Std.Ai.Provider
Types
type alias CustomImpl
type alias WireTransport
type alias WireImpl = { post : String -> String -> String -> Task Error String , model : String , stream : Maybe StreamTransport , measured : Maybe MeasuredTransport }
type alias MeasuredTransport
type alias StreamTransport
type Provider
type alias Message = { role : String , content : String }
type alias Usage = { promptTokens : Int , completionTokens : Int , totalTokens : Int }
type alias ChatResponse = { content : String , usage : Usage }
type alias Options
type alias RawMessage = { content : String }
type alias RawChoice = { message : RawMessage }
type alias RawUsage = { prompt_tokens : Int , completion_tokens : Int , total_tokens : Int }
type alias RawResponse = { choices : List RawChoice , usage : RawUsage }
type AnswerStream
type StreamStep = Piece String AnswerStream | Finished ChatResponse
type StreamEvent
type alias CallFacts
type alias Measured
type alias RuntimeFacts = { slot : Maybe String , tier : Maybe String , adapter : Maybe String , queueWaitMs : Maybe Int , startMs : Maybe Int , latencyMs : Maybe Int , firstTokenMs : Maybe Int , tokensPerSecond : Maybe Float , draftAcceptance : Maybe Float }
type alias Clock = { wallMs : Int , firstMs : Maybe Int }
type alias Timing a = { byRuntime : Bool , adapter : String , call : String -> Task Error ( a, Maybe Int ) , tokens : a -> Int }
type alias StreamClock = { stream : AnswerStream , first : Maybe Int }
type alias ToolSpec
type alias ToolSchema
type alias ToolRequest
type alias ResponseFormat
type ResponseShape a
type alias Structured
type alias ToolCall
type alias ToolResponse
type Turn = Say Message | Called (List ToolCall) | Returned ToolCall String
type alias PriceBook
Values
assistant : String -> Messagechat : Provider -> List Message -> Task Error ChatResponse
Send a chat and get the assistant text plus token usage. Fails with a
chatStream : Options -> Provider -> List Message -> (String -> Task Error ()) -> Task Error ChatResponse
`chatWith`, with each piece of the answer handed to `onPiece` as it is
chatStructured : ResponseShape a -> Options -> Provider -> List Message -> Task Error (Structured a)
`chatWith`, asking for an answer of a shape: the request carries the
chatToolRequest : Provider -> List Turn -> ToolRequest -> Task Error ToolResponse
`chatTools` with each tool's parameter schema and, optionally, the shape of
chatTools : Provider -> List Turn -> List ToolSpec -> Task Error ToolResponse
Like `chat`, but advertises `tools` and returns any `tool_calls` the model
chatWith : Options -> Provider -> List Message -> Task Error ChatResponse
`chat` with per-call `Options` (temperature, seed, max tokens, adapter).
closeStream : AnswerStream -> Task Error ()
Stop reading a stream before its end (the user left the page, the
compatible : String -> Secret -> String -> Provider
Any OpenAI-compatible endpoint — a gateway, vLLM, or a local llama.cpp server
cost : PriceBook -> Usage -> Result Error Money
The money cost of a usage under a price book: tokens/1000 × the per-1k rate,
custom : (List Message -> Task Error ChatResponse) -> Provider
Bring your own backend: wrap any `List Message -> Task Error ChatResponse`
customToolRequest : (List Message -> Task Error ChatResponse) -> (List Turn -> ToolRequest -> Task Error ToolResponse) -> Provider
Like `customTools`, but the provider receives the whole `ToolRequest`:
customTools : (List Message -> Task Error ChatResponse) -> (List Turn -> List ToolSpec -> Task Error ToolResponse) -> Provider
Like `custom`, but the provider ALSO implements native tool-calling: supply
decodeStructured : ResponseShape a -> String -> Result Error a
Decode a model's answer with the shape's codec. An answer that is not JSON
defaultOptions : Options
Every option unset: the server's defaults.
embed : Provider -> List String -> Task Error (List (List Float))
Embed each input text as a vector, in input order, through the server's
freePriceBook : Money.Currency -> PriceBook
A zero-cost book, for a local model or a test.
measure : Options -> Provider -> List Message -> Task Error (Measured ChatResponse)
`chatWith`, measured: the result (a failure is a value here, not a failed
measureStream : Options -> Provider -> List Message -> (String -> Task Error ()) -> Task Error (Measured ChatResponse)
`chatStream`, measured: each piece goes to `onPiece` as it is written,
measureToolRequest : Provider -> List Turn -> ToolRequest -> Task Error (Measured ToolResponse)
`chatToolRequest`, measured, for a native tool loop.
measureTools : Provider -> List Turn -> List ToolSpec -> Task Error (Measured ToolResponse)
`chatTools`, measured, for a native tool loop.
openStream : Options -> Provider -> List Message -> Task Error AnswerStream
Send a chat and start reading its answer as it is written.
openai : Secret -> String -> Provider
The hosted OpenAI chat-completions endpoint.
priceBook : Money.Currency -> String -> String -> Result Error PriceBook
Build a price book from decimal-string rates (e.g. "0.00015" per 1k input
readStream : AnswerStream -> Task Error StreamStep
Read the next piece of a stream.
responseFormat : ResponseShape a -> ResponseFormat
The `response_format` a shape sends.
responseShape : String -> Codec a -> ResponseShape a
The shape of an answer of type `a`, named `name` (letters, digits, `_`
router : (List Message -> Provider) -> Provider
A multi-provider ROUTER: a `Provider` that, per call, picks a concrete backend
shapeCodec : ResponseShape a -> Codec a
The codec that decodes a shape's answer.
system : String -> Messageuser : String -> Messagewire : WireTransport -> String -> Provider
An OpenAI-compatible provider over your own transport. The provider builds
wireMeasured : WireTransport -> StreamTransport -> MeasuredTransport -> String -> Provider
`wireStreaming`, with a transport that also reports what each call did
wireStreaming : WireTransport -> StreamTransport -> String -> Provider
`wire`, with a transport that streams an answer as it is written