Agentic.ModelRouter.Selector (agentic v0.2.2)

Scores and ranks catalog models based on an Analyzer.analysis() result and a user Preference.

Returns an ordered list of {model, score} tuples (lowest score first). Models with missing required capabilities receive heavy penalties so they sink to the bottom but remain available as fallbacks.

Summary

Types

ranked_model()

Functions

rank(analysis, preference, opts \\ [])

Rank all candidate models for a given analysis and preference.

select(request, preference, opts \\ [])

Analyze a request and return ranked models for the given preference.

select_all(request, preference, opts \\ [])

Like select/3 but returns all ranked models (not just the best), so the caller can walk the full list on failure.

top(analysis, preference, n \\ 3, opts \\ [])

Get the top N ranked models.

Types

ranked_model()

@type ranked_model() :: {Agentic.LLM.Model.t(), float()}

Functions

rank(analysis, preference, opts \\ [])

@spec rank(
  Agentic.ModelRouter.Analyzer.analysis(),
  Agentic.ModelRouter.Preference.preference(),
  keyword()
) :: [ranked_model()]

Rank all candidate models for a given analysis and preference.

Returns a list of {model, score} tuples sorted by score ascending.

select(request, preference, opts \\ [])

@spec select(String.t(), Agentic.ModelRouter.Preference.preference(), keyword()) ::
  {:ok, {Agentic.LLM.Model.t(), Agentic.ModelRouter.Analyzer.analysis()}}
  | {:error, term()}

Analyze a request and return ranked models for the given preference.

This is the main entry point combining analysis + ranking. When llm_chat is provided, uses LLM-based analysis; otherwise falls back to heuristic analysis.

select_all(request, preference, opts \\ [])

Like select/3 but returns all ranked models (not just the best), so the caller can walk the full list on failure.

top(analysis, preference, n \\ 3, opts \\ [])

@spec top(
  Agentic.ModelRouter.Analyzer.analysis(),
  Agentic.ModelRouter.Preference.preference(),
  pos_integer(),
  keyword()
) :: [ranked_model()]

Get the top N ranked models.