LeanRAG

A knowledge-graph RAG framework (AAAI 2026, KnowledgeXLab) that refines GraphRAG by attacking two failure modes of existing KG-based RAG: semantic islands and structure-unaware retrieval. It pairs a semantic aggregation step that wires the high-level summary nodes together with a hierarchical, structure-guided retrieval that climbs the graph from the bottom up. The wiki’s first peer-reviewed academic anchor for the GraphRAG thread (prior grounding was the T3 implementation graphrag-rs).

The two problems it names

Existing knowledge-graph RAG (GraphRAG and successors) builds aggregation/summary nodes over the corpus, but LeanRAG argues two gaps remain:

• Semantic islands — the high-level community/summary nodes are disconnected: each summarizes its own cluster, but there are no explicit relations between summaries, so reasoning that has to cross communities has no path to follow.
• Structure-unaware retrieval — retrieval is effectively flat: it searches by similarity and ignores the graph’s topology, which retrieves overlapping, redundant evidence and traverses the structure inefficiently.

The two fixes

• Semantic aggregation — clusters entities into “semantically coherent summaries” and constructs explicit relations among those summaries, forming a navigable aggregation-level network. This is the direct counter to semantic islands: the summary layer becomes a graph you can traverse, not a bag of disconnected reports.
• Hierarchical retrieval (bottom-up, structure-guided) — anchors the query to the most relevant fine-grained entities first, then traverses upward through the aggregation pathways to gather evidence along the graph’s actual semantic structure. Because retrieval follows topology instead of re-scoring flat chunks, it collects a more concise, less overlapping evidence set.

Reported results

• ~46% lower retrieval redundancy vs. flat (chunk-similarity) baselines.
• Across four QA benchmarks (Mix, Computer Science, Legal, Agriculture): average win rate 78.1% over GraphRAG and 71.9% over HiRAG (a hierarchical-RAG predecessor).
• Pipeline: 1024-token chunking with a 128-step sliding window → triple/entity extraction (via CommonKG or GraphRAG-style extraction) → graph construction + clustering → query-driven hierarchical retrieval → LLM generation.

Tier note (T2). Peer-reviewed (AAAI 2026) plus a reference implementation — stronger than the self-reported graphrag-rs (T3). But the comparative win-rates are author-reported on the authors’ own benchmark suite, so the margins over GraphRAG/HiRAG are motivated numbers; the method is the durable contribution.

Where it sits in the wiki’s KG/RAG thread

LeanRAG is a builder-querier like graphrag-rs — it builds the knowledge-graph and runs the retrieval over it — but its contribution is sharpening how the graph is queried. It does not open a sixth retrieval gap; it improves the mechanism for the fifth (whole-corpus / global synthesis, see retrieval-augmented-generation) and tightens factual-connection retrieval. It also supplies a critique of plain GraphRAG: community summaries alone are semantic islands unless you also model the relations between summaries and let retrieval follow them — a critique that lands directly on the Leiden per-community reports of the original GraphRAG paper. Author org KnowledgeXLab noted inline (thin; the technique is the value).

Related

graphrag · graphrag-rs · retrieval-augmented-generation · knowledge-graph · sift-kg · hybrid-retrieval-rag · temporal-knowledge-graph