The database gives you EXPLAIN ANALYZE; Context Runtime gives you this. For any request, watch the planner rank retrieval strategies by learned quality (not just cost), see the calibrated relevance of every candidate passage, and know exactly what it served — or why it declined to answer.
Decision — why this retrieval strategy
Every strategy the bandit can pick, ranked by learned reward. The quality and cost bars are tracked separately, so a genuinely better method wins on quality — it isn’t hidden behind a cheap-but-worse one.
Retrieval — what each method returned
The same query through every method, with calibrated P(relevant) per hit. A red bar flags a passage with a high raw score but low real relevance — the trap a raw-score ranking falls for, made visible.
Served & reward
This is the open-source planner — the same runtime powering the chat demo and the benchmarks. Reproduce EXPLAIN locally: python examples/explain.py