The Geneziz AI Page Grows Up: Streaming Answers, Vision, and Tasks That Run While You Sleep

A knowledge base that only answers when you poke it is a library. Geneziz is not a library: it's an assistant with a pulse. Three upgrades landed in this release, and together they change what "asking your second brain" feels like.

1. Answers that stream, live

Ask something and the answer starts building in front of you (word by word, the way you expect from a modern assistant) instead of a spinner followed by a wall of text. Streaming runs end-to-end: from the model (cloud or local) through the Geneziz engine into the chat UI, with sources cited when the answer draws on your knowledge base. If a connection can't stream, it degrades gracefully to a normal, complete answer. No errors, no dead ends.

Why it matters is subtler than "feels faster". Streaming makes long, source-rich answers usable: you start evaluating the first claim while the rest is still arriving, and you can tell within seconds whether the retrieval grabbed the right context or you should rephrase.

2. Vision: show, don't describe

You can now attach images to questions. A UI screenshot, an error dialog, a diagram, a photo of a whiteboard: the model reads it together with your question and your knowledge base context.

• On a cloud provider (your own key), image questions run on a multimodal model.
• Fully local setups work too, with local vision models.
• If the configured model can't see images (the built-in Geneziz AI reads text only), the request falls back to text-only instead of failing.

The practical win: "what is this error and did I save anything about this library?" is now one question with a screenshot attached, not a transcription exercise.

3. Scheduled tasks: the KB that taps you on the shoulder

This is the sleeper feature. On the AI page you can now schedule a task, a prompt that runs on a schedule, against your own knowledge base, headlessly:

• "Every morning at 9: summarize anything new I saved about the funding market."
• "Weekly: check my saved repos for the ones I tagged 'to evaluate' and give me a digest."

When a task runs, its answer is persisted and you get an OS notification. Open the app and the AI page shows each task's latest run and answer. No terminal, no cron, no duct tape: it's a first-class citizen of the app.

Point a task at built-in Geneziz AI and the whole briefing happens on your machine (retrieval, answer, notification) with no credits touched.

This is the inversion that matters: until now your knowledge base answered when asked. Now it initiates. Capture, ask, and briefings happen on a rhythm you define.

Your model, your rules

The AI page lets you switch between providers and models per conversation: built-in Geneziz AI, or your own keys: cloud providers when you want them, fully local models for the privacy-maximal setup.

Geneziz AI itself is local. It ships inside the installer (~1.6 GB), runs offline on your machine, and is free: it never touches credits. Cloud is an option you choose, not a requirement, and anything metered shows you the exact price before it runs.

Vision and streaming work across the lanes where the underlying model supports them, with graceful fallbacks everywhere else. The knowledge base itself never leaves your machine; only the AI call does, and only if you make one.

Putting it together

The current stack reads like a checklist for a serious research assistant:

The agents-remember story (the MCP memory loop) plus this release adds up to a simple pitch: Geneziz is where your knowledge works for you: captured automatically, answerable on demand, and briefed to you on schedule.

---

Geneziz is a local-first AI knowledge engine for macOS and Windows: bookmarks, stars, and web captures become a Markdown knowledge base you own, with full-text + semantic search and an MCP server included. The features above ship in the current release.

→ geneziz.app