Docs / Reference

Changelog

What changed in the tool, newest first.

November 2026

Local web app: ask, search, compare versions and browse rules in your browser, on your own machine.
Prompt-injection test: five planted-passage attacks against four prompt designs, ready to run.
The agent: routes card questions, checks its confidence, searches wider once when unsure. Traced in Arize Phoenix. Available as the ask tool.
Confidence score from measurable signals, with a calibration check that labels it uncalibrated if it fails.
Answer generation and an LLM judge, checked against hand labels before its numbers are used.
New rules versions are detected weekly, registered with one command, and warned about until they are searchable.
Two version-change labels corrected after Riot's Unleashed FAQ was found to have answered them earlier.
MCP server for Claude Desktop, matching the measured search exactly (224 of 224 questions).
Failure types: most search misses are the right rule ranked just below the cutoff.

Reranker tested: version changes improved but each question took over 10 seconds, so it was not adopted.
Hybrid keyword search tested and not adopted.
Qwen3-Embedding became the default embedder.
Date-aware search: outdated rules never outrank current ones.