Document-based question answering has been transformed by Retrieval-Augmented Generation, which blends information retrieval and generative models. Traditional methods often use fixed retrieval strategies and uniform chunking, which can limit performance across diverse document types. This paper introduces an Adaptive Hybrid RAG pipeline that adjusts preprocessing and retrieval according to document structure. The pipeline employs chunking classifiers, applying section-level segmentation to structured documents and paragraph-level segmentation to narrative texts. Semantic retrieval is conducted with FAISS, and responses are generated using Mistral-7B. Evaluations on DocuQA (20 PDFs) and SQuAD v2.0 (over 100,000 QA pairs) show strong semantic understanding (BERTScore 0.837 and 0.785) despite low exact-match BLEU scores (0.045 and 0.0012). Analysis indicates that retrieval quality and context continuity significantly influence answer accuracy, highlighting the importance of structure-aware chunking and multi-metric evaluation for practical RAG-based QA systems.