LLM-Driven Business Intelligence for Retail Digital Transformation: A Decision Support System Case Study

You’re listening to “LLM-Driven Business Intelligence for Retail Digital Transformation: A Decision Support System Case Study,” by X. Wang and Y. Zhang. Published in 2026.

Abstract.

The rapid digital transformation of the retail industry has generated increasing demand for intelligent business intelligence systems that can translate natural language managerial inquiries into actionable analytical and predictive decisions. However, existing enterprise analytics solutions either rely on rigid dashboard-based querying or isolated AI models, and thus struggle to jointly achieve knowledge grounding, executable analytical reasoning, and decision-oriented synthesis within a unified framework. RIDE constructs knowledge-grounded analytical contexts from enterprise repositories, employs multi-step reasoning–action planning to orchestrate tool-based analytics, generates grammar-constrained executable SQL for reliable database access, and integrates predictive forecasting to support decision synthesis.

Overall, this study establishes a new paradigm for large language model–driven business intelligence systems that bridge human-like inquiry and machine-executable analytics, providing a robust foundation for intelligent and trustworthy retail digital management.

INTRODUCTION.

The retail industry is undergoing rapid digital transformation driven by large-scale transaction data, diversified customer interactions, and increasingly complex supply chain networks. In such environments, business intelligence (BI) systems play a critical role in supporting managerial decision-making by transforming raw enterprise data into analytical insights and actionable recommendations. Traditional BI platforms, however, remain largely dependent on predefined dashboards, static reports, or manually constructed database queries, which limits their flexibility in responding to dynamic business questions and constrains the efficiency of data-driven decision processes.

Developing intelligent BI systems that can understand natural language managerial inquiries, automatically conduct analytical reasoning, and deliver decision-oriented outcomes has therefore become a key research topic for next-generation retail digital management.

Recent advances in large language models (LLMs) have demonstrated strong capabilities in natural language understanding, reasoning, and content generation, inspiring a growing interest in applying LLMs to enterprise analytics and decision support. Some studies have explored retrieval-augmented generation for factual grounding, text-to-Structured Query Language (SQL) translation for database querying, and tool-augmented agents for multistep reasoning. Parallel research in deep forecasting has achieved notable progress in retail demand prediction. These developments have established that LLMs and deep learning models can enhance individual components of BI pipelines. Nevertheless, current approaches typically address knowledge retrieval, analytical querying, reasoning, and predictive modeling in isolation.

As a result, they lack a unified mechanism to coordinate knowledge grounding, executable analytics, and predictive decision synthesis in a single closed-loop framework. This fragmentation raises a fundamental research question: Can LLMs be systematically integrated with enterprise knowledge, analytical tools, and forecasting modules to realize end-to-end, decision-oriented BI systems for retail digital management?

This study was motivated by the hypothesis that effective BI in modern retail environments requires not only accurate data retrieval or prediction but also coherent orchestration of knowledge grounding, analytical reasoning, executable querying, and decision synthesis under a unified control paradigm. The primary objective of this research was therefore to design an LLM-driven BI framework that transforms natural language business inquiries into verifiable analytical operations, predictive insight, and actionable recommendations within an integrated decision loop.

Achieving this objective involved several key challenges. First, vanilla LLMs lack enterprise-specific knowledge and an up-to-date organizational context, which often leads to hallucinated or noncompliant analytical conclusions. Second, mapping high-level business questions into correct and executable analytical operations requires multistep reasoning and reliable interaction with external tools, such as databases and forecasting engines. Third, unconstrained generation of analytical queries may produce syntactically invalid or semantically inconsistent SQL, undermining the verifiability of decision outcomes. Finally, most existing BI systems treat prediction as an isolated downstream module rather than an integral component of analytical reasoning and decision synthesis. Addressing these challenges is essential for building trustworthy and practical LLM-driven BI systems.

To overcome these limitations, we propose Retail Intelligence and Decision Engine (RIDE), and LLM-driven BI decision framework for retail digital management. The core idea is to integrate enterprise knowledge grounding, reasoning–action analytical planning, grammar-constrained executable SQL generation, and predictive forecasting into a unified closed-loop architecture; In this process, RIDE first constructs knowledge-grounded analytical contexts from enterprise knowledge repositories to ensure factual consistency. It then uses a reasoning–action planning mechanism to decompose complex business inquiries into structured analytical steps and orchestrate tool-based operations. A grammar-constrained SQL decoding module is introduced to guarantee syntactic correctness and schema-aware executable querying.

Finally, a forecasting component is embedded into the analytical loop to generate future-oriented predictions that directly support decision synthesis. Through this coordinated design, RIDE enables natural language business inquiries to be transformed into traceable analytical workflows, verifiable data operations, predictive insight, and actionable decisions.

The main contributions of this work are threefold. First, we propose a novel LLM-driven BI decision framework that unifies knowledge grounding, analytical reasoning, executable querying, and predictive decision synthesis in a single architecture for retail digital management. Second, we designed a reasoning–action planning and grammar-constrained analytical execution mechanism that improves the reliability, traceability, and verifiability of enterprise analytical workflows. Third, we conducted extensive experiments on interactive analytical querying and retail forecasting benchmarks that demonstrate that the proposed framework consistently outperforms state-of-the-art text-to-SQL models, tool-augmented reasoning agents, and forecasting baselines in analytical correctness, reasoning efficiency, predictive accuracy, and end-to-end decision quality.

Together, these contributions provide a practical foundation for next-generation intelligent, trustworthy, and decision-oriented BI systems in the retail industry.

Related Work

LLM-Based Enterprise Analytics and Knowledge Grounding

Recent advances in LLMs have stimulated growing interest in applying them to enterprise analytics and BI systems. Early studies explored LLM-based question-answering capacity as it related to enterprises’ documents, conversational data analysis assistants, and retrieval-augmented generation frameworks that inject external knowledge to improve factual consistency. Knowledge-grounded architectures have been shown to reduce hallucination and enhance domain adaptation by retrieving relevant organizational documents, policies, or knowledge bases before generating responses. Parallel efforts in enterprises’ decision-support systems have incorporated knowledge graphs and semantic retrieval to improve explainability and compliance in analytical reporting.

These studies demonstrate the potential of LLMs to facilitates enterprises’ understanding of questions and for knowledge-aware response generation. However, most existing approaches remain limited to text-level answering and report generation and do not include explicit mechanisms to perform verifiable analytical execution or to synthesize predictive decision outcomes. This limitation motivates the need for frameworks that integrate knowledge grounding with executable analytics and decision-oriented reasoning.

Executable Analytical Query Generation

Another closely related research direction focuses on translating natural language analytical requests into executable database queries. Text-to-SQL models, such as Structured Query Language–Pathways Language Model (SQL-PaLM; Sun et al., 2024), Parsing Incrementally for Constrained Auto-Regressive Decoding (PICARD; Scholak et al., 2021), and DAIL-Structured Query Language (DAIL-SQL; Gao et al., 2024), have significantly improved the accuracy of mapping natural language questions to structured SQL statements, and grammar-constrained decoding and execution-guided training further enhance query executability and schema awareness. Interactive benchmarks, such as BIRD-INTERACT, have enabled the systematic evaluation of multi-turn analytical query refinement. These advances establish strong foundations for natural language–driven database access.

Nevertheless, existing text-to-SQL models primarily address query generation in isolation; they typically do not incorporate enterprise knowledge grounding, multistep analytical reasoning, or downstream predictive modeling and decision synthesis. As a result, their applicability remains limited to isolated analytical querying rather than end-to-end BI decision-making.

Tool-Augmented Reasoning and Agentic Systems

Recent studies have introduced tool-augmented reasoning frameworks that enable LLMs to perform multistep reasoning by using external tools. Representative methods, such as Reasoning and Acting (ReAct; Yao et al., 2022) and Toolformer, demonstrate that reasoning– action loops allow LLMs to iteratively plan, call tools, observe results, and refine subsequent actions. Agentic LLM systems further extend this paradigm to complex task execution, information retrieval, and interactive problem-solving. Research has shown that LLMs can orchestrate external resources to solve multistep tasks beyond pure text generation; however, existing agentic frameworks are mostly designed for open-domain question-answering or general task automation.

They lack explicit integration with enterprise knowledge constraints, structured analytical objectives, and predictive decision pipelines that are central to BI scenarios. This gap highlights the necessity of designing reasoning–action architectures specifically tailored to executable enterprise analytics and decision-oriented workflows.

Predictive Analytics and Retail Forecasting

Deep learning–based predictive analytics has achieved notable success in retail demand forecasting and supply chain optimization, and related studies have also shown the effectiveness of advanced deep learning, transfer learning, and hybrid attention-based forecasting models in applied prediction tasks. Probabilistic forecasting models, such as Deep Autoregressive Recurrent Networks (DeepAR; Salinas et al., 2020) and temporal fusion transformers (TFTs; Lim et al., 2021) demonstrate strong performance on large-scale retail datasets, including the fifth Makridakis Forecasting Competition (M5) Forecasting benchmark. These methods provide accurate point and probabilistic forecasts that support inventory planning, replenishment control, and risk management.

Nevertheless, existing forecasting models operate as independent predictive modules, typically requiring manually designed data pipelines and lacking natural language analytical interfaces. They are not directly integrated with enterprise knowledge retrieval, analytical query generation, or reasoning-driven decision synthesis. As a consequence, current predictive analytics systems remain disconnected from natural language business inquiry and end-to-end BI workflows.

Overall, the existing research has advanced LLM-based enterprise question-answering, text-to-SQL query generation, tool-augmented reasoning, and retail forecasting as largely independent lines of study. However, a unified framework that coordinates knowledge grounding, reasoning– action planning, executable analytical querying, and predictive decision synthesis for retail digital management remains underexplored. This observation forms the foundation for the proposed RIDE framework, which aims to bridge these previously disjointed research directions into an integrated LLM-driven BI decision system.

METHOD

Method Overview

We propose an LLM-driven BI decision framework (denoted RIDE) to support natural language– based analytical inquiry, enterprise data reasoning, and decision-oriented synthesis in retail digital management. The key motivation is to bridge the long-standing gap between human-like business questioning and machine-executable analytical operations while ensuring factual grounding, query correctness, and predictive reliability.

Let q denote a natural language business query issued by an analyst or manager, such as “What is the projected inventory shortage risk for high-demand products next week?” The enterprise environment consists of (a) a structured database ​D,​containing transactional and operational data, and (b) an unstructured knowledge repository ​K, ​including business rules, policy documents, product descriptions, and management guidelines. The objective of the proposed framework is to generate a decision response y, which comprises analytical results, predictive insights, and actionable recommendations as shown in Eq.:

​y = F​(​q, D, K; Θ​)​ where ​F​(​ ⋅ ​)​ denotes the proposed end-to-end framework parameterized by Θ.

Unlike conventional BI systems, which rely on predefined dashboards or rigid SQL templates, RIDE decomposes the decision-making process into a sequence of tightly coupled modules:

• knowledge-grounded context construction,

• reasoning and tool-oriented analytical planning,

• natural language to executable SQL translation,

• predictive analytics, and

• decision synthesis.

The overall architecture of the proposed RIDE framework is illustrated in Figure 1, in which the data flow from natural language query to final decision output, as well as the interaction among knowledge retrieval, analytical reasoning, database querying, and predictive modeling components, can be clearly observed. This modular design enables the model to perform complex multistep analytical reasoning while maintaining traceable, verifiable, and enterprise-compliant operations.

Knowledge-Grounded Context Construction

A fundamental limitation of vanilla LLMs in enterprise decision-making is the lack of domain-specific and up-to-date organizational knowledge, which often results in hallucinated or noncompliant recommendations. To mitigate this issue, we introduced a knowledge-grounded context construction module inspired by retrieval-augmented generation paradigms. In practice, the enterprise knowledge repository can be constructed from internal business documents, such as policy files, product descriptions, operational guidelines, and management rules, and may be maintained through a combination of manual curation and automatic extraction or periodic updating processes.

Given an input query q, a dense retriever maps the query into a continuous semantic embedding, as shown in Eq.:

​​e​ q​ = ​f​ enc​(​q​)​​ where fenc(∙) denotes a transformer-based encoder. Similarly, each knowledge document ​d​ i​ ∈ K​is encoded as shown in Eq.:

​e​ ​d​ i​​ = ​f​ enc​(​d​ i​)​

The retriever identifies the most relevant enterprise knowledge documents using maximum inner product similarity, as shown in Eq.:

​Z = ​TopK​ ​d​ i​∈K​(​e​ q​ ⊤​ ​e​ ​d​ i​​)​ where ​Z​denotes the retrieved evidence set. The retrieved knowledge passages are concatenated with the original query to form a knowledge-enriched analytical context, as shown in Eq.:

​c​ 0​ = ​[​q; Z​]​

This context c0 provides factual and procedural grounding for subsequent reasoning and query planning. In practice, this mechanism allows the model to incorporate enterprise-specific business policies (e.g., replenishment rules, promotion constraints, supplier contracts) directly into its analytical reasoning, thereby improving factual consistency and organizational compliance. Moreover, the retrieved knowledge passages are also passed to the downstream SQL-generation stage as auxiliary grounding signals, where they help resolve schema linking ambiguities (e.g., business terms, attribute aliases, and entity descriptions) and support join-path selection when multiple relational paths are plausible in complex enterprise schemas.

The overall knowledge retrieval and context construction process, together with the subsequent constrained SQL generation mechanism, is illustrated in Figure 2, which visually depicts how external enterprise knowledge is retrieved, integrated, and supplied to downstream analytical modules.

Note. SQL = Structured Query Language; RIDE = Retail Intelligence and Decision Engine; LLM = large language model.

Reasoning and Tool-Oriented Analytical Planning

Retail business analysis often requires multistep operations, such as decomposing a high-level question into subqueries, retrieving intermediate statistics, invoking forecasting models, and integrating results before generating a final decision. To enable such structured workflows, RIDE adopts an interleaved reasoning–action planning strategy.

At time step t, the language model maintains an internal reasoning state st, computed as shown in Eq.:

​s​ t​ = LLM​(​c​ t−1​)​

On the basis of this reasoning state, the model predicts the next analytical action, as shown in Eq.:

​a​ t​ ∼ P(​a​ t​|​ ​s​ t​)​ where at represents a tool-oriented operation, such as issuing an SQL query, invoking a forecasting service, or requesting additional knowledge retrieval. Once an action is executed by the corresponding tool, an observation ot is returned as shown in Eq.:

​o​ t​ = ToolExec​(​a​ t​)​

The context is then updated iteratively as shown in Eq.:

​c​ t​ = ​[​c​ t−1​; ​a​ t​; ​o​ t​]​

This iterative reasoning–action loop continues until a termination condition is reached, that is, when the model predicts a special end-of-analysis action. In practice, the system is moderately sensitive to the maximum reasoning-step budget: Overly small limits may truncate complex analytical workflows, whereas excessively large limits mainly increase computational cost with limited additional benefit once the reasoning process has converged. This design allows RIDE to mimic human business analysts who alternate among reasoning, querying data systems, inspecting results, and refining subsequent analytical steps.

Compared with one-shot question answering, this planning mechanism provides two important advantages. First, it enables transparent and traceable analytical workflows, which are crucial in enterprise decision auditing. Second, it allows the system to dynamically adapt its analytical strategy on the basis of intermediate results, such as performing a deeper investigation when anomalies are detected in sales or inventory statistics.

The iterative reasoning–action mechanism adopted in RIDE is visually illustrated in Figure 3, where the interaction among reasoning state updating, tool invocation, observation acquisition, and context refinement is shown as a closed analytical loop.

Natural Language to Executable SQL Translation

A core capability of BI systems is transforming natural language analytical requests into executable database queries. In RIDE, once the reasoning module determines that structured data access is required, a specialized SQL generation component is activated.

Given the current context ct, the retrieved knowledge evidence Z, and database schema S, the SQL generator produces a candidate query as shown in Eq.:

​ ˆ y ​ sql​ = ​LLM​ sql​(​c​ t​, Z, S​)​

Here, ​Z​serves as an auxiliary semantic grounding source rather than a replacement for the structured schema, helping the model align user expressions with schema elements and choose more appropriate join paths in multi-table analytical queries.

However, unconstrained autoregressive decoding often leads to syntactically invalid or semantically inconsistent SQL statements. To address this, we introduced grammar-constrained decoding to enforce SQL correctness during generation. At decoding step j, the probability of generating token wj is adjusted as shown in Eq.:

​P ′ ​(​w​ j​)​ = ​{​P​(​w​ j​)​, ​w​ j​ ∈ ​V​ valid​(​y​ <j​)​​ 0 otherwise ​ where ​V​ valid​(​y​ <j​)​denotes the set of valid next tokens according to SQL grammar and schema constraints.

In practice, these constraints are constructed from two complementary parts: (a) a schema-independent SQL grammar that defines valid clause ordering and operator usage (e.g., SELECT, FROM, WHERE, GROUP BY, and aggregation patterns) and (b) schema-specific constraints dynamically instantiated from the target database metadata, including admissible table names, column names, and foreign-key relations. This design enables the decoder to adapt the valid token space to different database schemas while preserving SQL executability. When the database schema evolves, the framework can be adapted by refreshing schema metadata, reconstructing schema-specific decoding constraints, and updating schema-aware retrieval and linking signals so that newly added or modified schema elements can be incorporated into subsequent analytical generation.

The final SQL query is then executed on the enterprise database, as shown in Eq.:

​o​ sql​ = Exec​(​ ˆ y ​ sql​, D​)​

If execution fails because of runtime errors or empty result sets, execution feedback is returned to the reasoning module for query refinement, enabling self-correcting analytical loops. This execution-aware design ensures that all analytical conclusions and recommendations in the final response are grounded in verifiable database results. The schema-specific constraint set is automatically updated for each target database on the basis of its schema graph, ensuring that grammar-constrained decoding remains compatible across heterogeneous enterprise databases. The constrained SQL generation and execution feedback process is also reflected in Figure 2, where grammar-based token filtering and executable query formation are highlighted.

Predictive Analytics Module

BI decision-making in retail management requires not only a descriptive analysis of historical data but also predictive modeling to anticipate future demand fluctuations, inventory risks, and sales dynamics. To support forward-looking decisions, RIDE incorporates a predictive analytics module that estimates future business states on the basis of historical observations and contextual factors.

Let x1:T= {x1,x2...,xT} denote a sequence of historical business indicators, such as daily sales volume or inventory levels. The predictive module aims to estimate the future trajectory ​x​ T+1:T+H​over a horizon H. This process is formulated as shown in Eq.:

​ˆ x​ T+1:T+H​ = ​f​ pred​(​x​ 1:T​, z​)​ where fpred(∙) denotes a temporal forecasting network and z represents auxiliary covariates, including promotion events, holiday indicators, pricing adjustments, and store attributes.

Rather than producing point estimates only, the predictive module models a probabilistic distribution over future states, as shown in Eq.:

​p(​x​ T+1:T+H​| T+H ​ ​x​ 1:T​, z) = ​ ∏ t=T+1​ ​ p(​x​ t​ ∣ ​x​ 1:t−1​, z)​

This probabilistic formulation enables the system to quantify predictive uncertainty, which is essential for risk-aware decision-making in retail operations. The forecasting outputs are structured into machine-readable analytical observations, as shown in Eq.:

​​o​ pred​ = ​ˆ x​ T+1:T+H​, ​σ​ T+1:T+H​​ where σt denotes the predictive uncertainty at time step t. These predictive observations are then returned to the reasoning module for subsequent decision synthesis.

By decoupling numerical forecasting from natural language generation, RIDE ensures that all predictive results are computed by dedicated temporal models rather than implicitly inferred by the language model, thereby improving numerical accuracy and reliability. In the current framework, this predictive module is used as an external analytical component invoked within the reasoning loop rather than being jointly optimized end to end with the LLM.

Decision Synthesis and Response Generation

After acquiring knowledge evidence Z, database query results osql, and predictive observations opred, the final step is to synthesize these heterogeneous analytical outcomes into a coherent decision-oriented response.

The final response-generation process is formally defined as shown in Eq.:

​y = ​LLM​ gen​(​q, Z, ​o​ sql​, ​o​ pred​)​ where LLMgen(∙) denotes the response-generation component conditioned on the original query, retrieved enterprise knowledge, structured query results, and predictive analytics.

To ensure response faithfulness to analytical evidence, we introduced an evidence-conditioned generation constraint, as shown in Eq.:

∣y∣ ​P​(​y​)​ = ​∏ ​ P(​y​ i​ ∣ ​y​ <i​, q, Z, ​o​ sql​, ​o​ pred​)​ i=1​

This formulation guarantees that each generated token is conditioned on explicit analytical results, reducing the risk of unsupported recommendations. The response-generation process produces three categories of outputs:

1. Descriptive analytical summaries derived from database query results.

2. Predictive insights obtained from temporal forecasting.

3. Actionable recommendations aligned with enterprise knowledge policies.

This structured synthesis ensures that the final decision response is linguistically natural, analytically verifiable, and operationally actionable.

Unified Optimization Objective

Although RIDE consists of multiple functional modules, they are integrated into a unified optimization framework. Each module introduces a dedicated learning objective, and the overall system is trained by minimizing a weighted combination of these objectives.

Knowledge retrieval loss is obtained using Equation: exp​(​e​ q​ ⊤​ ​e​ ​d​ +​​)​ ​L​ ret​ = − log ​

​ ​∑ d∈K​ ​ exp(​e​ q​ ⊤​ ​e​ d​) where d+ denotes a relevant knowledge document. SQL generation loss is calculated using Equation:

​​L​ sql​ = − ​∑ j​ ​ logP(​y​ j​ sql​ ∣ ​y​ <j​ sql​, ​c​ t​, S)​

Predictive modeling loss is defined using Equation:

T+H ​ E [ ​(​x​ t​ − ​ ˆ x ​ t​)​ 2​]​ ​L​ pred​ = ​ ∑ t=T+1​

The overall optimization objective is formulated using Equation: where λ1, λ2, λ3 are hyperparameters controlling the contribution of each module. This unified objective encourages consistent optimization of knowledge grounding, query correctness, and predictive accuracy.

Computational Complexity Analysis

We analyzed the computational complexity of RIDE to demonstrate scalability in enterprise environments. Let N denote the number of knowledge documents, d the embedding dimension, L the SQL decoding length, T the historical sequence length, and h the hidden size of forecasting networks.

The retrieval operation requires O(Nd) for similarity matching in dense vector space. The reasoning and SQL decoding process follows autoregressive generation: O(L∙V), where V is the vocabulary size. The predictive module incurs O(T∙h2) because of recurrent or attention-based temporal modeling. Overall, the system achieves modular scalability, allowing independent deployment and parallel optimization of retrieval, querying, and forecasting components in real-world retail intelligence platforms.

EXPERIMENTAL SETUP

Datasets

To comprehensively evaluate the proposed RIDE framework in realistic retail BI scenarios, we conducted experiments on two representative public datasets covering natural language analytical querying and retail predictive decision-making, respectively.

For interactive natural language to SQL reasoning and executable BI evaluation, we used the BIRD-INTERACT dataset. This dataset contains multi-turn natural language analytical questions paired with executable SQL queries that address complex real-world database schemas, simulating practical enterprise-level BI environments. The dataset includes diverse database structures, cross-table relational reasoning requirements, and conversational query refinement patterns, which closely align with the reasoning–action analytical workflow targeted by the proposed framework. Its interactive setting allows evaluation of multistep query planning, tool invocation, and execution-aware refinement, thereby providing a rigorous testbed for assessing the effectiveness of knowledge-grounded reasoning, text-to-SQL translation, and iterative analytical planning in RIDE.

For predictive analytics and decision-oriented retail management evaluation, we used the M5 Forecasting (Walmart) dataset, a widely recognized dataset for retail demand and inventory forecasting. The dataset contains historical daily sales records across multiple stores and product categories, together with auxiliary attributes, such as promotions, pricing events, and calendar information. This dataset reflects realistic retail operational dynamics, making it suitable for evaluating probabilistic forecasting performance and prediction-driven decision support. Its large-scale and hierarchical structure provides a robust basis for assessing the scalability and predictive reliability of the forecasting module embedded in RIDE.

Following common practice in large-scale BI and forecasting evaluation, we randomly split each dataset into training, validation, and test subsets with a ratio of 80%/10%/10%, ensuring that model development, hyperparameter tuning, and final performance reporting were conducted on mutually exclusive data partitions. This unified data partition strategy guaranteed fair and unbiased comparison among baseline methods and the proposed framework.

Because we combined the BIRD-INTERACT dataset for interactive analytical querying and M5 Forecasting for retail predictive modeling, the experimental setting provides comprehensive coverage of the key functional capabilities required by LLM-driven BI decision systems, thus enabling rigorous and representative evaluation of the proposed RIDE framework.

Baselines

To rigorously evaluate the effectiveness of the proposed LLM-driven BI decision system for retail digital management, we compared our method with eight representative baselines selected from recent state-of-the-art research. These baselines jointly cover knowledge augmentation, reasoning and tool invocation, natural language database querying, and predictive analytics, which constitute the core functional modules of LLM-based BI systems.

For knowledge-enhanced text generation, we adopted retrieval-augmented generation (RAG), which integrates parametric language models with external document retrieval to provide evidence-grounded responses and reduce factual hallucinations. As a widely used paradigm for knowledge-intensive applications, RAG serves as a fundamental benchmark to assess the effectiveness of enterprise knowledge integration in our framework. To evaluate reasoning and tool-use capabilities, we include ReAct, which enables LLMs to interleave chain-of-thought reasoning with executable tool calls, thereby enhancing interactive decision-making performance.

In addition, we considered Toolformer an advanced baseline that allows language models to autonomously learn when and how to invoke external tools through self-supervised application programming interfaceapplication programming interfaceapplication programming interfaceapplication programming interfaceapplication programming interfaceapplication programming interfaceapplication programming interfaceapplication programming interfaceapplication programming interface (API) learning, providing a strong reference for assessing structured tool orchestration strategies.

For natural language database querying, we selected three representative text-to-SQL baselines: (a) SQL-PaLM, (b) PICARD, and (c) DAIL-SQL. SQL-PaLM is a holistic framework that adapts LLMs for text-to-SQL translation through prompt engineering and instruction fine-tuning, demonstrating strong performance on complex database schemas. PICARD introduces incremental constrained decoding to enforce syntactic and semantic correctness during SQL generation, significantly improving SQL executability and reliability. DAIL-SQL represents a recent integrated LLM-based text-to-SQL solution that systematically optimizes prompt design and execution-based refinement, achieving state-of-the-art execution accuracy (EX) on public benchmarks.

We adopted these baselines to comprehensively evaluate the accuracy and robustness of conversational analytical query generation in enterprise-scale data environments.

For predictive decision analytics, we used DeepAR, a probabilistic forecasting model based on autoregressive recurrent neural networks, which has been widely applied in retail demand and inventory prediction tasks. Moreover, we included TFT, an attention-based multi-horizon forecasting architecture that provides strong predictive performance together with interpretable temporal dynamics. We used these forecasting baselines to benchmark the accuracy and stability of the proposed prediction-driven decision-making module under realistic retail management scenarios. Through comparisons with these baselines, we evaluated the proposed framework against state-of-the-art methods in all key functional stages of LLM-driven BI systems, ensuring a fair and comprehensive experimental validation.

Evaluation Metrics

To rigorously assess the effectiveness of the proposed RIDE framework across interactive analytical querying and retail predictive decision-making tasks, we adopted a set of standardized evaluation metrics that measure SQL executability and correctness, multistep reasoning effectiveness, and predictive accuracy. These metrics ensure fair comparisons with state-of-the-art baselines and reflect practical performance requirements in enterprise BI systems.

With respect to interactive natural language to SQL analytical reasoning on the BIRD-INTERACT dataset, we evaluated query generation performance using EX and exact match accuracy (EM). EX measures whether the generated SQL query yields the same execution result as the ground-truth query, as shown in Eq.:

N ​ I(Exec(​ ˆ y ​ i​ sql​, D) = Exec​(​y​ i​ sql​, D​)​)​ ​ ​∑ ​EX = ​ 1 i=1​ N where ​ ˆ y ​ i​ sql​ and ​y​ i​ sql​denote the predicted and ground-truth SQL queries for the ith sample, respectively, and ​I​(​ ⋅ ​)​is the indicator function. This metric directly evaluates the executability and semantic correctness of analytical queries, which is critical for real-world BI applications.

EM further measures the syntactic equivalence between generated SQL and reference SQL, as shown in Eq.::

N ​ I(​ ˆ y ​ i​ sql​ = ​y​ i​ sql​)​ ​ ​∑ ​EM = ​ 1 i=1​ N

Although EM is stricter than EX, it provides complementary evaluation of query formulation precision and structural correctness.

To evaluate the effectiveness of the multistep reasoning–action analytical process in interactive query refinement, we additionally evaluated the turn-level success rate (TSR), which measures whether the system successfully completes the analytical task within a limited number of interaction turns, as shown in Eq.:

N ​ ​∑ ​TSR = ​ 1 ​ I(​T​ i​ ≤ ​T​ max​)​ i=1​ N where Ti denotes the number of reasoning–action iterations required to produce a correct executable query and Tmax is the maximum allowed interaction budget. This metric reflects the efficiency and stability of iterative analytical planning in enterprise conversational BI scenarios.

For retail predictive decision analytics on the M5 Forecasting dataset, we evaluated forecasting performance using mean absolute error (MAE), root-mean-square error (RMSE), and mean absolute percentage error (MAPE), which are widely adopted in retail demand and inventory forecasting evaluation, as shown in Eq.::

N ​ ∣ ​x​ i​ − ​ ˆ x ​ i​ ∣​ ​ ​∑ ​MAE = ​ 1 i=1​ N

​RMSE = ​√ N ​ ​(​x​ i​ − ​ ˆ x ​ i​)​ 2​ ​​ ​ ​∑ ​ 1 i=1​

N

​x​ i​ − ​ ˆ x ​ i​ N N ​ ​∑ ​MAPE = ​ 100% ​x​ i​ ​ ∣​ i=1​ ​ ∣ ​ where xi and ​ ˆ x ​ i​denote the ground-truth and predicted values, respectively, at time step i. MAE, RMSE, and MAPE quantify absolute prediction accuracy, penalize large forecasting errors, and reflect relative deviation in percentage terms, respectively. They collectively provide a comprehensive assessment of forecasting precision and robustness under real-world retail operational dynamics.

Together, the adopted evaluation metrics comprehensively measure the correctness of executable analytical queries, efficiency of multistep reasoning, and accuracy of predictive modeling. This unified evaluation protocol ensures rigorous and representative assessment of the proposed RIDE framework across all key functional components of LLM-driven BI decision systems.

Implementation Details

All models in our experiments were implemented using the PyTorch deep learning framework. For LLM components, we adopted a transformer-based architecture initialized from an open-source pretrained instruction-tuned model, which served as the backbone for reasoning, SQL generation, and response synthesis modules. The knowledge retriever was implemented using a dense dual-encoder structure with shared transformer encoders for query and document representation learning, and the predictive analytics module was implemented using temporal forecasting networks corresponding to the adopted baselines and the proposed framework.

During training, the maximum input sequence length was set to 4,096 tokens to accommodate the multi-turn analytical context, retrieved knowledge passages, SQL schema information, and intermediate tool observations. The hidden dimension of transformer layers was set to 1,024, and the number of attention heads was set to 16. For retrieval encoding, document embeddings were precomputed and indexed using Facebook AI Similarity Search (FAISS) for efficient similarity search. The number of retrieved knowledge documents was fixed to K = 5 in all experiments.

For the reasoning–action planning process, the maximum number of reasoning iterations was limited to six steps to balance analytical depth and computational efficiency. Grammar-constrained decoding was applied during SQL generation, where invalid tokens violating SQL grammar or schema constraints were masked at each decoding step. Beam search with a beam size of four was adopted to improve SQL generation stability while maintaining reasonable inference cost.

For predictive analytics, the forecasting horizon was set according to the standard setting of the M5 Forecasting dataset. The historical input window length was fixed to 90 time steps. The forecasting networks were trained using the Adam optimizer with an initial learning rate of 1e−4. Early stopping was applied on the basis of validation loss to prevent overfitting. Dropout with a rate of.1 was used in all transformer and forecasting layers to improve generalization.

The joint optimization objective combined retrieval loss, SQL generation loss, and forecasting loss with balancing coefficients λ1, λ2, and λ3 set to 1.0, 1.0, and 0.5, respectively. Models were trained with a batch size of eight for language-related modules and 64 for forecasting modules. Gradient clipping with a maximum norm of 1.0 was applied to stabilize training.

All experiments were conducted on a workstation equipped with NVIDIA A100 GPUs (80 GB memory), Intel Xeon CPUs, and 256 GB RAM. Training and inference were executed using Ubuntu (Version 20.04) with CUDA (Version 11.8). We repeated each experiment three times with different random seeds, and we report average performance to ensure stability. These implementation settings ensured a fair comparison with baseline methods while maintaining computational feasibility and reproducibility in large-scale enterprise BI scenarios.

EXPERIMENTAL ANALYSIS AND RESULTS

Overall Performance Summary

This experiment provides an overall evaluation of the proposed RIDE framework with respect to interactive analytical querying and retail predictive decision-making tasks. We compared RIDE with representative baselines on the BIRD-INTERACT dataset for text-to-SQL reasoning and on the M5 Forecasting dataset for retail demand prediction. The quantitative results are reported in Table 1, and the comparison of EX and TSR on the BIRD-INTERACT dataset is illustrated in Figure 4.

Note. EX = execution accuracy; EM = exact match accuracy; TSR = turn-level success rate; MAE = mean absolute error; RMSE = root-mean-square error; MAPE = mean absolute percentage error; RAG = retrieval-augmented generation; ReAct = Reasoning and Acting; SQL-PaLM = Structured Query Language–Pathways Language Model; PICARD = Parsing Incrementally for Constrained Auto-Regressive Decoding; DAIL-SQL = DAIL-Structured Query Language; DeepAR = Deep Autoregressive Recurrent Network; TFT = temporal fusion transformer; RIDE = Retail Intelligence and Decision Engine. Empty cells indicate that the metric is not applicable to the corresponding model because text-to-SQL and forecasting baselines were evaluated on different task types.

One can see in Table 1 and Figure 4 that RIDE achieved the highest executable query accuracy and reasoning success rate among all text-to-SQL and tool-augmented baselines, demonstrating more reliable interactive analytical reasoning. In addition, RIDE yielded lower forecasting errors than dedicated forecasting models on the M5 Forecasting dataset, indicating improved prediction-driven decision capability for retail management.

Interactive Text-to-SQL Performance on the BIRD-INTERACT Dataset

In this experiment, we evaluated the effectiveness of RIDE in interactive natural language analytical querying, with a particular focus on executable SQL generation and multi-turn query refinement. We compared RIDE with state-of-the-art text-to-SQL and tool-augmented reasoning baselines on the BIRD-INTERACT dataset. The detailed text-to-SQL results in terms of EX and EM are reported in Table 2, and the EX comparison is illustrated in Figure 5.

Note. All table values are percentages. EX = execution accuracy; EM = exact match accuracy; SQL-PaLM = Structured Query Language–Pathways Language Model; PICARD = Parsing Incrementally for Constrained Auto-Regressive Decoding; DAIL-SQL = DAIL-Structured Query Language; ReAct = Reasoning and Acting; RIDE = Retail Intelligence and Decision Engine.

Note. EX = execution accuracy; SQL-PaLM = Structured Query Language–Pathways Language Model; PICARD = Parsing Incrementally for Constrained Auto-Regressive Decoding; DAIL-SQL = DAIL-Structured Query Language; ReAct = Reasoning and Acting; RIDE = Retail Intelligence and Decision Engine.

One can see in Table 2 and Figure 5 that RIDE consistently outperformed all text-to-SQL and tool-augmented reasoning baselines in both EX and EM. The improvements indicate that integrating knowledge grounding, reasoning–action planning, and grammar-constrained SQL decoding effectively enhanced executable query correctness and structural precision in interactive BI scenarios.

SQL Executability and Error-Type Analysis

In this experiment, we further analyzed the executability of generated SQL queries and investigated typical error patterns in interactive analytical querying. Our objective was to evaluate whether RIDE improves practical SQL reliability beyond overall EX by reducing different categories of query generation errors. The SQL executability statistics and error-type distributions are summarized in Table 3, and the comparative error-type breakdown across models is depicted in Figure 6.

Note. All table values are percentages. SQL-PaLM = Structured Query Language–Pathways Language Model; PICARD = Parsing Incrementally for Constrained Auto-Regressive Decoding; DAIL-SQL = DAIL-Structured Query Language; ReAct = Reasoning and Acting; RIDE = Retail Intelligence and Decision Engine.

One can see in Table 3 and Figure 6 that RIDE achieved the highest SQL executability rate, and consistently lower error frequencies, across all major error categories. In particular, reductions in schema linking and join path errors indicate that the integration of knowledge grounding and grammar-constrained decoding effectively enhanced structural correctness and schema awareness in complex enterprise databases. These results demonstrate that RIDE generates more reliable and executable analytical queries for real-world BI applications.

Multi-Turn Reasoning Efficiency and Convergence Behavior

In this experiment, we evaluated the efficiency of multi-turn analytical reasoning in interactive BI scenarios. Our objective was to examine whether RIDE can reach correct analytical conclusions with fewer reasoning–action iterations compared with tool-augmented reasoning baselines. The TSR and average number of reasoning steps required to complete analytical tasks are reported in Table 4, and the success rate under different reasoning turn budgets is illustrated in Figure 7. Here, Success@k denotes the proportion of tasks successfully completed within at most k reasoning turns.

Note. SQL-PaLM = Structured Query Language–Pathways Language Model; PICARD = Parsing Incrementally for Constrained Auto-Regressive Decoding; DAIL-SQL = DAIL-Structured Query Language; ReAct = Reasoning and Acting; RIDE = Retail Intelligence and Decision Engine.

Note. Success@k = proportion of tasks successfully completed within at most k reasoning turns; BIRD-INTERACT is retained as the official dataset name because no official expanded form is provided; ReAct = Reasoning and Acting; SQL-PaLM = Structured Query Language–Pathways Language Model; PICARD = Parsing Incrementally for Constrained Auto-Regressive Decoding; DAIL-SQL = DAIL-Structured Query Language; RIDE = Retail Intelligence and Decision Engine.

One can see in Table 4 and Figure 7 that RIDE achieved the highest TSR while requiring fewer reasoning–action iterations to complete analytical tasks. The consistently higher success rates across different turn budgets indicate that the proposed reasoning–action planning mechanism enables faster convergence and more stable multistep analytical workflows in interactive BI environments.

Knowledge-Grounded Faithfulness and Evidence Attribution

In this experiment, we evaluated the faithfulness and evidence attribution capability of RIDE in knowledge-grounded analytical decision generation. Our objective was to assess whether the proposed framework produces BI responses that remain consistent with retrieved enterprise knowledge and grounded analytical evidence, compared with retrieval-augmented and tool-augmented baselines.

The quantitative results are reported in Table 5, which summarizes the evidence support rate, citation precision rate, and faithfulness scores (FS). A comparative illustration of these three metrics across different models is shown in Figure 8.

Note. All table values are percentages. ESR = evidence support rate; CP = citation precision; FS = faithfulness score; RAG = retrieval-augmented generation; ReAct = Reasoning and Acting; SQL-PaLM = Structured Query Language–Pathways Language Model; PICARD = Parsing Incrementally for Constrained Auto-Regressive Decoding; DAIL-SQL = DAIL-Structured Query Language; RIDE = Retail Intelligence and Decision Engine.

One can see in Table 5 and Figure 8 that RIDE consistently achieved higher evidence support, citation precision, and faithfulness scores than retrieval-augmented and tool-augmented baselines. These results indicate that integrating enterprise knowledge grounding with reasoning–action analytical planning effectively reduces hallucinated conclusions and improves traceable, evidence-consistent decision generation in BI scenarios.

Retail Forecasting Accuracy on the M5 Forecasting Dataset

In this experiment, we evaluated the effectiveness of RIDE in retail predictive analytics, with a focus on demand forecasting accuracy for decision-oriented inventory and supply management. We compared the forecasting module of RIDE with representative probabilistic forecasting baselines on the M5 Forecasting dataset. The quantitative forecasting results in terms of MAE, RMSE, and MAPE are reported in Table 6, and representative prediction curves against ground-truth sales trajectories are illustrated in Figure 9.

Note. M5 = fifth Makridakis Forecasting Competition; MAE = mean absolute error; RMSE = root-mean-square error; MAPE = mean absolute percentage error; DeepAR = Deep Autoregressive Recurrent Network; TFT = temporal fusion transformer; RIDE = Retail Intelligence and Decision Engine.

Note. M5 = fifth Makridakis Forecasting Competition; DeepAR = Deep Autoregressive Recurrent Network; TFT = temporal fusion transformer; RIDE = Retail Intelligence and Decision Engine.

One can see in Table 6 and Figure 9 that RIDE achieved lower forecasting errors than DeepAR and TFT across all evaluation metrics. The improved predictive accuracy indicates that integrating knowledge-aware analytical reasoning with forecasting modules enhances demand prediction reliability, providing stronger support for retail decision-making tasks such as inventory planning and stockout risk control.

Forecast Uncertainty Quality and Calibration

In this experiment, we evaluated the quality of predictive uncertainty estimation produced by RIDE for retail forecasting. Our objective was to assess whether the proposed framework generates well-calibrated probabilistic forecasts that can reliably support risk-aware business decision-making, compared with probabilistic forecasting baselines.

The uncertainty estimation performance in terms of prediction interval coverage probability, prediction interval normalized average width, and negative log-likelihood, is reported in Table 7, and the calibration behavior of different models is illustrated in Figure 10.

Note. M5 = fifth Makridakis Forecasting Competition; PICP = prediction interval coverage probability; PINAW = prediction interval normalized average width; NLL = negative log-likelihood; DeepAR = Deep Autoregressive Recurrent Network; TFT = temporal fusion transformer; RIDE = Retail Intelligence and Decision Engine.

Note. M5 = fifth Makridakis Forecasting Competition; DeepAR = Deep Autoregressive Recurrent Network; TFT = temporal fusion transformer; RIDE = Retail Intelligence and Decision Engine.

One can see in Table 7 and Figure 10 that RIDE achieved higher coverage probability with narrower prediction intervals and lower negative log-likelihood compared with DeepAR and TFT. The calibration curve of RIDE remained closer to the ideal diagonal line, indicating more reliable uncertainty estimation. These results demonstrate that integrating knowledge-aware reasoning with forecasting enhances risk-sensitive prediction quality, which is essential for decision-making in retail inventory and supply chain management.

End-to-End BI Decision Quality: Query → Forecast → Decision

In this experiment, we evaluated the end-to-end decision quality of the proposed RIDE framework in realistic BI scenarios. Our objective was to assess whether RIDE can effectively integrate analytical querying, predictive forecasting, and decision synthesis to support actionable retail management decisions compared with systems that perform only partial analytical or predictive functions.

We constructed a unified decision evaluation pipeline on the BIRD-INTERACT and M5 Forecasting datasets in which natural language analytical queries trigger executable SQL retrieval, followed by demand forecasting and final decision recommendation. To be specific, current-state indicators derived from executable SQL outputs are combined with future demand and uncertainty estimates from the forecasting module to derive decision outcomes. On the basis of these integrated signals, the stockout risk reduction (SRR) evaluates the reduction in predicted stockout risk, service-level improvement measures the improvement in service level, and the overstock cost penalty reflects the overstock cost associated with the recommended inventory action.

The decision quality results, measured in terms of these three metrics, are reported in Table 8, and the comparative decision outcomes are depicted in Figure 11.

Note. All table values are percentages. SRR = stockout risk reduction; SLI = service-level improvement; OCP = overstock cost penalty; SQL = Structured Query Language; LLM = large language model; RIDE = Retail Intelligence and Decision Engine.

Note. SRR = stockout risk reduction; SLI = service-level improvement; OCP = overstock cost penalty; SQL = Structured Query Language; LLM = large language model; RIDE = Retail Intelligence and Decision Engine.

One can see in Table 8 and Figure 11 that RIDE significantly outperformed partial-function systems in all decision-oriented metrics. Higher stockout risk reduction and service-level improvement demonstrate that RIDE produces more effective and actionable decisions, and lower overstock cost penalty indicates a better balance between supply sufficiency and inventory cost. These results confirm that the proposed framework successfully closes the loop from natural language business inquiry to predictive analytics and decision synthesis, providing practical value for retail digital management.

Scalability and Runtime Cost Analysis

In this experiment, we evaluated the computational scalability and runtime efficiency of the proposed RIDE framework in practical BI deployment scenarios. Our objective was to assess whether RIDE maintains acceptable response latency and computational cost compared with baseline systems while providing enhanced analytical and decision-making capabilities. The runtime performance, including average response latency, tool invocation overhead, and GPU memory consumption, is summarized in Table 9.

Note. SQL = Structured Query Language; LLM = large language model; GPU = graphics processing unit; GB = gigabyte; RIDE = Retail Intelligence and Decision Engine.

One can see in Table 9 that RIDE introduced moderate additional latency compared with single-function systems because of integrated reasoning, retrieval, SQL execution, and forecasting steps. However, the overall response time remained within practical BI requirements while achieving substantially stronger analytical and decision performance, as demonstrated in previous experiments. These results indicate that RIDE provides a favorable trade-off between decision quality and computational cost, supporting its feasibility for real-world enterprise deployment.

Robustness With Respect to Hard Queries and Sparse Signals

In this experiment, we evaluated the robustness of RIDE under challenging analytical and predictive conditions. Our objective was to examine whether the proposed framework maintains stable performance on complex database queries and sparse retail demand signals, which frequently occur in real-world BI scenarios.

For interactive analytical querying, we categorized BIRD-INTERACT test samples by query complexity on the basis of the number of joins and nested operations. For retail forecasting, we grouped M5 Forecasting time series samples into dense and sparse demand patterns. The results of this performance comparison under varying different difficulty levels are reported in Table 10, and the robustness trends across query and signal complexity are illustrated in Figure 12.

Note. All table values are percentages except MAE. Avg. = average; EX = execution accuracy; MAE = mean absolute error; RIDE = Retail Intelligence and Decision Engine.

Note. SRR = stockout risk reduction; SLI = service-level improvement; OCP = overstock cost penalty; SQL = Structured Query Language; LLM = large language model; RIDE = Retail Intelligence and Decision Engine.

One can see in Table 10 and Figure 12 that RIDE consistently outperformed baseline methods across all difficulty settings. The performance gap became more pronounced under complex SQL structures and sparse demand signals, indicating that the proposed knowledge-grounded reasoning and unified analytical–predictive architecture significantly enhanced robustness. These results confirm that RIDE remains reliable under challenging enterprise BI conditions.

At the same time, we observed a gradual decline in reasoning performance as query complexity increased, especially for queries with heavier join compositions and nested operations. Although no strict upper bound was defined in this study, these results suggest that extremely complex analytical queries remain a challenging case for all compared methods.

Ablation Experiments

In this experiment, we investigated the contribution of each key component in the proposed RIDE framework. Our objective was to verify whether knowledge grounding, reasoning–action planning, grammar-constrained SQL decoding, and forecasting integration individually contribute to the overall analytical and decision performance.

We constructed several ablated variants by removing one module at a time while keeping other components unchanged. The ablation results on the BIRD-INTERACT and M5 Forecasting datasets, including executable query accuracy (EX), multi- TSR, forecasting error (MAE), and end-to-end decision quality (stockout risk reduction), are summarized in Table 11.

Note. Empty cells indicate that the metric is not applicable to the corresponding variant. EX = execution accuracy; TSR = turn-level success rate; MAE = mean absolute error; SRR = stockout risk reduction; SQL = Structured Query Language; RIDE = Retail Intelligence and Decision Engine; w/o = without.

One can see in Table 11 that removing any key component led to consistent performance degradation. In particular, disabling knowledge grounding or reasoning–action planning resulted in the largest drop in executable query accuracy and reasoning success rate, confirming their critical roles in interactive analytical workflows. Removing forecasting integration significantly reduced end-to-end decision quality, verifying the necessity of unified analytical–predictive coupling in the proposed framework.

Significance Tests

In this experiment, we conducted statistical significance tests to verify whether the performance improvements achieved by RIDE over representative baselines are nonrandom. Our objective was to ensure that reported gains are statistically reliable across different test samples.

We perform paired t tests on text-to-SQL metrics (EX and TSR) and forecasting metrics (MAE) and applied bootstrap-based confidence interval testing for end-to-end decision metrics (SRR). All tests were conducted at the p <.05 significance level. The results are summarized in Table 12.

Note. p = probability value; EX = execution accuracy; TSR = turn-level success rate; MAE = mean absolute error; SRR = stockout risk reduction; RIDE = Retail Intelligence and Decision Engine; SQL-PaLM = Structured Query Language–Pathways Language Model; PICARD = Parsing Incrementally for Constrained Auto-Regressive Decoding; DAIL-SQL = DAIL-Structured Query Language; DeepAR = Deep Autoregressive Recurrent Network; TFT = temporal fusion transformer.

One can see in Table 12 that all p values were below the.05 threshold, confirming that the performance gains of RIDE over baseline methods in analytical querying, predictive forecasting, and end-to-end decision tasks were statistically significant.

DISCUSSION.

The experimental results provide consistent evidence that the proposed RIDE framework effectively supports LLM-driven BI decision-making in retail digital management. Across interactive analytical querying, predictive forecasting, and end-to-end decision synthesis, RIDE demonstrated superior performance over state-of-the-art text-to-SQL models, tool-augmented reasoning agents, and probabilistic forecasting baselines. These findings confirm that integrating knowledge grounding, reasoning–action planning, executable analytical querying, and prediction-driven decision synthesis within a unified architecture is a viable and effective solution for bridging human business inquiry and machine-executable analytical operations in modern retail environments.

From the perspective of existing research, prior studies have separately investigated retrieval-augmented generation for factual grounding, text-to-SQL translation for database querying, tool-augmented agents for multistep reasoning, and deep forecasting models for retail demand prediction. These directions have established that each component can improve specific aspects of enterprise intelligence systems. However, existing methods typically treat these capabilities as independent modules or pipelines that lack a unified mechanism to coordinate knowledge retrieval, analytical reasoning, executable querying, and predictive decision-making.

The results obtained in this study provide new evidence that jointly modeling these processes within a single reasoning-driven framework leads to consistent gains in executable query correctness, reasoning efficiency, forecasting accuracy, and final decision quality. This unified integration represents a new understanding that end-to-end BI systems benefit not merely from stronger individual components but also from their coordinated interaction under a reasoning–action loop.

Compared with other recent LLM-based analytical and agentic systems, RIDE exhibits distinct differences in design philosophy and functional scope. Text-to-SQL models, such as SQL-PaLM, PICARD, and DAIL-SQL, focus primarily on mapping natural language to executable SQL without explicitly incorporating domain knowledge retrieval or predictive analytics. Tool-augmented agents, such as ReAct and Toolformer, emphasize reasoning–action interaction but do not natively integrate enterprise knowledge grounding or downstream forecasting modules. Forecasting models, such as DeepAR and TFT, achieve strong predictive accuracy yet lack natural language analytical interfaces and decision synthesis capabilities.

In contrast, RIDE unifies these previously disjointed research threads into a coherent BI decision system that enables natural language inquiry, knowledge-aware reasoning, executable analytics, predictive modeling, and decision recommendations to operate within a single closed-loop framework. This distinction explains the consistent advantages observed in both analytical correctness and decision effectiveness.

The proposed framework also demonstrates several practical advantages for retail digital management. First, knowledge grounding substantially improves factual consistency and reduces hallucinated conclusions, which is essential for compliance-sensitive enterprise decision environments. Second, reasoning–action planning enables transparent and traceable analytical workflows, supporting auditability and managerial trust in automated BI systems. Third, integrating predictive forecasting into the analytical loop allows decisions to be driven not only by historical data retrieval but also by future-oriented risk-and-demand estimation, which is central to inventory planning, replenishment control, and supply chain optimization.

These advantages collectively suggest that LLM-driven decision systems can evolve beyond passive reporting tools into proactive and adaptive retail management assistants.

Despite these strengths, several limitations should be acknowledged. First, the RIDE framework relies on predefined enterprise knowledge bases and database schemas, which may limit its scalability and adaptability in rapidly changing business environments or highly heterogeneous data ecosystems; addressing such scenarios may require more adaptive schema evolution and dynamic knowledge updating mechanisms. Second, although runtime latency remains within practical ranges, the integrated multistep reasoning and tool invocation introduce additional computational cost compared with single-function systems, which may limit deployment in ultra–low-latency business environments.

Third, the decision evaluation in this study adopted standardized proxy metrics to ensure reproducibility; real-world deployment may require domain-specific utility functions and cost structures that vary across enterprises.

These limitations indicate directions for future research on adaptive schema learning, efficient model compression, and enterprise-specific decision utility modeling. In addition, practical enterprise deployment should consider data privacy and security issues, including access control, protection of sensitive business information, and secure interaction with internal databases and knowledge repositories.

Overall, the findings of this study suggest that LLM-driven BI systems can meaningfully enhance retail digital management by transforming natural language business inquiry into executable analytics, predictive insight, and actionable decisions within a unified framework. Although this study focused on retail digital management, the overall framework is also potentially applicable to other enterprise domains, such as finance and logistics, where knowledge-grounded analytics, executable data access, and prediction-driven decision support are equally important.

The evidence obtained across multiple benchmarks supports the conclusion that coordinated knowledge grounding, reasoning–action planning, and predictive integration are key enablers for next-generation intelligent decision systems in enterprise environments. This work thus provides both methodological guidance and empirical evidence for developing trustworthy, explainable, and decision-oriented LLM-based BI platforms for the retail industry.

CONCLUSION.

In this study, we proposed RIDE, an LLM-driven BI decision framework for retail digital management. By unifying knowledge-grounded context construction, reasoning–action analytical planning, grammar-constrained executable querying, and predictive forecasting within a closed-loop architecture, RIDE bridges the long-standing gap between natural language business inquiry and machine-executable analytical operations. In extensive experiments on interactive analytical querying and retail forecasting benchmarks, RIDE consistently outperformed state-of-the-art text-to-SQL models, tool-augmented reasoning agents, and probabilistic forecasting baselines in executable query correctness, multistep reasoning efficiency, predictive accuracy, and end-to-end decision quality.

These results highlight the effectiveness of coordinated knowledge grounding, analytical reasoning, and prediction-driven synthesis for next-generation intelligent retail management systems.

Despite these advantages, two limitations remain. First, the framework currently assumes relatively stable enterprise database schemas and curated knowledge repositories; adapting RIDE to continuously evolving data environments and heterogeneous corporate information systems requires further automatic schema adaptation and knowledge updating mechanisms. Second, the multistep reasoning and tool invocation pipeline introduces additional computational overhead compared with single-function systems, which may constrain deployment in latency-sensitive business scenarios.

Our future research will focus on developing adaptive schema-learning and knowledge-refreshing strategies to improve system scalability in dynamic enterprise settings as well as exploring lightweight model compression and scheduling mechanisms to reduce inference cost. In addition, integrating domain-specific utility optimization and reinforcement learning–based decision feedback loops represents a promising direction for aligning LLM-driven decision systems more closely with real-world business objectives.

This work demonstrates that LLMs can evolve from passive analytical interfaces into proactive, knowledge-aware, and decision-oriented BI systems. The proposed RIDE framework provides a practical and extensible foundation for advancing intelligent retail digital management and opens new opportunities for trustworthy, explainable, and actionable artificial intelligence–driven decision-making.

FUNDING STATEMENT

This research was supported by the Gansu Provincial Department of Education: Innovation Fund for College Teachers Project (Grant No. 2024A-147; “Research on the Development Difficulties and Digital Governance of Grassroots Government E-Government in Gansu Province”) and by the President’s Team Project of Hexi University (Grant No. CXTD2025015; “Research on the Whole-Chain Collaborative Mechanism of ‘Production-Cold Storage-Sales’ for Modern Cold and Arid Characteristic Agriculture in Gansu Province”).

COMPETING INTERESTS

The authors of this publication declare there are no competing interests.