API Reference

Completions (legacy)

Raw-prompt text completion, the pre-chat OpenAI shape, served for tooling that still speaks it.

Endpoint

POST https://inference.pearlresearch.ai/v1/completions

Unlike chat completions, the prompt string is handed to the model as-is, no chat template, no roles, no system prompt. The model simply continues your text. Streaming works the same way as on the chat endpoint.

Request body

Parameter	Type	Notes
model	string	Required. A catalog model ID from `GET /models`.
prompt	string	Required. Raw text the model continues.
max_tokens	integer	Cap on generated tokens.
sampling parameters	—	`temperature`, `top_p`, `frequency_penalty`, `presence_penalty`, `stop`, `seed` (plus `top_k`/`min_p` extensions) behave as on chat completions.
stream	boolean	Server-sent events with `choices[0].text` deltas, terminated by `data: [DONE]`.

Example

curl https://inference.pearlresearch.ai/v1/completions \
  -H "Authorization: Bearer $PEARL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zai-org/GLM-5.2",
    "prompt": "The three laws of robotics are:",
    "max_tokens": 200,
    "temperature": 0.7
  }'

{
  "id": "cmpl-…",
  "object": "text_completion",
  "created": 1785229039,
  "model": "zai-org/GLM-5.2",
  "choices": [
    {"index": 0, "text": " 1. A robot may not injure …", "finish_reason": "stop"}
  ],
  "usage": {"prompt_tokens": 9, "completion_tokens": 74, "total_tokens": 83}
}

The models in the catalog are instruction-tuned, and their quality comes from their chat formatting, prefer chat completions for anything conversational, and keep this endpoint for continuation-style tasks or legacy integrations that require the prompt/text shape.