ArchRAG: Attributed Community-based Hierarchical Retrieval-Augmented Generation скачать в хорошем качестве

ArchRAG: Attributed Community-based Hierarchical Retrieval-Augmented Generation

Paper: Notes: • Problem: GraphRAG communities low quality, single-granularity retrieval, extreme token waste from global traversal • Solution: Attributed Communities (ACs) + hierarchy; combine graph structure and node semantics, retrieve across levels • Offline KG build: chunk corpus; LLM extracts entities/relations per chunk; merge duplicates via LLM consolidation; nodes/edges carry text attributes • LLM-based hierarchical clustering loop: augment graph with similarity edges (KNN/CODICIL); weight edges by embedding cosine; run clustering (e.g., weighted Leiden); LLM summarizes each cluster • Collapse step: treat each cluster as a node; connect clusters if members linked; repeat until node count small or depth limit hit; yields multi-layer tree of ACs • Key design choice: summaries generated at every layer; higher layers = abstract themes, lower layers = concrete facts • Indexing problem: per-layer vector DB too costly; solution: C-HNSW, single unified ANN index spanning all layers • C-HNSW structure: nodes = entities + ACs; intra-layer links to M nearest neighbors; inter-layer link from each node to nearest parent-layer node • C-HNSW build: top-down insertion; reuse ANN search to wire neighbors; update inter-layer links if closer parent found; avoids quadratic NN search • Online query: embed question once; hierarchical search starts at top layer, greedy ANN; reuse best node as entry to next layer; collect top-k per layer • Retrieval behavior: abstract questions surface high-layer AC summaries; specific questions drill to entities + relations at bottom layer • Failure mode: long context “lost in the middle”; mitigation: adaptive filtering-based generation • Filtering step: LLM analyzes retrieved texts per layer, extracts relevant snippets, assigns relevance scores • Merge step: sort analyses by score; truncate to budget; final answer generated from condensed, ranked evidence • Intuition vs baselines: LLM bad retriever but good analyzer; push retrieval to ANN graph, reserve LLM for summarization and filtering; yields higher accuracy with 10–250× lower token cost This podcast is AI-generated. While we strive for accuracy, AI can hallucinate or misinterpret complex data. Please consult the original research via arXiv or Google Scholar to verify all technical findings.