Foresight.Evals.Labeling.PoolFusion (Foresight v0.1.0)

Re-rank a recorded candidate pool under a trial fusion configuration, offline.

This is what makes D2's coordinate-ascent affordable: scoring a candidate weighting costs a pure re-fusion of an already-recorded pool instead of a live recall run over the corpus.

It deliberately calls Foresight.Recall.RRF.fuse/2 — the real fuser — rather than reimplementing the scoring. A reimplementation would be free to drift from production, and a tuner that optimises against a drifted model of the system produces weights that do not transfer. The only work here is reconstructing the strategy_results shape the fuser expects from what the pool recorded.

RANK FIDELITY. RRF.fuse/2 derives each arm's rank from list position via Enum.with_index/2, but a recorded pool is not dense: candidates are dropped when they lack a stable eval_id, and per_query_cap truncates. Handing it a compacted list would silently promote candidates — a candidate recorded at rank 9 would score as if it ranked

So gaps are filled with placeholder entries that occupy the missing positions and are discarded from the result.

Summary

Functions

rank(query_pool, fusion_opts \\ [])

Rank one query pool's candidates under fusion_opts.

strategy_results(query_pool)

Rebuild the arm => ranked items map the live fuser consumes.

Functions

rank(query_pool, fusion_opts \\ [])

@spec rank(
  map(),
  keyword()
) :: [String.t()]

Rank one query pool's candidates under fusion_opts.

Returns memory_refs, most relevant first. fusion_opts is whatever Foresight.Recall.FusionConfig.resolve/2 produces.

strategy_results(query_pool)

@spec strategy_results(map()) :: %{required(String.t()) => [map()]}

Rebuild the arm => ranked items map the live fuser consumes.

Exposed so a test can assert the reconstruction itself, independently of ranking.