Notes
Notes
Written only when there is a measurement to publish. Numbers ship with the files to reproduce them.
If retrieval misses, RAG can't answer — in Korean, start with the tokenizer →
Nori on 183,240 Korean tender titles: a corpus dictionary took P@10 to 0.986.
When the table disappears, RAG states a wrong number with confidence — HWP and PDF table extraction measured on 214 Korean tender documents →
Tables in 214 Korean tender files: generic HWP extraction drops 85% of cells.