Every question used to carry the full catalogue (~39k tokens on this
library, 41% of it per-section sizes). Now the model gets one line per
course as a stable, cacheable index, and paths/tags/blurb only for the
~25 courses a local term match against the conversation picks out (plus
anything in progress or already cited). Typical prompts land at 7-10k
tokens and the model reads the relevant courses in depth.
Duplicate copies of a course (same folder key) are collapsed before the
prompt is built; progress on a dropped copy counts for the kept one and a
recommendation is steered back to the copy the learner is part-way
through. Finished courses become a single id list. Reasons are capped at
15 words / 120 chars, which also shrinks every replayed follow-up turn.
Thinking mode is off for advisor calls (ADVISOR_REASONING = 'off').
Claude-Session: https://claude.ai/code/session_011gkaBpxtBdCBwdLL3R6Pfo
Course metadata: a scan now gives every course a course_meta row from the
downloader's sidecar JSON when the folder has one, else from catalog_courses
matched by the folder name the downloader would create (folderKeyFor, a port
of ine.py's sanitize_name). tools/build-catalog-seed.php turns the INE export
into seeds/ine_catalog.sql, an idempotent patch that creates its own table and
upserts by id, so it can be applied to the live server before or after the
code and re-run freely. Course pages show the blurb, difficulty, vendor and
tags; search matches tags and blurbs; Settings reports how each course got its
details. Against the live library 163 of 179 courses match.
Course advisor, off by default: api/advisor.php answers "which course should I
take to learn X / what comes before Y" with courses from the library, in
order, plus vendor courses not yet downloaded. includes/ai_client.php is a
small curl wrapper for any OpenAI-format chat endpoint (DeepSeek V4 Flash by
default, no composer); the prompt is a byte-stable catalogue of the library so
the provider's prefix cache covers it. Every course id the model returns is
validated server-side, a bad JSON reply gets one retry then a plain-text
fallback, and a per-user daily cap with token counts lives in
advisor_requests. With AI_ENABLED false the app makes no outbound requests
and the box does not appear; users can also hide it per browser.
Claude-Session: https://claude.ai/code/session_012enquL2RZdjy3D5AcKP5yx
root