WHAT THIS IS: a ready-to-query semantic search index over Wikipedia - a FAISS index of passage vectors plus the passages themselves, so you can retrieve Wikipedia text by meaning rather than by keyword, offline. Vectors are intfloat/multilingual-e5-large (1024-d), so the query does not have to be in the corpus language.
THIS COMMIT: the working paper behind this release is now at version 1.1, DOI 10.5281/zenodo.22841107 (concept DOI 10.5281/zenodo.22782084 always resolves to the latest). Version 1.0 said the English rebuild could not be compared against the published English store; it could, and section 6.1 now carries that comparison (old MRR 25.1 against v11b 47.9, +22.8). A pre-deposit verifier rejected the first draft of the revision, and the corrections are in this commit too: the control's sampling filter is now disclosed (articles with 300 characters or less of old-store text are dropped, leaving about 3.88 million to stride over), and the English leak rate is now compared agai…
eval: English head-to-head against the published store (+22.8 MRR)
WHAT THIS IS: a ready-to-query semantic search index over Wikipedia. A FAISS index of passage vectors plus the passages themselves, so you can retrieve Wikipedia text by meaning rather than by keyword, offline. Vectors are intfloat/multilingual-e5-large (1024-d), so the query does not have to be in the corpus language.
THIS COMMIT: the English rebuild is now measured against the published English store, which the release note previously said could not be compared. The old store's cleaning script no longer exists, so its control questions are generated from that store's OWN chunk text - exactly what the old index encoded - over 400 articles sampled by uniform stride from the 5,328,397 titles present in both stores. Whole corpus, nprobe 128: old recall@1 19.5 / recall@5 33.5 / MRR 25.1 against v11b 39.2 / 59.5 / 47.9. That is +22.8 MRR while searching a corpus 2.9 times larger, and this set leaks less than the ja/zh ones (gold title verbatim in 3.0% of questions against 17.8% and 12.3%)…
eval: sync the final evidence set and release note
WHAT THIS IS: a ready-to-query semantic search index over Wikipedia. A FAISS index of passage vectors plus the passages themselves, so you can retrieve Wikipedia text by meaning rather than by keyword, offline, without running the embedding pass yourself. Vectors are intfloat/multilingual-e5-large (1024-d), so the query does not have to be in the corpus language.
THIS COMMIT: syncs eval/ with the final evidence set. Adds the result files for the title-query A/B and the small-pool question benchmark (eval_*.json, evalscale_*.json, evalqa_*.json), the offset-build and first re-encoding logs, and the working paper (10.5281/zenodo.22782085), and replaces an outdated copy of the release note. An earlier upload of this revision ran twice in parallel by mistake, which is why the previous two commits carry the same title.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
card: lead with what this index is, and link the other languages
Dataset card: add a plain-language opening (what this index is, what is in it, what you need to run it) and explicit cross-links to the other two language repositories.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
v11b: rebuilt from the 2026-05/06 Kiwix dumps with a render-faithful extractor and token-budget
chunking, and re-indexed as IVF4096,SQ4 with a residual-quantile range. 3,009,820 chunks.
Three defects in the previous revision, each measured before it was changed. Logs are in eval/.
1. Chunks were cut by character count. One ME5 token is 3.95 characters in English but 2.01 in
Japanese and 1.98 in Chinese, so 13.7% of the published Japanese chunks and 18.5% of the Chinese
ones ran past the 512-token window and lost 21.7% / 27.0% of their tokens at the encoder,
silently. v11b packs whole sentences to a 256-token budget.
2. The Kiwix licence footer, 184 characters of English legal text, sat inside 97.1% of the published
Japanese chunks and 92.9% of the Chinese ones. v11b extracts from id="mw-content-text" and
removes the footer and navboxes as markup.
3. IndexIVFPQ(m=64) returned 45-47% of the exact top-10 and cost 7.8 MRR in Japanese, 10.7 in
English. v11b uses IVF4096,…
v11b: rebuilt from the 2026-05/06 Kiwix dumps with a render-faithful extractor and token-budget
chunking, and re-indexed as IVF4096,SQ4 with a residual-quantile range. 3,009,820 chunks.
Three defects in the previous revision, each measured before it was changed. Logs are in eval/.
1. Chunks were cut by character count. One ME5 token is 3.95 characters in English but 2.01 in
Japanese and 1.98 in Chinese, so 13.7% of the published Japanese chunks and 18.5% of the Chinese
ones ran past the 512-token window and lost 21.7% / 27.0% of their tokens at the encoder,
silently. v11b packs whole sentences to a 256-token budget.
2. The Kiwix licence footer sat inside 97.1% of the published Japanese chunks and 92.9% of the
Chinese ones. v11b extracts from id="mw-content-text" and removes the footer and navboxes as
markup.
3. IndexIVFPQ(m=64) returned 47% of the exact top-10 and cost 7.8 MRR in Japanese, 10.7 in English.
v11b uses IVF4096,SQ4 with the quantiser range clipped to th…
Add Japanese Wikipedia ME5 index (1,550,503 vectors)
