual execution results and effect evaluation data from the decision cases serve as feedback signals to evaluate the quality of the decision logic tree and decision action plan. The evaluation is implemented through quantitative indicator calculations, culminating in a quality evaluation score. It can be defined as:

[0100]

[0101] in: It is the target alignment function, which measures the consistency between the actual result and the expected target. It is an execution efficiency function that measures resource consumption and time cost. It is a robustness function that measures tolerance to fluctuations in the execution environment. These are the weight coefficients of each function. For decision-making cases where the results do not meet expectations, analyze the reasoning flaws, knowledge gaps, or constraint violations in the decision-making logic tree, and analyze the operational errors or environmental mismatches in the decision-making action plan. The analysis process uses root cause analysis to trace the critical path from the decision-making logic to the execution failure, and identify nodes in the reasoning chain that incorrectly apply rules, missing key knowledge items, or operations in the action plan that do not adequately consider the dynamic nature of resources.

[0102] In practical implementation, the analysis results are used to selectively supplement, correct, or optimize the fine-tuning data, decision logic constraint graph, or executable atomic operation library of the large language model inference engine. For identified inference defects, erroneous inference cases and corrected correct inference chains are used as fine-tuning data pairs and added to the subsequent fine-tuning training set of the large language model inference engine. For identified knowledge gaps, the missing knowledge items and their relationships are added to the corresponding positions in the decision logic constraint graph. For identified improper operations, the risk conditions of the operation are marked in the executable atomic operation library, or alternative operation units are added. Optionally, the optimized model, graph, and operation library are applied to new decision request processing flows, forming a closed loop from decision to execution to feedback optimization. In the next IT operation and maintenance decision processing cycle, semantic normalization and knowledge enhancement processing will use the updated graph, the large language model inference engine will load the fine-tuned model parameters, and the assembly of the decision action plan will refer to the optimized operation library. The closed-loop feedback enables the system to continuously evolve based on historical IT operation and maintenance experience, improving the accuracy and efficiency of cloud server fault diagnosis, operation and maintenance scheduling, and other decisions.

[0103] See Figure 5 This is a dynamic stacked area chart of data collection volume during the decision-making and execution process, clearly showing the cumulative trends and percentage changes of different data types. By presenting the stacked area, the chart intuitively shows the proportion and dynamic changes of different data types, allowing for quick identification of core data sources and growth patterns, providing a basis for resource planning such as data storage and bandwidth allocation. The linear growth trend of the total collection volume verifies the stability of the decision-making and execution process; fluctuations or abnormal growth rates can promptly pinpoint execution bottlenecks or abnormal events. The approximately balanced proportions of the three data types reflect the rationality of the current data collection; for optimization, the collection frequency or granularity can be adjusted for data types with excessively high proportions or rapid growth. The approximately linear growth trend of the total collection volume directly reflects the stability of the decision-making and execution process. Fluctuations or interruptions in growth rates can quickly pinpoint execution bottlenecks, insufficient resources, or abnormal events, providing crucial evidence for operation and maintenance troubleshooting.

[0104] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An intelligent decision-making method based on a large model, characterized in that, Includes the following steps: Receive raw decision request streams from heterogeneous service sources, the raw decision request streams containing request identifiers, problem description text, associated context data, and expected output format indications; Semantic normalization and knowledge enhancement processing are performed on the original decision request stream to generate a structured decision query. The structured decision query includes a semantically aligned core question statement, key fact triples extracted from the associated context data, and an index of pre-set knowledge entries related to the question domain. The large language model inference engine is invoked, and the structured decision query is used as input to drive the large language model inference engine to generate a multi-step inference chain, thereby obtaining a preliminary decision inference chain. Based on a predefined decision logic constraint graph, consistency verification and completion operations are performed on the preliminary decision reasoning chain to generate a verified complete decision logic tree, specifically including: Load a predefined decision logic constraint graph corresponding to the problem domain in the structured decision query. The decision logic constraint graph represents the decision elements, logical rules and constraints in the domain in a graph structure. Traverse each reasoning step node in the preliminary decision reasoning chain and match the decision elements involved in the reasoning step node with the nodes in the decision logic constraint graph; For a successfully matched decision element, check whether the logical rules applied by the reasoning step node are consistent with the rules associated with the decision element in the decision logic constraint graph. If an inconsistency is found, the content of the inference step node is replaced or modified with consistent logical rules according to the decision logic constraint graph. If it is found that the preliminary decision reasoning chain is missing a decision element or reasoning branch that must be considered in the decision logic constraint graph, then a new reasoning step node is generated based on the decision logic constraint graph and inserted into the appropriate position of the preliminary decision reasoning chain. All the reasoning step nodes that have been verified, corrected and completed are reorganized into a tree structure according to their logical dependencies. The root node of the tree structure is the decision problem, the leaf nodes are the decision conclusions, and the intermediate nodes are the reasoning steps, forming the complete decision logic tree. Based on the complete decision logic tree, pattern matching is performed in a preset library of executable atomic operations to map and assemble an executable decision action plan, specifically as follows: The complete decision logic tree is analyzed to extract the decision intent and target state implied in all leaf nodes and key intermediate nodes; Traverse the executable atomic operation library, which stores a variety of predefined basic operation units with explicit inputs, outputs and execution interfaces; Each decision intention or target state is semantically matched with the basic operation units in the executable atomic operation library, and one or more basic operation units that can realize its function are found as candidates for each intention or state. Based on the logical order and dependencies between nodes in the complete decision logic tree, the execution timing and data flow between the selected basic operation units are determined. All selected basic operation units are logically connected according to the determined execution sequence and data flow direction, and specific input parameters for each basic operation unit are configured. The input parameters are derived from the context information of the corresponding node in the complete decision logic tree. The connected and configured sequence of basic operation units is encapsulated into the executable decision action plan, which includes a series of ordered operation instructions and the data transfer relationships between the operation instructions.

2. The intelligent decision-making method based on a large model according to claim 1, characterized in that, The step of performing semantic normalization and knowledge enhancement processing on the original decision request stream to generate a structured decision query specifically includes: The problem description text in the original decision request stream is parsed, and word segmentation, stop word removal, and stemming are performed to generate a standardized word sequence. Named entity recognition and relation extraction are performed on the standardized lexical sequence to extract entities, attributes and relationships between entities from the problem description text and associated context data, forming the key fact triples; By utilizing the core terms in the standardized word sequence, the domain knowledge graph is queried to obtain the concepts, attributes, and rules that are semantically associated with the core terms, thus forming the pre-set knowledge entry index; The standardized word sequence, the key fact triples, and the pre-defined knowledge item index are filled and combined according to a predefined structured query template to construct the data structure of the structured decision query. Add metadata to the structured decision query, the metadata including at least the request identifier, processing timestamp, and version identifier used in the semantic normalization process.

3. The intelligent decision-making method based on a large model according to claim 2, characterized in that, The process involves invoking the large language model inference engine, taking the structured decision query as input, and driving the engine to generate a multi-step inference chain to obtain a preliminary decision inference chain. Specifically: The core question statements, key fact triples, and pre-set knowledge item indexes in the structured decision query are constructed into a large language model prompt text according to a specific prompt template; The prompt text of the large language model is input into the large language model inference engine, which has been fine-tuned by instructions. The large language model reasoning engine, based on the prompt text of the large language model, uses chain-like thinking generation technology to gradually output reasoning steps, intermediate conclusions, and plans for subsequent steps, forming a preliminary reasoning chain composed of multiple reasoning step nodes. For each reasoning step node in the preliminary reasoning chain, its reasoning basis is extracted, and the reasoning basis points to the key fact triple or a specific item in the pre-set knowledge item index; The generation order, content, and corresponding reasoning basis of all reasoning step nodes in the preliminary reasoning chain are recorded and encapsulated together into the data structure of the preliminary decision reasoning chain.

4. The intelligent decision-making method based on a large model according to claim 3, characterized in that, It also includes steps for simulating and risk-assessing the executable decision-making action plan: Obtain a snapshot of the current system state, which contains the current values ​​of all environmental variables and resource states related to the decision action plan; In an isolated simulation environment, the system state snapshot is loaded as the initial state, and each operation instruction in the decision action plan is executed sequentially. It records in real time the output results of each operation command during the simulation execution, the modifications to the state of the simulation environment, and the simulation events triggered. The final system state achieved by the simulation is compared with the preset expected target state in the complete decision logic tree, and the target achievement deviation is calculated. Analyze the data recorded during the simulation execution to identify the sequence of operation instructions that leads to execution failure, resource conflicts, or violation of constraints, and mark them as potential risk points.

5. The intelligent decision-making method based on a large model according to claim 4, characterized in that, It also includes the step of optimizing and adjusting the executable decision-making action plan based on the simulation results: Based on the calculated deviation in target achievement, locate the decision logic tree node or operational instruction in the decision action plan that caused the main deviation. For the located node or instruction, backtrack and query its upstream dependent nodes and reasoning basis in the complete decision logic tree; Based on the analysis results of the potential risk points, reassess the sufficiency of the relevant reasoning or the applicability of the relevant operational instructions; Within the framework of the decision logic constraint graph, local re-reasoning is performed on the decision logic tree nodes that cause deviations, or the corresponding operation instructions in the decision action plan are replaced, parameters are adjusted, or sorting is optimized. Based on the re-inference results or the optimized sequence of operation instructions, an updated decision action plan fragment is generated and used to replace the corresponding part in the original decision action plan, thus forming an optimized decision action plan.

6. The intelligent decision-making method based on a large model according to claim 5, characterized in that, It also includes the steps of executing the optimized decision-making action plan and monitoring its actual effects: The optimized decision action plan is decomposed into discrete execution tasks, and the execution tasks are assigned to the corresponding execution agents or execution engines; The execution of each task is triggered sequentially, and before each task begins, its input conditions and preset premises are verified to match the current actual system state. During task execution, real-time data is collected on execution logs, output results, and the impact of the execution process on the actual system state. After all tasks are completed, the actual execution output is summarized and compared with the preset expected target state in the complete decision logic tree; The data collected during the actual execution process, the comparative analysis results of the final output and the expected results, along with the original decision request identifier, are stored in the historical decision case library.

7. The intelligent decision-making method based on a large model according to claim 6, characterized in that, It also includes the step of iteratively learning the decision-making process based on historical execution data: Decision cases are periodically extracted from the historical decision case library. The decision cases include the original decision request, the generated decision logic tree, the decision action plan, the actual execution results, and the effect evaluation data. The quality of the decision logic tree and decision action plan is evaluated by using the actual execution results and effect evaluation data in the decision case as feedback signals. For decision-making cases where the results did not meet expectations, analyze the reasoning flaws, knowledge gaps, or constraint violations in the decision-making logic tree, and analyze the operational inappropriateness or environmental mismatch in the decision-making action plan. Using the analysis results, targeted additions, corrections, or optimizations are made to the fine-tuning data of the large language model inference engine, the decision logic constraint graph, or the executable atomic operation library; The optimized model, graph, and operation library are applied to the new decision request processing flow, forming a closed loop from decision-making to execution and then to feedback optimization.

8. An intelligent decision-making system based on a large model, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent decision-making method based on a large model as described in any one of claims 1 to 7.

Images