Answer modes
The four settings controlling where an agent is allowed to source its answers, and how to choose between them.
Answer mode is the single most important guardrail setting. It decides where an agent is allowed to get its answers from.
Set it on the agent's Guardrails tab. The default is grounded only.
The four modes
Grounded only (default)
The agent answers strictly from retrieved knowledge and confirmed tool results. If neither contains the answer, it says so rather than guessing.
Use for: customer support, policy questions, anything where a wrong answer is costly.
Grounded with limited inference
As above, plus safe logical inference from what it retrieved. The agent must distinguish what came directly from a source and what it worked out.
Use for: internal assistants, and cases where a strictly literal answer is unhelpfully narrow.
Tools only
The agent answers exclusively from tool output. Knowledge sources are not used as an answer source.
Use for: agents whose whole job is looking things up in a live system — order status, account balance, stock levels.
In this mode, attaching knowledge sources will not make the agent answer from them. If your agent seems to be ignoring its knowledge, check whether this mode is selected.
General knowledge allowed
The agent may draw on the model's own general knowledge as well. It must still attribute what is grounded versus general, and must never fabricate specific facts or statistics.
Use for: general assistants, educational agents, anything conversational where being unable to answer common questions would be frustrating.
Choosing
| If you want the agent to… | Use |
|---|---|
| Never say anything not in your content | Grounded only |
| Reason a little from your content | Grounded with limited inference |
| Only report what a live system says | Tools only |
| Behave like a general assistant too | General knowledge allowed |
The tradeoff
Stricter modes produce fewer wrong answers and more "I don't know" responses. Looser modes answer more questions and carry more risk of a confident, wrong answer.
If your agent refuses too much, the fix is usually more knowledge, not a looser answer mode. Check what it is failing to answer, add that content to a knowledge source, and test again.