An import tariff superposition calculation method based on certainty lookup table and large language model synthesis

Abstract

This invention belongs to the field of tariff calculation technology, specifically involving a method for superimposed calculation of import tariffs based on a combination of deterministic table lookup and a large language model. The deterministic regulatory table lookup steps are as follows: calling a unified regulatory constant source, sequentially executing HTSUS basic tariff rate lookup, Section 122 Annex II exemption verification, anti-dumping / countervailing duty (AD / CVD) tariff rate lookup, and Chapter 99 footnote information extraction, outputting the original tax calculation facts data for all tax types; the regulatory constant source centrally stores all tariff regulations, tariff rates, exemption lists, and surcharge constant data. This invention delegates the deterministic numerical logic such as numerical query, tariff rate determination, and exemption verification to the program for table lookup. All tariff rate and exemption determination results are output by the unified regulatory constant source, without relying on a large language model for numerical calculations throughout the process; all original table lookup data, priority rules, and model output itemized tariffs are completely retained, and the calculation process can be traced and verified throughout the entire chain, meeting the customs compliance declaration verification requirements.

Description

Technical Field

[0001] This invention belongs to the field of tariff calculation technology, specifically relating to a method for superimposing import tariffs based on a combination of deterministic table lookup and large language model. Background Technology

[0002] In the import and export trade customs declaration process, import tariff calculation requires a comprehensive consideration of multiple rules, including basic tariff rates, various surcharges, trade remedy taxes, and special exemption clauses. The complex relationships of overlapping, exemptions, and mutual exclusions among different taxes constitute the core challenge of trade compliance accounting. Currently, the mainstream tariff calculation methods in the industry fall into two categories: manual calculation and estimation using purely large-scale language models.

[0003] The first method is a fully manual accounting model, in which customs brokers manually consult the basic tax rates corresponding to HTSUS tariff codes, search through Article 122 Annex II exemption list, additional tariffs on steel and aluminum / automobiles, anti-dumping / countervailing duties (AD / CVD), Chapter 99 special footnotes, and other regulatory documents, manually distinguish the overlapping logic of various taxes and summarize the taxable amount. This method relies entirely on manual operation and has two major drawbacks: First, tariff codes, trade remedy policies, and exemption lists are frequently updated, making manual document review time-consuming and resulting in low overall accounting efficiency; second, the overlapping levels of multiple taxes and the complexity of cross-exemption rules make it easy for manual operations to result in problems such as double taxation, omission of taxes and fees, and incorrect exemption determination, making it difficult to guarantee the accuracy of the accounting results. The second approach involves existing intelligent estimation software solutions. This method directly embeds all tariff rates and regulatory provisions into a large language model's prompts, relying on the model to autonomously complete rate queries, logical judgments, and tax calculations. While this solution avoids the inefficiency of purely manual operation, its numerical calculation reliability is poor. Large language models are essentially semantic generation models and lack stable and accurate numerical computation capabilities. When calculating multiple tax types and levels simultaneously, calculation errors are highly likely, resulting in results that do not meet customs compliance requirements. The cost of maintaining regulations is extremely high. When tax rates, exemption clauses, or additional tariff policies are adjusted, the corresponding regulatory text within the model's prompts needs to be manually modified one by one. This is a large workload and prone to omissions or errors in the prompts. In summary, existing manual calculation and pure large-model tariff estimation schemes cannot simultaneously ensure calculation accuracy, ease of regulation maintenance, adaptability to complex tax superposition logic, and system performance. The industry urgently needs an automated calculation scheme that can integrate deterministic program table lookup with large-model semantic reasoning and adapt to multi-level tariff superposition rules. Summary of the Invention

[0004] The purpose of this invention is to provide an import tariff superposition calculation method based on deterministic table lookup and large language model integration, which aims to solve the problems that existing manual calculation and pure large model tariff estimation schemes cannot simultaneously take into account the accuracy of calculation, the convenience of regulatory maintenance, the adaptability of complex tax superposition logic, and the system operation performance.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for superimposed calculation of import tariffs based on deterministic table lookup and large language model integration, comprising the following steps: Step 1, Deterministic Regulatory Lookup Steps: A unified regulatory constant source is invoked, and the following steps are executed sequentially in a fixed order: HTSUS basic tariff rate lookup, Section 122 Annex II exemption verification, anti-dumping / countervailing duty (AD / CVD) tariff rate lookup, and Chapter 99 footnote information extraction. The result is the output of all original tax calculation data for all tariff types. The regulatory constant source centrally stores all customs regulations, tariff rates, exemption lists, and surcharge constant data. Step 2, Tax Type Priority Modeling Steps: Construct a directed graph of tax type priority, define the superposition level, mutual exclusion relationship and exemption constraints between each tax type through the directed graph, and generate a standardized tax type priority rule set; Step 3: Single-time structured large model integration step: Input the tax calculation fact data output by S1 and the tax priority rule set generated by S2 into the large language model at once. The large language model completes the semantic compliance logic integration of multiple tax types and outputs the itemized amount of each tax type and the total amount of customs duties in accordance with the statutory declaration order of customs. Step 4: Loading the tariff data engine singleton: Set the HTSUS tariff database and AD / CVD anti-dumping and countervailing duty database as module-level singleton objects. The parsing and memory loading are completed only once when the system starts. Multiple import commodity tax calculation requests reuse this singleton data. Step 5: Update the regulations: When tariff regulations, tax rates, or exemption clauses change, only the corresponding data in the regulatory constant source needs to be updated. There is no need to modify the prompt words of the large language model. After the update, execute S1-S4 again to generate compliant tax calculation results.

[0006] As a preferred method for calculating import tariffs based on a combination of deterministic table lookup and large language model, in the S1 deterministic regulatory table lookup step, each lookup step adopts a fixed query logic with no random bias, and all tax rates and exemption determination results are output deterministically by the program without relying on the large language model for numerical calculation.

[0007] As a preferred method for calculating import tariffs based on deterministic table lookup and large language model, the directed graph of tax type priority is such that each node corresponds to a type of import tax item, and the directed edges are labeled with three types of constraint relationships: exemption, superposition, and mutual exclusion. This is used to distinguish between three statutory tax calculation logics: "a certain tax type is exempted from the upper-level tax type", "multiple tax types are forcibly superimposed for taxation", and "mutually exclusive tax types are selected for taxation".

[0008] As a preferred method for calculating import tariffs based on deterministic table lookup and large language model integration, the S3 single-step structured large model integration step adopts a one-input, one-inference architecture without splitting into multiple rounds of large model calls; the input data is in standardized structured JSON format, and the output is a structured compliant declaration message containing the name of each tax item, tax base, applicable tax rate, itemized tax amount, exemption explanation, and total tax amount.

[0009] As a preferred method for calculating import tariffs based on deterministic table lookup and large language model integration, the S4 engine singleton loading step includes a cache persistence mechanism to persist the parsed tariff and anti-dumping duty data to the memory cache. New tax calculation requests directly read the cached data, skipping the process of parsing duplicate files and querying the database.

[0010] As a proposed import tariff calculation method based on deterministic table lookup and large language model integration, the S5 regulation update step preferably supports two modes: incremental update and full update. Incremental update only replaces and changes the regulation entries, while full update refreshes the entire set of regulation constant sources. After the update is completed, single-instance data reload is automatically triggered without restarting the tariff calculation service.

[0011] As a preferred method for calculating import tariffs based on deterministic table lookup and large language model integration, the regulatory constant source is stored in an isolated database / configuration file, which is physically decoupled from the large language model prompts and business reasoning code. Regulatory changes only operate on the constant source storage layer without altering the model prompt logic and business calculation code.

[0012] Compared with the prior art, the beneficial effects of the present invention are: I. This invention delegates deterministic numerical logic, such as numerical lookup, tax rate determination, and exemption verification, to a program that performs table lookups. All tax rate and exemption determination results are output from a unified regulatory constant source, without relying on a large language model for numerical calculations. The large language model is only responsible for integrating the compliance semantic logic of multiple tax types, thus leveraging the strengths of both approaches. This avoids the errors inherent in pure large models while using them to uniformly organize multi-level regulatory semantic relationships. All original table lookup data, priority rules, and model output tax amounts are completely retained, allowing for full-chain traceability and verification of the accounting process, meeting the requirements for customs compliance declaration verification.

[0013] Second, this invention constructs a directed graph of tax type superposition priority, using nodes to represent various taxable items and directed edges to uniformly label three types of statutory taxation constraints: exemption, mandatory superposition, and mutually exclusive choice. It clearly defines the hierarchical priority and cross-exemption rules between different tax types. It can accurately distinguish complex statutory scenarios such as "upper-level tax exempting lower-level surtaxes," "simultaneous superposition of multiple taxes," and "selection of one of two taxes for the same type," eliminating taxation errors caused by confused superposition logic in traditional manual and simplified AI estimation schemes from the underlying rule level.

[0014] Third, this invention independently sets up a unified regulatory constant source, centrally storing all tariff codes, tax rates, exemption lists, and surcharge data, completely decoupled from the large language model prompts and business reasoning code. When tariff regulations, trade remedy policies, or exemption clauses change, only the corresponding data entries in the regulatory constant source need to be updated; there is no need to modify or rewrite the large model prompts. It also supports both incremental and full update modes, with updates taking effect automatically after completion, eliminating the need to restart the entire tariff calculation service. This significantly reduces the manual maintenance workload during policy iterations and solves the pain points of cumbersome and error-prone prompt modification in existing solutions. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating an embodiment of the present application. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 This invention provides the following technical solution: a method for superimposed calculation of import tariffs based on deterministic table lookup and large language model integration, comprising the following steps: Step 1, Deterministic Regulatory Lookup Steps: A unified regulatory constant source is invoked, and the following steps are executed sequentially in a fixed order: HTSUS basic tariff rate lookup, Section 122 Annex II exemption verification, anti-dumping / countervailing duty (AD / CVD) tariff rate lookup, and Chapter 99 footnote information extraction. The result is the output of all original tax calculation data for all tariff types. The regulatory constant source centrally stores all customs regulations, tariff rates, exemption lists, and surcharge constant data. Step 2, Tax Type Priority Modeling Steps: Construct a directed graph of tax type priority, define the superposition level, mutual exclusion relationship and exemption constraints between each tax type through the directed graph, and generate a standardized tax type priority rule set; Step 3: Single-time structured large model integration step: Input the tax calculation fact data output by S1 and the tax priority rule set generated by S2 into the large language model at once. The large language model completes the semantic compliance logic integration of multiple tax types and outputs the itemized amount of each tax type and the total amount of customs duties in accordance with the statutory declaration order of customs. Step 4: Loading the tariff data engine singleton: Set the HTSUS tariff database and AD / CVD anti-dumping and countervailing duty database as module-level singleton objects. The parsing and memory loading are completed only once when the system starts. Multiple import commodity tax calculation requests reuse this singleton data. Step 5: Update the regulations: When tariff regulations, tax rates, or exemption clauses change, only the corresponding data in the regulatory constant source needs to be updated. There is no need to modify the prompt words of the large language model. After the update, execute S1-S4 again to generate compliant tax calculation results.

[0018] Preferably, in the S1 deterministic regulatory lookup step, each lookup step uses a fixed query logic with no random bias, and all tax rate and exemption determination results are output deterministically by the program without relying on a large language model for numerical calculation.

[0019] In practical use, a standardized tax base and segmented table lookup calculation formulas are introduced: Let the CIF dutiable value of the goods be P (in US dollars). Obtain the four types of deterministic parameters sequentially from the regulatory constant source table: 1. HTSUS basic ad valorem tax rate r0; 2. Article 122 Additional tax rate r122, Annex II exemption flag Flag122∈{0,1} (0=no exemption, 1=exemption); 3. Anti-dumping duty rate rAD, countervailing duty rate rCVD, relief tax exemption flag FlagAD / CVD∈{0,1}; 4. Chapter 99 Special Surcharge Rates r99, Fixed Amount of Goods Handling Fee Ffee, Port Maintenance Fee Fport.

[0020] Single-stage deterministic tax calculation formula: Basic tariff: T0 = P × r0; Article 122 Supplementary Tax: IfFlag 122 =0, T 122 =P×r 122 ; IfFlag 122 =1, T 122=0; AD / CVD trade remedy tax: IfFlag_AD / CVD=0, T_AD / CVD=P×(r_AD+r_CVD); If Flag_AD / CVD=1, T_AD / CVD=0; Chapter 99 Special Taxes and Fees: T 99 =P×r 99 +F_fee+F_port; The table lookup outputs the original tax calculation facts dataset: D={P,r0,r 122 Flag 122 ,r_AD,r_CVD,Flag_AD / CVD,r 99 ,F_fee,F_port}. Example data: P = $10,000, r0 = 0.05, r 122 =0.075, Flag 122 =0, r_AD=0.12, r_CVD=0.03, Flag_AD / CVD=0, r 99 =0.02, F_fee=25, F_port=15; Result of table lookup: T0=500, T 122 =750, T_AD / CVD=1500, T 99 =240.

[0021] Preferably, in the directed graph of tax type superposition priority, each node corresponds to a type of import tax item, and the directed edges are marked with three types of constraint relationships: exemption, superposition, and mutual exclusion, which are used to distinguish the three statutory tax calculation logics: "a certain tax type is exempted from the upper-level tax type", "multiple tax types are forcibly superimposed for taxation", and "mutually exclusive tax types are selected for collection".

[0022] In practical use, set the tax type priority sorting (from high to low priority): Level 1: HTSUS Basic Tariff T0; Level 2: Article 122 Supplementary Tax T 122 ; Level 3: AD / CVD trade remedy tax T_AD / CVD; Level 4: Chapter 99 Special Taxes and Fees T 99 .

[0023] Mathematical formula for the constraint rules of directed graphs: Overlapping constraint: If the edges are overlapping, the total tax amount is added to this item: T_sum = T_sum + T x Exemption constraint: If a lower-level tax is exempted by a higher-level tax, T x =0; Decision formula: If the higher-level node → exemption → lower-level node, then the lower-level T x =0; Mutual Exclusion Constraint: When two taxes are mutually exclusive, only the item with the higher tax amount is retained: T x =max(T a ,Tᵦ),Tᵧ=0; Example rule: If Article 122 exempts part of the surcharge in Chapter 99, then r 99 =0,T 99 =F_fee+F_port=40.

[0024] Preferably, the S3 single-round structured large model integration step adopts a one-input, one-inference architecture without splitting into multiple rounds of large model calls; the input data is in a standardized structured JSON format, and the output is a structured compliant declaration message containing the name of each tax item, tax base, applicable tax rate, itemized tax amount, exemption explanation, and total tax amount.

[0025] In practical use, the comprehensive calculation formula for the large model is as follows: T_total=T0+T 122 +T_AD / CVD+T 99 ; Substituting the example data: T_total = 500 + 750 + 1500 + 240 = $2990.

[0026] If there are exemption constraints, the large model corrects the corresponding items according to the priority rules and then sums them up, and outputs the exemption text description at the same time. The output message carries the tax amount of each item, the basis for exemption judgment, and the total amount, which is directly adapted to the customs declaration interface. Preferably, the S4 engine singleton loading step includes a cache persistence mechanism, which persists the parsed tariff and anti-dumping duty data to the memory cache. New tax calculation requests directly read the cached data, skipping the process of parsing duplicate files and querying the database.

[0027] In practical use, set up quantitative metrics for comparing singleton loading time: Assume that the time taken to parse a complete HTSUS+AD / CVD tax database file is t_load=1200ms, and the time taken for a single database query is t_db=80ms; Without a singleton mechanism: each request performs parsing and querying, with a single request base time T_raw = t_load + t_db = 1280ms; the total time for 100 concurrent requests is 1280 × 100 = 128000ms. Singleton caching mechanism: The system loads only once at startup (1200ms), and each subsequent request reads only from the memory cache. The time taken for a single request is T_single=10ms; the total time taken for 100 concurrent requests is 1200+10×100=2200ms. Performance improvement rate calculation formula (direct copy): Rate=[(T_raw_all-T_single_all)÷T_raw_all]×100% Substituting the data: Rate=[(128000-2200)÷128000]×100%=98.28%.

[0028] Preferred: The S5 regulation hot update step supports two modes: incremental update and full update. Incremental update only replaces and changes the regulation entries, while full update refreshes the entire set of regulation constant sources. After the update is completed, single-instance data reload is automatically triggered without restarting the tariff calculation service.

[0029] In practical application, a quantitative comparison of update costs is provided: Traditional approach (modifying prompts): Changing one tax rate requires modifying three prompts, fine-tuning the model, and restarting the service, with a manual time consumption of H_old=120min; This invention provides incremental updates: only a single data entry in the regulatory constant source is modified, the cache is automatically refreshed, and the manual time consumption is H_inc=3min; Efficiency improvement formula: E_up=[(H_old-H_inc)÷H_old]×100% Substituting the data: E_up=[(120-3)÷120]×100%=97.5%; Full update is applicable to annual tax policy revisions, automatically reloads single-instance cache, ensures uninterrupted service, and eliminates business downtime.

[0030] Preferably, the regulatory constant source is stored in an isolated database / configuration file, which is decoupled from the large language model prompt words and business reasoning code. Regulatory changes only operate on the constant source storage layer and do not change the model prompt logic and business calculation code.

[0031] In practical use, the data isolation rules for a hierarchical storage structure are as follows: Constant storage layer (regulatory constant source): Data table tax_const, fields: code, tax rate, exemption mark, fixed fee, only stores values ​​and judgment identifiers; Model layer: The prompts only contain general tax calculation logic templates, without any hard-coded tax rates or exemption constants; Business code layer: Only defines general functions for table lookup and aggregation, without fixing any legal values; Decoupling criterion: Any tax rate r x Exemption Flag x The modification only operates on the tax_const table; the number of lines modified in the prompt words and business code files is 0.

[0032] Working principle: This invention adopts a dual-engine architecture of deterministic numerical computation and large-scale model semantic logic layered decoupling. It completes the calculation of compliant tariff superposition through five collaborative steps. The overall operation logic is divided into five layers: data layer, rule layer, AI integration layer, scheduling layer, and operation and maintenance update layer. Underlying data layer: After the system starts, step four is executed to load the massive HTSUS and AD / CVD tax databases into the memory cache in one go as a module single instance, eliminating the overhead of repeated parsing; all tax rates, exemptions, and surcharge constants are uniformly stored in an independent regulatory constant source, which is physically isolated from the model and business code, providing a stable and individually updatable numerical benchmark for the entire process.

[0033] Deterministic lookup layer (step 1): After receiving the CIF price and HTS code of the goods, the program is called to look up the table in a fixed serial order. Relying on the fixed and non-random query logic, the original sub-item data of basic tariff, 122 surtaxes, AD / CVD relief tax and 99 special taxes and fees are accurately calculated through the segmented tax calculation formula. The entire process does not rely on large models for numerical calculations, ensuring that the value of each tax and fee is unique and verifiable.

[0034] Rule Modeling Layer (Step 2): Construct a directed graph of tax type priority, define three types of taxation constraints—overlapping, exemption, and mutual exclusion—using textual judgment formulas, form a standardized rule set, solve the problems of double taxation and omission caused by the overlap and exemption of multiple tax types, and provide a legal taxation logic boundary for AI integration.

[0035] Large Model Structured Synthesis Layer (Step 3): All numerical facts and priority rules output from the table lookup are encapsulated into a structured JSON input to the large model at once. Multi-layered legal semantic fusion is completed in just one round of inference, applying T_total=T0+T according to the customs declaration order. 122 +T_AD / CVD+T 99 The system calculates the total tariff using a summary formula and outputs a standardized declaration message with itemized details and exemption explanations. It distinguishes between numerical calculation (program table lookup) and semantic logic analysis (large model), leveraging the strengths of the former and avoiding the distortion defects of numerical calculation in the large model.

[0036] Hot update operation and maintenance layer (step 5): When trade policies, tax rates, and exemption lists are adjusted, only incremental / full updates are performed on the regulatory constant source, and the memory single instance cache is automatically refreshed. There is no need to modify the prompt words of the large model or restart the computing service. After the update is completed, the table lookup, rule modeling, and AI integration process are re-executed to quickly output tax calculation results that comply with the latest regulations, which greatly reduces the cost of policy iteration and maintenance.

[0037] The entire process addresses five major industry pain points through four core mechanisms: layered division of labor, single-instance caching acceleration, decoupling of regulations and data, and priority rule constraints. It simultaneously solves these problems, including the inefficiency and error-proneness of traditional manual accounting, inaccurate tax calculations based on large-scale models, high cost of regulatory maintenance, poor concurrency performance, and chaotic logic of multiple tax types. It also balances calculation accuracy, operational performance, and ease of long-term maintenance.

[0038] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for superimposing import tariffs based on a combination of deterministic table lookup and large language model, characterized in that, Includes the following steps: Step 1, Deterministic Regulatory Lookup Steps: A unified regulatory constant source is invoked, and the following steps are executed sequentially in a fixed order: HTSUS basic tariff rate lookup, Section 122 Annex II exemption verification, anti-dumping / countervailing duty (AD / CVD) tariff rate lookup, and Chapter 99 footnote information extraction. The result is the output of all original tax calculation data for all tariff types. The regulatory constant source centrally stores all customs regulations, tariff rates, exemption lists, and surcharge constant data. Step 2, Tax Type Priority Modeling Steps: Construct a directed graph of tax type priority, define the superposition level, mutual exclusion relationship and exemption constraints between each tax type through the directed graph, and generate a standardized tax type priority rule set; Step 3: Single-time structured large model integration step: Input the tax calculation fact data output by S1 and the tax priority rule set generated by S2 into the large language model at once. The large language model completes the semantic compliance logic integration of multiple tax types and outputs the itemized amount of each tax type and the total amount of customs duties in accordance with the statutory declaration order of customs. Step 4: Loading the tariff data engine singleton: Set the HTSUS tariff database and AD / CVD anti-dumping and countervailing duty database as module-level singleton objects. The parsing and memory loading are completed only once when the system starts. Multiple import commodity tax calculation requests reuse this singleton data. Step 5: Update the regulations: When tariff regulations, tax rates, or exemption clauses change, only the corresponding data in the regulatory constant source needs to be updated. There is no need to modify the prompt words of the large language model. After the update, execute S1-S4 again to generate compliant tax calculation results.

2. The method for superimposing import tariffs based on deterministic table lookup and large language model integration as described in claim 1, characterized in that: In the S1 deterministic regulatory lookup step, each lookup step uses a fixed query logic without random bias. All tax rates and exemption determination results are output deterministically by the program and do not rely on large language models for numerical calculation.

3. The method for superimposing import tariffs based on deterministic table lookup and large language model integration as described in claim 1, characterized in that: In the directed graph of tax type superposition priority, each node corresponds to a type of import tax item. The directed edges are labeled with three types of constraints: exemption, superposition, and mutual exclusion. These constraints are used to distinguish between three statutory tax calculation logics: "a certain tax type is exempted from the upper-level tax type", "multiple tax types are forcibly superimposed for taxation", and "mutually exclusive tax types are levied by one".

4. The method for superimposing import tariffs based on deterministic table lookup and large language model integration as described in claim 1, characterized in that: The S3 single-round structured large model integration step adopts a one-input, one-inference architecture without splitting into multiple rounds of large model calls; the input data is in a standardized structured JSON format, and the output is a structured compliant declaration message containing the name of each tax item, tax base, applicable tax rate, itemized tax amount, exemption explanation, and total tax amount.

5. The method for superimposing import tariffs based on deterministic table lookup and large language model integration as described in claim 1, characterized in that: The S4 engine singleton loading step includes a cache persistence mechanism, which persists the parsed tariff and anti-dumping duty data to the memory cache. New tax calculation requests directly read the cached data, skipping the process of parsing duplicate files and querying the database.

6. The method for superimposing import tariffs based on deterministic table lookup and large language model integration as described in claim 1, characterized in that: The S5 regulatory hot update process supports two modes: incremental update and full update. Incremental update only replaces and changes regulatory entries, while full update refreshes the entire set of regulatory constant sources. After the update is completed, single-instance data reload is automatically triggered without restarting the tariff calculation service.

7. The method for superimposing import tariffs based on deterministic table lookup and large language model integration as described in claim 1, characterized in that: The regulatory constant source is stored in an isolated database / configuration file, which is decoupled from the large language model prompt words and business reasoning code. Regulatory changes only operate on the constant source storage layer and do not change the model prompt logic and business calculation code.

Images