LLAMA-Augmented Hierarchical Transformer for Fine-Grained Code Search

Code retrieval plays a pivotal role in intelligent software engineering by enabling accurate matching between natural language queries and relevant source code. However, existing retrieval models often struggle with preserving long-range context, effectively capturing multi-granular semantics, and integrating global knowledge from large-scale language models. To overcome these limitations, this work proposes HiTR-LLaMA, a hierarchical Transformer-based retrieval framework enhanced with LLaMA-based contextual augmentation. By incorporating hierarchical encoding with cross-layer feature aggregation, LLMguided semantic enrichment, and a multi-view residual reranking mechanism, the proposed model achieves fine-grained query-code semantic alignment in a unified architecture. The retrieval and reranking modules are optimized through a two-stage training strategy, supported by hard-negative mining and task-specific prompt engineering. Extensive experiments demonstrate that HiTR-LLaMA substantially improves retrieval quality and offers a scalable, adaptable solution for real-world code search tasks.