A normalised inference request.
Every provider receives an ExAthena.Request and is responsible for mapping
it to its own wire format. Keeping the shared shape here means callers don't
need to think about whether they're talking to Ollama, OpenAI, or Claude.
Summary
Functions
Build a request from a raw user prompt and options.
Types
@type reasoning_effort() :: :none | :minimal | :low | :medium | :high | :xhigh
How hard a reasoning model should think before answering.
The ladder is req_llm's canonical one, so a value set here means the same
thing to every provider that understands the concept. nil sends nothing
and leaves the model on its own default — which for Qwen3.8 is xhigh, the
setting that spends whole completion budgets circling (#198).
@type t() :: %ExAthena.Request{ max_tokens: pos_integer() | nil, messages: [ExAthena.Messages.Message.t()], metadata: map() | nil, model: String.t() | nil, provider_opts: keyword() | nil, reasoning_effort: reasoning_effort() | nil, response_format: :text | :json | map() | nil, stop: [String.t()] | String.t() | nil, system_prompt: String.t() | nil, temperature: float() | nil, timeout_ms: pos_integer() | nil, tool_choice: :auto | :any | :none | map() | nil, tools: [map()] | nil, top_p: float() | nil }
Functions
Build a request from a raw user prompt and options.
The prompt is prepended to opts[:messages] as a user message. Pass nil
to start from a pre-built message list.
Images shorthand
Pass images: [%{data: binary(), media_type: String.t()}] to attach inline
images to the trailing user message. Each entry may also use
%{url: String.t()} for remote image URLs. When a non-empty prompt is
given, a multimodal user message is built with the text first followed by
image parts. When prompt is nil or "", image parts are merged into the
last user message in :messages, or appended as a new user message.