open question · seed
Chunk, parent, or summary?
Where the retrieval unit should sit relative to the embedding unit — still unresolved.
Still unresolved for me: where the retrieval unit should sit relative to the embedding unit.
Embed chunkcheap, loses context
Embed summarybetter recall, more storage
Options I keep circling:
- Embed the chunk, return the chunk. Simple, and what everyone starts with. Breaks on documents where meaning lives across sections.
- Embed the chunk, return the parent. Small vectors, wide context. Costs context budget fast.
- Embed a generated summary, return the original. Retrieval quality goes up; you now own a generation step in your ingest pipeline that can drift from the source.
Suspicion: the right answer is document-type dependent, which means the real work is classifying documents at ingest, not tuning one global strategy.