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
- 1kv-cache paging in vLLMexact phrase match + close semantic0.955bm25 1.00 · vec 0.91
- 2paged attention memory managementstrong on both0.877bm25 0.87 · vec 0.88
- 3KV cache eviction policiesphrase-adjacent lexical hit0.739bm25 0.76 · vec 0.72
- 4attention mechanism intuitionweak lexically, strong semantically0.551bm25 0.24 · vec 0.86
- 5cache invalidation in CDNslexical lookalike, wrong domain0.514bm25 0.62 · vec 0.41
- 6batching strategies for inferencesemantic neighbor0.476bm25 0.17 · vec 0.78
- 7transformer KV-head arithmeticsemantic only0.458bm25 0.09 · vec 0.83
- 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.