Voice RAG · हिन्दी & मराठी · 241,572 chunks
Speak a question in Hindi or Marathi. Get an answer that is extracted, grounded and cited — in under 200 milliseconds. Nothing is invented; if the corpus can't support it, it says so.
Every figure below comes from a run you can reproduce — on the box that serves
this page. Raw JSON lives in data/reports/.
| Stage | P50 | P70 | P90 | P99 | P100 |
|---|---|---|---|---|---|
| run the live benchmark above to populate | |||||
Ask a question, then run every index over it. They disagree — that disagreement is the whole reason the ablation was worth running.
All 7 index subsets, 1,500 bilingual queries, paired bootstrap at 10k resamples. Every recall difference sat inside the noise; the 3-index ensemble was significantly behind a single index on MRR@10. We ship one — a third of the memory, a third of the search time.
ENSEMBLE − metadata_128 · R@10 +0.0047 [−0.0076, +0.0168] · n.s.
Python's \w drops vowel signs and the virama, so BM25 was indexing
consonant fragments — invisible in English, catastrophic in Hindi.
Fixing it gained +12% MRR, three times what all the ensemble
work gained.
दिल्ली → ['द','ल','ल'] Delhi → ['Delhi']
The latency tail tracked the longest sentence in a batch (r=0.84), not how many there were (r=0.02). Fixing the padding revealed that dynamic quantisation makes a wide batch less accurate, not more.
int8 Δ=1.03e-02 · fp32 Δ=4.66e-08
A single query is 2,717 characters — a translation artifact that loops — against a median of 32. It forced a full 512-token pass. Capping the query halved P100.
median 32 · p99 102 · max 2,717 chars