A slide carrying [1] markers is unreadable from the back of a room, and an
article that cites inline reads as a paper rather than as teaching material. The
model is now told explicitly not to cite in the body — no bracketed numbers, no
parenthetical "(Nelson, p. 2604)" inside sentences — and to put everything it
drew on in a References section at the end, which in a presentation is the final
slide.
Checked rather than assumed: a six-slide deck generated through the grounded
path contains zero in-text citation markers, and ends with a References slide.
The prose keeps the specificity that grounding is for — bilirubin produced at
two to three times the adult rate, conjugation immature until about two weeks,
thresholds in mg/dL — without a single marker interrupting it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
The grounding metadata went to the API response and the logs, which is no use to
someone holding the deck. A teaching resource shown to trainees should carry its
own provenance, so a grounded one now ends with a References section — the final
slide in a presentation — listing the library excerpts it actually used, by
title and page.
Restricted deliberately: only excerpts actually drawn on, nothing invented. That
was worth checking rather than trusting. Generated a deck and compared every
citation against the source metadata: "Kliegman R. Nelson Textbook of
Pediatrics, 22nd ed., 2024, p. 2604" against a stored title of "Kliegman R.
Nelson Textbook of Pediatrics 2-Volume Set 22ed 2024" at page 2604, and the same
for Fleisher & Ludwig, Rosen's, Understanding Pathophysiology and the AAP
compendium. The model reformatted filename-derived titles into readable
citations using only what it was given.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
Learning generated everything from the model alone. A deck on bronchiolitis was
whatever the model remembered about bronchiolitis, with no connection to the
documents this institution actually indexed — while the assistant had been
searching that corpus all along.
Same collection, deliberately. mcp_bge_m3_1024 is already embedded with
openrouter-bge-m3 at 1024 dimensions; a second index over the same documents
with the same embedder would be a copy that drifts. What differs is the budget:
a chat answer wants a few tight excerpts because the reader is waiting, a
teaching resource synthesises a whole topic. So learning.search_limit and
learning.context_chars default to 30 and 2500 against the assistant's 8 and
1400, and are separate keys so tuning one cannot move the other.
Not unbounded, though. "No limit" only moves the ceiling from a setting to the
model's context window, where overflow truncates the middle of the prompt
silently — the worst place to lose source material. 60 results and 8000
characters per excerpt are the caps.
Opt in per generation: a resource on something the library does not cover is
better written without it than padded with the nearest unrelated excerpts.
Retrieval never fails a generation — the resource is then written from the model
alone, which is what happened before this existed — and every response reports
what it was grounded on, so a caller can say "24 excerpts" or "the library had
nothing on this" rather than quietly serving ungrounded material.
Verified against the live corpus: bronchiolitis, neonatal jaundice and febrile
seizure each returned 12 excerpts and ~23k characters from Nelson, Rudolph and
the Pediatric Clinical Practice Guidelines. A deck generated through the full
chain came back with textbook specificity that is not general recall —
bronchiolar diameter, birth-weight thresholds, the full pathogen list.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
