Who wins where?

Accuracy per question category for each pipeline, scored against verified gold answers. Bars show the 95% bootstrap confidence interval.

RAG runGraphRAG runAgentic GraphRAG run

Small sample (rag: n=5, graphrag: n=5, agent: n=5). Confidence intervals are wide, so treat differences as indicative, not conclusive.

RAG

phase7-rag-dev · dev · n=5

Accuracy

60%

95% CI 20%–100%

Cost / q

Median latency

9.0 s

GraphRAG

graphrag-dev-20261007 · dev · n=5

Accuracy

60%

95% CI 20%–100%

Cost / q

Median latency

28.9 s

Agentic GraphRAG

phase7-agent-dev-nograph · dev · n=5

Accuracy

40%

95% CI 0%–80%

Cost / q

Median latency

57.0 s

Accuracy by category

Fraction of questions answered correctly, with 95% confidence intervals. Runs: RAG=phase7-rag-dev, GraphRAG=graphrag-dev-20261007, Agent=phase7-agent-dev-nograph.

Where each wins

The pipeline with the highest mean per category. A lead is clear only if the paired-difference interval excludes 0.

Category	Leader	Mean	Lead
Single fact	tie: RAG, GraphRAG, Agent	33%	tie on mean accuracy: rag, graphrag, agent
Unanswerable	tie: RAG, GraphRAG	100%	tie on mean accuracy: rag, graphrag
Overall	tie: RAG, GraphRAG	60%	tie on mean accuracy: rag, graphrag

Paired differences

Accuracy of A minus B on the questions both runs scored; positive means A is better.

A vs B	Pairs	A − B	95% CI	Verdict
GraphRAG vs RAG	5	+0.00	+0.00 to +0.00	no clear difference
Agent vs RAG	5	-0.20	-0.60 to +0.00	no clear difference
Agent vs GraphRAG	5	-0.20

Data behind the answers

Measured from the local extraction store.

Documents

14 parsed

Companies

Accepted facts

of 57 extracted

Entities

With a quoted source page

28.2%

measured, not assumed