The information loss is a context window problem; once the doc exceeds what the model can hold at once it starts dropping or summarising. Qwen 3.5 Plus has a large context so that's a reasonable call for bigger docs. For the PPTs specifically, converting to plain text first (strip formatting, remove slide structure) before feeding helps a lot; models handle dense structured slides poorly.

If any of your course material comes as actual textbooks in EPUB format rather than lecture slides, calice.app might be worth a look. We don't do RAG; under the hood it's a heterogeneous mixture-of-experts that routes between retrieval specialists (near-perfect recall) and reasoning specialists (for sota reasoning), so you get both without naive full-context inference costs. LLM sits split-screen alongside the reader and highlights the exact passages it references. Not self-hosted, and won't help with PPTs or diagram-heavy PDFs, but for dense text-based books it works well.