Streaming

Set stream: true to receive tokens as they are produced, rather than waiting for the complete response.

cURL

curl https://api.resetdata.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $RESETDATA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zai/glm-5.2",
    "messages": [{ "role": "user", "content": "Write a haiku about Sydney." }],
    "stream": true
  }'

Python

stream = client.chat.completions.create(
    model="zai/glm-5.2",
    messages=[{"role": "user", "content": "Write a haiku about Sydney."}],
    stream=True,
)

for chunk in stream:
    if not chunk.choices:
        continue
    delta = chunk.choices[0].delta.content
    if delta:
        print(delta, end="", flush=True)

JavaScript

const stream = await client.chat.completions.create({
  model: "zai/glm-5.2",
  messages: [{ role: "user", content: "Write a haiku about Sydney." }],
  stream: true,
});

for await (const chunk of stream) {
  const delta = chunk.choices[0]?.delta?.content;
  if (delta) process.stdout.write(delta);
}

The response

Server-sent events, each a chat.completion.chunk carrying an incremental delta, terminated by data: [DONE]. Any OpenAI-compatible SDK parses this for you.

Reassemble by concatenating choices[0].delta.content across chunks. The first chunk typically carries the role rather than content, and the final chunk carries finish_reason — so check for null content rather than assuming every chunk has text. Some models also emit a final chunk with an empty choices array (the usage/terminator chunk), so guard choices[0] before indexing it, as the examples above do.

Why streaming is worth using

Beyond appearing faster, streaming has a real operational advantage: once tokens are flowing you have continuous evidence the request is alive.

A non-streaming request to a busy model can sit silent for a long time — first waiting in the queue, then generating. Intermediate proxies are far more willing to cut a silent connection than an active one.

Streaming does not shorten the queue wait

Nothing is emitted until the model actually begins generating. Time spent queued is silent whether or not you asked for streaming, so your client timeout still has to accommodate the full wait.

Token usage while streaming

Ask for usage explicitly if you want it in the stream:

{ "stream": true, "stream_options": { "include_usage": true } }

This appends a final chunk carrying token counts. Support varies by model — check supported_parameters per Model parameters. Either way, the request is billed identically; this only affects whether the counts come back inline. Your authoritative record is always Usage.

Handling disconnects

If the connection drops mid-stream you keep the partial text but there is no way to resume — a retry starts a new generation, and is billed as one. For long generations, persist chunks as they arrive so a failure late in a response doesn’t discard everything before it.