graphrag-service

Ask a question about a small made-up company ecosystem and see two retrieval methods answer it side by side: ordinary document search (RAG) and GraphRAG, which follows relationships in a knowledge graph.

What you are looking at

RAG (retrieval-augmented generation) lets an AI answer from your documents: it first finds the most relevant passages, then writes an answer from them. Ordinary RAG finds passages that sound like the question. That works for "Where is Helios Robotics headquartered?" but struggles with questions that need several facts chained together, such as "Which university did the founder of Helios Robotics attend?" (the founder and the university are in different documents).

GraphRAG first turns the documents into a knowledge graph (people, companies and places as dots; relationships such as founded or acquired as lines) and then walks along the lines to collect exactly the chain of facts the question needs.

  1. Pick an example question (or type your own when the live service is running).
  2. Press Ask. The left panel is ordinary RAG, the right panel is GraphRAG.
  3. Compare what each retrieved, the evidence path, and whether the answer contains the expected facts. The graph at the bottom lights up the entities each method used.

The answer writer here is a simple deterministic stand-in that only copies the most relevant sentences it was given, so the difference you see comes purely from what was retrieved. Nothing is invented. checking…

1. Choose a question

The knowledge graph

Blue = what GraphRAG walked through for the last question. Orange ring = entities mentioned in the passages that plain RAG retrieved. Hover a dot for its name; lines are labelled when GraphRAG used them.

Evaluation: 17 questions, 4 pipelines

"Answer accuracy" = the answer contains every expected fact. "Doc recall" = share of the needed source documents that were retrieved. The questions were written by the same author as the graph retriever, so read the single-hop, multi-hop and global rows as a demonstration of when graph retrieval helps, not as a general benchmark. The paraphrase rows deliberately avoid the words the rule-based graph retriever knows, and that is where it breaks.

Show every question and answer