Method of Parallel Hybrid Search for Large-Scale Code Repositories

Context. Modern software systems contain extensive and growing codebases, making code retrieval a critical task for software engineers. Traditional code search methods rely on keyword-based matching or structural analysis but often fail to capture the semantic intent of user queries or struggle with unstructured and inconsistently documented code. Recently, semantic vector search and large language models (LLMs) have shown promise in enhancing code understanding. The problem – is designing a scalable, accurate, and hybrid code search method capable of retrieving relevant code snippets based on both textual queries and semantic context, while supporting parallel processing and metadata enrichment.Objective. The goal of the study is to develop a hybrid method for semantic code search by combining keyword-based filtering and embedding-based retrieval enhanced with LLM-generated summaries and semantic tags. The aim is to improve accuracy and efficiency in locating relevant code elements across large code repositories.Method. A two-path search method with post-processing is proposed, where textual keyword search and embedding-based semantic search are executed in parallel. Code blocks are preprocessed using GPT-4o model to generate natural-language summaries and semantic tags.Results. The method has been implemented and validated on a .NET codebase, demonstrating improved precision in retrieving semantically relevant methods. The combination of parallel search paths and LLM generated metadata enhanced both result quality and responsiveness. Additionally, LLM-post-processing was applied to the top-most relevant results, enabling more precise identification of code lines matching the query within retrieved snippets. Other results can be further refined on-demand.Conclusions. Experimental findings confirm the operability and practical applicability of the proposed hybrid code search framework. The system’s modular architecture supports real-time developer workflows, and its extensibility enables future improvements through active learning and user feedback. Further research may focus on optimizing embedding selection strategies, integrating automatic query rewriting, and scaling across polyglot code environments