Retrieval-Augmented Generation (RAG) has emerged as the dominant paradigm for grounding large language model (LLM) outputs in external knowledge, yet its systematic application to morphologically rich, low-resource languages remains largely unexplored. This paper addresses two under-studied problems for Kazakh-language RAG: the design of semantically coherent document chunking strategies and the acceleration of multi-stage inference on GPU hardware. We construct a corpus of 201,202 preprocessed news documents from the national newspaper Egemen Qazaqstan (2017–2025, 129 thematic categories) and propose three chunking strategies – Fixed-size (baseline), Semantic (cosine-similarity boundary detection), and Adaptive (joint semantic and length-regularized scoring) – evaluated under a four-dimensional quality framework measuring coherence, document coverage, boundary accuracy, and length regularity. The Adaptive strategy achieves an integrated  of 0.879, outperforming the Fixed-size baseline by 0.045 absolute, and improves retrieval F1@5 by 0.067 and end-to-end answer accuracy by 4 percentage points (92%) on a 50-query factual test set. For inference acceleration, we deploy an exact GPU-resident index (FAISS-GpuIndexFlatIP) on an NVIDIA H200 GPU, achieving 2.4 ms retrieval latency – a 70.8× speedup over ChromaDB – without accuracy degradation. Combined with INT4-quantised generation (), KV cache residency in HBM3, and an attention-based self-evaluation mechanism, the pipeline achieves a mean end-to-end response time of 0.92 ± 0.06 s under cold-cache conditions and approximately 0.20 s with full KV cache reuse (4.3× speedup), on a single GPU without external API dependencies. These results demonstrate the feasibility of high-accuracy, sub-second Kazakh RAG and establish reproducible baselines for future low-resource retrieval research.