Promptyard
Format

SSE / Streaming Response Viewer

Parse raw SSE bytes from any LLM streaming endpoint into events with reconstructed text.

Events
5
Parse errors
0
With JSON data
4
Reconstructed chars
13
Reconstructed text

Hello, world!

Events
  1. #0
    {
  "choices": [
    {
      "delta": {
        "role": "assistant"
      }
    }
  ]
}
  2. #1
    {
  "choices": [
    {
      "delta": {
        "content": "Hello"
      }
    }
  ]
}
  3. #2
    {
  "choices": [
    {
      "delta": {
        "content": ", world"
      }
    }
  ]
}
  4. #3
    {
  "choices": [
    {
      "delta": {
        "content": "!"
      }
    }
  ]
}
  5. #4
    [DONE]

What is SSE in LLM APIs

Server-Sent Events is a one-way streaming protocol where the server pushes a sequence of plain-text events to the client over a long-running HTTP response. Each event is one or more field: value lines terminated by a blank line. LLM streaming endpoints encode each delta as one event, with the JSON payload in the data: field.

Compared to WebSockets, SSE is simpler — uses regular HTTP, works through proxies, supports auto-reconnect via Last-Event-ID — at the cost of being uni-directional. Perfect fit for "model talks, client listens".

FAQ

Where do I get the raw SSE bytes?
curl -N https://api.openai.com/v1/chat/completions -d '{...}' or copy from your browser DevTools Network panel (right-click → Copy response). Paste into the textarea here.
How does it know to "reconstruct" the streamed text?
It probes for three known shapes: OpenAI choices[0].delta.content, Anthropic content_block_delta.delta.text, and Gemini candidates[0].content.parts[0].text. If your response uses a different shape, the events render fine but reconstruction stays empty.
Why does the [DONE] line cause no error?
Every parser handles it: it's the OpenAI sentinel for "stream complete". The viewer shows it as a non-JSON event without flagging it.
Are partial / truncated streams handled?
Yes. Events are split on blank-line boundaries; an incomplete final event without trailing blank line is parsed best-effort. Trailing partial JSON is shown as a parse error.

Common pitfalls

  • Splitting on single newlines instead of blank-line boundaries.
  • Forgetting that data: lines can repeat within one event — concatenate with newlines.
  • Passing the literal string [DONE] through JSON.parse and erroring out.
  • Not handling backpressure on the client; long streams can balloon memory if you store every chunk.

Related tools