Both. Visitors can ask one question and get a direct answer with a deep link, or open a back-and-forth conversation if they want to drill in. The widget defaults to one-shot mode for search behaviour, and follow-ups expand into a conversation when the visitor types another question. Either pattern works from the same install.

For most WordPress sites, yes - and at a fraction of the cost since you only pay your model provider rather than per-query SaaS pricing. Algolia still wins for very specific use cases like real-time facet filtering on millions of records, but for documentation, blogs, and most catalogs the AI search bar gives a better experience at lower cost.

Yes. Display conditions plus a custom WP_Query limit it to the docs, the blog, or any subset. You can also run multiple search bars on the same install via Multibot - one scoped to the docs, one to the shop, each with its own prompt and grounding. Different sections get different experiences without separate plugins.

Yes. The widget loads via JS so caching has no effect on its behaviour. WP Rocket, LiteSpeed, and standard object caching all coexist with the bar without configuration. The model request goes directly to your provider, bypassing the cached page entirely, so search results are always live.

Every element is themable via CSS variables and the widget supports a fully custom slot for teams that want pixel control. Most installs match the brand in five minutes by setting the brand colour and typography variables; deeper customisation lives in the slot, where you can render the entire UI yourself if needed.

Yes. Activate with cmd+k by default, or any custom key combination. Esc closes the search, arrow keys navigate between suggestions, and enter submits. The shortcut is configurable per site, so if cmd+k clashes with another tool you've embedded you can switch to a different combo without code changes.

The system prompt is where you set the policy. Most teams ask the bot to confirm the topic if the query is one or two words, and answer directly if it's a sentence. That avoids the failure mode where a single word matches three unrelated articles and the bot picks the wrong one. The behaviour is tunable, and the logs show whether the policy is working.

Yes, when configured. The system prompt can include a directive to summarise when the matched article is over a certain length, with the deep link still attached. That's useful for documentation portals where the canonical answer is buried halfway through a 4,000-word reference doc; the summary becomes the elevator pitch and the link delivers the full context.