Multi-vector retrieval improves searches that find semantically related but contextually incomplete passages. It associates several embedding vectors with one document, allowing different representations to match the same query while returning a larger synthesis chunk. section 8.4 section 8.8
The usual design has two layers: small retrieval chunks or summaries receive precise embeddings, while larger synthesis chunks remain available for the language model. Matching a lower-layer vector retrieves its linked upper-layer content. section 8.4
A practical implementation follows this sequence:
- Query the MultiVectorRetriever so a focused match expands into the larger linked document. section 8.4.3
Summary embeddings for travel content
For destination pages, embed summaries of coarse chunks and retrieve the corresponding larger chunks. A direct search over summaries returns concise, focused text; multi-vector retrieval returns broader context that is more useful for answer synthesis. section 8.4.3
