Fast GraphRAG swapped LLM summarization for PageRank to make graph RAG 6x cheaper

Retrieval-augmented generation answers questions from documents, but questions that chain several facts strain plain retrieval, and the graph-based answer that became popular in 2024 — index entities and their relations so connected facts are reachable — paid for its power with heavy LLM usage in indexing and answering. Circlemind's open-source Fast GraphRAG took the same graph idea and removed the expensive part of the query: instead of spending more model calls to summarize what the graph holds, it lets a 1990s algorithm walk the graph.

Fast GraphRAG stores knowledge as an entity graph that is human-navigable, queryable and updatable, revising it incrementally as data evolves rather than rebuilding. At query time it runs personalized PageRank — the rank propagation behind web search — seeded by semantic retrieval, so the algorithm explores the graph and finds the pieces most relevant to the question. The README prices the difference: the same job over The Wizard of Oz costs $0.08 with Fast GraphRAG versus $0.48 with GraphRAG, a sixfold saving that grows with data size and insertions.

The move was not invented blind: the README points to the HippoRAG paper, published at NeurIPS 2024, which showed that combining LLMs, knowledge graphs and personalized PageRank can match or beat iterative LLM retrieval on multi-hop questions while being 10–30 times cheaper and 6–13 times faster. Fast GraphRAG's twist was to ship that research as a promptable library — typed, asynchronous, MIT-licensed — with none of the complexity of building and maintaining agentic workflows, which is exactly the overhead the README names as its enemy.

Swap LLM summarization for PageRank at query time

Personalized PageRank is deterministic and nearly free to run, so one graph algorithm replaces the many LLM calls a rival pays at query time.

The graph itself stays the source of truth — human-navigable and debuggable — so answers remain explainable instead of emerging from opaque model calls.

Incremental graph updates keep the index current without full rebuilds, lowering the cost of keeping knowledge fresh.

The design rested on peer-reviewed evidence that single-step PageRank retrieval matches iterative LLM retrieval at a fraction of the cost, so the team was betting on published results, not folklore.

MIT licensing plus a managed service with 100 free requests a month let developers adopt the library without permission or upfront spend.

When a rival's edge comes from expensive per-use LLM calls, find the one step a cheap deterministic algorithm can do just as well — keep the model where judgment matters, move everything else to math.

The Show HN post — "Better RAG using good old PageRank" — drew 457 points and 119 comments in November 2024, with maintainers in the thread describing plans to diverge from HippoRAG toward weighted edges and negative PageRank. The repository has since grown to roughly 3.9k stars and 250 forks while staying MIT-licensed, and Circlemind sells a managed version with a free monthly tier. The case became a template for a wider idea: when a newer, more expensive technique wins on quality, the winning move can be finding which part of it a cheap classical algorithm can perform.