Skip to main content
Kyara supports the chat completions endpoint for text generation. Any OpenAI-compatible client can connect, but the capabilities below are not available.

​
Summary

  • Function and tool calling: tools and tool_choice are silently ignored, so the model never sees them.
  • Structured outputs and JSON mode: response_format is silently ignored.
  • Vision: image inputs are rejected. Message content must be plain text.
  • Audio: no audio input or output, speech synthesis, or transcription.
  • Embeddings: no /v1/embeddings endpoint.

​
Available endpoints

These are the endpoints available to API clients. Every other OpenAI API path returns 404 Not Found in the standard OpenAI error envelope.

​
Request parameters

Chat completions requests are validated against an allowlist. Every parameter falls into one of three groups: forwarded to the model, silently ignored, or rejected with a 400 error.

​
Supported

These parameters are accepted and passed through to the model:
  • model, messages
  • temperature, top_p, top_k, min_p, presence_penalty, frequency_penalty, repetition_penalty, seed
  • max_tokens, max_completion_tokens, stop
  • n (only 1 is accepted; any other value is rejected with a 400 error)
  • stream, stream_options (include_usage, include_obfuscation)
  • logprobs, top_logprobs, user, session_id
  • reasoning (object with effort, max_tokens, enabled, and exclude)
  • reasoning_effort (folded into reasoning.effort; an explicit reasoning.effort wins)

​
Silently ignored

The following parameters do not produce an error. For client compatibility, Kyara removes unrecognized top-level parameters from the request instead of rejecting it. Sending tools or response_format returns a normal text response. The model never sees those fields, so tool calls and structured output simply will not happen.
  • tools, tool_choice, functions, parallel_tool_calls
  • response_format (JSON mode and structured outputs)
  • logit_bias, prediction
  • modalities, audio, web_search_options
  • store, metadata, service_tier

​
Rejected with an error

The following request shapes return a 400 validation error. The error body follows the OpenAI envelope; message explains the problem and param names the offending field, for example messages.0.tool_calls.
  • Message content that is not a plain string. Multimodal content arrays containing image_url or audio parts are rejected.
  • Message roles other than system, developer, user, or assistant. There is no tool role.
  • Unknown fields inside messages, stream_options, or reasoning objects. Message objects accept only role, content, name, reasoning, reasoning_content, and reasoning_details.
  • A session_id longer than 256 characters.

​
Notes

  • top_k, min_p, repetition_penalty, and session_id are Kyara extensions beyond the OpenAI standard. See the chat completions reference for details.
  • Requests without max_tokens or max_completion_tokens get a default output limit of 8192 tokens. Explicit values above the model’s maximum output length are lowered to that maximum.
  • To control reasoning, send "reasoning": { "effort": "medium" }. The top-level reasoning_effort also works and is folded into reasoning.effort, with an explicit reasoning.effort winning.
  • Support for sampling parameters varies per model, so check the catalog.
This list can shrink. Unsupported today does not mean unsupported forever. Check back here as the platform evolves; newly added capabilities will move off this page and into the API reference.