What is Kothai?

Kothai is a self-hosted link manager. You drop in links, and it stores everything locally in a single SQLite file. It classifies your links automatically and lets you search by meaning, not just keywords. Ask it a question and it pulls up the relevant cards from your collection to answer. One container, no cloud account, runs on your own hardware.

The AI part is optional. Turn it on and Kothai will extract metadata from links, describe their thumbnails, tag them by topic and embed everything for semantic search. If you'd rather not run models locally you can point it at any LLM endpoint instead. Or turn AI off entirely and use it as a fast, private bookmark manager. Your data stays on your machine unless you choose to send inference elsewhere or connect Telegram capture.

In case you're wondering, Kothai is Bengali for "where". It's the question you end up asking when you're trying to find something, which is the problem this is built around.

Features

• Two-phase saving. A card appears the moment you hit Enter. Models catch up in the background.
• Ask your archive. Hybrid retrieval (cosine + keyword), cited answers, no hallucination beyond your data.
• Links turn into cards. Metadata, YouTube captions, and full article text get fetched and indexed.
• Spaces. Collections with optional tag rules and a freeform canvas view.
• Swappable models. Three slots (language, embedding, vision), all changeable live.
• Tunable memory. Each model role can be always on, on demand, or off. Turn all three off for a plain bookmark manager on 1 GB.
• Runs on modest hardware. Raspberry Pi 5 handles the small models. A lite image (475 MB) sends inference to any OpenAI-compatible endpoint instead.
• Your data stays yours. One SQLite file, JSON export in a click, hot backups with no downtime.