Hybrid Layered Retrieval and Task-Aware Embeddings for Efficient Persian RAG Systems

Abstract

This study introduces task-injected layered hybrid retrieval-augmented generation (TILHR-RAG), a framework specifically designed for Persian to address the scarcity of native-language resources and the limitations of English-centric approaches. The architecture combines task-aware query augmentation, a layered retrieval strategy, and a hybrid semantic-lexical retriever, all supported by a multi-stage pipeline that includes preprocessing, document chunking, question generation, and embedding. A novel mechanism for injecting task-specific vectors directs retrieval toward domain intent while preserving comparability across documents. The layered design operates in three stages: per-task frequently asked questions (FAQ) retrieval, hybrid document search using FAISS semantic similarity combined with BM25 keyword matching, and a fallback response generated by a large language model (LLM). This structure ensures both precision and robustness. Comprehensive experiments across five progressively refined configurations demonstrate that TILHR-RAG achieves the best balance among accuracy, efficiency, and scalability, reaching 89.67% semantic accuracy with moderate latency and memory consumption on NVIDIA A100 hardware. Further evaluations on low-resource graphics processing units (GPUs) confirm that accuracy remains stable under hardware constraints, although latency increases significantly. Moreover, multilingual E5 embedding models substantially improve retrieval and generation quality for Persian – outperforming ParsBERT and Sentence-BERT (SBERT) – by mitigating challenges such as orthographic variation and complex compound word structures. Taken together, these findings establish task-injected layered hybrid retrieval-augmented generation as a practical, reproducible, and resource-efficient blueprint for Persian question answering, advancing retrieval-augmented generation for low-resource languages without requiring costly large language model fine-tuning, while also offering adaptable strategies for broader multilingual applications.