Microsoft's open-source GraphRAG framework, plus the Neo4j/LangChain ecosystem built around it. An LLM reads a text corpus, extracts entities and typed relationships into a graph, runs Leiden community detection, and writes an LLM summary per community — enabling both narrow "local search" and broad "global search" across an entire corpus.
Multi-hop reasoning across many documents that flat vector-RAG can't do — traversing relationships to connect facts no single chunk contains — plus a navigable, human-browsable bird's-eye view of a whole corpus via community summaries. This is the category-defining feature of the whole knowledge-graph school.
Both store typed entities and edges rather than raw chunks alone, and both aim to reason across a corpus rather than just retrieve the nearest passage. The injection point in Mimir's serving core — lenses riding along with query results — is architecturally similar to a GraphRAG payload carrying community summaries alongside search hits.
A compulsory governance layer — GraphRAG is a retrieval architecture with no concept of a standard being force-injected into an answer. Mimir also has a designed capture path that GraphRAG simply doesn't: GraphRAG indexes a corpus that already exists, it doesn't capture new thoughts as they occur. Row-level privacy tiers are likewise absent from GraphRAG itself, which inherits whatever the underlying graph database is configured with.
The actual graph reasoning. GraphRAG's community detection and global search are shipped, documented, and proven at real multi-hop reasoning; Mimir's entities and edges exist and are populated (1,868 entities, 1,634 edges) but entity resolution is exact-name-only today and there is no community-summary layer at all — Mimir's graph is real but shallow next to GraphRAG's. The trade-off cuts both ways: the research documents GraphRAG's own indexing cost as roughly $20–500 for a typical corpus versus $2–5 for vector RAG, and its query cost as capable of hitting 40,000-token prompts on global search — real, documented costs Mimir's simpler lexical-plus-facets approach doesn't carry.
research/09-graphrag.md