Retrieval-Augmented Generation(RAG) has established itself as a critical framework for mitigating hallucinations and knowledge gaps in Large Language Models(LLM). While the evolution toward "Agentic RAG" has enabled autonomous planning and complex rea...
Retrieval-Augmented Generation(RAG) has established itself as a critical framework for mitigating hallucinations and knowledge gaps in Large Language Models(LLM). While the evolution toward "Agentic RAG" has enabled autonomous planning and complex reasoning, indiscriminately applying such high-cost workflows to every user query creates a bottleneck, leading to prohibitive computational costs and latency.
To resolve this inefficiency, this study proposes a Lightweight Adaptive Orchestration framework that dynamically determines execution strategies based on query complexity. We utilized Google’s Gemma-3-1B model, fine-tuned via QLoRA(Quantized Low-Rank Adaptation) for 4-bit quantization and parameter efficiency. The system classifies queries into three strategic pathways: “DIRECT” for simple fact verification, “PARALLEL” for multi-faceted analysis, and “SEQUENTIAL” for complex reasoning. This architecture ensures that computational resources are allocated only where necessary, preventing redundant operations.
Experiments were conducted using a reconstructed dataset derived from Google Research’s Natural Questions. The proposed sLM-based orchestrator achieved a routing accuracy of approximately 98.7%, a performance comparable to GPT-4 based few-shot classification and significantly superior to heuristic rule-based or scoring-based methods. In terms of efficiency, the adaptive approach reduced end-to-end latency by roughly 29% compared to a baseline that applies high-cost strategies uniformly, while maintaining practical response quality.
Notably, ablation studies revealed that the 1B model outperformed its 4B counterpart in routing accuracy. We interpret this result as a manifestation of optimization imbalance or inverse scaling, rather than simple overfitting. The larger 4B model struggled to converge on the specific decision boundaries of simpler tasks, whereas the 1B model demonstrated superior task-specific adaptation.
These findings offer significant implications across academic and practical domains. Academically, this work extends adaptive RAG research by demonstrating that dynamic workflow selection is a viable alternative to binary retrieval decisions. Technically, it proves that purpose-built sLMs can outperform larger general-purpose models in specialized decision-making tasks, suggesting a promising direction for on-device AI. Practically, the framework presents a production-ready solution that simultaneously optimizes user experience and operational costs, suitable for deployment in large-scale, domain-specific applications such as legal or medical QA systems.