Modes
Every assistant operates in one of four modes that control how it behaves during conversations. You choose the mode when creating or editing an assistant.
Open / General
The default mode. The assistant has no constraints on how it converses — it can discuss any topic, use any enabled tool, and respond freely. This is ideal for general-purpose assistants like content writers or research helpers.
Knowledge Only
In this mode, the assistant is required to search the project knowledge base before answering any factual question. If the knowledge base doesn’t contain relevant information, the assistant will tell the user it doesn’t have that information rather than guessing or using general knowledge. The assistant will also decline creative requests (poems, songs, stories, etc.) that are unrelated to its knowledge domain.
This is useful when you need the assistant to stay strictly within the boundaries of your documented information — for example, a customer support agent that should only reference your actual product documentation.
Technical Details
How modes work — Each mode injects specific behavior rules into the system prompt. In Open/General mode, no constraints are added. In Knowledge Only mode, boundary instructions require the assistant to search the knowledge base before answering factual or domain-specific questions.
Knowledge Only guardrail — Beyond prompt instructions, Knowledge Only mode has an active enforcement mechanism. If the assistant generates a substantive response (more than 50 characters) without having called the search_knowledge tool first, the system intercepts the response before delivering it. It injects a correction message reminding the assistant to search knowledge, then re-runs the LLM call. This prevents the assistant from relying on its general training data when it should be consulting your knowledge base.
Guided modes — Interview and Discovery modes inject the full guided script (questions, field keys, required flags, follow-up notes), the current progress tracking state (which questions have been answered, which are pending), goal description, and completion action instructions into the system prompt on every turn. This gives the assistant continuous awareness of where it is in the conversation flow.
Configuring Knowledge Search
All knowledge search settings — boundary instructions, similarity threshold, and source type filtering — are configured through the search_knowledge tool’s configuration page. Go to the Tools tab on the assistant detail page and click Configure next to search_knowledge.
See Tool Configuration for details on the available settings.
Guided - Interview
Interview mode turns the assistant into a structured interviewer. You define a set of questions in the guided script, and the assistant must collect answers to all required questions before concluding the conversation.
The assistant can ask the questions naturally and in any order — it doesn’t have to read them verbatim. But it will keep circling back to unanswered required questions until everything is covered.
This is ideal for intake forms, onboarding flows, or data collection scenarios.
Guided - Discovery
Discovery mode is a more flexible version of interview mode. The assistant uses the guided script questions as a conversation guide but can adapt freely based on what the user says. It can skip irrelevant questions or explore tangents if they’re productive.
Once the assistant has gathered enough context, it will automatically use the primary tool you’ve configured — for example, triggering a product search or generating a recommendation.