GraphRAG

Standard vector RAG retrieves isolated text chunks that are semantically similar to a user's query. That works well when the answer lives inside a single paragraph, but fails on two classes of enterprise questions: multi-hop reasoning (where connecting Entity A to Entity C requires traversing an intermediate relationship in Document B) and global sense-making ('What are the top systemic supply-chain risks across all 2,000 board minutes from the past three years?').

A GraphRAG pipeline adds a knowledge-graph extraction and community-clustering layer during indexing. Using a language model or specialised NER extractor, the pipeline identifies nodes (people, companies, contracts, systems, statutes) and typed edges (`OWNS`, `SUPPLIES_TO`, `AMENDS`, `GOVERNED_BY`) across the corpus, storing them in a graph database such as Neo4j or alongside vector indexes, and optionally pre-computing hierarchical community summaries. At query time, the retriever combines vector search with graph traversal (`Cypher` or subgraph expansion) to pull both direct text chunks and the structural neighbourhood around matched entities.

For example, a London financial crime investigation unit uses GraphRAG across corporate filings, sanctions notices, and correspondence to trace beneficial ownership chains across shell companies that never appear in the same document.

In domains defined by dense relational structure, such as anti-money laundering, pharmaceutical drug discovery, complex M&A due diligence, and enterprise dependency mapping, GraphRAG surfaces multi-document connections that flat vector retrieval consistently misses.

However, GraphRAG carries a steep indexing and maintenance price tag. Extracting entities, resolving duplicates ('Acme Ltd' vs 'Acme Holdings UK'), and computing community summaries with an LLM across a large corpus can cost ten to fifty times more in indexing tokens than standard vector embedding, and schema drift complicates incremental updates. Teams should only adopt GraphRAG after proving with an eval dataset that hybrid vector + BM25 search with reranking cannot answer their target queries.