The semantic index decides what counts as relevant. When Copilot grounds a prompt, it searches Microsoft Graph through the semantic index rather than with a plain keyword search. Because it works on meaning, a question about "the client who was unhappy" can match an email that says "frustrated" or "disappointed". Whatever it turns up is appended to your prompt before the model writes the answer.
You never use the semantic index directly; it runs behind every grounded prompt. It also explains why answers can vary: it ranks what looks most relevant right now, so a renamed or recently active file can change what gets picked. Like the rest of grounding, it only surfaces content you have permission to see.