Summarization-based retrieval addresses a core RAG failure: relevant meaning may emerge across chunks, while chunk retrieval can miss the broader answer or produce fragmented context. section 4.7.2
The technique creates synthetic retrieval targets by summarizing clusters of source content. In a tree-organized design, retrieval can move between specific leaf facts and broader summary nodes, connecting a query with answers expressed at the right level of abstraction. section 10.0 section 10.0
Its main design choices are:
- Hierarchy preserves relationships between segments instead of treating the corpus as a flat collection. section 10.0
- Soft clustering lets one paragraph inform multiple higher-level summaries when it bridges topics. section 10.0
- Summary nodes target broad questions, while leaf nodes support retrieval of specific facts. section 10.0
- Compared with chunk-based retrieval, summarization-based retrieval can match long-context performance in question answering while retaining retrieval value for dialogue and general queries. section 4.7.2
A broad question over a complex document
A flat retriever may return several locally relevant chunks whose combined context is disjointed. A summary node can instead state the higher-level answer directly, while leaf nodes remain available for supporting details. section 10.0 section 4.7.2
