GraphRAG is retrieval-augmented generation (RAG) extended with a knowledge graph: while ingesting documents, the system extracts entities (companies, people, products, chemicals, contracts…) and the relations between them, and stores them as a graph alongside the usual vector index. At query time the graph lets the AI follow connections across documents instead of only matching similar text passages.
Why plain vector RAG hits a wall
Classic RAG embeds document chunks as vectors and retrieves the passages most similar to the question. That works for “point lookups” — what does clause 7.2 say? — but fails on questions whose answer is distributed over several documents: which of our suppliers are affected by the 2025 penalty-clause changes? No single chunk contains that answer, so no similarity search can retrieve it.
How GraphRAG works
- Entity & relation extraction. During ingestion, an LLM extracts typed entities and relations, validated against an ontology (a fixed schema of allowed entity types and relations) so the graph stays clean.
- Entity resolution. Mentions like “Nestlé SA”, “Nestlé” and “NESN” are merged into one canonical node, so facts about the same thing connect.
- Community summaries. Clusters of related entities are pre-summarised, giving the system a “table of contents” view for broad questions.
- Graph-aware retrieval. A query first finds seed entities, then traverses their relations to pull in facts from documents that pure vector search would never surface — enabling multi-hop reasoning.
GraphRAG vs. vector-only RAG
| Vector RAG | GraphRAG | |
|---|---|---|
| Single-document lookups | Excellent | Excellent |
| Multi-hop questions across documents | Weak | Strong |
| “Compare X across the portfolio” | Misses documents | Traverses relations |
| Broad “what do we know about…” | Fragmentary | Community summaries |
| Ingestion cost | Low | Higher (extraction) |
GraphRAG in SovraRAG
SovraRAG combines multilingual semantic and keyword retrieval with a knowledge graph built from your documents. Complex questions additionally trigger agentic reasoning that researches the question step by step across the graph and the corpus. In internal multi-hop benchmarks (HotpotQA-style, full-corpus retrieval), the agentic GraphRAG pipeline improved exact-match accuracy by roughly a third over a single-pass vector baseline — while lowering median answer latency. The entire pipeline, including graph extraction, runs on Swiss infrastructure.