Graph RAG Playground
Build a simple knowledge graph from your document and see which extra passages it finds. Free, online, no model needed. Nothing is uploaded.
- 1Paste text or upload a PDF
- 2Click an entity in the graph
- 3Ask a question and compare
Document
Graph
namenumberrepeated wordA simple graph, built without a language model. A link means two entities are mentioned in the same sentence. Graphs built by your own model, with named relationships, are coming later.
Ask
What it is for
- Learn what a knowledge graph is by seeing one built from a document you know.
- See how following links between entities brings in passages that plain RAG leaves out.
- Find questions in your own material that need facts from more than one place.
- Decide whether a graph is worth trying before paying for a model to build one.
How it works
Plain RAG retrieves the passages that look most like the question. A graph adds a second route: from the things named in the question, along links, to passages that mention connected things.
This playground builds a small graph without a language model. It keeps three kinds of entity: names (words written with a capital letter inside a sentence), numbers with their unit, and words that are repeated. Two entities are linked when they appear in the same sentence, and the link is stronger the more often that happens.
When you ask a question, the entities it mentions are found in the graph. Their neighbours, and optionally the neighbours of those, are followed, and the passages that mention them are offered in addition to what plain RAG retrieved. Each added passage shows the chain of entities that led to it.
Full GraphRAG uses a language model to decide what the entities are and to name each relationship, and it also writes summaries of groups of entities. That gives a cleaner graph than this one. Building a graph with your own model is planned for this page.
Questions
Is this the same as Microsoft GraphRAG?
No. It shows the same idea, a graph of entities used for retrieval, in a much simpler form. Entities and links are found by rules, not by a language model, and there are no community summaries or global search.
What does a link mean here?
Only that the two entities were mentioned in the same sentence. Click an entity to read the sentence behind each link. A model-built graph would label the link, for example "leads" or "hosted in".
Why are some entities ordinary words?
Without a language model the page cannot tell a noun from a verb, so any word repeated in the document can be kept. Names and numbers are usually the useful ones. Lower "Entities to keep" to drop the weakest.
Why did the graph add a passage that does not help?
Being mentioned together is a loose connection, so some added passages are only related, not relevant. Real systems have the same problem and use a model to filter. Set "Follow links" to 1 step for fewer, closer additions.
When does a graph help most?
When the question names one thing and the answer is in a passage about something connected to it. In the example, the question names a person and the answer is in a sentence about the app her team builds.
Can I get a written answer?
Not on this page yet. It shows retrieval only. Connecting your own model for answers and for building the graph is coming later.
Is my text uploaded?
No. Everything runs in your browser tab. Your documents and questions never leave your device.
More
Try RAG on your own document · Chunking Visualizer · Token Counter & Cost Estimator · Compare Two RAG Setups · RAG Cost Calculator · What is RAG? · How to choose a chunk size for RAG · What are embeddings? · What are tokens, and what do they cost? · Hybrid search: words plus meaning · How to test RAG retrieval · Seven common RAG mistakes · What is GraphRAG?