mnem-graphrag 0.1.0

LLM-free GraphRAG algorithms over mnem's AdjacencyIndex: Leiden community detection (E1) + extractive summarization, Centroid + MMR (E4).
Documentation

mnem-graphrag

GraphRAG primitives for mnem.

Scope (E4 T2)

Extractive community summarization: Centroid + MMR.

  • Score(s_i) = alpha * cos(s_i, centroid) + beta * cos(s_i, query) + gamma * centrality(i)
  • Diversity via MMR with tunable lambda.
  • No LLM, no BM25, no new heavy deps.
  • Reuses the existing mnem-embed-providers::Embedder trait.

Non-goals (this crate, this turn)

  • Leiden communities (E1 will add community::leiden in a separate turn).
  • Personalized PageRank (E2 will replace the degree-centrality fallback).
  • Abstractive summarization (explicitly forbidden: no LLM in this path).