Behaviour for re-ranking search results.

Re-rankers improve retrieval quality by scoring chunks based on their relevance to the question, then filtering and re-sorting by score.

No reranker runs unless you configure one. There is no implicit default: with neither config :arcana, reranker: ... nor a :reranker option, Arcana.search/2 returns its results unreranked, even when an LLM is configured for Arcana.ask/2.

config :arcana, reranker: Arcana.Reranker.LLM        # every search
Arcana.search(query, reranker: Arcana.Reranker.LLM)  # this search

A module, or a {module, opts} tuple, or a three-arity function. There is no :llm shortcut - unlike :embedder and :chunker, the reranker spec defines none, so name the module.

Reranking filters as well as reorders

Worth knowing before you enable one: the callback returns chunks filtered by :threshold and sorted by score, so a chunk that was in the results can be absent afterwards rather than merely lower down. :threshold is therefore a recall-for-precision trade, not a display preference - if a relevant chunk disappears when you turn reranking on, lower it (or set it to 0 to reorder only).

Built-in Implementations

• Arcana.Reranker.LLM - uses your LLM to score relevance. Costs an LLM round trip per search, which is seconds rather than milliseconds - see its moduledoc before using it on an interactive path
• Arcana.Reranker.CrossEncoder and Arcana.Reranker.ColBERT - local models, no LLM call

Custom Implementations

Implement the rerank/3 callback:

defmodule MyApp.CrossEncoderReranker do
  @behaviour Arcana.Reranker

  @impl Arcana.Reranker
  def rerank(question, chunks, opts) do
    # Your custom logic
    {:ok, scored_and_filtered_chunks}
  end
end

Or provide a function directly:

Pipeline.rerank(ctx, reranker: fn question, chunks, opts ->
  {:ok, my_rerank(question, chunks)}
end)