Every local AI feature starts with the same question, and it isn’t “which model is best?”. It’s “what kind of data do I have, and what do I need to know about it?”. The answer narrows the choice to one or two kinds of model before you compare any of them.

​
Most models read text

Language models, embedding models and decision models all take text as input. That makes text the common language of an AI pipeline. Anything that isn’t text, an image, a scan, a recording, usually goes through a conversion step first. Once you have text, the second question is what you want from it.

​
OCR or a vision model?

Both read images, but they answer different questions. A vision-language model is good at “what’s in this picture?” and at reasoning about layout or diagrams. OCR is good at “what exactly does it say, and where?”. If you need the precise characters, for search, for matching a pattern, or for covering them, OCR is the better fit. If you need positions, OCR is often the only fit: its boxes are in the image’s own pixels, while a vision model’s sense of “where” is approximate. OCR models are also much smaller and faster, which matters when everything runs on a laptop. A vision model is worth its size when the meaning isn’t in the words: a photo of a rash, a chart’s trend, whether a form is signed.

​
Speech: transcription first

The same logic applies to audio. If you need to act on what was said, transcribe it and work with the text. If you need to know when something was said, to cut a clip or jump to a moment, choose a transcription model that returns timestamps.

​
Then pick the smallest model that works

Within a kind, size is a trade: bigger models usually handle harder cases, but they take more disk, memory and time on every run. Start small, measure on real examples of your data, and only move up when the small one fails in ways that matter. QVAC’s assess model fit checks whether a model fits your machine before you download it.

​
How Redact chose

Redact’s input is screenshots and scanned PDFs, so the first step is turning images into text. It needs exact characters and exact positions, because it paints boxes over specific pixels, so it uses OCR rather than a vision model. Its question about each line is fixed in advance, “is this a name, an email, an account number?”, so it uses a decision model, Laya, rather than a language model.