Interactive Re-Ranking Playground
One manual ranking change → a precise, localized boost in Vespa’s multi-vector rank profile. Similar queries inherit the improvement; dissimilar ones stay untouched.
Multi-vector rankingKeyword + polynomial boostSession-scoped / resettableVespa attribute updates
What this demo shows
A single, manual re-ranking becomes a small, explainable, and reversible bias in vector space. We boost only the selected document—and afterwards it's embedding primarily affects similar searches—leaving most searches untouched.
Rank profile (localized boost)
rank-profile ann_fine_tune_1 inherits default {
inputs {
query(query_embedding) tensor<bfloat16>(d0[384])
}
first-phase {
expression: closeness(field, title_Embedding_1)
+ pow(max(0, closeness(field, key_Embedding_1)),
attribute(weight1))
* attribute(weight2)
}
}The polynomial term adds a localized “bump” near the learned keyword vector. weight1 controls sharpness; weight2 controls magnitude.
2-step flow
- Step 1 — Re-Ranking UI: Search. Drag to reposition exactly one result, then click Apply Fine-Tune. Backend computes minimal attribute update and persists it for this session.
- Step 2 — Compare View: See Before/After for the original query, a similar query (shows transferable boost), and a dissimilar query (unchanged).
Session-scoped: closing/reloading resets updates so demos start clean.
Mathematical intuition
Let: • q · d₁ = base similarity (query to the moved document) • q · k = similarity to the learned keyword vector Additional term in score: (max(0, q · k))^a × b a = weight1 (sharpness) b = weight2 (magnitude) Effect: Only near-neighbor queries inherit the uplift; others are unchanged.
Viewport height
0:00 / 0:00
Security and Application Monitoring!
Enterprise Grade Protections and Logging for Persistent Service!