Making AI Answer From Your Documents, Not Its Training

The problem it solves

A general model knows a great deal about the world and nothing about your prices, your policies or your customers. Asked about them, it will produce something plausible and wrong.

Retrieval fixes that by finding the relevant passages from your own material first, then asking the model to answer using only those.

How it works, in four steps

1. Split your documents into passages of a few hundred words
2. Index them so similar meaning can be found, not just matching words
3. Retrieve the handful most relevant to the question asked
4. Answer using those passages, with a citation for each claim

Nothing is trained and nothing is fine-tuned. Update a document and the next answer reflects it, which is what makes this practical for a live business.

Why it beats fine-tuning for facts

Where it goes wrong

• Passages split badly, so a table is separated from its heading
• Stale documents retrieved alongside current ones
• Retrieval returning nothing, and the model answering anyway
• No permission filter, so restricted content surfaces

Every one of those is fixable and every one is common in first attempts.

The refusal rule

When retrieval finds nothing relevant, the correct answer is “I do not have information on that”, with a route to a person.

Systems that answer regardless are the ones that destroy trust, because their confident wrong answers are indistinguishable from their confident right ones.

Frequently asked questions