Evaluation & Advanced Retrieval
GraphRAG & Structured Retrieval
Vector search can't answer corpus-level questions
'Summarize the main themes across all documents.' 'Who is connected to whom in this corpus?' 'What changed between the 2022 and 2024 policies?' — these are global questions whose answers live in the shape of the whole dataset, not in any retrievable chunk.
Top-k retrieval structurally can't answer them: the answer isn't IN a chunk, it's distributed across hundreds. GraphRAG builds the missing layer: extract an entity/relation graph at index time (who, what, related-to), cluster it into communities, summarize each community — then global questions are answered over the summaries, not over raw chunks.
insight
Top-k retrieval finds the needles. GraphRAG describes the haystack — themes, clusters, and connections that only exist at corpus zoom level.