Parent-Child and Multi-Vector Retrieval
Small child chunks can be indexed for precise matching while a larger parent section is returned for context. Multi-vector systems can represent titles, tables, text, or image descriptions separately. Preserve lineage and remove duplicate representations during retrieval. Every derived vector must be updated and deleted with its source.
These patterns add complexity and should be adopted when they address measured failures such as missing table context.
Implementation example
Search child chunks -> resolve parent section -> expand selectively
-> deduplicate -> generate with original source citations💡 Derived representations need source IDs and consistent lifecycle operations.
Key Takeaways
- Inspect the relevant pipeline stage instead of guessing.
- Preserve source provenance, versions, and access scope.
- Measure quality on representative queries.
Knowledge Check
What must happen when a source with multiple vectors is deleted?