RAG vs GraphRAG vs Agentic GraphRAG vs cost-aware routing, over 100 questions on the TigerGraph Agentic GraphRAG corpus.
Every number on this page is computed from the raw run files in artifacts/ at build time.
Up and to the left is better. The agent reaches the top; routing reaches the same height for roughly 13× fewer tokens.
RAG does not degrade gracefully on aggregation and superlative questions. Those need a median of 11 documents and up to 43, so the evidence does not fit in a top-10 retrieval whatever the prompt says.
Solid bars are answer accuracy; pale bars are completeness, the share of the question's gold documents the pipeline actually retrieved. A wide gap means answers are being produced without the evidence that settles them.
| pipeline | accuracy | completeness | evidence precision | tokens / question | tokens / correct answer | steps | sec / q |
|---|---|---|---|---|---|---|---|
| RAG | 63% | 76% | 31% | 3,259 | 5,174 | 2.0 | 2.5 |
| GraphRAG | 97% | 98% | 98% | 838 | 864 | 2.0 | 2.1 |
| Agentic GraphRAG | 100% | 99% | 99% | 890 | 890 | 2.1 | 2.4 |
| OCCAM (routed) | 100% | 99% | 99% | 64 | 64 | 2.1 | 0.2 |
Comparing the agentic run with single-pass GraphRAG on the same 100 questions:
| questions the agent rescued | 3 (3%) |
| questions the agent broke | 0 |
| extra tokens spent in total | 5,221 |
| extra tokens per question rescued | 1,740 |
Each question starts at the cheapest tier and climbs only when a tier reports what it was missing.
| tier | questions | accuracy | tokens / question |
|---|---|---|---|
| tier0_rule_graph | 98 | 100% | 0 |
| tier1_disambiguated | 1 | 100% | 263 |
| tier3_document_fallback | 1 | 100% | 6,161 |
Ground truth for these is not published, so only coverage and cost can be reported here - accuracy is scored by the organisers. Raw answers and full agentic traces are in artifacts/results_hidden.jsonl.
| pipeline | answered | tokens / question | steps |
|---|---|---|---|
| RAG | 45/50 | 3,332 | 2.0 |
| GraphRAG | 47/50 | 835 | 2.0 |
| Agentic GraphRAG | 50/50 | 952 | 2.2 |
| OCCAM (routed) | 50/50 | 142 | 2.3 |