bashar-alasadi/StudyAI

AI-powered study assistant for converting lectures to text # StudyAI StudyAI is an Arabic-first assistant with one focused workflow: ```text Upload → Full transcript → Full summary → Questions ``` It is designed so multi-hour audio/video lectures do not depend on one browser request and are never silently treated as complete when a segment is missing. ## Architecture ```text Browser (chunked upload + status polling) │ ▼ Flask blueprints ──► SQLite (users, uploads, jobs, segments, results) │ ▼ Redis / RQ ──► dedicated worker │ ├─► FFprobe validation ├─► FFmpeg video/audio → mono 16 kHz FLAC ├─► 30-minute segments with 5-second overlap ├─► Gemini transcription, one recoverable segment at a time ├─► strict ordered completeness check and overlap removal └─► token-aware full summary and full-coverage questions ``` Flask, Jinja, vanilla JavaScript, and SQLite remain intentionally: the product is a focused modular monolith and does not need a SPA, ORM, or microservices. Redis stores delivery state; SQLite is the durable source of truth. RQ is smaller than Celery and supplies `SpawnWorker` for Windows. ## Reliability guarantees - Browser lifetime and processing lifetime are independent. - Uploads use server-generated 128-bit IDs and configurable chunks (8 MB by default). - Public YouTube links and direct audio/video URLs can enter the same transcription, summary, and question pipeline. Playlists, private videos, and private-network URLs are rejected. - Duplicate identical chunks are idempotent; conflicting duplicates are rejected. - The default maximum lecture upload is 5 GB with disk-space reserve checks. - Video is converted to speech-focused lossless FLAC rather than sent to Gemini as video. - Completed segment transcripts survive retries and worker restarts. - A transcript can be assembled only when indexes are exactly `0..N-1`, all successful and non-empty. - Summary/questions use the complete transcript when it fits; otherwise every segment contributes to a hierarchical reduction chosen with Gemini token counting. No …