Fusion blending model
Choosing between semantic and lexical results is a classic search challenge. Instead of relying on a fixed heuristic like RRF, our Fusion model learns the decision-making process. By training on the historical outcomes of your dense and sparse retrievers, it knows when to trust semantic breadth or lexical precision for any given query—handing the reranker a superior candidate set.
Dense and sparse candidates are merged with rich features—scores, rank deltas, overlap, metadata—and trained against historical satisfaction signals.
The fusion model compresses the candidate set without sacrificing recall, keeping latency and downstream token usage lean.
Fusion outputs arrive calibrated, with score bands that plug straight into guardrails, dashboards, and feedback loops.
Learned fusion chooses the right balance of dense and sparse signals for every query, handing the reranker a calibrated shortlist while keeping latency and token spend low.
Intelligently blends dense and sparse candidates instead of relying on fixed heuristics.
Learns from historical dense and sparse outcomes so every blend reflects your domain.
Ranks by learned relevance, not raw similarity, giving the reranker a stronger shortlist.
Produces scores with clear thresholds so downstream systems know when to trust the set.
We replace Reciprocal Rank Fusion (RRF)—a fixed, query-agnostic heuristic—with a learned model that produces a calibrated fused score before reranking. Concretely, for a query q and candidate d, we learn:
s_fuse(q, d) = g([s_dense(q, d), s_sparse(q, d), Δrank, lex_overlap, meta, φ(q)])
Inputs to g include:
Training optimizes pairwise or listwise losses (e.g., NDCG, MRR) so the fusion score mirrors real customer satisfaction outcomes while keeping downstream latency and token budgets in check.