Custom hybrid rerankers receive two lists of candidate maps: vector results with normalized _distance (smaller is better), and text results with normalized _score (larger is better). Both include _rowid, independently of with_row_id, and the requested data columns or computed projections.

Pass a two-argument function or a module implementing rerank/2 through reranker:. Return a list of maps containing _rowid and _relevance_score, or {:ok, list} / {:error, reason}. String or atom keys are accepted. Extra fields are ignored: returned data always comes from the original candidates. Row IDs must be unique and belong to either channel; scores must be finite float32-compatible numbers. An empty list produces no results.

The returned order is authoritative. Sort by your scores in the callback if desired. Explicit order_by overrides this order; pagination follows reranking and ordering. Candidates are limited to limit + offset per channel.

Callbacks run in a separate process. Exceptions, throws and exits become error tuples; the query timeout includes candidate retrieval, callback execution, and final result processing. Streaming invokes the callback on enumeration and gathers both candidate sets before yielding rows.