This study investigates a hybrid Question Answering (QA) architecture that synergizes the determinism of structured database queries with the semantic flexibility of Large Language Models (LLMs). We address the limitations of monolithic approaches by introducing a lightweight Semantic Router that dynamically dispatches natural language questions to the most appropriate processing pipeline: a deterministic Text-to-SQL module for analytical queries or a DirectLLM module for open-ended reasoning. Our design minimizes entity misinterpretation and improves robustness by combining schema grounding with a dual-path execution flow. Experimental results on the Spider 1.0 benchmark demonstrate that our pipeline with fully open-source LLMs achieves 79.2% Execution Accuracy ensuring full reproducibility and significantly reducing inference costs. Furthermore, the routing mechanism attains 99.8% classification accuracy, proving that a non-parametric scoring system can effectively decouple analytical constraints from conversational flexibility without introducing latency overhead.