language model engines

The insight stage — titles, summaries, lenses — is run by one of two engines on your machine. You don't pick between them; yapyap decides for each request:

• if Ollama is running, yapyap uses it. It's the setup you manage, and it may already have the model.
• otherwise, yapyap's built-in engine runs the same model itself — downloading it first only if no copy exists anywhere on your machine.

You get the same model and the same results whichever one answers, so switching between them never changes anything mid-conversation.

the built-in engine

The built-in engine needs zero setup and takes care of itself:

• it installs itself — the first time it's needed, yapyap fetches a small one-off download and reuses it from then on.
• it's private — it only ever runs locally and is never reachable from your network.
• it starts when needed and stops when idle — the first use after a break takes a moment to load the model, and after about 15 minutes of inactivity it shuts down so it isn't holding gigabytes of memory between recordings.
• it uses your graphics card when it can, and your processor otherwise.

very long recordings

Every model has a limit on how much text it can consider at once. A transcript longer than that limit is caught with a clear message rather than being quietly cut short — pick a model that handles more (Qwen handles the most) or run the lens on a shorter recording.

why Ollama wins when it's there

Running Ollama is entirely optional. But when it's up, yapyap prefers it: it's the install you manage, it holds your own library, and updates and graphics settings stay under your control. The built-in engine is simply the fallback that makes Ollama optional.