The real cost of GraphRAG: indexing, queries and keeping the graph fresh

The first question I get about GraphRAG from anyone who holds a budget is not about accuracy. It is: what will this cost to run? The answers in circulation range from "about the same as vector search" to "thousands of dollars to index a few books", and both can be true, depending on the method, the model and how often the data changes.

This post separates the three costs, indexing, queries and freshness, puts the published measurements next to each other, and turns them into a cost model you can fill in with your own numbers.

The short answer: indexing means reading your whole corpus with an LLM, and with today's small models the token bill is modest; the model you choose changes it by a factor of fifty or more. Query cost depends far more on the search mode than on the graph: global search can use hundreds of times more prompt tokens per question than plain RAG. Keeping the graph fresh is the cost most teams forget, and the open-source tools only partly solve it.

Indexing: you pay an LLM to read everything

Vector RAG only embeds your text, which is cheap. GraphRAG and its relatives also ask an LLM to extract entities and relations from every chunk, and Microsoft's GraphRAG then writes summaries of each community of entities. Microsoft's documentation estimates that graph extraction is about 75% of indexing cost (GraphRAG indexing methods). The original paper took 281 minutes to index a one-million-token podcast corpus with GPT-4 Turbo (Edge et al.).

GraphRAG-Bench measured the tokens each method needs to index the same book of about 56,000 tokens:

Tokens become money through the model's price. Taking Microsoft GraphRAG's measured tokens and scaling them linearly to a million tokens of corpus gives an illustrative indexing bill at current list prices (OpenAI pricing, October 2026):

Two conclusions follow. With a small model, the token bill for indexing even a large document base is a rounding error next to engineering time. And the model choice matters more than the method: the same pipeline costs about sixty times more with the model the original paper used. Cheaper models extract worse, though, so test extraction quality on a sample before you pick the cheapest one. Microsoft's fast indexing method, which replaces LLM extraction with classic NLP, is cheaper again, at the price of a noisier graph.

Queries: the search mode decides the bill

At query time, the cost driver is how much text goes into the prompt. The original paper already showed the range inside GraphRAG itself: answering from the top-level community summaries needed 9 to 43 times fewer tokens than answering from lower levels or from the source text (Edge et al.). GraphRAG-Bench measured the average prompt per question across methods:

Microsoft has worked on this. Dynamic community selection, which lets a cheap model skip irrelevant community reports, cut the average token cost of global search by 77% in its tests and shipped in GraphRAG 0.5 in November 2024 (Microsoft Research). LazyGraphRAG goes further: Microsoft reports that it matched GraphRAG global search on global questions at more than 700 times lower query cost, and indexes at the cost of vector RAG (Microsoft Research). That is Microsoft's own test, on one news dataset, and the post does not say which models were used or how cost was counted; I treat it as promising, not proven. As of version 3.3, LazyGraphRAG is not part of the open-source GraphRAG library.

The practical lesson is routing. Most questions in an enterprise are lookups that plain or hybrid retrieval answers well; graph methods earn their cost on multi-hop and whole-corpus questions (GraphRAG vs RAG: what the evidence actually says). Send only those to the expensive mode.

Freshness: the cost nobody budgets

Documents change. In vector RAG you re-embed the changed chunks. In a graph, a changed document can alter entities, relations and the summaries built on top of them. The LightRAG authors estimated that, at the time, updating a GraphRAG index on a five-million-token legal corpus meant regenerating every community report, about 14 million tokens, while their own method merges new entities and relations into the existing graph (Guo et al. 2024).

GraphRAG has since added incremental indexing (version 0.4, November 2024) and a graphrag update command that writes a new index next to the old one (GraphRAG CLI, changelog). In the current release, an update processes new documents and appends their communities rather than re-clustering the whole graph. That keeps updates cheap, but the graph drifts from what a full rebuild would produce. Before relying on it, check how your version handles documents that are edited or deleted, and plan a periodic full rebuild.

A cost model you can fill in

Ways to keep the bill down, roughly in order of impact:

1. Route by question type. Lookups to hybrid retrieval, multi-hop to local graph search, whole-corpus to global search.
2. Pick the extraction model on quality and price together. Test two or three on a labelled sample.
3. Use batch processing for indexing. It halves the token price and indexing is not interactive.
4. Prefer dynamic or lazy global search where your stack supports it.
5. Update incrementally, rebuild on a schedule, and measure how far answers drift between rebuilds.

Whether a graph is worth these costs at all is a separate decision: five questions before a knowledge graph for RAG help decide, and A GraphRAG business case, done the evidence-based way puts the cost model next to a value range and a pilot threshold.

How much weight this evidence deserves

Sources

• Edge et al. (2024). From Local to Global: A Graph RAG Approach to Query-Focused Summarization. arXiv:2404.16130.
• Xiang et al. (2025, v3 2026). When to use Graphs in RAG (GraphRAG-Bench). arXiv:2506.05690.
• Guo et al. (2024). LightRAG: Simple and Fast Retrieval-Augmented Generation. arXiv:2410.05779.
• Microsoft Research (November 2024). LazyGraphRAG: Setting a new standard for quality and cost.
• Microsoft Research (November 2024). GraphRAG: Improving global search via dynamic community selection.
• Microsoft GraphRAG documentation: Indexing methods.
• Microsoft GraphRAG documentation: CLI.
• Microsoft GraphRAG changelog.
• OpenAI API pricing (accessed 26 October 2026).