mirror of
https://github.com/tiennm99/DocsGPT.git
synced 2026-10-24 12:16:14 +00:00
Query-side: composes ClassicRAG (rephrase/budget/fallback). Per source: pgvector entity-name seed -> bounded subgraph -> networkx personalized PageRank (IDF hub down-weight) -> rank graph_node_chunks -> chunk text -> shared token budget; no LLM at query time. Citations derived from chunk metadata (matches ClassicRAG). Falls back to ClassicRAG per-source when no graph rows / unavailable / error (retrieval never breaks). get_chunk_texts queries the co-located documents table by configured table/column names (parameterized). Registered 'graphrag'. Unit G5.