Retrieval fused a bi-encoder and BM25 by reciprocal rank. A bi-encoder embeds a
document long before the question exists, so the two never meet: it is good at
"same topic" and mediocre at "answers this". A cross-encoder reads the pair.
The proxy already serves three — `cohere-rerank-v4.0-pro` is the default and
measurably better than the fast variant. Query text goes exactly where the
embeddings already go, and nothing new was signed up for.
It found a defect nobody was looking for. In AI Mode each finder scored
`1/(1+rank)` *within its own corpus*, so the best article, section, question and
card all scored 1.0 and the shortlist was a meaningless round-robin. A
cross-encoder is the first thing in this system that can compare a question
with a section. Candidates per kind widened so it can select rather than merely
reorder.
Measured against labels neither ranker produced. Questions, 60 disease tags:
precision@3 0.394 → 0.483. Sections, 60 article titles: 0.772 → 0.833.
"Management of bronchiolitis" led with influenza transmission and a pregnancy
question; "when do you image a first febrile seizure" returned the definition
rather than the sentence saying imaging is unnecessary.
And the honest negative, in docs/reranking.md: board vignettes are written
*not* to name their diagnosis, so on "what causes croup" it prefers a question
that says the word in passing over the barking-cough vignette that never says
it. Some of the bi-encoder's strength is traded away.
Not on the typeahead. A page of results is a choice being made and worth a
third of a second; a typeahead is a word being finished, runs on every
keystroke, and has nothing to judge yet.
The three-state thresholds stay on cosine, argued at the constant: a reranker
only ever sees a shortlist and structurally cannot answer the corpus-wide
question those numbers ask, and whether an answer claims to come from the
library is a promise that must not depend on a network hop.
Every failure returns None and leaves the order alone — unconfigured, no proxy,
connect error, bare 502, timeout, non-JSON, a duplicate or out-of-range index,
a non-numeric score, a list the wrong length. Verified against the running site
with a bogus model name: same results, fused order, no error to the reader.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TqXevQJhxFrM7jJg82cgZN
The voice picker was a dropdown in the quiz player, beside the question — the
one control on that screen with nothing to do with answering it, and one a
learner sets once and never touches. It is a setting now, on the user rather
than in a Redis blob, with a play button beside each voice because a voice is
worth hearing before it is chosen. Choosing nothing stays a real choice: it
means whatever an administrator marked default, so a site that changes its
default reaches everybody without a row being edited.
The tutor reads figures from `question_media` rather than the two legacy path
columns. Those agree exactly today, so nothing was being lost — the first
question given a second figure in the editor would have been the one that
broke it, silently and only for the tutor. The legacy columns remain as a
fallback for anything not projected into that table yet.
And the retrieval thresholds are written down in docs/retrieval-thresholds.md:
the three answers, the sixteen queries they were measured against, why they are
deliberately not the retrieval floor, and how to re-measure when the corpus
grows. Worth keeping the headline in mind — "discuss love" scores 0.491,
alongside "tell me a joke". A number in the 0.4s is noise, not a weak signal.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TqXevQJhxFrM7jJg82cgZN
Retrieval could not say "nothing". `hybrid_ids` fuses two rankers by reciprocal
rank and throws the distances away, and it returns the union — so the shortlist
was never empty, the "nothing matches" branch never fired, and a question about
photosynthesis came back with six paediatric sources and an instruction to
answer only from them.
So the fix is not more scenarios in the prompt. It is one calibrated number,
and three short prompts chosen by it in code. Asking a model to work out which
situation it is in is the part that does not work, and it is also the part that
makes prompts long.
Measured against this corpus with the bodies now embedded — eight clearly
on-topic questions and eight clearly off-topic:
off-topic 0.339 – 0.499 the French revolution … photosynthesis
on-topic 0.586 – 0.740 what causes croup … posterior urethral valves
The thresholds sit in the gap. They are deliberately not the retrieval floor:
that one decides what is worth putting in a list, where a weak hit costs a
reader a glance. These decide whether an answer claims to come from the
library, and a wrong claim costs them their trust in every other answer.
Above 0.55 the answer is sourced and cited, as before. Between 0.50 and 0.55 it
says nothing covers this directly, names what the closest material is, and
marks which parts came from where. Below, it says so in one line and then helps
anyway from general knowledge, citing nothing — refusing outright reads as a
broken assistant rather than a careful one, and the shortlist is not handed to
a model that has just been told the library does not cover the question.
An unmeasurable closeness is not a low one. No vector database or a downed
encoder returns None, and retrieval still found its rows by other means, so
those are still cited; dropping every citation because the ruler is missing
would be the worse failure.
Also: only published articles are indexed now. A draft is unfinished by
definition and has no business in a search result or in that shortlist. The
index follows publication both ways, and the fifteen-minute sweeper drops rows
whose article has been deleted or unpublished — an article that is never edited
again would otherwise keep its rows for good.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TqXevQJhxFrM7jJg82cgZN
**Linking.** A question could be tied to an article only from the article, by
typing the question's number into a box — so opening a question you had just
linked showed no sign of the link, and there was no control to add one. Both
ends now search: find the article by title from the question, find the question
by stem from the article, pick which section of the article the link lands on,
and see what is already linked. One shared finder, so the two ends of one
relationship cannot describe it differently. `GET /questions/{id}/articles`
mirrors the endpoint that already existed the other way, and `GET
/articles/linked` is retired — it answered this question by shipping the whole
prose of every linked article to the quiz player for a list of titles.
"Practise this topic" is a reader's control and no longer appears on an editing
screen.
**The player.** The rail was a bordered card floating in the page with a
scrollbar of its own, so a session had two scrollbars side by side and a
collapse handle tucked inside the card's padding. It is a column now: flush,
full height, its own background rather than its own border, the handle on the
boundary it moves, and a progress bar under the count. The bar at the foot is
the bottom edge of the window — three flush segments, no gaps, no pills —
because Exit as a small grey pill beside a large blue Next made leaving look
like the accident.
Study mode no longer asks whether you are sure. Leaving suspends: every answer
is saved, nothing is graded, and it is waiting where you left it — so the
dialog asked permission for something reversible, under a name for something
that does not happen. An exam still asks once, because a block has a clock, and
it now says what it is: "Leave this block?", not "End Session".
Options are lettered. The explanations already are — a stem extracted from a
board PDF says "Preferred Response: E" — so numbering them 1 to 5 left the
reader translating between two labellings of the same five lines. The tutor is
told the same letters, and the answer key is marked against its own option and
declared authoritative, so a model that would have answered differently cannot
tell a student the marked answer is wrong.
"Preferred response" and "Source page 518" are gone: the first labelled a block
that is obviously the answer, the second named a page of a book the learner
does not have. The clocks moved out of a grey strip across the explanation,
where they read as part of the answer, to the foot of the rail with everything
else about the session.
**AI Mode.** Sources are headed and counted at the end, where evidence belongs,
with the practise button after them rather than above. That button appears only
when there is something to build from and says what it will build — it used to
sit under "how can I help you today?" offering to make a session out of
nothing. A cited question opens in place: `/questions/:id` is the editor, so
following one dropped a learner into a form for changing the question they had
just been told about. And a session built from a chat is named like every other
session, rather than after the chat — asking "hi" produced "hi — practice".
Also: two test questions with raw `<p> </p>` in their stems were live in
the bank; retired. And 36 article summaries were written as a table of contents
with the colons filed off — "Peanut allergy prevention and management: LEAP
guidelines by risk tier, risk stratification, and anaphylaxis treatment" — every
noun phrase sounding informative and none of them saying anything. Rewritten as
claims, with the rule added to the prompt that produced them.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TqXevQJhxFrM7jJg82cgZN
AI Mode could cite an article and link to it; it could not do the other
half of the job. POST /ai/conversations/{id}/practice turns an answer
into a study session, built from what that answer actually cited: a
question it named first, then questions filed under the category of an
article it named, then retrieval on the learner's own words. Everything
goes through the bank's visibility rules on the way out — a chat is not a
route to questions a learner could not otherwise reach. Study mode, never
exam: this is reading followed by practice, not a paper.
Two false alarms on the Settings page, both visible in a screenshot:
The STT test called /model/info on the LiteLLM proxy. Our virtual key is
scoped to llm_api_routes and cannot, so a working transcription model
reported a red 403. It now falls back to /v1/models, which the key may
call, and says plainly that the proxy would not confirm what the model is
for — presence, not suitability.
And the TTS test raised a 400 carrying an instruction ("use the Preview
button"), which the page rendered in red with a ✗. That is not a failure.
It answers, and Preview stays the way to hear a voice.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TqXevQJhxFrM7jJg82cgZN
The design settled earlier, built as described: retrieval decides what may be
cited, and the server enforces it.
The model is handed a shortlist of at most fourteen sources from the learner's
own library and told to cite them by marker. Afterwards every citation it wrote
is checked against that shortlist and anything else is deleted before it is
stored or shown. A hallucinated citation is not unlikely here, it is impossible
— surviving is not a decision the model gets to make. A URL it invents is not a
citation either: only the marker form counts, so a plausible-looking link stays
in the prose citing nothing.
Retrieval reuses the hybrid search already in place, and each corpus keeps its
own visibility rules — the bank predicate and exam scope for questions, the
draft rule for articles, deck ownership for cards. A question source carries the
stem only: a chat that printed the answer would hand away the practice it exists
to prepare you for.
Curated links do the job they were built for. A retrieved row an educator tied
to another retrieved row is boosted, because two things somebody already linked
surfacing for one query is evidence rather than coincidence. Nothing is stored
for this; the boost lives only in that ordering, and the answer marks those
sources so the reader knows which claim rests on an educator's judgement rather
than on a ranking.
Citations are stored with the answer as filtered, so reopening a thread shows
the links it showed at the time rather than a fresh retrieval that may now rank
differently. In the page the markers become numbers and each number opens its
source; a section citation deep-links into that section.
Two smaller decisions worth naming: a question appears in the thread the moment
you send it and is handed back to the input if the answer fails, because typed
words are not something to lose on a 502; and someone else's thread returns 404
rather than 403, since whether it exists is not your business either.
182 backend, 206 frontend green — 16 of the backend tests are the citation
contract and the retrieval boundary.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TqXevQJhxFrM7jJg82cgZN
Daniel