1.15.12

Merge pull request #14 from EudaLabs/copilot/update-ai-to-gemini-2-5-…

v1.5.11 — Lighthouse re-audit (100 / 100 / 100 on three categories) + 100 % code documentation

Highlights

• Lighthouse re-audit sweep — re-ran the production preview build after adding robots.txt, swapping the eager recharts preload for per-route lazy chunks, enabling sourcemaps, and preview-proxying /api. Scores moved from the Sprint 5 baseline of 93 / 100 / 96 / 91 (Perf / A11y / BP / SEO) to 91 / 100 / 100 / 100 — +4 Best Practices, +9 SEO, kept Accessibility at 100.
• JSDoc + docstring coverage is now 100 % on both sides. Frontend JSDoc went from ~64 % to 38 / 38 = 100 %; backend interrogate went from 23.9 % to 188 / 188 = 100 %. Tests stayed green (191 / 191).
• Week 11 jury showcase placeholder — new wiki page with the required-deliverables checklist, live surfaces, 10-min deck outline, FE prep checklist, and risk register.

Notable code changes

• frontend/public/robots.txt added (was missing — SEO 91 → 100).
• frontend/vite.config.ts: build.sourcemap: true; modulePreload.resolveDependencies drops vendor-charts from the HTML preload list; new preview.proxy so Lighthouse audits against pnpm preview can reach /api/*.
• frontend/src/App.tsx: Step 2/3/4/7 all lazy (recharts no longer eagerly loaded on Step 1 landing — ≈70 KiB less initial JS); loading skeleton gets min-height: 60vh.
• frontend/src/components/NavBar.tsx: navbar uses /logo-192.png with fetchPriority="high" + decoding="async" (~35 KiB less).
• frontend/src/styles/globals.css: nav-dropdown trigger uses black-alpha background so the white label clears 4.5 : 1 WCAG AA on the green nav.
• frontend/src/pages/Step1ClinicalContext.tsx: drops opacity: 0.85 on the active-step description (was blending to 4.34 : 1).
• +143 backend docstrings across app/** (package modules, Pydantic schemas, routers, services, specialty registry). +10 frontend JSDoc blocks across the App shell, components, Step 4, glossary, and the legend colour map.

Evidence artifacts

• docs/reports/Sprint5_Lighthouse.report.html / .json — re-audit (21 Apr).
• docs/reports/Sprint5_Lighthouse.report.baseline.html / .json — baseline snapshot (20 Apr, pre-re-audit).
• docs/reports/Sprint5_Lighthouse_Report.png — re-audit screenshot.
• docs/reports/Sprint5_Lighthouse_Report.baseline.png — baseline screenshot (for before/after).
• docs/reports/coverage/backend-docstring-coverage.txt + backend-docstring-badge.svg — interrogate output.
• docs/reports/coverage/frontend-jsdoc-coverage.txt — per-directory JSDoc scan.

Wiki: Sprint-5 page now shows both Lighthouse screenshots side by side, the Reports table carries the coverage + re-audit artifacts, the Releases table lists v1.5.11, and the new [[Final Submission]] page collects the Week 11 jury deliverables.

Not user-facing

• 191 / 191 pytest passed after the docstring sweep. No runtime code changed — only new docstrings / JSDoc blocks, frontend lazy-load plumbing, and the Vite/CSS tweaks called out above.

v1.5.10 — Step 7 label fix

UI fix

• Step 7 AI Clinical Assessment card badge now correctly shows Gemma 4 instead of the stale Gemini 2.5 Flash label. The default provider switched to Gemma 4 in v1.5.8 but this label was missed.

v1.5.9 — Step 7 AI insights reliability

Bug fix

• Step 7 AI Clinical Assessment no longer goes blank after the spinner.
  Gemma 4 reasoning calls occasionally exceeded the old 45s httpx timeout and surfaced as a silent ReadTimeout(''), leaving the card empty.

Changes

• Backend per-call LLM timeout: 45s → 200s
• Backend now retries transient failures (ReadTimeout, TransportError, 429, 5xx) once with jittered exponential backoff before falling back to the template.
• Backend logs exception repr so the real error class is visible.
• Frontend axios timeout: 120s → 450s
• Step 7 now renders a "assessment unavailable, reload to retry" warning instead of a blank space when the LLM falls back to the template.

Verification

5 consecutive /api/insights/<model_id> calls all returned source=gemini for all three tasks (ethics_insight, case_studies, eu_ai_act_insights). Times: 87s, 291s, 79s, 90s, 99s.

v1.5.8 — Gemma 4 default + MIT License

Highlights

Default LLM provider switched to Gemma 4 via Google AI Studio. Clinical insight generation (Step 7 Ethics, EU AI Act enrichment, case studies) now uses gemma-4-26b-a4b-it — a 26B Mixture-of-Experts model with only 4B active parameters per token, balancing clinical-reasoning quality with free-tier throughput.

Repository licensed under MIT. The informal "Academic / all rights reserved" note is replaced with a standard MIT LICENSE file so the project is usable as a portfolio reference and contributions are unambiguously welcome.

Changes

LLM / Insights

• backend/app/services/insight_service.py: default GEMINI_MODEL → gemma-4-26b-a4b-it
• Response parser now filters out thought=true parts (Gemma 4 returns chain-of-thought in a dedicated part) so the final answer is extracted correctly. Works transparently for any reasoning model, including future Gemini variants with thinking enabled.
• System instructions (clinical safety framing) confirmed to work on Gemma 4 via AI Studio.
• .env.example documents the new default; users can still pin a specific model via GEMINI_MODEL.

Licensing

• Added LICENSE (MIT) at repo root. GitHub now detects the license as "MIT" (previously "Other").
• README badge: License-Academic → License-MIT.
• frontend/package.json declares "license": "MIT".

Runtime requirements (HF Space)

Set GEMINI_API_KEY as a Secret in the HuggingFace Space settings (Variables and secrets → New secret). Without it, the provider falls back to static templates — no behaviour regression, only AI-generated narrative disappears.

Release details

Upstream diff: v1.5.7...v1.5.8

v1.5.7 — Sprint 5 Polish: Brand Identity, Docker, A11y 100, JSDoc

Sprint 5 Polish

Brand identity

• New HealthWithSevgi logo wired up: navbar mark (44×44 white rounded tile over green header), favicon family (16/32/48 ICO + 16/32 PNG + 180 Apple Touch), OpenGraph + Twitter previews at 192px
• Logo is alpha-centroid-centred (not just bbox-centred) so the heart + figure sit visually in the middle of the tile even with the leaf flourish on the right
• site.webmanifest (PWA-ready) with theme colour #1a7a4c

Accessibility

• Lighthouse Accessibility: 91 → 100
• Fixed 5 color-contrast violations: removed the blanket opacity: 0.45 that flattened locked wizard step labels to 1.79:1. Opacity is now scoped to the step-number circle; labels use --text-secondary (5.85:1). Session-privacy footer uses --text-secondary too.
• Added missing <main> landmark (landmark-one-main)
• Full WCAG math + before/after in docs/wiki/Accessibility-Log.md

Docker

• docker-compose.yml pulls ghcr.io/eudalabs/healthwithsevgi:latest when available, falls back to local multi-stage build
• Added healthcheck (probes /api/specialties every 10 s), container_name, and restart: unless-stopped
• Measured startup: 8 seconds — well inside the Sprint 5 ≤30 s target

Code documentation

• Added JSDoc to every frontend export in api/, pages/, and components/charts/ — 100%+ coverage (37 exports, 38 JSDoc blocks), hits the Sprint 5 ≥80% target

Sprint 5 planning

• docs/seng430-sprints/sprint-5-task-distribution.md: 24-task breakdown across BE (Efe+Berat), FE (Batu+Burak), QA (Berfin) with a 10-day schedule

Reports (new)

• docs/reports/Sprint5_Lighthouse_Report.png (+ .report.html, .report.json)
• docs/reports/Sprint5_Docker_Running.png
• docs/reports/Sprint5_Logo_Navbar.png

No runtime changes to

• Backend API contract
• ML pipelines
• Training / certificate flows

v1.5.6 — Sprint 4 Final: EU AI Act Checklist Fix

Sprint 4 Final

Fix

• EU AI Act compliance checklist now has exactly 8 items with 2 pre-checked on load, matching Sprint 4 spec (removed entry that caused 9-item/3-pre-checked state)

v1.5.5 — PDF Certificate Text Overflow Fix

Fixed

• PDF text overflow: EU AI Act checklist items (e.g. "Dataset licensing verified — 18/20 datasets...") now wrap properly across lines using ReportLab Paragraph objects instead of plain strings

Improved

• F1 Score added to subgroup fairness table (was on web UI but missing from PDF)
• Metric values color-coded (green/amber/red) in the Value column, matching the Status column
• Footer version updated to v1.5
• Page numbers added at bottom of every page

Verified

• 191/191 backend tests passing

v1.5.4 — Fix Learn More links + popover positioning

Fixed

• Learn More links now clickable: Portal popover's click-outside handler was treating clicks inside the popover as "outside" (since the portal is not a DOM child of the wrapper). Added popoverRef to the check so popover interactions work correctly.
• Popover edge clamping: Popovers near left/right screen edges now clamp their position so they don't overflow the viewport.

v1.5.3 — Fix popover clipping + broken links

Fixed

• InfoTip popover clipping: Popover now renders via React Portal at document.body level with position: fixed — no longer clipped by overflow: hidden containers (charts, column mapper, badge areas)
• 2 broken Learn More links: Fixed Google ML Crash Course URLs for "Overfitting" and "Train/Test Split" that were returning 404