The models, and the brain behind them.

A small piece of text. Prices and context limits are usually counted in tokens, not pages.

How much the model can consider at once. A larger window does not guarantee a better answer.

The trained numbers inside a model. Open weights can be downloaded; closed weights stay on the provider’s servers.

Local runs on your machine. Cloud runs elsewhere through an app or API. Local is not automatically private; every connection still matters.

A documented way for software to send a job to a model and receive the result.

A model-size measure—not a quality score. Test models on your real job instead of choosing the biggest number.

Turning a sentence into numbers so similar ideas sit near each other, even if they do not share the same words.

A store of those numbers. Ask in your own words; it brings back notes that mean the same thing.

A back-of-the-book index. It finds a name, a date, or a model number exactly as written.

An optional map of who is tied to what. Useful for “who decided this?” — not required for ordinary search.

Retrieve first, then generate: the program finds notes, then the model writes the answer from those notes.

A plug that lets a model use a tool (files, mail, a calendar) instead of only talking about it. Credentials stay on the host.