Voice RAG · 14 Indic languages

MSMARCO-XI · bge-m3 dense retrieval · extractive reader · measured numbers & known limits
Ask
Chunking, live
The Indic bug
The same query, run through every chunking strategy right now. Top-k is the protocol everyone uses; it is confounded, because a strategy emitting more, smaller chunks gets more shots at the same passage (corr with chunks/doc = −0.820). Equal budget gives every strategy the same 400 words of context — correlation drops to −0.152 and the ranking inverts.
Every token-level metric in QA runs text through re.sub(r"[^\w\s]", " ", s). In Python, \w is str.isalnum(), which is False for Unicode Mn/Mc — the combining marks carrying the vowels in every Brahmic script. Paste any Indic sentence and watch it shatter.