model

Documentation

Overview

Package model provides built-in LLM metadata and dual-model routing.

The catalog map holds known model limits (context length, max output, features). Chat agent and context management resolve token budgets from this catalog.

Index

• Constants
• func ApplyDefaultRouter(cfg msg.Config) msg.Config
• func ContextWindowFor(modelName string) int
• func HasFeature(modelName string, feature Feature) bool
• func MaxContextWindow(modelNames ...string) int
• func RegisterTestMetadata(t *testing.T, meta Metadata)
• type Feature
• type Metadata
  • func Lookup(id string) (Metadata, bool)
• type Router
  • func NewRouter(chatModel, toolModel string) *Router
  • func (r *Router) ApplyToContext(ctx *msg.Context, afterToolExecution bool)
  • func (r *Router) PrepareNextTurnHook() msg.PrepareNextTurnFn
  • func (r *Router) Select(afterToolExecution bool) string

Constants

const DefaultContextWindow = 128000

DefaultContextWindow is the fallback input token budget for unknown models.

Functions

func ApplyDefaultRouter

func ApplyDefaultRouter(cfg msg.Config) msg.Config

ApplyDefaultRouter injects dual-model PrepareNextTurn when chat and tool models are set.

func ContextWindowFor

func ContextWindowFor(modelName string) int

ContextWindowFor returns the input token budget for a model name.

func HasFeature

func HasFeature(modelName string, feature Feature) bool

HasFeature reports whether a known model supports the given feature.

func MaxContextWindow

func MaxContextWindow(modelNames ...string) int

MaxContextWindow returns the largest input token budget among the given model names.

func RegisterTestMetadata

func RegisterTestMetadata(t *testing.T, meta Metadata)

RegisterTestMetadata adds or overrides a catalog entry for the duration of a test.

Types

type Feature

type Feature string

Feature names a capability or modality supported by a model.

const (
	// CapChat marks conversational chat support.
	CapChat Feature = "CapChat"
	// CapFunctionCall marks native tool / function calling support.
	CapFunctionCall Feature = "CapFunctionCall"
	// CapJsonMode marks structured JSON output mode support.
	CapJsonMode Feature = "CapJsonMode"
	// ModalityTextIn marks text input modality support.
	ModalityTextIn Feature = "ModalityTextIn"
	// ModalityTextOut marks text output modality support.
	ModalityTextOut Feature = "ModalityTextOut"
	// ModalityImageIn marks image input modality support.
	ModalityImageIn Feature = "ModalityImageIn"
	// ModalityFileIn marks file input modality support.
	ModalityFileIn Feature = "ModalityFileIn"
)

type Metadata

type Metadata struct {
	// ID is the provider-facing model identifier.
	ID string
	// Name is a human-readable display name.
	Name string
	// Description summarizes model capabilities for UI and docs.
	Description string
	// ContextLength is the maximum input token budget.
	ContextLength int
	// MaxOutput is the maximum completion token budget.
	MaxOutput int
	// Features lists supported capabilities and modalities.
	Features []Feature
}

Metadata describes a known LLM and its operational limits.

func Lookup

func Lookup(id string) (Metadata, bool)

Lookup returns catalog metadata for a model ID when known.

type Router

type Router struct {
	ChatModel string
	ToolModel string
}

Router selects between chat and tool models for dual-model strategies.

func NewRouter

func NewRouter(chatModel, toolModel string) *Router

NewRouter creates a router with explicit chat and tool model names.

func (*Router) ApplyToContext

func (r *Router) ApplyToContext(ctx *msg.Context, afterToolExecution bool)

ApplyToContext updates the agent context model field using router defaults.

func (*Router) PrepareNextTurnHook

func (r *Router) PrepareNextTurnHook() msg.PrepareNextTurnFn

PrepareNextTurnHook returns a turn-boundary hook that routes to chat or tool models.

func (*Router) Select

func (r *Router) Select(afterToolExecution bool) string

Select chooses the model for the next provider request.