Deploy: corpus current to 2026-08-31, --force across all ingesters
Beemer+1: Claude Opus 5 (1M context)
0fb8dd3
Squashed deploy of eight commits (998172e..24f6069 plus the three earlier
unpushed ones) onto the reclaimed-storage base. Squashed deliberately: the
intermediate commits carry a 93 MB pre-float16 embeddings.npz that would
otherwise be pinned in this Space's storage forever.
Corpus:
- 5 Acts re-ingested bilingually to consolidation 2026-06-21 (A-8.8,
C-38.8, C-46, E-14.1, F-11.6); Criminal Code +184 en / +190 fr sections.
- CIU Regulations refreshed off the PDF the union reposted 2026-08-25 --
a re-export, not an amendment: 4 of 47 chunks differ, all PDF text-layer
spacing artifacts.
- Re-embedded: 49,678 sections -> 55,417 vectors, float16.
- refresh --check clear: stale 0, errors 0, 59/59 legislation, 35/35 union.
- Eval ALL n=258 Hit@1 0.77 / Hit@5 0.97 / MRR 0.85, holdout MRR 0.84.
Code:
- All 12 ingesters now take --force and thread it to every fetch_cached
call. Without it a document reposted in place at the same URL could
never be picked up, and refresh --check's STAL…
Squash history to reclaim LFS storage (1 GB cap); tree unchanged
