← All questions
Retrieval & EmbeddingsMedium· Propositions· Recall· Pooling

Why does indexing single-fact propositions instead of p...

A passage vector is a blend of multiple facts, which dilutes the similarity score for any single fact. By indexing individual propositions, you ensure that the vector represents a single, focused unit. This requires decontextualization—rewriting the unit to stand alone—rather than just cutting the text shorter, which is why simple sentence splitting often fails to achieve the same recall gains.

Code playground (Python)