GraphRAG learning lab

Understand GraphRAG through five graph types.

A guided EU AI Act launch decision shows why GraphRAG is not one graph, but a modelling choice across documents, domains, knowledge, conversations, and time.

Test demo now Read the story

Story Graph

read only

Structured reasoning instead of an isolated text answer.

The method map

Five graph types, five different jobs

Each type explains a different part of the GraphRAG architecture. The demo overview links into the dedicated learning pages.

Wissensgraph

Semantische Basis

Ein Wissensgraph modelliert Dinge, Begriffe, Rollen, Eigenschaften und Beziehungen so, dass Bedeutung explizit abfragbar wird. Er hilft, wenn ein System nicht nur Text finden, sondern verstehen soll, wie Konzepte zusammenhängen.

Domänen-Graph

Fachliche Realität

Ein Domänen-Graph modelliert die fachliche Realität eines klar begrenzten Arbeitsbereichs. Er bildet ab, welche Objekte, Rollen, Prozesse, Regeln und Verantwortlichkeiten in dieser Domäne relevant sind.

Dokument-Graph

Korpus und Quellen

Ein Dokument-Graph modelliert Quellen als Netz aus Dokumenten, Abschnitten, Chunks, Claims, Erwähnungen und Belegen. Er macht sichtbar, woher eine Aussage stammt und welche Textstellen zusammengehören.

Conversational Graph

Interaktion und Memory

Ein Conversational Graph modelliert Gesprächsverlauf, Nutzerziele, Annahmen, Korrekturen, Präferenzen und offene Fragen als strukturierten Kontext. Er macht Memory prüfbar statt nur Chatverlauf anzuhäufen.

Dynamischer Graph

Zeit und Veränderung

Ein dynamischer Graph modelliert, was wann galt, was sich geändert hat und welcher Zustand zu einem bestimmten Zeitpunkt relevant ist. Er ergänzt Beziehungen um Zeit, Versionen und Ereignisse.

Guided learning path

One question, five lenses

The demo is designed like a machine-room tour. Each page opens a different part of the GraphRAG pipeline and shows what changes inside the answer.

Start with evidence

Document Graph

Start with evidence: which source, chunk, claim, and citation is actually relevant?

Knowledge Graph

Then clarify meaning: which concepts, roles, and obligations must stay separate?

Domain Graph

Translate the law into a launch decision: product, market, responsibility, risk, and release.

Conversational Graph

Make follow-up questions useful by showing what the system remembers, rejects, or asks back.

Dynamic Graph

Add time: what is valid now, what changes later, and which answer expires?

Blackbox to whitebox

The answer is not the demo. The path to the answer is.

A normal chat hides retrieval, context selection, and uncertainty. GraphRAG Lab makes these steps inspectable.

Blackbox chat

Plausible, but opaque

The LLM answers directly.
Evidence and graph path stay hidden.
The user cannot see what was ignored.

Whitebox trace

Visible and debuggable

User intent is separated from graph context.
The prompt package is visible.
The answer shows its boundary.

Concrete example

One EU AI Act launch decision as the red thread

The same question is viewed through each graph type: Can an AI feature be launched for EU customers, and what evidence, roles, obligations, memory, and timing matter?

What becomes visible

What the user contributes and controls
What is stored in the graph
What context package the LLM receives
Where the answer is evidence-bound

FAQ

Common questions before you start

Is this a legal advice tool?

No. The EU AI Act is used as a realistic learning example. The demo explains GraphRAG modelling patterns, not binding legal assessments.

Why five graph types instead of one graph?

Because document evidence, domain responsibility, semantic meaning, conversation memory, and time validity solve different problems.

Does the demo need live AI calls?

The learning mode is simulated on purpose. That makes the intermediate steps visible and avoids depending on external services.

Test the method map now.

Start with the overview, pick a graph type, and inspect how user, graph, and LLM interact.

Test demo now