Skip to main content

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

ai calculator · free

Hybrid Search Mixer

Fuse two ranked lists with a weighted mixer and see where each shines.

What alpha blends BM25 and vectors best for this query?

Engine 1.0.0 · calculation
0.5

1.0 = pure lexical (BM25). 0.0 = pure vector similarity. Middle = hybrid.

Query: “how does vLLM page the KV cache?” — BM25 normalized to 0–1 before blending.

Pure BM25 splits on
"cache"
Pure vector splits on
meaning
Trap cases
2 lexical lookalikesCDN & CPU caches
  1. 1kv-cache paging in vLLMexact phrase match + close semantic0.955bm25 1.00 · vec 0.91
  2. 2paged attention memory managementstrong on both0.877bm25 0.87 · vec 0.88
  3. 3KV cache eviction policiesphrase-adjacent lexical hit0.739bm25 0.76 · vec 0.72
  4. 4attention mechanism intuitionweak lexically, strong semantically0.551bm25 0.24 · vec 0.86
  5. 5cache invalidation in CDNslexical lookalike, wrong domain0.514bm25 0.62 · vec 0.41
  6. 6batching strategies for inferencesemantic neighbor0.476bm25 0.17 · vec 0.78
  7. 7transformer KV-head arithmeticsemantic only0.458bm25 0.09 · vec 0.83
  8. 8cache coherence protocols in CPUslexical lookalike, wrong domain0.454bm25 0.58 · vec 0.33

Hybrid ranking punishes one-sided matches: the fused list prefers documents strong on BOTH signals.

Method

  • Normalizes BM25 to 0–1 and blends with the vector score: fused = α·bm25 + (1−α)·vec. α is yours to move.
  • The baked set includes two lexical 'trap' documents (CDN and CPU caches) so you can watch pure keyword ranking get fooled.