Building user-facing AI features that feel fast, reliable, and trustworthy. Streaming UIs, abort controls, error states, and the front-end patterns that turn a raw LLM call into a real product.
Stream by default for any answer longer than a single sentence. Users perceive streaming as faster than synchronous responses of identical content, and abandonment rates drop noticeably.
SSE for one-way LLM streaming — simpler, works over normal HTTP, automatic reconnection. WebSockets only when you need bidirectional real-time (e.g. live collaboration with AI suggestions).
Use a streaming-tolerant markdown parser (react-markdown with a remark plugin, or marked with sanitize). Render incomplete code blocks as plain text until the closing fence arrives. Memoise rendered chunks so React only updates the tail.
Almost always: no abort propagation. Users close tabs mid-response; fetch keeps consuming; OpenAI keeps generating until max_tokens. Add AbortController on unmount and pass req.signal to the provider — the fix is usually 6 lines.
Three states: idle, streaming-with-partial-text, and error. Render a typing cursor at the end of the partial text while streaming. On error, keep what was already streamed visible and show a "retry from here" button — never wipe the partial response.
Build a ChatGPT-style streaming text interface in React that displays LLM responses token by token as they arrive — complete with error handling, loading states, and clean component design.
Wire a React chat interface directly to a local Ollama instance using server-sent events. The result is a private, zero-cost ChatGPT-style UI that streams tokens as fast as your GPU produces them.