SleekAI opens a server-sent events connection from the PHP request to the upstream provider with stream: true in the request body. PHP flushes each chunk to the browser as it arrives. The chat widget consumes the EventSource on the client and appends tokens to the active assistant bubble until the [DONE] sentinel arrives.

Yes. Anthropic exposes streaming on the messages API with stream: true, and SleekAI handles the slightly different event shape (content_block_delta versus chat.completion.chunk) transparently. From the WordPress side the configuration is identical to a streaming OpenAI bot.

Some WordPress hosts run a WAF or reverse proxy that buffers responses and breaks SSE. SleekAI sends the standard headers (Content-Type: text/event-stream, Cache-Control: no-cache, X-Accel-Buffering: no) to tell Nginx and most CDNs to pass chunks through. If your stack still buffers, disable buffering on the chat endpoint route.

Yes. Each chatbot has a streaming toggle. Turn it off when the upstream provider does not support SSE, when you are debugging a structured output use case, or when a downstream tool expects a full JSON blob rather than a chunked response.

Most streaming providers send usage data in the final chunk before the [DONE] marker. SleekAI captures that and writes prompt and completion token counts into wp_sleek_ai_chats alongside the full reply text. On providers that omit usage, SleekAI estimates with the encoding for the configured model.

If the visitor disconnects mid-stream, SleekAI still writes a row to wp_sleek_ai_chats with whatever text was streamed so far and the tokens consumed up to that point. The row is marked as a partial response so you can filter for it when auditing failed conversations.

Object caching is fine, but full-page cache plugins can intercept the chat endpoint and cache the SSE response, which obviously breaks streaming. SleekAI registers the chat endpoint as non-cacheable through standard WP filters, and the docs include an exclude list snippet for WP Rocket, W3 Total Cache, and Litespeed.

Yes. The chat widget shows a typing indicator from the moment the request fires until the first token arrives, then switches to the streaming bubble. The indicator returns briefly during long inference pauses if the upstream stops sending tokens for more than two seconds.