Skip to main content

Simulation run on this device. Scenario values are not sent anywhere — the engine is fully local.

ai studio · free

Reranker Workbench

Blend lexical and embedding scores and watch the top-20 reshuffle.

Which candidate deserves rank 1 — semantic or BM25?

Engine 1.0.0 · studio
20
0.55

1 − this is BM25 weight.

0.15
Top-1 (blend)
doc#3score 0.91
BM25→blend
was BM25 #2
Top-5 overlap dense
2/5
Blend
45% BM25 / 55% dense

Balanced hybrid — top-k should reshuffle less violently than either extreme.

#1 doc3 blend=0.91 bm25=0.99 dense=0.84
#2 doc2 blend=0.88 bm25=0.88 dense=0.88
#3 doc1 blend=0.84 bm25=0.94 dense=0.75
#4 doc0 blend=0.82 bm25=1.15 dense=0.55
#5 doc4 blend=0.72 bm25=0.78 dense=0.67
#6 doc12 blend=0.70 bm25=0.53 dense=0.85
#7 doc11 blend=0.69 bm25=0.44 dense=0.89
#8 doc10 blend=0.65 bm25=0.50 dense=0.76

Method

  • Synthetic BM25 vs dense candidate scores blended by weight; top-1 and top-5 overlap show the reshuffle.
  • Numbers are teaching approximations, not production capacity guarantees.