Crate llm_api 

Canonical model invocation and continuation contract. No Agent or transport implementation.

Re-exports

pub use service::Image;

Modules

service

Structs

BackendCapability

What one backend protocol can express, independent of any model. These are wire facts owned by the adapter, not catalog facts: a protocol that cannot carry a field makes the field unusable even when the model supports it.

Completion

Provider-neutral completion returned to Runtime.

CompletionRequest

One logical LLM invocation.

Continuation

Provider-issued replay data attached to exactly one completed model message. Adapters own the format and compatibility key. Callers must not synthesize or edit payloads.

EffectiveGeneration

Generation parameters that may reach the wire. Fields left None are omitted from the payload, which leaves the provider default in effect.

GenerationControls

What a settings UI may expose after intersecting catalog facts with the backend protocol. This is the same judgment resolve uses; a hidden control cannot appear on the wire, and a shown control is one the current backend can actually carry.

GenerationParameters

Deployment-owned generation settings, never supplied through Agent constraints. None leaves the setting unspecified. These are desired preferences. Adapters omit unsupported fields and clamp reasoning intensity onto the model’s list. Syntax validation does not establish model/provider support.

GenerationSupport

Confirmed per-model parameter support from catalog metadata.generationSupport. Missing information never establishes support; it is never inferred from preferences. Unknown catalog keys are ignored so a newer catalog cannot break an older client.

Message

Provider-neutral conversation message.

ModelCapabilities

Confirmed model capabilities. Missing information never establishes support.

ModelConstraints

Caller-declared requirements. Listed input modalities and true flags require confirmed support; an empty input list and false flags impose no requirement. Adapters must not infer or override these declarations from message content.

ModelMode

Logical model mode selected by the product, independent of provider and model identifiers.

ModelProfile

Provider-neutral limits for one frozen logical model route.

TokenUsage

Normalized token accounting reported by an implementation.

ToolCall

Tool call returned by an LLM.

ToolDefinition

A callable function exposed to the model. Execution policy belongs to the caller.

UseCase

Logical model use case resolved by the deployment’s LLM implementation.

Enums

ContentPart

One typed part of a model message.

FinishReason

Why a model invocation stopped.

MessageRole

Role of a message in a model conversation.

Constants

DEFAULT_MAX_OUTPUT_TOKENS

Internal per-response output cap. Not a user setting. Adapters omit it when the model snapshot says maxTokens is false.

INPUT_AUDIO

Audio input.

INPUT_FILE

Generic file input.

INPUT_IMAGE

Closed vocabulary for non-text model input. text is implied and must not be listed.

INPUT_VIDEO

Video input.

MESSAGE_SCHEMA

Offline normative schema for persisted model messages.

REASONING_EFFORT_LADDER

Reasoning intensity ordered from least to most. Turning thinking off is the thinking switch, never a listed intensity, so none is not on this ladder.

Functions

controls

Settings that can take effect on this backend for this model. Callers must not infer controls from catalog fields alone.

input_modality_for_kind

Closed-vocabulary token for an artifact kind.

input_modality_for_mime

Closed-vocabulary token for a MIME type. Non-media types map to file.

normalize_capability_input

Normalizes catalog capability tokens. text is ignored; unknown tokens fail.

normalize_constraint_input

Normalizes caller-declared constraint tokens. text is illegal here.

normalize_input_token

Maps one catalog/constraint token. text is dropped. vision and unknown values fail.

payload_input_modalities

Distinct payload modalities required by Artifact and Image parts.

resolve

The single resolution of desired generation parameters against model and protocol facts. Every caller resolves exactly once, against facts frozen for the request, so replaying a frozen route always yields the same wire payload.