Study Assistant

Inspiration

Students drown in lecture slides, PDFs and textbooks. General chatbots answer fast but invent formulas, cite nothing, and can't be trusted before an exam. We wanted a tutor that teaches only from your material, shows exactly where each fact came from, and doesn't make up math.

What it does

Upload your course material: PDFs, PowerPoint slides, Markdown or text notes, and slide images. A whole textbook is fine.
Ask questions and get a structured, teaching-style answer with numbered citations [1] [2]. Each citation names the file and page, and the cited PDF page is rendered under the answer.
Choose the source: Notes (only your material), Web (academic web search via Tavily) or Both. Grounding is always explicit, and if your notes don't cover a question, the tutor says so instead of guessing.
Checked math: equations in the answer are verified with SymPy. When a step is confirmed, a green "Symbolically Verified (CAS)" card appears, with a runnable Python check.
Formulas, graphs and tables: KaTeX rendering, 2D/3D plots on request, and comparison tables.
Recall mode: flashcards generated from your notes for active recall.
Memory: chats persist, and each visitor gets a private workspace.

How we built it

Nebius Token Factory + NVIDIA Nemotron, routed per task:
- Nemotron 3 Nano rewrites the question (multi-query and HyDE), reranks retrieved passages and makes flashcards, with thinking turned off for speed.
- Nemotron 3 Super writes the everyday tutoring answers.
- Nemotron 3 Ultra takes derivations, proofs and calculations, with automatic fallback to Super.
Retrieval: NVIDIA llama-nemotron-embed-vl-1b-v2 embeddings (NIM), ChromaDB with cosine HNSW, query variants searched in parallel, then LLM listwise reranking.
Backend: FastAPI with server-sent events streaming. Frontend: Next.js / React, built as static files and served by the same process.
Deployment: one Docker container on Render's free tier (512 MB), with per-visitor workspaces, rate limits and upload caps.
Quality: 280+ automated tests on GitHub Actions CI, plus live end-to-end tests on the deployed site.

Challenges we ran into

Reasoning tokens ate the budget: on short helper calls, Nemotron spent all of max_tokens thinking and returned nothing. We turn thinking off for the fast role and filter <think> blocks out of the stream, even when a tag is split across chunks.
Citation format: Nemotron often wrote 【1】 instead of [1]; we normalize it before parsing citations.
Model ID casing: we guessed the Ultra ID from Super's naming and got a 404; listing /v1/models gave the real ID. Our fallback to Super had silently hidden the problem until a live test.
Embedding stalls: the hosted embedding endpoint intermittently hangs a request. Short timeouts with retries, plus searching query variants in parallel, cut typical answer time from about a minute to roughly 10–30 s.
Real textbooks broke ingestion: math fonts in PDFs extract as broken Unicode characters that crashed indexing part-way through a book. We now clean the text and the titles before indexing.
512 MB of RAM: indexing a full textbook pushed the free server over its memory limit. Limiting Python's memory pools (MALLOC_ARENA_MAX=2) cut peak memory by about 60 MB, enough to index a 2,600-chunk book live.

Accomplishments that we're proud of

Math that is checked symbolically, not just trusted.
Three Nemotron sizes, each doing the job it is best at, all behind one Token Factory key.
A full machine-learning textbook indexed and answered with page citations, on a free 512 MB instance.

What we learned

Grounding is a product decision as much as a model decision. Showing the source page, admitting "not in your notes" and verifying math built more trust than a smarter-sounding answer. We also learned that model routing pays off: small models for plumbing, large ones only where reasoning matters.