pig - v0.5.0

pig/provider

Streaming-first provider boundary for pig.

A Provider owns how an inference is started; an Inference owns its cancellable event stream. The coordinator below is the only layer that decides inference terminality, so exactly one Finished event crosses the boundary even when a source races cancellation or emits late events.

Types

Inference

opaque </>

An opaque, cancellable inference handle.

pub opaque type Inference

InferenceEvent

</>

Events emitted by an inference. Finished is emitted exactly once.

pub type InferenceEvent {
  Delta(inference.InferenceDelta)
  Finished(Result(inference.InferenceResult, error.AiError))
}

Constructors

InferenceMetadata

</>

Metadata returned by the provider alongside the message.

pub type InferenceMetadata =
  inference.InferenceMetadata

InferenceRequest

</>

The messages, tools, and settings for one provider call.

pub type InferenceRequest {
  InferenceRequest(
    messages: List(message.Message),
    tools: List(tool_definition.ToolDefinition),
    settings: InferenceSettings,
  )
}

Constructors

InferenceResult

</>

Result of a provider call - the message plus metadata from the API response.

pub type InferenceResult =
  inference.InferenceResult

InferenceSettings

</>

Inference settings passed to a provider.

pub type InferenceSettings {
  InferenceSettings(thinking: ThinkingSetting)
}

Constructors

Provider

opaque </>

An opaque provider that starts asynchronous inference work.

pub opaque type Provider

ThinkingSetting

</>

The thinking configuration for one inference request.

pub type ThinkingSetting {
  UseProviderDefault
  UseThinkingLevel(thinking.ThinkingLevel)
}

Constructors

Values

cancel

</> 
pub fn cancel(inference: Inference) -> Nil

Cancel an inference. Repeated cancellation is harmless.

collect

</> 
pub fn collect(
  inference: Inference,
  timeout_ms: Int,
) -> Result(inference.InferenceResult, error.AiError)

Collect an inference for transitional synchronous callers. Timeout cancels.

default_metadata

</> 
pub fn default_metadata() -> inference.InferenceMetadata

Create an InferenceMetadata with all fields set to None.

default_settings

</> 
pub fn default_settings() -> InferenceSettings

Return settings that defer thinking configuration to the provider.

from_buffered

</> 
pub fn from_buffered(
  call: fn(InferenceRequest) -> Result(
    inference.InferenceResult,
    error.AiError,
  ),
) -> Provider

Build a provider from the existing buffered provider shape.

from_message

</> 
pub fn from_message(
  msg: message.Message,
) -> inference.InferenceResult

Wrap a bare Message into an InferenceResult with default metadata.

from_streaming

</> 
pub fn from_streaming(
  start: fn(InferenceRequest, fn(InferenceEvent) -> Nil) -> Nil,
) -> Provider

Build a provider from a source that may emit normalized events.

from_streaming_with_timeout

</> 
pub fn from_streaming_with_timeout(
  start: fn(InferenceRequest, fn(InferenceEvent) -> Nil) -> Nil,
  update_timeout: fn(Int) -> Provider,
) -> Provider

Build a provider with a provider-specific timeout update.

receive

</> 
pub fn receive(
  inference: Inference,
  timeout_ms: Int,
) -> Result(InferenceEvent, Nil)

Receive the next inference event from its handle.

run

</> 
pub fn run(
  provider: Provider,
  request: InferenceRequest,
  timeout_ms: Int,
) -> Result(inference.InferenceResult, error.AiError)

Start and collect one inference for transitional synchronous callers.

settings_from_string

</> 
pub fn settings_from_string(
  value: String,
) -> Result(InferenceSettings, Nil)

Decode the stable provider-neutral representation of inference settings.

settings_to_string

</> 
pub fn settings_to_string(settings: InferenceSettings) -> String

Encode inference settings using the stable provider-neutral representation.

start

</> 
pub fn start(
  provider: Provider,
  request: InferenceRequest,
) -> Inference

Start inference and return immediately with an opaque handle.

with_default_thinking_level

</> 
pub fn with_default_thinking_level(
  provider: Provider,
  level: thinking.ThinkingLevel,
) -> Provider

Set the provider fallback thinking level. Request-level settings still win.

with_input_tokens

</> 
pub fn with_input_tokens(
  meta: inference.InferenceMetadata,
  tokens: Int,
) -> inference.InferenceMetadata

Set input_tokens on metadata.

with_output_tokens

</> 
pub fn with_output_tokens(
  meta: inference.InferenceMetadata,
  tokens: Int,
) -> inference.InferenceMetadata

Set output_tokens on metadata.

with_response_id

</> 
pub fn with_response_id(
  meta: inference.InferenceMetadata,
  id: String,
) -> inference.InferenceMetadata

Set response_id on metadata.

with_response_model

</> 
pub fn with_response_model(
  meta: inference.InferenceMetadata,
  model: String,
) -> inference.InferenceMetadata

Set response_model on metadata.

with_stop_reason

</> 
pub fn with_stop_reason(
  meta: inference.InferenceMetadata,
  reason: stop_reason.StopReason,
) -> inference.InferenceMetadata

Set stop_reason on metadata.

with_thinking_level

</> 
pub fn with_thinking_level(
  level: thinking.ThinkingLevel,
) -> InferenceSettings

Return settings requesting the given thinking level.

with_timeout

</> 
pub fn with_timeout(
  provider: Provider,
  timeout_ms: Int,
) -> Provider

Apply a provider-specific HTTP timeout when the provider supports it.

✨ Search Document