Addresses real gaps found in testing (a Macedonian question on an English
page returned nothing):
- Detect the language of the QUESTION (Cyrillic -> MK; Slovene markers -> SI;
else the page language) and load that language's index + curated answers.
Fixes cross-language questions.
- Expand curated multilingual intents (now incl. enrolment/eVŠ, deadlines,
health, "what is MSOS", etc.) with stem keywords for better paraphrase
matching. MK/SI answers machine-drafted — proofread before launch.
- Suggested-question chips on open (localized) to guide users.
- Privacy-safe analytics: fire a GA4 'assistant_query' event ONLY after the
visitor accepts cookies (q text, result type, question vs page language) so
the team can see real FAQs and unanswered questions.
- Move the widget to bottom-left on pages with the floating "Become a member"
CTA so they never overlap. Design polish (chips, typing dots, launcher pulse).
- CI: rebuild the search index on every deploy (setup-python + build script)
so the assistant stays in sync with content automatically. Build script is
now path-portable (no hardcoded path).
Verified in a headless browser: MK-on-EN now answers correctly, chips work,
analytics event fires only with consent, become-a-member offset applies,
zero console errors.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A privacy-friendly help widget that answers visitor questions from the
site's own content and links to the full page — runs 100% in the browser,
no API, no server, no per-token cost.
- tools/build_search_index.py: parses every EN/MK/SI page into ~720
heading-level chunks per language -> assistant/index-<lang>.json
(lazy-loaded only when the chat is opened; gzips to ~80 KB).
- assistant.js: dependency-free BM25 search + multilingual stopword
filtering + a curated short-answer layer for common questions, plus an
accessible chat widget (role=dialog, keyboard, aria). CSP-safe (self only).
- assistant.css: scoped .msa-* styles in the site's teal palette;
responsive (full-screen panel on mobile).
- Injected on 208 pages (blog WIP excluded; picks it up when that lands).
Verified in a headless browser across EN/MK/SI: curated answers, search
fallback with correct ranking (rare terms win), correct deep links,
Cyrillic + Slovene diacritics, and zero console errors. Curated MK/SI
answers are machine-drafted — proofread before launch.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
popovskik