Experiments

Evaluations and experiments comparing the different ways the knowledge base is queried. Pages here are regenerated from data in eval/results/ by the eval/build_report.py script — they do not live inside the curated knowledge base itself.

• RAG vs. Agentic — benchmarks pure-vector RAG against the PydanticAI agent across 30 train-split questions (lookup, list, synthesis, definition).
• Traversal explorer — pick a question and see what each retrieval strategy actually touched: the agent’s tool-call path across the wiki graph next to RAG’s chunk-retrieval footprint, on the same graph backbone.