Home / News & Blog

What “local-first AI” means, and what it does not

We use the phrase “local-first” on this website, so we owe readers a precise account of what we mean by it. The phrase gets used loosely, and a buyer who takes it at face value can end up with something quite different from what they expected.

What it means

In a local-first system, the work happens on hardware the customer controls unless the customer decides otherwise. For an AI feature, that covers three things: where the documents are stored, where the index built from them is stored, and where the model that reads them runs.

In Sasly Knowledge all three are on the customer’s server. The documents never leave the shared drive, the index sits beside the application, and the language model runs as a separate process on the same machine with no network port.

What it does not mean

It does not mean offline

A local-first product can still have features that reach the internet, such as checking for updates or, if an administrator allows it, calling a cloud model. What matters is that these are settings the customer can see and turn off, and that the product keeps working when they are off.

It does not mean the local model is as capable as the largest cloud models

A model that fits on a single server is smaller. It is good at summarizing a passage, answering a question from a handful of documents, and classifying text. It is weaker at long chains of reasoning. We would rather say that plainly than have someone discover it during a pilot.

Redacted does not mean anonymous

Some products, ours included, can remove names and identifiers from a message before sending it to a cloud model. That lowers the risk. It does not remove it. A message about a specific property dispute is still about that dispute after the names are gone, and someone who knows the context may recognize it. Redaction is a control, and it should be described as one.

Questions to ask any vendor

These are the questions we would ask if we were on the other side of the table. We try to answer each of them in writing before a project starts.

  • Where are my documents stored, and does the product keep its own copy?
  • Where does the model run, and who operates that machine?
  • If a cloud model can be used, what exactly is sent, and can I see a sample?
  • Does the model provider keep what is sent, and for how long?
  • Can I turn AI off entirely and still use the product?
  • Is every AI request recorded somewhere I can review?

Why we work this way

Most of our delivered work is for municipal government. A clerk or a records manager is personally accountable for what happens to the records in their care, and “the vendor’s cloud handles it” is not an answer they can give to a council or to the public. Running the software where the records already are is the simplest arrangement to explain, and the easiest one to audit.

It is not the right arrangement for everyone. An organization without a server room, or without anyone to look after one, may be better served by a hosted product in a Canadian data centre. That is a legitimate choice, and it is one we plan to offer.