How grounding works

When an assistant answers from your documents rather than the model’s general knowledge, that’s grounding (often called RAG, retrieval-augmented generation). Here’s the idea, without the jargon.

The problem it solves

A model on its own only knows what it learned in training. It doesn’t know your handbook, your contracts, or last week’s policy update, and if asked, it may guess. Grounding fixes that by giving the model the relevant text to work from.

How it works

1. When you add a document, it’s split into passages and indexed so it can be searched by meaning, not just keywords.
2. When you ask a question, Ajutant retrieves the passages most relevant to it.
3. Those passages are given to the model along with your question, so it answers from your content.
4. The answer cites the passages it used, so you can check the source.

Why it matters

• Answers reflect your documents, and stay current as you update them.
• Citations let you verify, rather than trust blindly.
• Nothing is sent to train a public model; your content stays in your tenant.

Retrieval settings (how many passages, how strict the match) and the embedding model used for indexing can be tuned per assistant, under its knowledge options.