Dialect implementation for the OpenAI Chat Completions wire format.
See Omni.Dialect.Text for the behaviour specification and delta types.
Compat flags
The Chat Completions format is spoken by many providers, most with small
deviations from OpenAI's own API. With no flags, the dialect behaves
exactly like OpenAI's API; each flag describes one deviation. Set them on
a provider's endpoint for this dialect,
dialects: %{Omni.Dialects.OpenAICompletions => [compat: [...]]} (see
Omni.Provider).
| Flag | Default (OpenAI) | Setting it |
|---|---|---|
:max_tokens_field | "max_completion_tokens" | sends :max_tokens under this field |
:thinking_effort_field | "reasoning_effort" | sends the effort level here (dots nest: "reasoning.effort") |
:thinking_budget_field | no budget sent | sends a token budget here, for models that take one and list no effort values |
:thinking_replay_format | earlier reasoning not sent back | how reasoning is sent back on assistant messages (below) |
:thinking_toggle_format | off is reasoning_effort: "none"; no on-switch | how thinking is switched on and off (below) |
:output_format | :json_schema | :json_object: JSON mode, with the schema in the system prompt |
:attachment_format | :file | :file_url: non-image files as file_url content parts |
:omit | [] | never sends these: :cache, :thinking, :strict_schema |
:extra_body | %{} | merges these literal fields into every request body |
:thinking_toggle_format values:
:thinking_type-thinking: {type: "enabled" | "disabled"}:enable_thinking-enable_thinking: true | false:reasoning_enabled-reasoning: {enabled: true | false}
:thinking_replay_format values:
:none- reasoning is not sent back:reasoning_content- sends the message's thinking text in areasoning_contentfield on that assistant message:reasoning- the same, in areasoningfield:reasoning_details- capturesreasoning_detailsfrom responses and sends them back on that assistant message; streamed fragments are merged into one entry per reasoning block when the stream completes:content_chunks- sends the message's thinking as one leading{type: "thinking", thinking: [{type: "text", ...}]}part of acontentarray, before its text parts
Models that list effort values (see Omni.Model) take the effort, even
when a toggle format is set; the toggle switches them off, and switches on
models without effort values. Likewise a budget is sent only to models
without effort values.