Datasets:

---

dmpantiu

/

copernicus-rag-core

like 1

Fixes from running this server against the published dataset copy, cleared with Dmitrii
in July, then hardened by an adversarial review round (codex + an independent reviewer):

• get_dataset_code: md_path in the registry names notebook_harvest/parsed/... and
  eqc_qa/notebooks_code/..., neither of which ships; both families are published under
  originals_md/notebooks/. With the fallback, 186 of 186 listed notebook ids resolve.
  The family md_path implies is searched first, so a basename repeated across families
  would resolve to its own copy.

• read_publication: papers.jsonl records absolute paths from the build machine, so on
  any other deployment 0 of 12,411 md_paths resolve — while publications_md.tar.gz in
  this same dataset ships the verbatim originals for 11,209 of them. The same
  registry-vs-published-layout mismatch, the same fix: resolve
  /originals_md/publications_md/parsed_md//vlm/.md (or
  $PUBS_MD_ROOT) before anything else.

• Only when no original exists anywhere (1,210 of 12,411) is the text reassembled from
  the publications collection's chunks — and the reply is honest about what that is:
  full_text=false, text_fidelity="reconstructed-lossy", every 2500-char loader cap
  marked in-line in the text (measured: 12.8% of corpus characters fall past that cap),
  registry metadata (authors/abstract/linked_products) merged back in, and a store
  failure reported as an error instead of masquerading as "not parsed yet". Ordering
  prefers an explicit chunk_index payload field (the newer content-hash chunkers) over
  the numeric chunk_id tail (the currently served corpus); chunks with neither sort
  last and are counted in order_uncertain. The chunk cap truncates a sorted TAIL,
  never holes in the middle; per-paper scroll results are cached so pagination does not
  re-move the paper once per page.