System GraphRAG Lab

System GraphRAG Sandbox

Test decision architecture live.

Follow in real time how an isolated question turns into evidence-backed, systemic decision support.

The systemic use case

Why system thinking? Complex organizations operate through connected dependencies. Plain text alone is not enough to make bottlenecks and leverage points auditable. This demo shows how GraphRAG weaves isolated documents into a logical decision network.

What is evaluated

  • Transparent context binding instead of a RAG black box
  • Complete evidence path for every derived claim
  • Measurable stability across iterative follow-up questions

Start pipeline

Send a question into the system.

Observe the dynamic node selection and the LLM prompt in real time.

Go to evaluation

  • Derive

    The pipeline path makes visible how a resilient answer emerges from question, context, and evidence.

  • Substantiate

    Core evidence and references show transparently what supports the AI argumentation.

  • Execute

    Structured next steps translate analytical insight into directly actionable measures.

Answer flow

Guided Mode

Learning path visible

  1. 1. Question

    Frame the system problem

  2. 2. Context

    Prioritize context

  3. 3. Graph

    Connect nodes

  4. 4. Synthesis

    Derive the answer

  5. 5. Action

    Use next steps

Quality Gate

No analysis yet. Submit a question to inspect quality signals, references, and context budget.

Node selection for the LLM

Context selection transparency

The system first identifies fitting concepts through embedding match and graph score, adds related evidence, and expands to direct neighbors when needed. Only this selected set is sent to the LLM as structured context.

No nodes selected yet. After the first run you will see the actual selection including score and hop.

RAG vs GraphRAG

3 core differences
  • Context shape: classic RAG mainly returns text chunks; GraphRAG also returns explicit relations between nodes.
  • Traceability: with GraphRAG the reasoning path across nodes and edges stays visible, not just text snippets.
  • Multi-hop logic: GraphRAG can intentionally include neighbors across hops and structure cause chains more clearly.

LLM-only vs GraphRAG

Why the graph helps

LLM only

No request sent yet. Submit a question to compare both variants.

GraphRAG

No request sent yet. Submit a question to compare both variants.

Prompt inspector (read only)

LLM-only prompt

[SYSTEM]
You are a helpful assistant. Give a clear answer in plain English.

[USER]
Question: Where do we lose time day to day because work keeps bouncing between teams?

Respond only as JSON with the fields main, coreRationale, nextSteps.

main: 120-220 words.

coreRationale: short rationale without source references.

nextSteps: 2-4 concrete steps.

GraphRAG Prompt

[SYSTEM]
You are a system-thinking assistant. Answer clearly in plain English, but only on the basis of the provided context.

[USER]
Question: Where do we lose time day to day because work keeps bouncing between teams?

References:
Keine Referenzen verfügbar.

Context summaries:
Keine Kontextzusammenfassungen verfügbar.

Use **only** the references and context information above. Do not add external facts.

Respond only with valid JSON containing the fields "main", "coreRationale", and "nextSteps".

"main": 120-220 words, easy to understand, with 3 parts: situation, explanation of relations, concrete consequence in everyday work.

"coreRationale": briefly explain traceability with references [1], [2], [3] to the sources listed above.

"nextSteps": array with 2-4 concrete, actionable next steps.

Important: The graph context may be written in German. Still answer in English.

GraphRAG context payload for the LLM

{
  "query": "Where do we lose time day to day because work keeps bouncing between teams?",
  "references": [],
  "contextSummaries": []
}

Session Memory

stored locally

No local session history available yet.

Answer

Current question: Where do we lose time day to day because work keeps bouncing between teams?

Ready

Short answer

Ready for analysis. Submit a question to start the evaluation.

Context budget: 0 tokens (estimated usage of selected context).

What does this change now?

next steps
  1. 1No next steps available yet. Submit a question to generate concrete actions.

Why this answer holds

evidence-based rationale

The concise P0 core evidence will appear here once an answer is available.

Context and tools

Graph view

Learning view

Learning view: this is how a traceable answer emerges from question, concept, and evidence.

Legend
Question

Drag to explore and scroll to zoom. Open Graph Explorer to change the layout.

Reference concepts

Waiting for answer

After a successful answer we will show up to three reference concepts here together with the concrete tools.

Derivation details

contextual depth

After a successful answer, the most relevant context summaries and sources will appear here.