[Paper Review] CBR-to-SQL: Rethinking Retrieval-based Text-to-SQL using Case-based Reasoning in the Healthcare Domain

Hung T. Nguyen, Hans Moen+1 more|arXiv (Cornell University)|Mar 5, 2026

Biomedical Text Mining and Ontologies0 citations

TL;DR

CBR-to-SQL introduces a two-stage retrieval framework that uses masked case templates and a separate entity grounding step to translate natural language questions into SQL in healthcare, achieving state-of-the-art logical form accuracy and robust generalization, especially under data scarcity.

ABSTRACT

Extracting insights from Electronic Health Record (EHR) databases often requires SQL expertise, creating a barrier for healthcare decision-making and research. While a promising approach is to use Large Language Models (LLMs) to translate natural language questions to SQL via Retrieval-Augmented Generation (RAG), adapting this approach to the medical domain is non-trivial. Standard RAG relies on single-step retrieval from a static pool of examples, which struggles with the variability and noise of medical terminology and jargon. This often leads to anti-patterns such as expanding the task demonstration pool to improve coverage, which in turn introduces noise and scalability problems. To address this, we introduce CBR-to-SQL, a framework inspired by Case-Based Reasoning (CBR). It represents question-SQL pairs as reusable, abstract case templates and utilizes a two-stage retrieval process that first captures logical structure and then resolves relevant entities. Evaluated on MIMICSQL, CBR-to-SQL achieves state-of-the-art logical form accuracy and competitive execution accuracy. More importantly, it demonstrates higher sample efficiency and robustness than standard RAG approaches, particularly under data scarcity and retrieval perturbations.

Motivation & Objective

• Motivate text-to-SQL in the healthcare domain where SQL expertise is a barrier for EHR data access.
• Propose a Case-based Reasoning (CBR) framework that decouples logical structure retrieval from entity grounding.
• Show that masked case templates enable reusable patterns and improve sample efficiency and robustness.

Proposed method

• Convert question-SQL pairs into abstract case templates by masking schema-specific entities (Case Retain).
• During inference, retrieve similar masked templates (Template Construction) and generate a draft SQL template with an LLM.
• Ground placeholder entities against a schema-aware lookup table (Source Discovery) and fill in concrete schema entities to produce executable SQL.
• Use a two-stage retrieval approach to separate logical structure from entity resolution, leveraging separate data sources for patterns and EHR grounding.
• Evaluate on MIMICSQL with Complete and Incomplete Database settings, plus brittleness metrics to assess sensitivity to retrieved cases.

Figure 1: Overview of the CBR-to-SQL architecture.

Experimental results

Research questions

• RQ1 How does a two-stage, case-based retrieval framework perform compared to standard RAG in text-to-SQL for healthcare?
• RQ2 Does masking and templating improve sample efficiency and robustness under data scarcity and retrieval perturbations?
• RQ3 What is the impact of decoupling logical structure retrieval from entity grounding on accuracy and brittleness?
• RQ4 How does CBR-to-SQL generalize to incomplete data regimes compared to RAG-based baselines?

Key findings

• CBR-to-SQL achieves higher Acc_EX and higher Acc_LF than RAG-to-SQL in the Complete Database setup (0.882 vs 0.855 for Acc_EX; 0.828 vs 0.811 for Acc_LF).
• CBR-to-SQL attains state-of-the-art Acc_LF and competitive Acc_EX on MIMICSQL, outperforming several baselines (e.g., TREQS, Coarse2Fine, PtrGen, SQLNet).
• In the Incomplete Database setup, CBR-to-SQL maintains a strong lead over RAG-to-SQL (0.842 vs 0.777 Acc_EX; 0.780 vs 0.747 Acc_LF).
• CBR-to-SQL shows lower brittleness than RAG-to-SQL under both Ex and LF metrics in both Complete and Incomplete settings (e.g., Δbrittle_EX: 0.047 vs 0.065 in CDB; Δbrittle_EX in IDB: 0.049 vs 0.068).
• Ablation studies reveal that removing Source Discovery drastically reduces performance, while masking-based Template Construction provides resilience against noise.

Figure 2: CBR cycle adapted from aamodt1994case .