LLM-Driven Query Planning for Intelligent Access to Structured and Semi-Structured Data

This paper addresses intelligent access to structured and semi-structured data using large language models (LLMs). We develop an approach to LLM-driven query planning in which a user’s information need is handled as a sequence of interdependent operations: query analysis, grounding in data sources and the schema, plan construction, execution, validation of intermediate results, and, when needed, replanning. The study draws on a comparative analysis of the open benchmarks BIRD, Spider 2.0, and HybridQA and of representative published approaches, from Text-to-SQL pipelines such as DIN-SQL and CHASE-SQL to hybrid methods such as BlendSQL and Weaver. The proposed architecture separates planning from execution and allows SQL operations to be combined with semantic processing by an LLM. We formalize the plan as a directed acyclic graph of typed SQL, semantic, and lookup operations with validation predicates, define the conditions for partial replanning, and illustrate the procedure with a worked example. Our analysis of existing experimental results indicates that decomposing complex queries, combining heterogeneous sources, and monitoring intermediate state are each associated with higher accuracy or lower cost in multi-step data access tasks. The evidence supports assigning the LLM the role of a control component that coordinates the sequence of actions, although the proposed architecture still needs to be evaluated quantitatively under uniform experimental conditions. To that end, we propose an evaluation protocol with baselines, ablation variants, and falsifiable predictions.