How do community summaries and selective traversal differ as graph scaling strategies?
This explores two opposite answers to the same problem — graphs are too big to fit in an LLM's context — where one strategy pre-digests the whole graph into summaries (GraphRAG) and the other learns to walk only the parts that matter (Graph-O1).
This explores two opposite answers to the same problem: an entity graph built from a corpus is far too large to hand an LLM whole, so something has to give. Community summaries and selective traversal resolve that tension in mirror-image ways — one pays the cost up front and reads everything once, the other pays at query time and reads as little as possible.
The community-summary strategy works ahead of time. Can community detection enable RAG systems to answer global corpus questions? partitions the graph into clusters with Leiden community detection, pre-generates a summary for each cluster, then answers a question by map-reducing over those summaries. The whole graph is effectively 'read,' just compressed — which is exactly why it can answer *global* questions ('what are the main themes across this corpus?') that ordinary chunk-retrieval RAG can't, because no single chunk contains the answer. The trade is precomputation cost and staleness: the summaries are only as fresh as the last time you rebuilt them.
Selective traversal flips every one of those choices. Can learned traversal policies beat exhaustive graph reading? (Graph-O1) never ingests the whole graph at all — it navigates step by step using Monte Carlo Tree Search and a reinforcement-learned policy, expanding only the nodes that look promising for *this* query. That fits comfortably inside a context window and adapts to the question, but it trades certainty for efficiency: you get decisions under uncertainty about a graph you never fully saw, rather than a guaranteed-complete (if lossy) global view. So the real axis isn't 'better vs. worse' — it's *completeness with compression* versus *selectivity with adaptivity*.
What's worth knowing is that these two aren't the only points on the map, and the corpus hints at a third move: skip the persistent graph entirely. Can query-time graph construction replace pre-built knowledge graphs? (LogicRAG) builds a small query-specific graph at inference time, dodging both the precomputation bill of community summaries and the navigation overhead of traversing a giant fixed structure. And Can building a document map first improve retrieval over long texts? (MiA-RAG) shows the summary-first instinct generalizes beyond graphs — summarizing a document before retrieving from it recovers the same kind of global structure that community summaries capture, without any graph at all.
If you want to push further: Can hypergraphs capture multi-hop reasoning better than graphs? changes what the nodes-and-edges even represent (binding three-plus entities into one relation), which reshapes both strategies — bigger, richer structures make pre-summarizing more valuable but selective traversal harder. The deeper pattern across all of these is a single recurring lever in scaling: do you compress everything once, or select sharply each time you're asked?
Sources 5 notes
GraphRAG uses Leiden community detection to partition entity graphs into modular groups with pre-generated summaries, enabling map-reduce answering of global questions that pure RAG and prior summarization methods cannot handle efficiently.
Graph-O1 replaces whole-graph ingestion with step-by-step agentic navigation using Monte Carlo Tree Search and reinforcement learning. This approach fits within LLM context windows while learning domain-specific traversal policies, though it trades certainty about the full graph for decision-making under uncertainty.
LogicRAG constructs directed acyclic graphs from queries at inference time rather than pre-building corpus-wide graphs, eliminating construction overhead, avoiding staleness, and enabling query-specific retrieval logic without sacrificing multi-hop reasoning capability.
MiA-RAG inverts standard RAG by summarizing documents first, then conditioning retrieval on that global view. This approach recovers discourse structure that bag-of-chunks retrieval destroys, making scattered evidence findable by their document role rather than surface similarity alone.
HGMem organizes retrieved evidence as hyperedges rather than flat lists or binary graphs, allowing three or more entities to bind into single relations without decomposition. This structure accumulates coherent knowledge across retrieval steps, trading representational complexity for constraint expressiveness.