Jylus posts 84.54 NDCG@10 on 1,000 temporal questions over 100k docs, launches MCP tools
jylusdev · reddit · 2026-10-08
Developer jylusdev benchmarked Jylus, a temporal retrieval engine, on 1,000 questions over 100,000 real documents, scoring 84.54 NDCG@10. Category scores: historical state 90.37 (200 q), changes between periods 88.46 (200 q), chronological ordering 80.09 (503 q), late corrections 88.31 (65 q), superseded facts 84.48 (29 q). Median latency on a 100-question subset dropped from 6.06s to 1.99s (-67.13%) with rankings unchanged. The engine is exposed via three read-only MCP tools — getcontextpack, resolvestate, getevidence — so your existing model reasons over source-backed Context Packs. Author notes this is a self-run benchmark, not official TEMPO, and NDCG@10 measures ranking, not answer accuracy.
More from coding & agent
- Tavily, LangChain and Nebius host SF Tech Week event with 100K credits for startups — LangChain · 2026-10-08
- Dev ditches OpenClaw for Grok bot, citing speed and free live X access — haltakov · 2026-10-08
- Obsidian Starter Kit v5 automates archiving, todos and AI context so you stop doing them by hand — dSebastien · 2026-10-08
- Hamel Husain: same model as judge is usually fine — verify human-label alignment first — HamelHusain · 2026-10-08
- Mitchell Hashimoto's Rex: a terminal replacement built for AI coding agents — ricklamers · 2026-10-08
- DAIR.AI curates 31 papers on recursive self-improvement, from Gödel Machines to today — omarsar0 · 2026-10-08