A semantic arrangement system and method based on a large language model

Abstract

The application discloses a semantic arrangement system and method based on a large language model, and the system comprises an intention understanding module, a hybrid retrieval module and a dynamic arrangement engine module; the intention understanding module is used for obtaining a natural language instruction, performing semantic analysis on the natural language instruction through a large language model to obtain an analysis result, and performing verification to obtain a structured intention representation; the hybrid retrieval module is used for obtaining a hybrid retrieval rule, performing retrieval on the structured intention representation through the hybrid retrieval rule to obtain a target skill package, performing parameter binding on initial parameters of the structured intention representation according to the target skill package to obtain an executable business plan; and the dynamic arrangement engine module is used for performing topological construction on the executable business plan according to the target skill package to obtain a directed acyclic graph, sequentially calling the target skill package according to a topological order of the directed acyclic graph to obtain an execution result. The natural language instruction is analyzed, retrieved and arranged, and accurate response is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a semantic arrangement method, system, terminal, and computer-readable storage medium based on a large language model. Background Technology

[0002] Enterprise business systems (such as ERP (Enterprise Resource Planning), CRM (Customer Relationship Management), contract management, and financial and invoicing systems) typically drive business operations through graphical interfaces or fixed forms. Existing "natural language-driven business execution" solutions mainly employ intent recognition and static interface binding: user input is mapped to predefined APIs through keyword matching or intent classification, and then single-step or small-step calls are made. This type of solution is usable in scenarios such as simple queries and fixed template updates, but it still has shortcomings in complex business scenarios.

[0003] Traditional natural language-driven business execution solutions (such as keyword matching and intent classification binding to predefined APIs) suffer from limitations in semantic understanding, such as shallow matching, lack of complete modeling of business semantics, inability to dynamically orchestrate multi-step processes, and lack of compliance verification loops. This results in low efficiency in accurately parsing user intent and skill positioning in complex business scenarios, making it impossible to generate executable dynamic processes. This has become a problem that urgently needs to be solved. Summary of the Invention

[0004] The main objective of this invention is to provide a semantic orchestration method, system, terminal, and computer-readable storage medium based on a large language model, aiming to solve the problem that existing natural language-driven business execution methods are difficult to accurately parse user intent and have low efficiency in skill positioning, resulting in the inability to generate executable dynamic processes.

[0005] To achieve the above objectives, the present invention provides a semantic arrangement system based on a large language model, wherein the semantic arrangement method based on a large language model includes the following steps: The intent understanding module and the hybrid retrieval module are respectively connected to the dynamic orchestration engine module; The intent understanding module is used to acquire natural language instructions input by the user, perform semantic parsing on the natural language instructions through a large language model, obtain the parsing results, verify the integrity of the parsing results, and obtain a structured intent representation. The hybrid retrieval module is used to obtain hybrid retrieval rules, retrieve the structured intent representation through the hybrid retrieval rules to obtain the target skill package, and bind the initial parameters of the structured intent representation according to the target skill package to obtain an executable business plan; The dynamic orchestration engine module is used to construct the topology of the executable business plan based on the target skill package to obtain a directed acyclic graph, and to call the target skill package sequentially according to the topological order of the directed acyclic graph to obtain the execution result.

[0006] Optionally, in the semantic orchestration system based on a large language model, the intent understanding module includes a semantic preprocessing unit, a semantic parsing unit, and a structured intent generation unit, wherein the semantic preprocessing unit, the semantic parsing unit, and the structured intent generation unit are sequentially connected: The semantic preprocessing unit is used to obtain the natural language instructions input by the user, perform redundant word removal and standardization on the natural language instructions, and obtain the processed natural language instructions. The semantic parsing unit is used to perform semantic parsing on the processed natural language instructions using a large language model to obtain the parsing result; The structured intent generation unit is used to convert the parsing result according to the JSON format to obtain an initial structured intent representation, and to perform integrity verification on the initial structured intent representation according to preset verification rules to obtain a structured intent representation.

[0007] Optionally, in the semantic orchestration system based on a large language model, the semantic parsing unit includes a semantic recognition subunit, a data mapping subunit, and a parameter extraction subunit, wherein the semantic recognition subunit, the data mapping subunit, and the parameter extraction subunit are sequentially connected: The semantic recognition subunit is used to perform semantic recognition on the processed natural language instructions through a large language model to obtain the core verbs and target entities; The data mapping subunit is used to map the core verb and the target entity to obtain the operation type and the standard entity type; The parameter extraction subunit is used to extract key parameters from the operation type and the standard entity type to obtain the parsing results.

[0008] Optionally, in the semantic orchestration system based on a large language model, the hybrid retrieval rules include vector retrieval and rule retrieval.

[0009] Optionally, in the semantic orchestration system based on a large language model, the target skill package includes a first target skill package, a second target skill package, and a third target skill package.

[0010] Optionally, in the semantic orchestration system based on a large language model, the hybrid retrieval module includes a user first target skill pack matching unit, a second target skill pack matching unit, a weighted calculation unit, and a data integration unit, wherein the first target skill pack matching unit, the second target skill pack matching unit, the weighted calculation unit, and the data integration unit are connected sequentially. The first target skill pack matching unit is used to obtain vector retrieval and rule retrieval, and retrieve the structured intent representation through the embedding model and the vector retrieval to obtain the first target skill pack; The second target skill pack matching unit is used to match the structured intent representation according to the entity type matching strategy, behavior type matching strategy and parameter signature matching strategy retrieved by the rules, so as to obtain the second target skill pack; The weighted calculation unit is used to perform weighted processing on the first target skill pack and the second target skill pack to obtain the third target skill pack; The data integration unit is used to map, complete context, and perform follow-up processing on the initial parameters of the structured intent representation based on the first target skill package, the second target skill package, and the third target skill package to obtain an executable business plan.

[0011] Optionally, in the semantic arrangement method based on a large language model, the data integration unit includes a parameter mapping subunit, a parameter completion subunit, a data probing subunit, and a data arrangement subunit, wherein the parameter mapping subunit, parameter completion subunit, data probing subunit, and data arrangement subunit are sequentially connected: The parameter mapping subunit is used to perform direct mapping and semantic mapping on the initial parameters of the structured intent representation based on the first target skill pack, the second target skill pack, and the third target skill pack to obtain the target semantics; The parameter completion subunit is used to perform context completion on the initial parameters based on the target skill pack to obtain the target parameters; The data follow-up subunit is used to obtain natural language follow-up statements, and to follow up on the initial parameters based on the natural language follow-up statements to obtain the target ratio; The data orchestration subunit is used to perform semantic parsing on the target semantics, the target parameters, and the target proportion to obtain structured data, and to orchestrate the structured intent representation based on the structured data to obtain an executable business plan.

[0012] Optionally, in the semantic orchestration method based on a large language model, the dynamic orchestration engine module includes a process parsing unit, a topology construction unit, and a destructured processing unit, wherein the process parsing unit, the topology construction unit, and the destructured processing unit are connected sequentially. The process parsing unit is used to perform process parsing on the executable business plan based on the target skill package to obtain the contract number, master record, clause record and related record; The topology construction unit is used to obtain the node types and edge relationships of the executable business plan, and to construct the topology of the contract number, the master record, the clause record and the associated record according to the node types and edge relationships to obtain a directed acyclic graph; The inverse structuring unit is used to sequentially call the target skill pack according to the topological order of the directed acyclic graph to obtain a structured response, and then perform inverse semantic restoration on the structured response to obtain the execution result.

[0013] Furthermore, to achieve the above objectives, the present invention also provides a semantic arrangement method based on a large language model for a semantic arrangement system, wherein the semantic arrangement method based on the large language model includes: The intent understanding module is used to acquire natural language instructions input by the user, perform semantic parsing on the natural language instructions through a large language model, obtain the parsing results, verify the integrity of the parsing results, and obtain a structured intent representation. The hybrid retrieval module is used to obtain hybrid retrieval rules, retrieve the structured intent representation through the hybrid retrieval rules to obtain the target skill package, and bind the initial parameters of the structured intent representation according to the target skill package to obtain an executable business plan; The dynamic orchestration engine module is used to construct the topology of the executable business plan based on the target skill package to obtain a directed acyclic graph, and to call the target skill package sequentially according to the topological order of the directed acyclic graph to obtain the execution result.

[0014] The semantic orchestration method based on a large language model, wherein the dynamic orchestration engine module constructs a topology for the executable business plan based on the target skill package to obtain a directed acyclic graph, and sequentially calls the target skill package according to the topological order of the directed acyclic graph to obtain the execution result, and then further includes: Determine whether the execution result is a successful test; Obtain historical call records. If the execution result is a pass, determine triples based on the data of the directed acyclic graph. Update the weights of the historical call records based on the triples to obtain the target weights. The target weight is used to improve the accuracy of target skill pack matching.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a semantic orchestration program based on a large language model, and the semantic orchestration program based on the large language model, when executed by a processor, implements the steps of the semantic orchestration method based on the large language model as described above.

[0016] In this invention, the semantic orchestration system based on a large language model includes: an intent understanding module, a hybrid retrieval module, and a dynamic orchestration engine module. The intent understanding module acquires natural language commands input by the user, performs semantic parsing on the natural language commands using a large language model, obtains the parsing results, verifies the completeness of the parsing results, and obtains a structured intent representation. The hybrid retrieval module acquires hybrid retrieval rules, retrieves the structured intent representation using the hybrid retrieval rules, obtains a target skill package, and binds the initial parameters of the structured intent representation according to the target skill package to obtain an executable business plan. The dynamic orchestration engine module constructs a topology for the executable business plan based on the target skill package, obtains a directed acyclic graph (DAG), and sequentially calls the target skill package according to the topological order of the DAG to obtain the execution result. This invention achieves accurate response by parsing, retrieving, and orchestrating natural language commands. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the principle of the semantic arrangement system based on a large language model according to the present invention. Figure 2 This is a schematic diagram of the static structure of a preferred embodiment of the semantic arrangement system based on a large language model according to the present invention; Figure 3 This is a schematic diagram illustrating another specific principle of the semantic arrangement system based on a large language model according to the present invention; Figure 4 This is a schematic diagram of the dynamic process of a preferred embodiment of the semantic arrangement system based on a large language model of the present invention; Figure 5 This is a schematic diagram of the processing flow in a preferred embodiment of the semantic arrangement method based on a large language model in the semantic arrangement system of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, 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 of the invention and are not intended to limit the invention.

[0019] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0020] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0021] One embodiment of the semantic orchestration system based on a large language model according to a preferred embodiment of the present invention includes, as follows: Figure 1 and Figure 2 As shown, the semantic orchestration system based on a large language model includes: an intent understanding module 10, a hybrid retrieval module 20, and a dynamic orchestration engine module 30; the intent understanding module 10 and the hybrid retrieval module 20 are respectively connected to the dynamic orchestration engine module 30.

[0022] The intent understanding module 10 is used to acquire natural language commands input by the user, perform semantic parsing on the natural language commands using a large language model, obtain parsing results, verify the integrity of the parsing results, and obtain a structured intent representation. The hybrid retrieval module 20 is used to acquire hybrid retrieval rules, retrieve the structured intent representation using the hybrid retrieval rules, obtain a target skill package, bind the initial parameters of the structured intent representation according to the target skill package, and obtain an executable business plan. The dynamic orchestration engine module 30 is used to construct a topology for the executable business plan based on the target skill package, obtain a directed acyclic graph, and sequentially call the target skill package according to the topological order of the directed acyclic graph to obtain an execution result.

[0023] For example, such as Figure 2As shown, the user inputs a natural language command: "Help me create a contract with a test client, totaling X thousand yuan, payable in 3 installments." The system uses the intent understanding module to call a large language model to perform deep semantic analysis on this statement, identifying the user's business operation intent as "create," the target business object as "contract," and the specific action as "create contract" (Contract.Create). The system further extracts key parameter information from the statement: the client name is "test client," the total contract amount is 5,000,000 yuan, and the payment installments are 3. This intent does not contain explicit business constraints. Simultaneously, the system automatically injects context information, including the current user ID as "U12345" and the request timestamp as "2026-01-23T10:30:00Z." Finally, the above semantic analysis results are structured into a standard JSON format intent representation, serving as the input for subsequent mixed retrieval, parameter binding, and execution orchestration, realizing the conversion from fuzzy natural language to precise machine-understandable instructions.

[0024] like Figure 3 and Figure 4 As shown, another specific embodiment of the semantic orchestration system based on a large language model in this invention includes: an intent understanding module 10, a hybrid retrieval module 20, and a dynamic orchestration engine module 30; the intent understanding module 10 and the hybrid retrieval module 20 are respectively connected to the dynamic orchestration engine module 30.

[0025] Specifically, the intent understanding module 10 includes a semantic preprocessing unit 101, a semantic parsing unit 102, and a structured intent generation unit 103. The semantic preprocessing unit 101 and the semantic parsing unit 102 are connected in sequence. The module acquires the natural language instructions input by the user, performs redundant word removal and standardization processing on the natural language instructions (receiving the original text input and performing preprocessing (removing redundant words and standardizing numerical expressions)) to obtain the processed natural language instructions, performs semantic parsing on the processed natural language instructions through a large language model to obtain the parsing result, converts the parsing result according to the JSON format to obtain the initial structured intent representation, and performs integrity verification on the initial structured intent representation according to preset verification rules to obtain the structured intent representation (the integrity verification of the structured intent identifies missing necessary parameters).

[0026] For example, the core verb "Create" is identified as the operation type "create", the target entity "Contract" is identified as the standard entity type "Contract", and the key parameters "Test Customer" is extracted as "customer_name" and "5 million" are extracted as the numerical value standardized to 5,000,000, with the unit automatically identified as RMB. The core verb "Query" is identified as the operation type "query", the target entity "Contract" is identified as the standard entity "Contract", and the behavior identifier is "Contract.Query" (predefined according to the skill package library). The core verb "Issue Invoice" is identified as the operation type "create", the target entity "Invoice" is identified as the standard entity "Invoice", and the behavior identifier is "Invoice.Create" (predefined according to the skill package library).

[0027] Specifically, the semantic parsing unit 102 includes a semantic recognition subunit 1021, a data mapping subunit 1022, and a parameter extraction subunit 1023, wherein the semantic recognition subunit 1021, the data mapping subunit 1022, and the parameter extraction subunit 1023 are connected sequentially. The processed natural language instruction is semantically recognized through a large language model to obtain the core verb and the target entity. The core verb and the target entity are mapped to obtain the operation type and the standard entity type. Key parameters are extracted from the operation type and the standard entity type to obtain the parsing result.

[0028] In this embodiment, the core verb is "query" → operation type query, the target entity is "contract" → mapped to the standard entity Contract, the behavior identifier is Contract.Query (predefined according to the skill package library), and the key parameters are extracted as follows: customer_name: "test customer", time_range: "last year" → needs to be converted to a specific date, amount_filter: "over 1 million" → parsed as total_amount>1000000. Phase 2: Intent structuring, converting natural language into a structured intent representation in a unified JSON format.

[0029] Specifically, the hybrid retrieval module 20 includes a first target skill package matching unit 201, a second target skill package matching unit 202, a weighted calculation unit 203, and a data integration unit 204. The first target skill package matching unit 201, the second target skill package matching unit 202, the weighted calculation unit 203, and the data integration unit 204 are sequentially connected. Vector retrieval and rule retrieval are obtained, and the structured intent representation is retrieved using an embedding model and the vector retrieval (using a pre-trained embedding model (e.g., text-embedding-ada-002) to generate query vectors, performing approximate nearest neighbor search (ANN) in a vector database (e.g., Milvus or Faiss) to return Top-K candidate skill packages). The first target skill package is obtained (ContractService.Create corresponds to the contract creation function, and OrderService.Create corresponds to the order creation function). The function InvoiceService.Create corresponds to the invoice creation function, and the first target skill package is the encapsulated function / interface. Based on the entity type matching strategy retrieved according to the rules (entity type matching target_entity=="Contract" to only ContractService.Create fully matching), behavior type matching strategy (action_type=="Create"), and parameter signature matching strategy (whether total_amount and payment_terms_count parameters are supported), the structured intent representation is matched to obtain the second target skill package. The first and second target skill packages are weighted to obtain the third target skill package. Based on the first, second, and third target skill packages, the initial parameters of the structured intent representation are mapped, context-completed, and followed up to obtain an executable business plan.

[0030] In this embodiment, the weighting formula is as follows: ; in, Represents semantic weight, Indicates the verification weight of the identifier. This represents historical weights; if semantic matching errors are frequently found, the weights are reduced. If a more reliable rule match is found, then improve... If a skill set consistently performs well, then its stats will be increased. Influence, VectorSimilarity, calculation method: Using a pre-trained embedding model (such as BERT, text-embedding-ada-002) to transform user intent and skill package description into high-dimensional vectors, and calculating cosine similarity. For example, user input: "Create a new sales order", skill package description: "Used to generate new sales documents", vector similarity: 0.87. RuleMatchScore, calculation method: Based on structured fields, perform precise or fuzzy matching, including: entity type matching (such as target_entity=="Order"), behavior type matching (such as action_type=="Create"), parameter signature matching (such as input parameters containing amount, customer_id). For example, complete match: 1.0, partial match (such as matching only the entity but not the behavior): 0.5~0.7, no match: 0.0. HistoricalWeight, calculation method: Dynamically calculated based on historical execution data, for example: .

[0031] Specifically, the data integration unit 204 includes a parameter mapping subunit 2041, a parameter completion subunit 2042, a data probing subunit 2043, and a data arrangement subunit 2044. The parameter mapping subunit 2041, parameter completion subunit 2042, data probing subunit 2043, and data arrangement subunit 2044 are sequentially connected. Based on the first target skill package, the second target skill package, and the third target skill package, the initial parameters of the structured intent representation are directly mapped and semantically mapped (direct mapping: total_amount ← 5000000, payment_terms_count ← 3; semantic mapping: ...) `customer_name`: "Test Customer" → needs to be converted to `customer_id` (skill package requires foreign key reference). Call LLM to determine: "Customer Name" and "customer_name" are semantically equivalent → trigger database query), obtain the target semantics, perform context completion on the initial parameters according to the target skill package to obtain the target parameters, obtain natural language follow-up statements, and follow up on the initial parameters according to the natural language follow-up statements to obtain the target ratio. Perform semantic parsing on the target semantics, the target parameters, and the target ratio to obtain structured data. Arrange the structured intent representation according to the structured data to obtain an executable business plan.

[0032] In this embodiment, context auto-completion is used: created_by ← current user U12345 (extracted from context), created_at ← 2026-01-23T10:30:00Z (system timestamp), contract_status ← default value "draft" (from skill pack definition). Furthermore, executing an SQL query: SQL, will directly auto-completion is used.

[0033] As an example, in the target percentage generation process, the system generates a natural language follow-up question: "Please provide the specific percentage for each payment period, for example: return to the user, waiting for supplementary input, "First period 30%, Second period 30%, Third period 40%", after which the system enters the parameter completion process.

[0034] Specifically, the dynamic orchestration engine module 30 includes a process parsing unit 301, a topology construction unit 302, and a destructured processing unit 303. The process parsing unit 301, topology construction unit 302, and destructured processing unit 303 are sequentially connected. Based on the target skill package, the executable business plan is parsed (the user-supplemented content is parsed into structured data) to obtain the contract number, master record, clause record, and associated record. The node types and edge relationships of the executable business plan are obtained. Based on the node types and edge relationships, the contract number, master record, clause record, and associated record are processed (contract number generated automatically, contract master record created (Contract.Create), i.e., the master record, payment...). The topology of the payment term record (PaymentTerm.BatchCreate, i.e., the term record, which is associated with the contract and payment terms, i.e., the associated record) is constructed to obtain a directed acyclic graph. The target skill pack is called sequentially according to the topological order of the directed acyclic graph to obtain a structured response. The structured response is then reverse-semantically restored to obtain the execution result (3 contracts that meet the conditions are found: 1. CT-20250315-001, signing date 2025-03-15, total amount 1.5 million; 2. CT-20250708-002, signing date 2025-07-08, total amount 2.8 million; 3. CT-20251120-003, signing date 2025-11-20, total amount 5 million).

[0035] Furthermore, the dynamic orchestration engine calls the rule check triggers: Rule_PaymentTermSumCheck, rule type: validation rule, trigger condition: when the contract is created, validation logic: SUM(payment_terms.percentage)==100, current value: 30+30+40=100, then pass; Rule_AmountPositive, validation logic: total_amount>0, current value: 5,000,000>0, then pass; Rule_CustomerActive, checks whether the customer status is active, query result: status='active', then pass. All preceding rules pass, and execution is allowed.

[0036] In this embodiment, all operations are incorporated into a single database transaction; a rollback point is set, and the entire system rolls back upon any failure. Service interfaces are called sequentially according to the topology: NumberGeneratorService is called to generate a serial number; ContractService.create() is called to insert data into the main table; PaymentTermService.batchCreate() is called to insert installment records; and RelationService.link() is called to establish a relationship. Upon successful execution, the system generates a structured response and returns it to the user in natural language: "Contract CT-20260123-001 has been successfully created. The client is a test client, the total amount is 5 million yuan, payable in three installments (30% / 30% / 40%), and the current status is draft."

[0037] Furthermore, based on Figure 1 The semantic arrangement system based on a large language model shown in the present invention, and the semantic arrangement method based on a large language model in the preferred embodiment of the present invention, are as follows: Figure 5 As shown, the semantic arrangement method based on a large language model includes the following steps: Step S10: The intent understanding module obtains the natural language instruction input by the user, performs semantic parsing on the natural language instruction through a large language model, obtains the parsing result, verifies the completeness of the parsing result, and obtains a structured intent representation.

[0038] Step S20: The hybrid retrieval module obtains hybrid retrieval rules, retrieves the structured intent representation through the hybrid retrieval rules to obtain a target skill package, and binds the initial parameters of the structured intent representation according to the target skill package to obtain an executable business plan.

[0039] Step S30: The dynamic orchestration engine module constructs a topology for the executable business plan based on the target skill package to obtain a directed acyclic graph. The target skill package is then called sequentially according to the topological order of the directed acyclic graph to obtain the execution result.

[0040] Furthermore, after step S30, the method further includes determining whether the execution result is a successful detection; obtaining historical call records; if the execution result is a successful detection, determining triples based on the data of the directed acyclic graph, updating the weights of the historical call records based on the triples, and obtaining the target weight; wherein, the target weight is used to improve the accuracy of target skill pack matching.

[0041] For example, the check results are as follows: Payment stage 1 amount: 1,500,000, invoice amount: 1,600,000. Result: mismatch. If the verification fails, the system will interrupt execution and return a structured error message: "Operation failed: Invoice amount (1.6 million) must equal the first installment payment amount of the contract (1.5 million). Please modify the amount or contact the administrator." The system automatically records this rule error event, forming a structured log, stored in a dedicated table `rule_conflict_log`, with an index: (rule_id, contract_id, payment_stage), for subsequent trend analysis and rule optimization. Querying historical similar errors: Result: Occurred 5 times, current cumulative error count: 5 times. System threshold: When the cumulative number of similar rule errors is ≥10 times → automatic generation of optimization suggestions. Currently, it is the 5th time → no formal suggestions are generated, only marked as "high-frequency error". Although the threshold has not been reached in this stage, the system has begun to pay attention to the adaptability of this rule.

[0042] Furthermore, assuming similar errors continue to occur, accumulating to 10 times, the system will generate formal optimization suggestions. Problem analysis: The following scenarios exist in actual business: invoicing in batches: the amount of a single invoice is less than the amount payable; allowable error: slight differences due to rounding or exchange rate fluctuations; reissue invoices: reissue historical differences. In the past three months, similar errors have occurred 10 times, mainly concentrated in the finance department and sales department. Users generally report that "the system is too rigid".

[0043] As examples, Option 1 allows a ≤ relationship, supports batch invoicing, and the cumulative invoiced amount cannot exceed the total amount payable in the current period. Option 2 introduces a tolerance mechanism, allowing reasonable errors. For example, an error of ±75,000 is allowed for 1.5 million, meaning that 1.425 million to 1.575 million can pass the verification. Option 3 uses conditional rule triggering, modifying the trigger condition to: IF contract.type=='standard' THEN strict match, ELSE allows flexible handling. Subsequent operation suggestions: It is recommended to adjust this rule to "Option 1" or "Option 2". The new rule logic can be verified in a test environment first. After updating the rule, the system will automatically reload the rule configuration and retain the old version for rollback.

[0044] Furthermore, automatic rule updates (authorization required) are implemented. Administrators log in to the system, review optimization suggestions, and if approved, select "Option Two": Introduce a 5% tolerance, implement the rule update process, generate a new version rule file `Rule_InvoiceAmountExact_v2.json`, and submit it to the rule configuration center; trigger a hot-reload mechanism to dynamically update the runtime rule engine; the version control system records the change: Old version: v1.0 (exact match), New version: v2.0 (5% tolerance), Operator: system / optimizer (authorized by the administrator), Time: 2026-02-01T09:00:00Z, Subsequent similar requests (invoices of 1.6 million); 1.6 million vs 1.5 million → difference of 100,000, tolerance limit: 1.5 million × 5% = 75,000, 100,000 > 75,000 → still not approved, if invoices of 1.57 million (difference 70,000 < 75,000) → approved, → system flexibility is significantly improved while retaining necessary controls.

[0045] In this embodiment, no objection was raised by the user → recorded as "successful execution", execution time was 412ms, within the expected range, and the triple was recorded: (Intent: "Create Contract + Customer + Amount + Installment", Skill Package: ContractService.Create, Result: success). Based on the triple, HistoricalWeight was updated: increased from 0.88 to 0.89, added to the successful sample library for subsequent retrieval model training and rule error monitoring. No rule failure was generated this time, no optimization suggestions were generated, the cumulative number of successful cases increased, and the system's confidence in matching this type of request was enhanced.

[0046] In summary, this invention provides a semantic orchestration system and method based on a large language model. The method includes: acquiring natural language commands input by a user; performing semantic parsing on the natural language commands using a large language model to obtain parsing results; verifying the completeness of the parsing results to obtain a structured intent representation; acquiring hybrid retrieval rules; retrieving the structured intent representation using the hybrid retrieval rules to obtain a target skill package; binding initial parameters of the structured intent representation according to the target skill package to obtain an executable business plan; constructing a topology for the executable business plan based on the target skill package to obtain a directed acyclic graph (DAG); and sequentially invoking the target skill package according to the topological order of the DAG to obtain an execution result. This invention achieves accurate response by parsing, retrieving, and orchestrating natural language commands.

[0047] 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, method, article, or terminal system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal system that includes that element.

[0048] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0049] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A semantic arrangement system based on a large language model, characterized in that, The semantic orchestration system based on a large language model includes: an intent understanding module, a hybrid retrieval module, and a dynamic orchestration engine module; The intent understanding module and the hybrid retrieval module are respectively connected to the dynamic orchestration engine module; The intent understanding module is used to acquire natural language instructions input by the user, perform semantic parsing on the natural language instructions through a large language model, obtain the parsing results, verify the integrity of the parsing results, and obtain a structured intent representation. The hybrid retrieval module is used to obtain hybrid retrieval rules, retrieve the structured intent representation through the hybrid retrieval rules to obtain the target skill package, and bind the initial parameters of the structured intent representation according to the target skill package to obtain an executable business plan; The dynamic orchestration engine module is used to construct the topology of the executable business plan based on the target skill package to obtain a directed acyclic graph, and to call the target skill package sequentially according to the topological order of the directed acyclic graph to obtain the execution result.

2. The semantic arrangement system based on a large language model according to claim 1, characterized in that, The intent understanding module includes a semantic preprocessing unit, a semantic parsing unit, and a structured intent generation unit, wherein the semantic preprocessing unit, the semantic parsing unit, and the structured intent generation unit are connected sequentially. The semantic preprocessing unit is used to obtain the natural language instructions input by the user, perform redundant word removal and standardization on the natural language instructions, and obtain the processed natural language instructions. The semantic parsing unit is used to perform semantic parsing on the processed natural language instructions using a large language model to obtain the parsing result; The structured intent generation unit is used to convert the parsing result according to the JSON format to obtain an initial structured intent representation, and to perform integrity verification on the initial structured intent representation according to preset verification rules to obtain a structured intent representation.

3. The semantic arrangement system based on a large language model according to claim 2, characterized in that, The semantic parsing unit includes a semantic recognition subunit, a data mapping subunit, and a parameter extraction subunit, wherein the semantic recognition subunit, the data mapping subunit, and the parameter extraction subunit are connected sequentially: The semantic recognition subunit is used to perform semantic recognition on the processed natural language instructions through a large language model to obtain the core verbs and target entities; The data mapping subunit is used to map the core verb and the target entity to obtain the operation type and the standard entity type; The parameter extraction subunit is used to extract key parameters from the operation type and the standard entity type to obtain the parsing results.

4. The semantic arrangement system based on a large language model according to claim 1, characterized in that, The hybrid retrieval rules include vector retrieval and rule retrieval.

5. The semantic arrangement system based on a large language model according to claim 4, characterized in that, The target skill pack includes a first target skill pack, a second target skill pack, and a third target skill pack.

6. The semantic arrangement system based on a large language model according to claim 5, characterized in that, The hybrid retrieval module includes a user first target skill pack matching unit, a second target skill pack matching unit, a weighted calculation unit, and a data integration unit, wherein the first target skill pack matching unit, the second target skill pack matching unit, the weighted calculation unit, and the data integration unit are connected sequentially. The first target skill pack matching unit is used to obtain vector retrieval and rule retrieval, and retrieve the structured intent representation through the embedding model and the vector retrieval to obtain the first target skill pack; The second target skill pack matching unit is used to match the structured intent representation according to the entity type matching strategy, behavior type matching strategy and parameter signature matching strategy retrieved by the rules, so as to obtain the second target skill pack; The weighted calculation unit is used to perform weighted processing on the first target skill pack and the second target skill pack to obtain the third target skill pack; The data integration unit is used to map, complete context, and perform follow-up processing on the initial parameters of the structured intent representation based on the first target skill package, the second target skill package, and the third target skill package to obtain an executable business plan.

7. The semantic arrangement system based on a large language model according to claim 6, characterized in that, The data integration unit includes a parameter mapping subunit, a parameter completion subunit, a data probing subunit, and a data arrangement subunit, wherein the parameter mapping subunit, parameter completion subunit, data probing subunit, and data arrangement subunit are connected sequentially: The parameter mapping subunit is used to perform direct mapping and semantic mapping on the initial parameters of the structured intent representation based on the first target skill pack, the second target skill pack, and the third target skill pack to obtain the target semantics; The parameter completion subunit is used to perform context completion on the initial parameters based on the target skill pack to obtain the target parameters; The data follow-up subunit is used to obtain natural language follow-up statements, and to follow up on the initial parameters based on the natural language follow-up statements to obtain the target ratio; The data orchestration subunit is used to perform semantic parsing on the target semantics, the target parameters, and the target proportion to obtain structured data, and to orchestrate the structured intent representation based on the structured data to obtain an executable business plan.

8. The semantic arrangement system based on a large language model according to claim 7, characterized in that, The dynamic orchestration engine module includes a process parsing unit, a topology construction unit, and a destructured processing unit, wherein the process parsing unit, the topology construction unit, and the destructured processing unit are connected sequentially. The process parsing unit is used to perform process parsing on the executable business plan based on the target skill package to obtain the contract number, master record, clause record and related record; The topology construction unit is used to obtain the node types and edge relationships of the executable business plan, and to construct the topology of the contract number, the master record, the clause record and the associated record according to the node types and edge relationships to obtain a directed acyclic graph; The inverse structuring unit is used to sequentially call the target skill pack according to the topological order of the directed acyclic graph to obtain a structured response, and then perform inverse semantic restoration on the structured response to obtain the execution result.

9. A semantic arrangement method based on a large language model, based on the semantic arrangement system based on any one of claims 1-8, characterized in that, The semantic arrangement method based on a large language model includes: The intent understanding module acquires the natural language instructions input by the user, performs semantic parsing on the natural language instructions through a large language model, obtains the parsing results, verifies the completeness of the parsing results, and obtains a structured intent representation. The hybrid retrieval module acquires hybrid retrieval rules, retrieves the structured intent representation using the hybrid retrieval rules to obtain a target skill package, and binds the initial parameters of the structured intent representation according to the target skill package to obtain an executable business plan. The dynamic orchestration engine module constructs a topology for the executable business plan based on the target skill package to obtain a directed acyclic graph. The target skill package is then called sequentially according to the topological order of the directed acyclic graph to obtain the execution result.

10. The semantic arrangement method based on a large language model according to claim 9, characterized in that, The dynamic orchestration engine module constructs a topology for the executable business plan based on the target skill package, obtaining a directed acyclic graph (DAG). It then sequentially calls the target skill package according to the topological order of the DAG to obtain the execution result. The module further includes: Determine whether the execution result is a successful test; Obtain historical call records. If the execution result is a pass, determine triples based on the data of the directed acyclic graph. Update the weights of the historical call records based on the triples to obtain the target weights. The target weight is used to improve the accuracy of target skill pack matching.

Images