Configure grounded Chat AI
Prepare grounded AI conversations with appropriate knowledge sources and a clear escalation path.
On this page
Overview
When AI is enabled and Live Chat mode is ai_first or hybrid, the first tab is Chat AI. It answers factual support questions only from retrieved allowed Fulltext + Chunk text. Internal KB and internal notes are excluded. retrieval_status no_evidence, search_error, or handoff renders a non-answer plus a Talk to human chip if handoff is on. After N unresolved AI responses (live_chat_ai_failed_turns, default 3) the widget can offer or request a human. Hybrid lets a human take over while AI may still have assisted. Human only hides Chat AI and uses the Chat tab for people.
When to use this
Visitors get 24/7 deflection on documented topics without letting the model invent policy. Failed turns and no-evidence are how you protect the brand.
Where to find the controls
Widget Chat AI tab. Provider: Settings → AI. Mode and failed-turns: Live Chat settings. Corpus: #/kb + RAG index. Agents watch Inbox Live Chat when a session needs a human.
Who can make these changes
Visitors. Agents do not need use_ai to take over a session. manage_ai configures models. use_ai is for admin Copilot, not a requirement for the public widget.
Before you begin
AI feature on, provider tested, RAG indexed, mode ai_first or hybrid, documents enabled if you want the Document tab alongside answers, handoff enabled if humans should be reachable, rate limits left on. Falls back to human only when AI is unavailable.
Follow these steps
- Index KB.
- Set mode.
- Open the widget.
- Ask a documented question — expect a grounded answer.
- Ask an undocumented question — expect no_evidence, not a novel policy.
- Click Talk to human.
- Confirm waiting/handoff UI and Inbox.
- Repeat after toggling hybrid vs AI first to train agents on takeover.
Settings reference
Mode, retrieval_status, citations if shown, failed-turns counter, Talk to human chip, Ask AI chip, rate-limit errors, attachment rejection (files are not a way to bypass grounding).
Example workflow
AI first: visitor asks shipping times documented as “3–5 days”. Answer cites that chunk. Next question about a secret discount code has no chunk — no_evidence + human chip. After three weak turns, the widget prompts handoff automatically.
Good practices
Point visitors to the Document tab for long procedures. Keep prompts from instructing the model to “be helpful even without sources”. Test in Vietnamese and English if you author both KB languages.
Important considerations
Factual Chat AI without documentation is not supported in grounded mode. Do not paste internal notes into public articles to “help the bot”. Rate limits returning 429 are success of security, not a broken bot. Do not add unofficial engines.
If something does not work
AI no answer: RAG, internal article, mode human only, feature off, provider test failing, or rate limit. Hallucinated-looking text: treat as a bug if retrieval_status was no_evidence — contact support with the session id and status, not prompts that include PII.