Learned Hybrid Fusion — Train a Fusion Model on Your Query Data

I will train a learned fusion model for your hybrid search — using your historical query-click data to train a small model that predicts optimal fusion weights per query type rather than applying static weights across all queries, producing a hybrid system whose fusion is calibrated specifically to your users' actual relevance judgements. Static fusion weights are an average over all query types that may be optimal for neither — conceptual queries benefit from heavy semantic weighting, exact-match queries benefit from heavy keyword weighting, and a learned fusion that detects query type and applies calibrated weights per type outperforms any static configuration on a mixed query distribution.

The training covers query-click data preprocessing, query feature extraction (length, term specificity, question form), fusion weight labelling from click data, lightweight model training, online serving integration, and quality comparison against static fusion.