How your assistant answers questions

Your agent doesn't just guess from general knowledge — it uses retrieval-augmented generation (RAG): when someone asks a question, the platform first searches your knowledge base for the most relevant passages, then hands those passages plus the question to the underlying AI model to compose an answer.

That has a few practical consequences:

• Answers are grounded in what you uploaded. If your knowledge base doesn't contain the information, the agent may say it doesn't know, or fall back to general knowledge depending on its instructions — it won't invent facts about your specific documents.
• What you upload matters more than how you phrase questions. If answers are missing or wrong, the most common fix is improving the source document, not the agent's instructions.
• Freshly uploaded or edited content isn't instant. New files are processed and indexed in the background before they're searchable — see the note in Build a knowledge base.
• Recordings are answered from their transcript. Audio and video are turned into text first, with a timestamp at the start of each paragraph. The agent sees the words that were spoken, not tone, music or anything on screen.
• Instructions shape tone and boundaries, not facts. An agent's custom instructions control how it responds (tone, what it should refuse, formatting) — the facts it draws on still come from the knowledge base.