Multi-Granular Embeddings
Maintaining separate embeddings for entities, chunks, and relations enables different matching strategies at query time.
Entity embeddings support exact-match-then-expand patterns. Chunk embeddings enable traditional semantic similarity. Relation embeddings allow matching on the connections themselves, not just the things being connected.
This separation mirrors how expertise works: sometimes you’re looking for a specific thing (entity), sometimes for content about a topic (chunk), sometimes for how things connect (relation). A unified embedding collapses these distinctions.
Related: 07-molecule—vectors-vs-graphs, 02-molecule—cascaded-retrieval-pattern