RAG-based harnesses, Markdown conversion top student LLM setups

Standard chat interfaces fail on 100-page university documents because they ignore the "garbage in, garbage out" problem of raw PDF parsing.

• –Moving from 35B local models to RAG pipelines (Open WebUI, AnythingLLM) is the only way to maintain detail without hitting context limits or "middle-loss" hallucinations.
• –High-fidelity parsing via Marker or MinerU is the real secret sauce; standard text extraction misses the tables and diagrams where university exam material actually lives.
• –Qwen 2.5 Plus is a strong backbone, but it requires semantic chunking to ensure the model focuses on technical nuances rather than general summaries.
• –Local setups using DeepSeek V4 (Flash) are becoming viable for students with high-end consumer GPUs (RTX 3090/4090) due to KV cache optimizations.