OCCAM runs RAG, GraphRAG and Agentic GraphRAG on the same 100 questions, then adds a router that sends each question to the cheapest tier that can truly answer it. Same accuracy as the agent, a fraction of the tokens.
▶ Watch the four pipelines race SourceThe hackathon guidebook asks whether the extra reasoning and retrieval steps are worth the extra complexity and token cost. Most submissions will report accuracy and stop there.
"How many cycling events had more than 30 competitors?" needs up to 43 documents. A top-10 retrieval cannot hold that evidence, so aggregation accuracy is 10%.
When the graph returns a shortlist instead of an answer, a single pass gives up. It lost 3 of 100 questions that way.
Agentic GraphRAG reaches 100%, but spends ~890 tokens per question, including the 97 that did not need it.
Pick a real benchmark question and run it. Each lane replays the recorded trace of that pipeline: every step, every token, every verdict, straight from artifacts/results_public.jsonl.
Replay of recorded runs, not a fresh LLM call: step durations are the measured latencies from the benchmark.
Accuracy against mean tokens per question, log scale. The agent reaches the top. Routing reaches the same height for ~14× fewer tokens.
RAG collapses on aggregation (10%) and superlatives (60%).
Escalation is never a guess about difficulty. Every tier reports its gap, and that gap is recorded in the trace, so the controller cannot silently skip evidence.
We report answer accuracy and evidence recall separately. RAG's 63% against 76% recall is the whole argument for doing both.
Every agent tool is an installed GSQL query on Savanna. 78 of 100 questions were verified identical to the reference store, 0 mismatches.
Tier 0 is a rule cache. On six paraphrased questions no rule matches, tier 1 answered all six correctly. Single model, single run, no variance bars.