Hybrid search

Running keyword and meaning search side by side, then fusing two rankings that share no common scale.

Part of the Memory and truth track on lAItest.

Now you have two search systems that disagree, and their scores cannot be added. One says 14.2, the other says 0.83.

Reciprocal rank fusion throws the scores away entirely.

It keeps only the position. A document at rank one contributes one divided by a constant plus one; at rank two, one divided by the constant plus two, and so on down. Add up each document's contributions across both lists. Anything both systems liked climbs to the top, anything only one system liked still gets a hearing, and nothing needs calibrating, because ranks compare honestly even when scores do not.

A common misconception

Commonly believed: Hybrid search is an advanced optimisation to add later.

Actually: It is the default now, and every major vector store ships it. The reason is not sophistication, it is coverage: the two methods fail on different queries. Keyword search misses paraphrases, meaning search misses exact strings. Fusing them removes both failure modes for the cost of one extra query.

Why does reciprocal rank fusion use positions rather than the two systems' own scores?

Answer: The scores come from different systems and share no scale. A BM25 score and a cosine similarity measure different things on different scales and neither is a probability, so adding them is meaningless. Position in a ranking is the one output both produce that can be compared fairly. It is also why fusion needs no retuning when you swap either component out.

In one sentence

When two rankers disagree, trust their ordering and ignore their numbers.