Chunk Lab
Change how the documents are split and watch how often the answer survives in one piece. Scored live in your browser on 25 documents and 47 questions with a known answer passage.
Used on RAG-13 and RAG-15
Evidence recall@3
…
Chunks in index
…
Bars start at 0.70 so the differences are visible. A hit needs one of the top 3 chunks to hold the whole answer passage.
Missed questions (…)
How is cut (0 chunks)
Each shaded band is one chunk. With overlap, the end of one band repeats at the start of the next.
What to notice
Fixed 128 tokens misses 11 of 47 answers, and 6 of those are splits: the boundary cut the answer in half. At 512 tokens no miss is a split.
Surprise
Overlap made fixed chunks worse here. Near-duplicate chunks pushed the right one out of the top 3.
Try next
Open SYN-TEL-FIBERBOX-v3.2 at fixed 128, then at by headings, and compare where the error code table is cut.