Overview
DeepNote turns textbooks, lecture transcripts, and exam papers into searchable, conversational study guides. Instead of skimming hundreds of pages, students can ask specific questions, generate verified answer keys for old tests, and build review materials pulled directly from their course materials.
Architecture and pipeline
- Document processing. Uploaded PDFs save to Supabase Storage. A Flask service extracts the text, splits passages into semantic chunks, and creates vector embeddings using Google Gemini.
- Vector search. Pinecone indexes each embedding with document IDs, page numbers, and section titles. When a user asks a question, the retrieval pipeline runs a hybrid search to pull the most relevant passages from the original text.
- Generation engine.
- Exam answer keys. The model analyzes exam questions, locates matching course material, and writes step-by-step solutions linked to source pages.
- Flashcards. The generator pulls out key formulas, definitions, and distinct concepts, formatting them for spaced repetition study.
- Document chat. The assistant grounds every response in the uploaded files. If a fact is missing from the reading, it says so rather than guessing.
Technical details
- Streaming responses. Text streams token by token through Server-Sent Events, so students read answers immediately rather than waiting for an entire generation to finish.
- Inline citations. System prompts require exact page and chapter citations for each assertion, keeping model output grounded in the text.