Artifipedia
GlossaryMapBlogMy Library
Home/Language & LLMs/GraphRAG
Language & LLMs

GraphRAG

Retrieval over a knowledge graph the model built from your documents — better at questions about the whole corpus, and expensive enough that most projects shouldn't.

Reviewed July 16, 2026Stable
Reading level: Curious
Pick your depth ↓

When not to use it

  • Lookup questions. If the answer lives in a chunk, standard RAG finds it faster and for a fraction of the cost.
  • Frequently changing corpora. The index is expensive to build and doesn't update incrementally in any pleasant way.
  • Before you've built standard RAG. You don't yet know whether your users ask global questions, and they probably don't.

Reach for something else instead

  • Standard RAG answers most real questions and costs an embedding per chunk.
  • Reranking is the cheapest large improvement to a RAG system and should be exhausted first.
  • Hierarchical summarisation without a graph may capture much of the benefit — this is the untested ablation.

Sources & further reading

  • Edge et al. (2024), From Local to Global: A Graph RAG Approach to Query-Focused Summarization — the Microsoft Research paper; graph extraction, community detection, hierarchical summaries.
  • Lewis et al. (2020), Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks — the baseline it extends and the thing to build first.
  • Liu et al. (2023), Lost in the Middle: How Language Models Use Long Contexts — why stuffing the whole corpus into context isn't the alternative it appears to be.

Primary sources, listed so you can check the claims on this page rather than take them on trust.

Where people go wrong

  • Building it before knowing your query mix. It's an expensive answer to a question most users don't ask.
  • Underestimating entity resolution. Duplicate entities silently split the connections you built the graph for, and nothing errors.
  • Assuming the graph is what's helping. Multi-level summarisation may be doing the work, and the ablations that would tell you are largely missing.

At a glance

FieldLanguage & LLMs
CostLLM calls across the whole corpus at index time
Wins onglobal, corpus-wide questions
Loses oneverything else
DifficultyAdvanced
Flashcards for this concept · study, save or share them →
Question
—
Answer
—
1 / 4

Related concepts

Language & LLMs

Retrieval-Augmented Generation (RAG)

Letting an AI answer from a specific set of documents by looking them up as it responds — instead of relying only on what it memorized.

Tools & Ecosystem

Knowledge Graph

Facts as a network of entities and relationships — symbolic AI's one commercial survivor, and RAG is rediscovering it.

Language & LLMs

Chunking

Cutting documents into retrievable pieces — the least glamorous decision in RAG, and the one that most often decides whether it works.

Language & LLMs

Reranking

A second, slower pass that reorders retrieved results by actually reading them — usually the cheapest large improvement available to a RAG system.

Language & LLMs

Token

The small piece of text an AI reads and writes — usually a chunk of a word, not a whole word.

Language & LLMs

Large Language Model (LLM)

An AI trained on enormous amounts of text to predict the next piece of writing — the technology behind chatbots like ChatGPT and Claude.