The rapid proliferation of digital educational re-sources has created a critical information retrieval challenge for students and researchers. Standard keyword-based search mechanisms are increasingly inadequate for navigating dense, highly technical academic corpora, while raw Large Language Models (LLMs) suffer from severe knowledge cutoffs and hal-lucinatory tendencies when queried on specific, localized study materials. This paper introduces the architecture and empirical evaluation of the NoteLeech AI platform, a highly optimized, cross-platform Retrieval-Augmented Generation (RAG) system engineered specifically for academic context extraction and synthesis. By integrating a multi-stage ingestion pipeline with hierarchical semantic chunking and a quantized dense vector database, NoteLeech AI dynamically bridges the gap between unstructured academic notes and generative AI. We propose a hybrid retrieval mechanism that combines Hierarchical Naviga-ble Small World (HNSW) dense vector search with BM25 sparse keyword matching, deployed alongside a lightweight Llama-3-8B generative endpoint. The system architecture is uniquely tailored to process highly technical datasets, including complex engineering mathematics and computer science syllabi. Extensive experimental evaluations utilizing a specialized corpus of GATE CSE preparation materials demonstrate that the NoteLeech RAG pipeline achieves a Recall@5 of 92.4% while restricting end-to-end question-answering latency to under 450ms. Furthermore, we present comprehensive ablation studies on semantic chunk sizes and embedding dimensionality, proving that hybrid retrieval drastically reduces hallucination rates by over 87% compared to zero-shot LLM baselines. The findings establish that optimally tuned RAG architectures provide a robust, highly accurate, and scalable solution for personalized educational technologies, fundamentally altering how students interact with unstructured knowledge bases.

Index Terms—Retrieval-Augmented Generation, NoteLeech AI, Large Language Models, Vector Databases, Educational Technology, Natural Language Processing, Semantic Search.