Improving LLM-based code completion using LR parsing

Code completion is a crucial feature in modern IDEs, improving programming efficiency. Traditional systems rely on prefix filtering and static ranking but often overwhelm users with lengthy, alphabetically sorted lists. Recent research has introduced LR-parsing-based approaches that derive completion candidates from language syntax and compute their ranks using open-source programs; however, these methods only suggest structural candidates, requiring manual refinement into complete code. To address this, we propose a hybrid method that integrates LR parsing with LLMs to enhance accuracy and usability. Our approach refines structural candidates using LR parsing into textual code suggestions via an LLM, referencing a database of ranked candidates from open-source programs. This combines the syntactic precision of LR parsing with the generative capabilities of LLMs. This study examines whether LLMs benefit from LR structural candidates in code completion. By comparing completions with and without these candidates, we assess their impact. Building on prior research, we also explore how leveraging top-ranked structural candidates can effectively enhance LLM-based code completion precision. We also demonstrate our method through VSCode extensions for Microsoft Small Basic and C. As a language-agnostic solution, our system applies to any language with a defined LR grammar. Our findings suggest that integrating LR parsing with LLM-based completion improves both accuracy and usability, paving the way for more effective code completion in modern IDEs.