Retrieval-Augmented Strategies for LLM-Based Text-to-SQL Systems: A Systematic Literature Review

This OSF project documents the protocol, methodological audit trail, and reproducibility materials for the systematic literature review “Retrieval-Augmented Strategies for LLM-Based Text-to-SQL Systems: A Systematic Literature Review.” The review characterizes how retrieval-augmented strategies are designed and empirically evaluated in large language model (LLM)-based Text-to-SQL systems. It focuses on approaches that dynamically retrieve contextual information—such as database schemas, values, natural language–SQL examples, documentation, domain knowledge, or other structured knowledge—and use the retrieved information to support SQL generation, reasoning, decomposition, refinement, validation, or correction. The review covers peer-reviewed journal articles published in English within a search window from January 2022 to August 19, 2026. Systematic searches were conducted in Scopus, Web of Science Core Collection, and IEEE Xplore. Eligible studies were required to investigate natural-language-to-SQL generation, directly involve an LLM in the Text-to-SQL pipeline, incorporate a retrieval-augmented mechanism according to the predefined operational definition, and report an empirical evaluation. The review addresses four research dimensions: (1) the retrieval strategies and types of contextual information used; (2) the LLMs, datasets, database domains, and experimental configurations employed; (3) the evaluation metrics, baselines, and reported effects of retrieval on effectiveness and efficiency; and (4) recurrent limitations, implementation challenges, and open research gaps. The definitive searches identified 110 database records. After cross-source deduplication, 61 unique records entered title-and-abstract screening. Thirty records were retained for full-text assessment, all of which were successfully retrieved. Full-text eligibility assessment resulted in 24 included studies and 6 exclusions. Study selection was conducted using predefined inclusion and exclusion criteria, with independent reviewer assessment and reconciliation procedures documented in the project audit materials. Pilot searches and calibration activities were conducted before the definitive review stages to refine the search strategy and eligibility framework. These pilot activities are maintained separately from the formal study-selection dataset and are not included in the reported review counts. The resulting evidence synthesis characterizes retrieval according to four complementary analytical dimensions—WHAT is retrieved, HOW retrieval is performed, WHERE retrieval intervenes in the Text-to-SQL pipeline, and WHY retrieval is introduced. The project also documents the review’s data-extraction framework, quality assessment, retrieval-specific evidence synthesis, limitations, and reproducibility materials. Because the definitive database searches had already been executed when the OSF project was made public, this project should not be interpreted as a fully prospective preregistration of all review stages. It serves as a public protocol, methodological audit trail, and repository for the review’s reproducibility materials.