EduTube Tasks

Generate learning tasks from YouTube video transcripts and get AI feedback on answers. Uses Hugging Face models (Inference API) and RAG over the transcript.

---

Contents

• Technologies & features
• Usage examples
• Run with Docker
• Local setup
• Configuration
• Development

---

Technologies & features

• Transcript: YouTube URL → transcript (and optional language detection) via youtube-transcript-api and a small LangChain chain.
• RAG: Transcript split into chunks, embedded with sentence-transformers (multilingual MiniLM), stored in FAISS; retriever used for evaluation and Q&A.
• Task generation: LLM (Hugging Face Inference API) produces structured tasks: comprehension, multiple choice, fill-in-the-blanks, vocabulary, reflection.
• Answer evaluation: User answer + task + RAG context → same LLM → short feedback and suggestions.
• Agent: Follow-up questions answered using transcript search + LLM (no separate tool-calling agent).
• UI: Streamlit app: paste URL → generate workbook → answer tasks → submit for feedback; optional follow-up question per task.

---

Usage examples

Generate workbook and get feedback

1. In the sidebar, paste a YouTube URL (video must have captions).
2. Set number of tasks and language (e.g. en, ru).
3. Click Generate. After loading, tasks appear on the main page.
4. Use ← Previous / Next → to switch between tasks. Type your answer and click Submit answer to receive feedback.
5. Use the Follow-up question field and click Ask to get an answer based on the video transcript.

---

Run with Docker

Build and run the Streamlit app in a container:

docker build -t edutube-tasks .
docker run -p 8501:8501 -e HUGGINGFACEHUB_API_TOKEN=your_token edutube-tasks

Open http://localhost:8501. The image runs flake8 at build time; tests are not run in the default CMD (run them in CI or locally).

---

Local setup

1. Clone the repo and create a virtual environment:

   python -m venv .venv
   source .venv/bin/activate   

2. Install dependencies:

   pip install -r requirements.txt

3. Copy env and set your token:

   cp .env-example .env

4. Run the app (from project root):

   PYTHONPATH=. streamlit run src/app.py

---

Configuration

All optional; see .env-example.

Variable	Description
HUGGINGFACEHUB_API_TOKEN	Required for task generation and evaluation (Hugging Face Inference API). Alternative env names: HUGGINGFACE_API_TOKEN, HF_TOKEN.
EDUTUBE_LLM_MODEL_API	LLM model ID (default: HuggingFaceH4/zephyr-7b-beta).
EDUTUBE_LLM_PROVIDER	Optional Inference API provider (e.g. together, fal-ai).
EDUTUBE_MAX_TRANSCRIPT_CHARS	Max transcript length for generation (default: 12000).
EDUTUBE_EMBEDDING_MODEL	Sentence-transformers model for RAG (default: multilingual MiniLM).
EDUTUBE_RAG_CHUNK_SIZE, EDUTUBE_RAG_CHUNK_OVERLAP, EDUTUBE_RAG_RETRIEVER_K	RAG splitter and retriever settings.

---

Development

• Lint: flake8 src (see setup.cfg).
• Tests: PYTHONPATH=. pytest tests/ -v
• CI: GitHub Actions runs flake8 and pytest on push/PR to main/master.