Retrieval grounds every answer in your indexed content: the visitor's question is matched against your embeddings, the best chunks are handed to the model, and the reply cites its sources. If nothing in your content matches well enough, the chat says so instead of guessing. This is the type for a docs, support, or product assistant.
Chat is a direct conversation with the model — no retrieval, no citations, just the model plus your system prompt. Use it when you want a general-purpose assistant whose answers don't need to come from your site.
Agent Network (Wyregraf Pro) runs the conversation through a pipeline of agents — for example: rewrite the question, retrieve, fall back to conversation history — for multi-step behavior beyond a single retrieval pass.
You pick the type per chat on its Main tab, so a site can run all three side by side. See Creating a Chat.
