Loading local engines...
Debug the evidence before the ranker
Label the phrases your demo must retrieve. This bench proves whether chunk boundaries preserve them whole and prices the overlap repair in duplicate tokens.
Step = 8 terms
Ctrl + Enter compiles from either editor. Inputs never leave this page.
Chunk-boundary gate
Chunk window ruler
Use an explicitly synthetic twelve-term sequence Alpha through Mu, with evidence Delta Epsilon at terms four through five. Chunk size four and zero overlap produce windows one through four, five through eight and nine through twelve. Rectangles share fifty pixels per term along the same token axis. The highlighted two-term evidence crosses the first boundary, so no whole chunk contains it. This is exact normalized span retention, not a model retrieval, semantic ranking or generation benchmark. Keep the synthetic twelve terms, chunk size four and evidence terms four through five. Overlap one means stride three. Windows become one through four, four through seven, seven through ten and ten through twelve. Rectangles share fifty pixels per term on the same axis; the second window now contains both evidence terms. Token instances sum four plus four plus four plus three, fifteen, so three duplicates mean twenty five percent overhead against twelve original terms. No semantic quality or provider billing equivalence follows. With the same synthetic sequence and chunk size four, overlap two gives stride two and five windows starting at terms one, three, five, seven and nine. Each has four terms, so twenty token instances versus twelve original means eight duplicates, sixty six point seven percent overhead rounded. Bars share twenty pixels per instance: two hundred forty original and four hundred repeated-instance pixels. The evidence remains retained, but retained spans alone cannot prove semantic retrieval, ranking, answer quality, citations, latency or cost. Compromise terms are not a provider billing tokenizer.