Smart Segments

Smart Segments split one agent's knowledge into named areas — Products, Support, Docs — routed by the page the visitor is chatting from. A visitor on /products/… gets product answers first; a visitor in your help center gets support answers first. The rest of the knowledge base stays reachable: segments prioritize, they never wall off.

How a conversation gets its segment

1. The visitor opens the chat on some page. The page's URL path is matched against every segment's path list — the most specific (longest) match wins; unmatched pages go to the default segment.
2. The winning segment is stored on the conversation once, at creation. It owns the conversation's presentation and behavior — persona, greeting, starter questions, and the per-segment behavior overrides — and never changes mid-conversation, so the chat stays one coherent thread.
3. Knowledge follows the visitor. Each answer searches the segment of the page the visitor is on right now — the current page is the strongest signal of what a question is about, and a visitor who navigates from one category to another gets the new category's knowledge immediately, with the full conversation history intact. A turn without a page signal (API clients, inbound channels) uses the conversation's own segment.
4. Follow-ups keep their subject. A visitor who asks "Where was Jan Roel born?" and then "And in which year?" is really still asking about Jan Roel — but that bare follow-up carries no subject to search on. Before retrieving, the assistant widens a short or "and …"-style follow-up with the previous question so the search still finds the right facts. (The reply itself always sees the full conversation; only the behind-the-scenes search query is widened.)
5. The current page is preferred, but the rest of your knowledge base is never walled off. The home segment's chunks lead every answer. Any remaining room is then filled with the strongest matches from other segments — one slot held open per distinct other segment that has a match, filled one-per- segment first. So when the answer is split across two different pages (say the birth place lives on one page and the birth year on another), a chunk-rich page can no longer crowd the other one out: both facts reach the answer instead of a false "I couldn't find that". A clearly stronger match from another segment leads outright. When the home segment has nothing at all, a single boosted fallback pass answers from the whole base — and that pass is segment-diversified too, so one busy page can't fill every slot.
6. Only when the whole knowledge base has no grounded answer does Strict Knowledge Mode defer — a segment miss alone never triggers the fallback message.
7. Visitors can see the active context. When the current page maps to a non-default segment, the open panel's header shows a subtle "Context: <segment name>" chip, so visitors understand why the assistant is answering with knowledge specific to that page. The chip mirrors the exact scope retrieval uses (it disappears on default-segment pages and for agents without segments), and the segment's name is what renders — name segments something visitor-friendly.

URL path patterns

• Patterns match the URL path only — query strings and #fragments are ignored; matching is case-insensitive.
• A path matches itself and everything under it: /products also covers /products/red-pen, but never /products-old (boundaries are respected).
• A trailing /* is accepted as the same section match: /docs/* ≡ /docs.
• * alone matches every page — useful as an explicit catch-all on a non-default segment.
• The longest matching pattern wins across all segments: with /products (Products) and /products/manuals (Docs), a visitor on /products/manuals/v2 lands in Docs.

Setting it up

1. Open your agent → Smart Segments and add segments. The first one you create becomes the default (you can move the star later). Give each segment the URL paths that should route into it; the default usually needs none.
2. On the Sources page, assign each knowledge source to a segment with the per-row picker. Assignment is instant — membership lives in the database, so no re-crawl or re-index happens when you move a source.
3. Sources you never assign behave as default-segment knowledge, so an agent that predates segments keeps answering exactly as before.
4. Prefer moving an existing source with the per-row picker over re-adding the same URL. If you do add a page that another source already indexed (a common one: a page the Auto-indexed from visitors source grabbed first), the new source now crawls its own copy so it can be scoped to your segment — it won't show up empty. Reindex the source if it was added before this behavior shipped.
5. Rehearse in the Playground: the Scenario tab gets a Knowledge segment override so you can ask "what would a /products visitor get?" without opening the live site. The Diagnostics tab shows whether the reply came from the home segment or via the fallback.

Category chrome — persona, starter questions, greeting

Each segment can present itself as a category specialist. All fields live in the segment's editor on the Smart Segments page and are optional — an empty field inherits the agent-level value, so a segment with no chrome behaves exactly like the plain agent.

• Starter questions — the suggestion chips a fresh conversation shows. A Products visitor can see "Which sizes do you have?" while a Support visitor sees "How do I cancel?".
• Persona name + avatar — "Jordin, the TV specialist" in the panel header and launcher for this segment's conversations.
• Greeting bubble — a small, dismissible speech bubble that invites the visitor ("Hi, I'm Jordin — can I help?"). Shows after a short delay while the thread is still empty — also above the minimized launcher, so it reaches visitors who closed the widget. A dismissal is remembered per conversation per message: it never nags, but editing the greeting text gives the new message one fresh showing to everyone. Clear conversation (widget menu) resets the whole cold-start chrome — greeting and starter questions — which is also the easiest way to preview what a new visitor sees.
• Disclaimer — every chat panel carries a one-line "can occasionally be wrong, like a human" notice. Hide or reword it per agent via theme.disclaimer (false or a custom string).

Inline embed block

Besides the floating bar, you can place a chat invite inside page content — e.g. halfway down a category page. Add an empty container anywhere in the page's HTML:

<div data-pitchbar-inline></div>

The widget renders a compact card there (persona, greeting, starter chips, input). Clicking a chip or the input opens the main chat with that question — one conversation, no matter how many blocks the page hosts.

Product context — "ask about this product"

On product pages there is no segment per product (a shop has thousands). Instead the widget reads the page's Schema.org Product JSON-LD — which WooCommerce, Shopify themes, and most shop platforms emit by default — and sends name, price, availability, SKU, and brand to the backend with each message. The prompt then carries a "Current product" block, so "does this fit a 15-inch laptop?" is answered about that product.

• Visitors on a detected product page get an "Ask about this product" chip as the first suggestion.
• Product fields are whitelisted and length-capped server-side, and injected as data — page content can never override the assistant's instructions.
• Rehearse it in the Playground: pick the Shopify product page template on the Page-context tab (or paste your own JSON-LD).

Per-segment behavior — instructions, threshold, leads, workflows

• Category instructions — extra system-prompt text for this segment only ("You are the storage specialist; always mention unit sizes"). Appended after the agent's own instructions; it can never override the grounding rules or Strict Mode.
• Confidence threshold override — a stricter or looser retrieval gate per segment. Empty inherits the agent's threshold.
• Lead capture — per segment: inherit, ask for contact details as usual, or never ask (e.g. a support portal where a lead form feels wrong).
• Workflows per segment — on the workflow editor, scope a flow to one segment: it then only fires for conversations routed to that category. Unscoped workflows keep firing everywhere.
• Analytics & alerts — the Conversations list shows a segment badge per conversation, and new-lead emails name the segment the lead came from.

What segments do NOT change

• Curated answers, workflows, and the human-handoff / booking shortcuts run before retrieval and ignore segments entirely.
• The widget embed is unchanged — routing happens server-side from the page URL the widget already reports. Nothing to re-install.
• Agents with zero segments behave byte-identically to before the feature existed.
• Deleting a segment re-points its sources at the default segment (deleting the last one turns the feature off); conversations pinned to it simply fall back to whole-knowledge-base retrieval.

Plans

The per-plan Smart Segments per agent limit controls the feature: blank means unlimited, 0 hides it, a number caps how many segments each agent may define. Platform admins set it under Admin → Plans.