Alpha-Blended Fusion, Revisited

Where we left off

Lesson 7 showed weighted-sum fusion's fragility with hand-rolled scores. Now that both retrievers are real libraries (Lesson 21), this lesson revisits alpha-blending one more time, and puts it head-to-head against RRF on the same two queries, so the tradeoff is concrete rather than abstract.

The code, piece by piece

distances = results["distances"][0]
dense_scores = {doc_id: 1 - distance for doc_id, distance in zip(ids, distances)}

chromadb returns distance (lower means closer), the opposite convention from this course's own cosine similarity (higher means closer). Flipping it with 1 - distance isn't exact cosine similarity, but it restores the "higher is better" direction min-max normalization and blending both assume, the same care Lesson 10 took with score conventions before combining anything.

blended = sorted(names, key=lambda n: ALPHA * dense_norm[n] + (1 - ALPHA) * sparse_norm[n], reverse=True)
fused = reciprocal_rank_fusion([dense_ranking, sparse_ranking])

Both fusion methods run on the exact same two underlying rankings, so this lesson isn't comparing different retrievers, only different ways of combining the same two opinions.