Retrieval decides what the assistant can say. These three settings decide how it says it. They live on the knowledge base, under Knowledge Base → Settings → AI Persona & Instructions, and are prepended to the system prompt on every answer — widget, API, bots, MCP and the Ask panel alike.

​
Persona

One or two sentences naming the assistant and the product it speaks for:
Naming the assistant matters more than it looks: without one the model refers to itself as “the documentation” or “I” inconsistently across a conversation.

​
Tone

Four values, applied as “Respond in a tone.” Default sends nothing at all, leaving the base prompt’s voice — which is already short and neutral, and is the right choice for most developer documentation.

​
Custom instructions

Standing rules that apply to every answer, whatever was retrieved:
This is the place for escalation paths, calls to action, terminology preferences and topics to decline.
Custom instructions steer presentation, not knowledge. They cannot add facts the index does not hold, and they cannot override grounding: when retrieval finds nothing, the abstention gate withholds the excerpts and citations before the model is called, so no instruction can talk it into an answer. That is the point — an instruction is a request, and abstention is enforced.

​
What the assistant already knows about itself

Independent of these settings, every system prompt carries a summary of the subjects present in the index — built from the documents themselves, not from the knowledge base’s name. A knowledge base called docs-v2 or testing-proj is an internal label that tells a reader nothing, so it is not used. This is what lets the assistant answer “what can you help me with?” by describing the corpus as a whole rather than whichever three pages this turn happened to retrieve.
In a multi-knowledge-base question, persona, tone and instructions come from the primary knowledge base only. The Ask panel and the portal API can search several at once, and there is no principled way to blend two personas — so the first knowledge base in the request supplies all three. The corpus summary spans every knowledge base being searched.This mirrors bring your own model, where a multi-knowledge-base question likewise ignores per-knowledge-base overrides.

​
From the API

Any of the three can be set with PATCH /api/v1/knowledge-bases/:id. Send null to clear one back to the default:

​
Getting it right

Changing how the assistant speaks is not measurable by reading a few answers — a persona that reads well on three questions can degrade the tenth. Score the change the way you would score a retrieval change, with an eval set run before and after.