An AI-based system for structured data operations based on natural language with deterministic parsing

Abstract

A Kl-based system (100) for performing structured data operations based on natural language inputs using lightweight models and deterministic parsing methods, comprising: a computing device (102) having a processor (104) and a memory (106) configured to store one or more instructions executable by the processor (104), wherein the processor (104) is configured to execute a plurality of modules (108) to receive natural language queries from users via a user interface (120) hosted on user devices (124), generate structured intermediate representations, analyze and normalize conditional expressions, distinguish operational and filter conditions, and apply requested aggregation and selection operations to structured data sources (118); wherein the computing device (102) communicates with the structured data sources (118), a database (122), and the user devices (124) via a network (116); wherein the plurality of modules (108) comprises: a semantic interpretation module (109) configured to process the natural language query using a lightweight language model and generate a structured JSON output, wherein the structured JSON output comprises an operation to be performed, a main column to which the operation is applied, a target column to be returned if applicable, a reference to the relevant data source, and a list of filter conditions in their original natural language form; a deterministic parsing module (110) configured to analyze the list of filter conditions and extract the column, the relational operator and the value(s) for each condition and standardize different natural language formulations into logical filter rules, wherein the parsing module (110) is configured to identify the first range-based condition as the operative column if no explicit main column is specified and treat the remaining conditions as filters; an execution module (112) configured to apply the analyzed filter rules to the data set stored in the structured data sources (118), retain rows that meet all conditions, and apply the specified aggregation and selection operation to the filtered part; and a communication module (114) configured to transmit the structured result to the user on the user devices (124) via the user interface (120).

Description

TECHNICAL FIELD OF THE INVENTION

[0001] The present disclosure generally relates to the technical field of data processing and natural language interfaces, and more particularly to an AI-based system for performing structured data operations on tabular datasets driven by natural language input. Resource-efficient models and deterministic parsing are used to efficiently interpret, analyze, and execute user queries. BACKGROUND OF THE INVENTION

[0002] Modern data-driven applications increasingly rely on structured datasets such as tables, spreadsheets, and relational databases to gain actionable insights. As these datasets grow in both volume and complexity, users seek intuitive ways to query and analyze data without requiring knowledge of formal query languages ​​such as SQL. Traditional approaches to querying structured data often rely on static graphical user interfaces, manual filter configurations, or fixed dashboards. While useful, these methods remain rigid and require an understanding of the underlying data schema, limiting accessibility for non-technical users.

[0003] Natural language-based interfaces have emerged to make data query more user-friendly; however, many existing systems rely on large, parameter-rich language models. These models incur high computational costs, large memory requirements, and latency, making their use in resource-constrained environments such as embedded systems, browser-based applications, and edge devices difficult.

[0004] Furthermore, traditional systems often lack robust processing of complex, multi-conditional natural language queries. When a user enters a query with multiple filters, range conditions, or aggregation requirements, existing solutions are often unable to correctly parse them, incorrectly identify operational columns, or require predefined schema modeling and domain-specific retraining. For example, US20190123456A1 discloses a system that translates user questions into SQL queries using a large neural network trained on question-answer pairs. While effective in certain domains, it requires extensive training data tailored to each specific database schema, limiting portability and scalability.

[0005] Similarly, CN110987654B describes a hybrid system that combines a lightweight language model with rule-based parsing to process user queries. However, this system relies on manually defined templates and schema mappings, limiting flexibility in processing free-form queries with multiple conditions. Furthermore, it lacks deterministic logic to distinguish the operational column from pure filter conditions when multiple range filters are present. While some systems generate structured intermediate representations such as JSON, they often rely on post-processing with large models or manual schema mapping, increasing latency and hardware requirements. Such systems also struggle with ambiguous queries containing multiple range-based conditions, leading to unpredictable or erroneous results.

[0006] To solve all of the above-mentioned problems, there is a need for an AI-based system for performing structured data operations on tabular datasets driven by natural language input. This system uses resource-efficient models and deterministic parsing to efficiently interpret, analyze, and execute user queries. There is also a need for an AI-based system that uses a lightweight language model to convert freely formulated natural language queries into structured JSON output suitable for further processing. Furthermore, there is a need for an AI-based system that deterministically analyzes multiple natural language conditional expressions, extracts the logical components, and standardizes them into unified filter rules.In addition, a system is needed that can distinguish between operational conditions and filter conditions and specifically identifies the first range-based condition to determine the operational column if it is not explicitly specified by the user.

[0007] Furthermore, there is a need for an AI-based system that supports the execution of aggregation and selection operations on filtered datasets without predefined schema mapping. Likewise, there is a need for a system suitable for deployment in resource-constrained or latency-sensitive environments such as embedded systems, browser-based applications, and edge devices by minimizing computational and memory requirements. Furthermore, there is a need for an AI-based system that presents query results to the user in structured and visual formats, including JSON output, tables, charts, and dashboards. There is also a need for an approach that reliably transforms complex, multi-conditional natural language queries into standardized operations over structured datasets.Furthermore, there is a need for an approach that integrates semantic interpretation and deterministic parsing modules to enable domain-independent structured data queries without retraining on new datasets. Finally, there is a need for an AI-based system that can process synonyms, variants, and ambiguous expressions in natural language input by mapping them to standardized operational terms. Such a system should improve user experience by reducing the need to understand formal query languages ​​or database schemas, thus enabling intuitive and flexible natural language querying. SUMMARY OF THE INVENTION

[0008] The present disclosure proposes an artificial intelligence (AI)-based system for performing structured data operations based on natural language with deterministic parsing. A simplified summary is presented below to provide a basic understanding of some aspects of the claimed subject matter. This summary does not constitute a comprehensive overview. It is not intended to identify essential elements or define the scope of the claimed subject matter. Its sole purpose is to present some concepts in a simplified form as an introduction to the more detailed description below.

[0009] To overcome the above-mentioned deficiencies of the prior art, the technical objective of the present disclosure is to provide an artificial intelligence (AI)-based system that performs structured data operations on tabular datasets driven by natural language input. Resource-efficient models and deterministic parsing are used to efficiently interpret, analyze, and execute user queries.

[0010] According to one aspect, the invention provides an AI-based system for performing structured data operations on tabular data sets driven by natural language, using lightweight models and deterministic parsing. In one embodiment, the AI-based system comprises a computing device, a network, structured data sources, a database, and user devices.

[0011] In one embodiment, the computing device comprises a processor and a memory storing one or more instructions executable by the processor. In another embodiment, the computing device communicates with the structured data sources, the database, and the user devices via the network.

[0012] In one embodiment, the processor is configured to execute a plurality of modules to receive natural language queries from users, generate structured intermediate representations, parse and normalize conditional expressions into standardized logic, distinguish between operational and filter conditions, and apply requested aggregation and selection operations to structured data sources. The plurality of modules includes a semantic interpretation module, a deterministic parsing module, an execution module, and a communication module.

[0013] In one embodiment, the semantic interpretation module is configured to process natural language queries using a lightweight language model to generate structured output. This structured output includes an operation to be performed, a main column for the operation, an optional target column to be returned, a reference to the relevant data source, and a list of filter conditions in their original natural language formulation. The semantic interpretation module is further configured to process synonyms and variants in user queries by mapping them to standardized operation terms.

[0014] In one embodiment, the user interface supports interaction modalities, including graphical and conversational interfaces hosted on user devices such as smartphones, tablets, laptops, and desktop computers, to enable accessible and intuitive access to the AI-based system. The user interface is configured to allow users to enter queries in natural language and receive structured results, including tables, charts, or dashboards—without knowledge of formal query languages ​​or database schemas.

[0015] In one embodiment, the deterministic parsing module is configured to parse the list of filter conditions and extract the column, relational operator, and value(s) for each condition. Different natural language formulations are standardized into logical filter rules. The deterministic parsing module supports logical operators such as equality, inequality, greater than, less than, and range expressions. It identifies the first range-based condition as the operative column unless a main column is explicitly specified and treats remaining conditions as filters.

[0016] In one embodiment, the execution module is configured to apply the analyzed logical filter rules to the data set stored in the structured data sources by retaining rows that meet all conditions and performing the specified aggregation and selection operations (such as sum, average, count, minimum, maximum, top-n, or last-n) on the filtered portion of the data set. The execution module processes data sets without a predefined schema mapping, thus enabling schema-free operation.

[0017] In one embodiment, the communication module is configured to deliver the structured results to the user devices via the user interface. The communication module generates visual dashboards and interactive graphical outputs such as tables and charts to display the query results to improve usability and accessibility.

[0018] In one embodiment, the semantic interpretation module uses a transformer-based or recurrent neural network model with a reduced number of parameters to enable low-latency deployment in resource-constrained environments such as embedded systems, browser-based applications, and edge devices. This lightweight model ensures efficient processing of natural language queries while maintaining sufficient semantic understanding.

[0019] In one embodiment, the semantic interpretation module stores the structured output in the database for logging, review, and later reuse to enable traceability and system optimization. The database supports the storage of user queries and intermediate representations to increase operational efficiency.

[0020] In one embodiment, the AI-based system integrates semantic interpretation, deterministic parsing, execution, and communication modules to provide a schema-free, low-latency pipeline for processing complex, multi-conditional natural language queries on structured tabular datasets. This improves user accessibility and operational efficiency across diverse application domains.

[0021] This describes a procedure for executing structured data operations from natural language input using the AI-based system, using lightweight models and deterministic parsing. In one step, the semantic interpretation module receives the user's natural language input via the user interface displayed on the user device. In a further step, the semantic interpretation module processes the natural language input using a lightweight language model to generate a structured JSON output containing the operation to be performed, a main column for applying the operation, a target column (if different), a reference to the relevant data source, and a list of filter conditions in their original linguistic form.In the next step, the deterministic parsing module analyzes the list of filter conditions, extracts the column, the relational operator, and the value(s) for each condition, and standardizes the different expressions into logical filter rules.

[0022] In one step, the deterministic parsing module identifies the first range-based condition as the defining operational column if no main column is explicitly specified, and treats the remaining conditions as filters. In a further step, the execution module applies the analyzed logical filter rules to the dataset stored in the structured data sources, retaining only rows that meet all conditions. The execution module then performs the specified aggregation and selection operations on the filtered dataset. Finally, in a further step, the communication module transmits the structured result to the user device via the user interface.

[0023] Further objects and advantages of the present invention will become apparent from the following description, claims and accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The accompanying drawings, which form a part of this specification, illustrate an embodiment of the invention and, together with the description, explain the basic principles of the invention. Fig. 1 shows a block diagram of an artificial intelligence (AI)-based system for performing structured data operations from natural language input using lightweight models and deterministic parsing according to an exemplary embodiment of the invention. Fig. 2 shows a flowchart of a procedure for executing structured data operations from natural language input by the AI-based system according to an exemplary embodiment of the invention. DETAILED DISCLOSURE OF THE INVENTION

[0025] Various embodiments of the present invention will be described with reference to the accompanying drawings. Wherever possible, the same or similar reference numerals are used throughout the drawings and the description to refer to the same or similar parts or steps.

[0026] The present disclosure was developed with the goal of solving the above-described problem of the prior art. The objective of the present invention is to provide an artificial intelligence (AI)-based system that performs structured data operations on tabular datasets based on natural language input, using resource-efficient models and deterministic parsing to efficiently interpret, analyze, and execute user queries.

[0027] Fig.Figure 1 shows a block diagram of an artificial intelligence-based system 100 for performing structured data operations from natural language input using lightweight models and deterministic parsing. In one embodiment, the invention relates to an AI-based system 100 for performing structured data operations on tabular data sets driven by natural language input. Lightweight models and deterministic parsing are used to efficiently interpret, analyze, and execute user queries. The AI-based system 100 uses a lightweight language model to convert freely formulated natural language queries into a structured JSON output suitable for downstream processing. In one embodiment, the AI-based system 100 comprises a computing device 102, a network 116, structured data sources 118, a database 122, and user devices 124.

[0028] In one embodiment, computing device 102 includes a processor 104 and a memory 106 configured to store one or more instructions executable by processor 104. The instructions enable AI-based system 100 to perform functions such as processing natural language queries, deterministically parsing filter conditions, and executing structured data operations. Processor 104 acts as a central processing unit responsible for coordinating tasks, performing computations, and controlling the plurality of modules 108 by retrieving and decoding instructions from memory 106.

[0029] In one embodiment, memory 106 serves as a storage unit for holding executable instructions and operational data required by processor 104. The data may include user queries, structured JSON output, intermediate representations, and configuration parameters required by AI-based system 100 to process natural language input and produce structured results. The interaction between processor 104 and memory 106 enables AI-based system 100 to perform deterministic parsing, apply filtering rules, and perform aggregation or selection operations on structured data sources 118.

[0030] In one embodiment, the computing device 102 represents the backend or processing unit of the AI-based system 100. The computing device 102 may operate as a server, embedded processing unit, or cloud-based instance and communicates with the user devices 124 on which the user interface 120 is hosted.

[0031] In one embodiment, user interface 120 is a graphical or conversational interface displayed on user devices 124, such as smartphones, tablets, laptops, and desktop computers. User interface 120 allows users to enter queries in natural language and display structured output such as JSON data, tables, charts, or dashboards. User interface 120 allows users to query structured data sets naturally, without requiring knowledge of formal query languages ​​or database schemas.

[0032] In one embodiment, computing device 102 communicates with user devices 124 hosting user interface 120 via network 116. Network 116 enables natural language queries to be transmitted from user devices 124 to computing device 102 and the structured results to be returned. Network 116 may use wired or wireless technologies to support flexible and remote access to AI-based system 100.

[0033] In one embodiment, network 116 may include a local area network (LAN), wide area network (WAN), cellular network, intranet, virtual private network (VPN), or wireless networks using radio frequency (RF) or infrared (IR) technologies. Network 116 enables AI-based system 100 to remain accessible from various user devices 124 and locations to enhance user experience and convenience.

[0034] In one embodiment, processor 104 is configured to execute a plurality of modules 108 to receive natural language queries from users via user interface 120 hosted on user devices 124, generate structured intermediate representations, parse conditional expressions and convert them into standardized logic, distinguish between operational and filter conditions, and apply requested aggregation and selection operations to structured data sources 118. Computing device 102 communicates with structured data sources 118, database 122, and user devices 124 via network 116.

[0035] In one embodiment, the plurality of modules 108 includes a semantic interpretation module 109, a deterministic parsing module 110, an execution module 112, and a communication module 114. In one embodiment, the semantic interpretation module 109 is configured to process the natural language query using a lightweight language model and generate a structured JSON output. The structured JSON output includes an operation to be performed, a main column to which the operation is applied, a target column to be returned (if different), a reference to the relevant data source, and a list of filter conditions in their original natural language formulation.

[0036] In one embodiment, the semantic interpretation module 109 uses a transformer-based or recurrent neural network model with a reduced number of parameters to enable low-latency deployment. The semantic interpretation module 109 stores the structured JSON output in the database 122 for logging, review, and later reuse. The semantic interpretation module 109 processes user queries with synonyms or linguistic variations by mapping them to standardized operational terms.

[0037] In one embodiment, the deterministic parsing module 110 is configured to analyze the list of filter conditions and extract the column, relational operator, and value(s) for each condition, as well as standardize different natural language formulations into logical filter rules. The deterministic parsing module 110 is also configured to recognize the first range-based condition as the operational column, unless an explicit main column is specified, and to treat the remaining conditions as filters.

[0038] In one embodiment, the deterministic parsing module 110 supports logical operators such as equality, inequality, greater than, less than, and range-based expressions. The semantic interpretation module 109 and the deterministic parsing module 110 process natural language queries with multiple conditions without requiring a predefined schema mapping of the dataset.

[0039] In one embodiment, the execution module 112 is configured to apply the parsed filter rules to the data set stored in the structured data sources 118, retain rows that match all conditions, and perform the specified aggregation or selection operation on the filtered data set. The execution module 112 supports multiple aggregation operations such as sum, average, count, minimum, and maximum.

[0040] In one embodiment, the communication module 114 is configured to deliver the structured result to the user via the user interface 120 on the user devices 124. The communication module 114 generates visual dashboards and interactive graphical outputs that display the query results via the user interface 120.

[0041] In one embodiment, the invention enables the performance of analytical operations on structured tabular data using natural language inputs. The AI-based system 100 uses a lightweight language model to semantically interpret user queries and generates a structured intermediate representation, followed by a deterministic, rule-based processing stage that applies the logic defined in the structured output directly to the data set.

[0042] In one embodiment, the invention comprises two main stages. In the first stage, the semantic interpretation module 109 processes the natural language input using the lightweight language model and generates a structured JSON output. The JSON output contains the operation to be performed, the main column to which the operation is applied, the target column to be returned (if different), a reference to the relevant structured data source, and a list of filter conditions expressed in natural language. The condition strings are retained in their original form for processing in the next stage.

[0043] In one embodiment, the second stage uses the deterministic parsing module 110 to process the list of condition strings and extracts the logical components of each condition. For each condition, the module identifies the relevant column, the relational operator, and the filter value(s). Supported operators include equality, inequality, relational operators such as greater than and less than, and range-based expressions. The parser maps diverse natural language formulations to these operators using standardized logical rules.

[0044] In one embodiment, the invention handles multiple range-based conditions by using only the first such condition to define the operational column, unless the user has explicitly specified a primary column. All subsequent range-based conditions are applied solely as filters. After parsing all conditions, the execution module 112 applies the logical filter rules sequentially to the structured dataset. Each row in the dataset is checked against the full set of parsed conditions, retaining only those rows that satisfy all conditions.

[0045] In one embodiment, after filtering, execution module 112 performs the specified aggregation or selection operation, such as calculating an average, maximum, or minimum, or applying selection operations such as Top-N, Last-N, or Between to the defined column of the filtered data set. The final result is returned to the user in a structured format via user interface 120.

[0046] In one embodiment, the invention enables structured querying of tabular data sets using natural language inputs, designed for use in lightweight environments. The invention includes converting free language into a structured intermediate JSON format, parsing and normalizing diverse natural language conditional expressions into standardized logical rules, distinguishing between operational and filter conditions, and performing the desired aggregation or selection operation only after deterministic filtering.

[0047] In one embodiment, the novelty lies in the systematic integration of the semantic interpretation module 109 and the deterministic parsing module 110 within a modular architecture. This integration enables the processing of complex natural language queries with multiple conditions without relying on large, parameter-intensive language models or predefined schema mappings, thus ensuring robustness and flexibility across diverse datasets.

[0048] In one embodiment, the AI-based system 100 is suitable for use in scenarios where low latency and resource efficiency are critical, such as embedded systems, browser-based tools, or edge platforms. By minimizing dependence on complex models and applying deterministic logic after interpretation, the AI-based system 100 achieves scalability, robustness, and applicability across different domains without domain-specific retraining.

[0049] In one embodiment, the AI-based system 100 processes a user's natural language input to generate a structured intermediate representation for subsequent parsing and execution. For example, a user may enter a query requesting the top 20 values ​​of a specific attribute (e.g., number of base shoots) from a dataset (e.g., finger millet), sorted by another attribute (e.g., inflorescence length), with conditions such as a minimum value for one attribute (e.g., longest finger length greater than 60) and a specific value for another (e.g., fruit color = ragi-brown). The semantic interpretation module 109 interprets this query and generates a structured output containing the operation type, the dataset identifier, the primary sort column, the target column for returning the values, and a list of filter conditions.

[0050] In one embodiment, the structured output includes fields specifying the analytical operation (e.g., selecting the top N rows), the data set to be searched, the column for sorting or aggregation, an optional target column whose values ​​are returned, and natural language conditions. A correction layer validates this output to ensure logical consistency, e.g., by checking whether the column used for sorting is numeric, if necessary. For example, if the user's query implies sorting on a non-numeric column, the AI-based system 100 can swap the main and target columns to comply with the query intent and ensure correct execution.

[0051] In one embodiment, the semantic interpretation module 109 uses a lightweight language model to transform natural language queries into structured intermediate representations. The model is a pure decoder-transformer model with autoregressive language understanding, designed for efficient language understanding and structured output generation at low latency.

[0052] In one embodiment, the model architecture includes transformer-decoder blocks with multi-head self-attention, rotary positional embeddings, SwiGLU activation functions, grouped-query attention for memory efficiency, and LayerNorm before the attention and MLP layers. The model contains approximately 12.4 billion parameters, achieving a balance between expressiveness and usability. In one embodiment, Table 1 shows the memory footprint of the lightweight language model under various precision settings to enable deployment in resource-constrained environments. Table 1: precision Memory requirements (VRAM / RAM) Use case FP16 ∼ 24 GB Inference with full precision BF16 ∼ 20-22 GB Hardware-accelerated deployment 8-bit ∼ 12-14 GB Hardware-accelerated deployment 4-bit ∼ 6-8 GB Quantized for this system

[0053] In one embodiment, the model achieves a prompt-to-generation latency of 200-500 milliseconds for typical queries ( ~50 tokens) for 4-bit quantized inference with a throughput of 30–50 tokens per second on commercial GPUs (e.g., RTX 3060 / 4060). The model was selected due to its low latency, reduced hardware requirements, sufficient semantic understanding for structured query extraction, offline compatibility, and ease of deployment without complex infrastructure. In one embodiment, Table 2 shows the rationale for selecting the lightweight model over larger alternatives. Table 2: Reason Justification Low latency Response times under one second are crucial for interactive systems Hardware requirements Runs on <10 GB VRAM and supports edge and offline deployment Semantic understanding 12B parameters sufficient for structured query extraction No external APIs Enables local inference and thus reduces data protection risks and costs Avoids overfitting Smaller models reduce hallucinations through rule-based post-processing Simplified deployment No need for complex infrastructure (e.g., distributed inference, vector DBs) "equal" / "equals" / "=" == "unequal" / "unequal" != Raw symbols (eg, “!=", “>=") Used as is

[0054] In one embodiment, the parser extracts the column name, operator, and values ​​from each condition using fuzzy matching to handle variations or typos. The parser identifies the first range-based condition (e.g., "between" or "top-n") as the operative column if no explicit main column is specified, and treats subsequent conditions as filters to resolve ambiguities.

[0055] In one embodiment, the AI-based system 100 processes structured tabular data sources 118, including Excel files, CSV files, and Pandas DataFrames, which are converted into a unified DataFrame interface for analysis purposes. In one embodiment, Table 3 shows the supported data formats and their compatibility. Table 3: format Support level Notes Excel (.xlsx, .xls) Fully supported Primary test format CSV (.csv) Fully supported Drop-in compatible Pandas DataFrames Indigenous Internal processing structure

[0056] In one embodiment, the AI-based system 100 operates without a predefined schema mapping, using fuzzy matching to resolve column names and dynamic type inference (e.g., string, numeric, date) to adapt to arbitrary data sets. This schema-free approach supports plug-and-play functionality across different domains.

[0057] In one embodiment, the efficiency of the system is evaluated using a lightweight language model in a CPU-based environment, testing queries of varying complexity. In one embodiment, Table 4 shows the performance benchmarks for different query types. Table 4: Query type LLM inference time(s) Filtration + Operating time (s) Memory delta (MB) Complex query (target + main + filter) 22.41 0.03 +8.42 Query with multiple conditions (no target) 20.48 0.06 +3.90 Query with one condition (no target) 9.77 0.10 -1.73 Operation-only query (no filters) 11.53 0.01 +0.44

[0058] In one embodiment, the AI-based system 100 has a memory footprint of approximately 250 MB, with a runtime memory change of +0 to +9 MB, and achieves a filtering and operation latency of less than 0.1 seconds, demonstrating its suitability for resource-constrained environments. In one embodiment, Table 5 shows a comparison of system efficiency compared to conventional LLM+SQL agents. Table 5: aspect This system Traditional LLM+SQL agent Schema required NO Yes Inference model Light, local Often cloud-based, slower Memory usage ~ 250 MB, +0 to +9 MB during execution 1-2 GB typical Operating latency < 0.1s post-LLM Higher due to SQL translation / retrieval Offline compatibility Fully supported Mostly in the cloud

[0059] In one embodiment, the system's low memory footprint and schema-free operation enable deployment in resource-constrained environments such as browser-based tools, embedded systems, and enterprise servers. Table 6 illustrates practical deployment scenarios and their feasibility in one embodiment. Table 6: Deployment type feasibility Why it works Browser tools (Pyodide / Streamlit) Yes Small memory, fast logic, no schema required Raspberry Pi / Jetson Nano Yes Low RAM usage, no cloud required Local desktop apps (exe) Yes Single binary possible Corporate server Yes Secure, schema-free Android / iOS (with quantized LLMs) With tuning Mobile backends with optimization possible

[0060] In one embodiment, the AI-based system supports 100 applications such as browser-based analytics assistants for instant dashboard creation, offline field devices for data query in remote areas, desktop applications for research labs, secure local servers for regulated industries, and mobile-integrated backends for voice or text input.

[0061] In one embodiment, the KI-based system 100 integrates robust error handling to ensure execution integrity. Invalid filter conditions, such as type inconsistencies (e.g., numeric comparison on a string column), are skipped with warning logging, allowing valid filters to be applied without interrupting the process.

[0062] In one embodiment, the schema-free and domain-independent design of the AI-based system 100 enables the processing of datasets from various domains such as agriculture (e.g., crop yields), medicine, and environmental data without the need for retraining or schema adaptation.

[0063] Fig.Figure 2 shows a flowchart 200 of a procedure for executing structured data operations based on natural language input by the AI-based system 100 using lightweight models and deterministic analysis. In step 202, the semantic interpretation module 109 receives the natural language input from the user via the user interface 120 displayed on the user device 124. In step 204, the semantic interpretation module 109 processes the language input with a lightweight language model and generates a structured JSON output containing an operation to be performed, a main column for the operation, a target column to be returned (if any), a reference to the relevant data source, and a list of filter conditions in their original linguistic form.

[0064] In step 206, the deterministic parsing module 110 analyzes the list of filter conditions and extracts the column, relational operator, and value(s) for each condition, standardizing the different expressions into logical filter rules. In step 208, the deterministic parsing module 110 identifies the first range-based condition as the main column (unless one is explicitly specified) and treats all further conditions as filters. In step 210, the execution module 112 applies the logically parsed filter rules to the data set stored in the structured data sources 118 to retain only rows that meet all conditions. In step 212, the execution module 112 performs the specified aggregation or selection operation on the filtered data. In step 214, the communication module 114 delivers the structured result to the user device 124 via the user interface 120.

[0065] Numerous advantages of the present disclosure are apparent from the above description.

[0066] According to the present disclosure, an AI-based system 100 for natural language-driven structured data processing with deterministic analysis is disclosed.

[0067] The proposed invention provides an AI-based system 100 that processes tabular data sets based on natural language input, efficiently employing lightweight models and deterministic analysis to interpret, process, and execute user queries. The AI-based system 100 uses a lightweight language model to convert free language input into structured JSON output suitable for downstream processing.

[0068] The system deterministically analyzes multiple natural language conditional expressions, extracts the logical components, and standardizes them into consistent filter rules. It distinguishes between operational conditions and filter conditions and specifically identifies the first range-based condition as the main column if it has not been explicitly specified by the user.

[0069] The AI-based system 100 supports the execution of aggregation and selection operations on filtered data sets without predefined schema mapping. It enables deployment in resource-constrained or latency-sensitive environments such as embedded systems, browser-based applications, and edge devices by minimizing computational and memory overhead. It provides the user with query results in structured and visual formats such as JSON, tables, charts, and dashboards. The approach processes complex, multi-conditional natural language queries and reliably transforms them into standardized operations over structured data sets. The method integrates semantic interpretation and deterministic parsing modules to enable domain-independent data queries without retraining on new data sets. The AI-based system 100 can process synonyms, variations, and ambiguous expressions in speech input by mapping them to standardized surgical terms. It improves the user experience by eliminating the need to understand formal query languages ​​or database schemas and enables intuitive and flexible natural language queries.

[0070] It will be readily apparent that numerous changes and adaptations can be made to the measures described above without departing from the spirit underlying the principles of the invention. Reference number list 100 Cl-based system 102 Computer device 104 processor 106 Reminder 108 variety of modules 109 Semantic Interpretation Module 110 Deterministic parsing module 112 Execution module 114 Communication module 116 Network 118 Structured Data Sources 120 User Interface 122 Database QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited patent literature

[0000] US 20190123456A1

[0004] CN 110987654B

[0005]

Claims

[1] A Kl-based system (100) for performing structured data operations based on natural language inputs using lightweight models and deterministic parsing techniques, comprising: a computing device (102) having a processor (104) and a memory (106) configured to store one or more instructions executable by the processor (104), wherein the processor (104) is configured to execute a plurality of modules (108) to receive natural language queries from users via a user interface (120) hosted on user devices (124), generate structured intermediate representations, analyze and normalize conditional expressions, distinguish operational and filter conditions, and apply requested aggregation and selection operations to structured data sources (118); wherein the computing device (102) communicates with the structured data sources (118), a database (122), and the user devices (124) via a network (116); wherein the plurality of modules (108) comprises: a semantic interpretation module (109) configured to process the natural language query using a lightweight language model and generate a structured JSON output, wherein the structured JSON output comprises an operation to be performed, a main column to which the operation is applied, a target column to be returned if applicable, a reference to the relevant data source, and a list of filter conditions in their original natural language form; a deterministic parsing module (110) configured to analyze the list of filter conditions and extract the column, the relational operator and the value(s) for each condition and standardize different natural language formulations into logical filter rules, wherein the parsing module (110) is configured to identify the first range-based condition as the operative column if no explicit main column is specified and treat the remaining conditions as filters; an execution module (112) configured to apply the analyzed filter rules to the data set stored in the structured data sources (118), retain rows that meet all conditions, and apply the specified aggregation and selection operation to the filtered part; and a communication module (114) configured to transmit the structured result to the user on the user devices (124) via the user interface (120). [2] The AI-based system (100) according to claim 1, wherein the semantic interpretation module (109) executes a transformer-based or recurrent neural network model with a reduced number of parameters to enable low-latency delivery. [3] The Cl-based system (100) of claim 1, wherein the deterministic parsing module (110) supports logical operators including equality, inequality, greater than, less than, and range-based expressions. [4] The AI-based system (100) of claim 1, wherein the semantic interpretation module (109) and the deterministic parsing module (110) process multi-conditional natural language queries without requiring a predefined schema mapping of the dataset. [5] The AI-based system (100) of claim 1, wherein the semantic interpretation module (109) stores the structured JSON output in the database (122) for logging, auditing, and reuse purposes. [6] The AI-based system (100) of claim 1, wherein the execution module (112) supports multiple aggregation operations including sum, average, count, minimum, and maximum. [7] The AI-based system (100) of claim 1, wherein the semantic interpretation module (109) processes user queries with synonyms or variants of natural language by mapping them to standardized operational terms. [8] The AI-based system (100) of claim 1, wherein the communication module (114) generates visual dashboards and interactive graphical outputs representing the query results via the user interface (120).

Images