Getting a RAG tutorial online for free, without it crashing, is the hard part.

Free hosts give you 512 MB of RAM, and PyTorch-based embeddings crash it. Old LangChain code breaks on 1.x. The Groq model most tutorials use was shut down in August 2026. Free vector DBs get suspended after a week idle.

This kit is the tested path from PDF to a public URL, on free tiers only.

What you build
A document Q&A API that answers with sources, and clearly refuses when the answer isn’t in your documents.
Stack: Flask · Groq gpt-oss-20b · Qdrant Cloud Inference · LangChain 1.x · Render

The 22-page guide
• Architecture, plus every free limit in one table
• The full pipeline, explained line by line
• Deploying on Render’s 512 MB, and beating both “sleep timers”
• A 10-question evaluation and a test suite

The starter code (.zip)
✅ Flask API with clear error responses
✅ One-click render.yaml deploy
✅ Weekly keep-alive so Qdrant isn’t suspended
✅ eval.py + 14 offline tests
✅ Every package version pinned and tested together

Who it’s for: students and developers who know basic Python and want a deployed AI project for a portfolio or hackathon.
Not for: heavy traffic. Free limits allow roughly 280 questions a day.

Limits verified November 2026.