In Routing you decide which collections are searched first, using keywords, domain rules, intention, semantic/keyword balance, and relevance thresholds.

  • Collections: Organize your knowledge into semantic search collections such as support, products, billing, or orders. In Routing, the Title column shows the collection name.
  • Keywords: Add keywords to collections so matching queries prioritize the right collection. Keywords affect which collection is checked first; they do not replace semantic matching.
  • Balance: For each collection, use the Balance control to mix meaning-based (semantic) search with keyword search. Move toward Semantic for broader meaning matches; move toward Keywords when exact phrases matter more.
  • Domain routing: You can tell a collection to do nothing special, be ignored for a domain, become the priority for a domain, or be used only when the page domain exactly matches a keyword.
  • Collection intention: Use intention types carefully. For example, an order collection can trigger order-related prompting behavior before order lookup.
  • Relevance: Raise the Relevance threshold for stricter matching. Lower it if you want broader recall. Default is often around 30%.
  • Current page: When the chat widget is open on a product or article page, the bot uses that page address to find matching knowledge and to understand vague questions such as “is it a good product?” as referring to that page.
  • Priority order: Reorder collections so the most important sources are checked first.
  • Conversation logs: On a real turn, open Logs for the conversation to see what reached the model, per-collection scores, retrieval method (semantic / keyword), and the Rout. match column (routing phrases and intention keywords found in chunk text).

Intention Logic (How Routing Decides)

Before knowledge retrieval, the system runs a lightweight intention classification for each user message. The result is one of four labels:

  • human: User asks to speak with a human/agent/support.
  • order: User asks about their invoices, orders, balance, or other personal order/account financial details.
  • eshop: User asks about a product.
  • none: Any other general query.

The classifier uses strict priority human > order > eshop > none if multiple labels could apply.

  • If intention = human: knowledge search is bypassed and the support escalation flow is triggered (according to your escalation settings in Widget).
  • If intention = order or eshop: routing gives first priority to collections with matching intention. If matching collections exist, they are searched first.
  • If intention = none: normal routing applies (keywords, domain routing, collection priority, semantic/keyword balance, and relevance thresholds).
  • For follow-up turns: the previous assistant turn stores internal routing tags, so order/eshop context can continue naturally in the next message.

Table of contents · Previous: Add Knowledge Sources · Next: Customize the Visitor Experience