Global Sensemaking vs. Local Retrieval

Retrieval tasks divide into two fundamentally different shapes.

Local retrieval asks: “Find me the passage that answers this question.” The answer exists somewhere; the system’s job is to locate it. Vector similarity excels here, match the query to the content that resembles it.

Global sensemaking asks: “Synthesize information scattered across multiple documents to answer this.” No single passage contains the answer. The system must gather fragments and compose them.

Query-Focused Summarization is the canonical global sensemaking task: “What are all the factors that affect X across this corpus?”

The mismatch explains why vanilla RAG disappoints on certain questions. Vector search optimizes for local retrieval. When you throw global sensemaking questions at it, you get partial answers from whichever single chunk scored highest, missing the synthesis that the question actually needs.

Related: 05-atom—multi-hop-reasoning, 07-atom—directness-comprehensiveness-tradeoff