A Chat is the assistant itself: which content it answers from, how it speaks, what it refuses, and how it looks. A site can run several — a public docs assistant on one page, a members-only helper on another — each with its own settings.
Go to Wyregraf → New Chat, give it a title, and work through the tabs.
The Main tab #
- The title identifies the chat in your admin; visitors don't see it.
- Chat Type is the big decision:
- Retrieval — answers are grounded in your indexed content, with citations. This is the type to use for a docs or support assistant.
- Chat — a plain conversation with the model, no retrieval.
- Agent Network (Pro) — the chat is driven by a network of agents for multi-step behavior.
- Namespace scopes which content the chat searches. Embeddings and chats tagged with the same namespace form a collection — so a "Products" chat can search product content only, while a "Docs" chat searches documentation.
The Context tab #
The Context tab shapes how the assistant behaves and what visitors first see.
- The Retrieval System Prompt is the assistant's standing instruction — who it is, what tone to take, what to do when the content doesn't cover a question. You pick it from your Prompts library (Wyregraf → All Prompts); Wyregraf ships solid defaults there.
- The Welcome Message greets visitors before their first question. The tab also sets the input placeholder and submit button text, plus the chat's language and tone.
For retrieval chats, this tab also holds the accuracy switches — hybrid search, entity safety, planning and reranking, plus citations and answer verification. Hybrid search is on by default; the others are off and safe to ignore until the basics work.
The Restrictions tab #
Restrictions control what the chat may use and how much visitors may ask.
- Top most relevant document (k) is how many matching chunks are handed to the model per question. More context can improve thorough answers but costs more tokens.
- Cosine Similarity Threshold is the minimum match quality a chunk needs to be used. Raise it and the chat sticks strictly to well-covered topics (answering with the no-results message otherwise); lower it and the chat attempts more questions with looser grounding. Turn on the profile's Tuning option to see live scores while you calibrate.
- Logged In limits the chat to signed-in users — the strongest way to keep API costs predictable. The tab's History Limit (how much prior conversation is re-sent per question) and Message Limit (maximum characters per message) bound the token cost of each exchange.
- Disable Citing Sources hides the source citations under answers, if you prefer a cleaner reply.
The Profiles and Style tabs #
- The Embeddings & Vector Profile and
- the LLM Profile default to the site-wide defaults you set in Options — override them here to give this one chat a different model or index.
The Style tab picks the theme, accent color, logo, chat names, and avatars, and the Launcher tab turns the chat into a floating bubble — both are covered in Chat Appearance and the Floating Launcher. The Restrictions tab's burst and usage-limit fields are covered in Usage and Rate Limits.
What to do next #
Publish the chat, then place it on a page.
