TDS Virtual TA

A virtual teaching assistant for IIT Madras' Tools in Data Science course. Ask it a question the way you'd ask on the forum, and it answers from the course material and the archive of past Discourse threads — with links back to the threads it used.

Coursework project. Built to answer the questions a TA answers forty times a term.

How it works

scrape_course.py     course pages  ─┐
scrape_discourse.py  forum threads ─┴─→ preprocess.py ─→ knowledge_base.db (embeddings)
                                                              │
                                            app.py (FastAPI) ─┘  → answer + source links

• Retrieval, not recall. Answers are grounded in retrieved chunks and cite the threads they came from, so a wrong answer is traceable to its source.
• Images accepted. Students paste screenshots of errors; the endpoint takes an optional base64 image alongside the question.
• Local vector store. Embeddings live in SQLite (knowledge_base.db) — no vector database service to run or pay for.

Run it

pip install -r requirements.txt
export API_KEY=...          # embedding + completion provider
uvicorn app:app --reload

POST /api/ with {"question": "...", "image": "<optional base64>"} returns an answer and the links it drew from.

Rebuilding the knowledge base

python scrape_course.py      # course content
python scrape_discourse.py   # forum threads
python preprocess.py         # chunk + embed → knowledge_base.db

Honest limits

It answers from what it retrieved. If the archive never covered a topic, it will say so rather than invent a course policy — the failure mode a TA bot must not have.