merge

feat: US-014 - Update to Gemini 3 Flash Preview with model indicator
feat: US-013 - Self-host ONNX embedding model
2026-02-15 23:20:24 +00:00 · 2026-02-15 21:02:52 +00:00 · 2026-02-15 20:59:03 +00:00 · 2026-02-15 20:48:00 +00:00 · 2026-02-15 20:43:48 +00:00
20 changed files with 32578 additions and 48 deletions
@@ -31,7 +31,8 @@
      "Bash(timeout /t 3 /nobreak)",
      "Bash(jq:*)",
      "Bash(git stash:*)",
-      "Bash(npx tsc:*)"
+      "Bash(npx tsc:*)",
      "mcp__context7__resolve-library-id"
    ]
  }
 }
@@ -0,0 +1,46 @@
 # Repository Guidelines
 ## Project Structure & Module Organization
 - Core app code lives in `src/`:
  - `src/components/` for UI components (`PascalCase.tsx`)
  - `src/hooks/` for custom hooks (`useX.ts`)
  - `src/lib/` for utilities and integrations (search, embeddings, Gemini)
  - `src/contexts/`, `src/types/`, and `src/data/` for state, typing, and static data
 - Static/public assets live in `public/` (including `public/models/`), while build output is generated in `dist/`.
 - Utility scripts live in `scripts/` (for example, `scripts/generate-embeddings.ts`).
 - Design references and experiments are in top-level folders such as `designs/`, `References/`, and `LogoAnimation/`.
 ## Build, Test, and Development Commands
 - `npm run dev` starts the Vite development server.
 - `npm run build` runs TypeScript project builds and creates a production bundle.
 - `npm run preview` serves the production build locally.
 - `npm run lint` runs ESLint across the repo.
 - `npm run typecheck` runs TypeScript checks without emitting files.
 - `npm run generate-embeddings` regenerates semantic-search embeddings.
 ## Coding Style & Naming Conventions
 - Language stack: TypeScript + React 18 + Vite.
 - Follow ESLint (`eslint.config.js`) and TypeScript strictness before opening PRs.
 - Use 2-space indentation and trailing commas where existing files do.
 - Naming conventions:
  - Components: `PascalCase` (`DashboardLayout.tsx`)
  - Hooks: `useCamelCase` (`useFocusTrap.ts`)
  - Utilities/data files: lowercase or kebab-style by domain (`semantic-search.ts`, `consultations.ts`).
 ## Testing Guidelines
 - There is currently no committed automated test framework (`*.test.*` / `*.spec.*` not present).
 - Minimum validation for each change: `npm run lint`, `npm run typecheck`, and `npm run build`.
 - For UI changes, include manual verification notes (route/flow tested, responsive behavior, accessibility impact).
 ## Commit & Pull Request Guidelines
 - Follow the existing history style: Conventional Commit prefixes (`feat:`, `chore:`) plus optional story IDs (for example, `feat: US-014 - ...`).
 - Keep commits focused and atomic; avoid mixing refactors with feature behavior.
 - PRs should include:
  - concise summary and motivation
  - linked task/story ID when available
  - screenshots/GIFs for visual changes
  - confirmation that lint, typecheck, and build passed.
 ## Security & Configuration Tips
 - Store secrets in `.env`; never hard-code API keys.
 - Do not commit local env files or generated artifacts outside intended tracked data.
@@ -0,0 +1,276 @@
 {
  "project": "Portfolio — Semantic Search & AI Chat",
  "branchName": "ralph/semantic-search",
  "description": "Replace Fuse.js command palette search with client-side semantic vector search (ONNX model), then add a Gemini Flash-powered AI chat widget.",
  "userStories": [
    {
      "id": "US-001",
      "title": "Install @xenova/transformers and add generate-embeddings script skeleton",
      "description": "As a developer, I need the Transformers.js dependency installed and a runnable script scaffold so subsequent stories can generate and use embeddings.",
      "acceptanceCriteria": [
        "npm install @xenova/transformers",
        "Create scripts/generate-embeddings.ts with a main() function that imports the pipeline from @xenova/transformers",
        "Script loads the all-MiniLM-L6-v2 model and embeds a single test string, logging the vector length to confirm it works",
        "Add npm script: \"generate-embeddings\": \"npx tsx scripts/generate-embeddings.ts\"",
        "Running npm run generate-embeddings prints the vector length (384) and exits cleanly",
        "Typecheck passes"
      ],
      "priority": 1,
      "passes": true,
      "notes": "Use @xenova/transformers (not @huggingface/transformers — the Xenova fork has better Node.js ONNX support). The model ID is 'Xenova/all-MiniLM-L6-v2'. Pipeline type is 'feature-extraction'. tsx is already available via npx for running TypeScript scripts."
    },
    {
      "id": "US-002",
      "title": "Build rich text representations for each palette item",
      "description": "As a developer, I want each palette item to have a natural-language paragraph for embedding that captures deep context, not just the title.",
      "acceptanceCriteria": [
        "New function buildEmbeddingTexts() in src/lib/search.ts that returns Array<{ id: string, text: string }> for all palette items",
        "Consultation items include: role, org, duration, history narrative, examination bullets, coded entry descriptions",
        "Skill items include: name, category, frequency, proficiency percentage, years of experience",
        "KPI items include: value, label, explanation, story context and outcomes",
        "Investigation items include: name, methodology, tech stack list, results",
        "Education items include: title, institution, type, research detail",
        "Quick Action items include: title and subtitle (short text is fine)",
        "Achievement items include: title, subtitle, and linked KPI story context if available",
        "Each text is a readable natural-language paragraph, not a keyword dump",
        "Typecheck passes"
      ],
      "priority": 2,
      "passes": true,
      "notes": "This function will be used by both the build script (to generate embeddings) and potentially by the chat widget (for context). Import the raw data files (consultations, skills, kpis, investigations, documents) to access the full data beyond what buildPaletteData() surfaces. The id must match the PaletteItem id so embeddings can be correlated."
    },
    {
      "id": "US-003",
      "title": "Generate and commit embeddings.json",
      "description": "As a developer, I want the generate-embeddings script to produce a complete embeddings.json file using the rich text representations.",
      "acceptanceCriteria": [
        "scripts/generate-embeddings.ts imports buildEmbeddingTexts() from src/lib/search.ts",
        "Script embeds each item's text using the all-MiniLM-L6-v2 model via @xenova/transformers pipeline",
        "Outputs src/data/embeddings.json as an array of { id: string, embedding: number[] }",
        "Each embedding is a 384-dimensional float array",
        "Running npm run generate-embeddings regenerates the file successfully",
        "The JSON file is valid and parseable",
        "Typecheck passes"
      ],
      "priority": 3,
      "passes": true,
      "notes": "The pipeline returns a Tensor — use .tolist() or .data to extract the raw float array. Mean-pool across the token dimension (dim 1) to get a single 384-d vector per input. Process items sequentially to avoid OOM in Node. The output file will be ~200KB for ~40 items with 384 floats each."
    },
    {
      "id": "US-004",
      "title": "Preload ONNX model during boot sequence",
      "description": "As a visitor, I want the semantic search model to download in the background during the boot/ECG/login phases so it's ready when I reach the dashboard.",
      "acceptanceCriteria": [
        "New src/lib/embedding-model.ts module that exports: initModel(), embedQuery(text: string), and isModelReady()",
        "initModel() loads the all-MiniLM-L6-v2 pipeline from @xenova/transformers and stores it in a module-level variable",
        "embedQuery() returns a Promise<number[]> (384-d vector) for a given text string",
        "isModelReady() returns boolean indicating if the model has finished loading",
        "initModel() is called in App.tsx useEffect on mount (during boot phase) — fire and forget, no await",
        "If initModel() fails (network error, etc.), isModelReady() remains false — no error thrown or shown",
        "Model is cached by @xenova/transformers in IndexedDB — subsequent page loads are near-instant",
        "Boot/ECG/login animations are not affected by model loading",
        "Typecheck passes"
      ],
      "priority": 4,
      "passes": true,
      "notes": "Use pipeline('feature-extraction', 'Xenova/all-MiniLM-L6-v2') which auto-downloads and caches the ONNX model. The module-level pattern (let pipelineInstance = null) avoids React re-render issues. embedQuery should mean-pool the tensor output the same way as the build script. Wrap initModel() in a try/catch that silently swallows errors."
    },
    {
      "id": "US-005",
      "title": "Implement cosine similarity search module",
      "description": "As a developer, I need a semantic search function that compares a query embedding against pre-computed item embeddings and returns ranked results.",
      "acceptanceCriteria": [
        "New src/lib/semantic-search.ts module",
        "Exports semanticSearch(queryEmbedding: number[], embeddings: Array<{ id: string, embedding: number[] }>, threshold?: number): Array<{ id: string, score: number }>",
        "Uses cosine similarity: dot(a,b) / (magnitude(a) * magnitude(b))",
        "Results sorted by score descending",
        "Optional threshold parameter filters out low-relevance results (default 0.3)",
        "Exports loadEmbeddings() that imports embeddings.json and returns the parsed array",
        "Typecheck passes"
      ],
      "priority": 5,
      "passes": true,
      "notes": "Keep the cosine similarity implementation simple — no libraries needed for 384-d vectors over ~40 items. The loadEmbeddings function can use a dynamic import or direct import of the JSON file (Vite handles JSON imports natively)."
    },
    {
      "id": "US-006",
      "title": "Integrate semantic search into command palette",
      "description": "As a visitor, I want the command palette to use semantic search when available, falling back to Fuse.js otherwise.",
      "acceptanceCriteria": [
        "CommandPalette.tsx checks isModelReady() from embedding-model.ts",
        "When model is ready and query is non-empty: call embedQuery(query), then semanticSearch() against loaded embeddings, then map result IDs back to PaletteItem objects",
        "When model is NOT ready: use existing Fuse.js search (current behavior preserved exactly)",
        "Search is debounced by ~200ms to avoid calling embedQuery on every keystroke",
        "Results maintain existing groupBySection() grouping and section ordering",
        "Existing keyboard navigation, action routing, and UI unchanged",
        "Typecheck passes",
        "Verify in browser: search 'data analysis' surfaces analytics-related roles/skills not just items with 'data' in title"
      ],
      "priority": 6,
      "passes": true,
      "notes": "The debounce is important — embedQuery takes ~20-50ms per call. Use a useRef + setTimeout pattern or a simple debounce hook. The mapping from semantic search results (id + score) back to PaletteItems should use a Map for O(1) lookup. Keep the Fuse.js imports and buildSearchIndex — they're the fallback path."
    },
    {
      "id": "US-007",
      "title": "Chat widget — floating button component",
      "description": "As a visitor, I see a floating chat button at the bottom-right of the dashboard that I can click to open a chat panel.",
      "acceptanceCriteria": [
        "New src/components/ChatWidget.tsx component",
        "Renders a 48px circular button, fixed position, bottom: 24px, right: 24px",
        "Uses teal accent background (var(--accent)), white MessageCircle icon from lucide-react",
        "Shadow: var(--shadow-md). Hover: var(--shadow-lg) + scale(1.05) transition",
        "Button has a subtle entrance animation: fade + translateY(8px) → translateY(0), delayed ~1s after mount",
        "Respects prefers-reduced-motion (no animation, just visible)",
        "z-index above dashboard content but below command palette overlay (z-index 90)",
        "onClick toggles an isOpen state (panel rendering comes in next story)",
        "Mounted in DashboardLayout.tsx",
        "Typecheck passes",
        "Verify in browser using dev-browser skill"
      ],
      "priority": 7,
      "passes": true,
      "notes": "Use framer-motion for the entrance animation to match the rest of the app's motion patterns. The button should use font-ui for any text. On mobile (<640px), button is 40px and positioned bottom: 16px, right: 16px. The VITE_GEMINI_API_KEY env var check can wait until the Gemini integration story — for now just render the button unconditionally."
    },
    {
      "id": "US-008",
      "title": "Chat widget — panel UI with message display",
      "description": "As a visitor, I want a chat panel that opens above the floating button where I can type questions and see responses.",
      "acceptanceCriteria": [
        "Chat panel renders when isOpen is true, positioned above the floating button (bottom: 88px, right: 24px)",
        "Panel dimensions: 380px wide, max-height 480px, with overflow-y auto for messages",
        "Header: title text ('Ask about Andy'), close button (X icon)",
        "Message area: user messages right-aligned in teal-tinted bubbles, assistant messages left-aligned in light gray bubbles",
        "Input area at bottom: text field with placeholder 'Ask me anything...', send button (Send icon)",
        "Enter key submits message, Shift+Enter for newline",
        "Panel entrance animation: scale(0.95) + opacity(0) → scale(1) + opacity(1), 200ms ease-out",
        "Panel exit animation: reverse of entrance",
        "Respects prefers-reduced-motion",
        "Responsive: on mobile (<640px), panel is full-width (left: 0, right: 0, bottom: 0) with rounded top corners only",
        "Messages are stored in component state as Array<{ role: 'user' | 'assistant', content: string }>",
        "Submitting a message adds it to state and shows it in the UI (no API call yet — assistant response is a placeholder)",
        "Typecheck passes",
        "Verify in browser using dev-browser skill"
      ],
      "priority": 8,
      "passes": true,
      "notes": "Use the design system tokens: var(--surface) for panel bg, var(--border-light) for borders, var(--text-primary) for text, var(--accent) for user bubble bg at 10% opacity, font-ui for body text, font-geist for timestamps. The placeholder assistant response can be a static string like 'AI chat coming soon — this is a preview of the chat interface.' This lets us verify the full UI before wiring up Gemini."
    },
    {
      "id": "US-009",
      "title": "Chat widget — Gemini Flash integration",
      "description": "As a visitor, I can ask natural language questions and get intelligent, streamed answers about Andy's experience.",
      "acceptanceCriteria": [
        "New src/lib/gemini.ts module that exports sendChatMessage(messages: ChatMessage[], cvContext: string): AsyncGenerator<string>",
        "Calls Google Gemini Flash API (gemini-2.0-flash) using the REST API with fetch (no SDK needed)",
        "API key sourced from import.meta.env.VITE_GEMINI_API_KEY",
        "System prompt includes structured CV context built from buildEmbeddingTexts() output",
        "System prompt instructs the model to answer questions about Andy's professional experience accurately and concisely",
        "System prompt instructs the model to include relevant palette item IDs in its response as a JSON array at the end",
        "Responses are streamed using the Gemini streaming endpoint",
        "ChatWidget.tsx wires up real messages: on submit, calls sendChatMessage and streams tokens into the assistant message bubble",
        "Loading state shown (typing indicator) while waiting for first token",
        "If VITE_GEMINI_API_KEY is not set, chat button is still visible but panel shows 'Chat is currently unavailable' message",
        "If API call fails, show error message in chat: 'Sorry, I couldn't process that. Please try again.'",
        "Conversation history (last 10 messages) passed to API for multi-turn context",
        "Typecheck passes"
      ],
      "priority": 9,
      "passes": true,
      "notes": "Gemini REST streaming endpoint: POST https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash:streamGenerateContent?alt=sse&key=API_KEY. The response is SSE (server-sent events) — parse each 'data:' line as JSON and extract candidates[0].content.parts[0].text. The system prompt with CV context will be ~2-3K tokens — well within Gemini Flash limits. For the palette item IDs, instruct the model to end its response with a line like [ITEMS: id1, id2, id3] which can be parsed client-side."
    },
    {
      "id": "US-010",
      "title": "Chat widget — clickable portfolio item cards in responses",
      "description": "As a visitor, I want AI chat responses to include clickable portfolio items so I can drill into relevant sections.",
      "acceptanceCriteria": [
        "After parsing the assistant response, extract referenced palette item IDs from the [ITEMS: ...] suffix",
        "Render matched items as compact clickable cards below the answer text in the assistant bubble",
        "Cards reuse icon/color mapping from CommandPalette (iconByType, iconColorStyles)",
        "Cards show item title and subtitle in a compact horizontal layout",
        "Clicking a card triggers the same action routing as command palette via handlePaletteAction in DashboardLayout",
        "If no items are referenced or IDs don't match, no cards are shown (just the text answer)",
        "Typecheck passes",
        "Verify in browser using dev-browser skill"
      ],
      "priority": 10,
      "passes": true,
      "notes": "The action routing needs to flow from ChatWidget up to DashboardLayout. Add an onAction prop to ChatWidget (same pattern as CommandPalette). DashboardLayout passes handlePaletteAction to ChatWidget. Export iconByType and iconColorStyles from CommandPalette (or extract to a shared module) so ChatWidget can reuse them."
    },
    {
      "id": "US-011",
      "title": "Mobile full-screen chat panel",
      "description": "As a mobile visitor, I want the chat panel to be a full-screen overlay so it's easy to use on small screens.",
      "acceptanceCriteria": [
        "Below md breakpoint (768px), chat panel renders as full-screen overlay using position: fixed; inset: 0 with 100dvh height",
        "Full-screen mode has the existing header with close button (no visual change needed, just full-width)",
        "Floating chat button is hidden (display: none or opacity: 0) while panel is open on mobile (<768px)",
        "Above 768px, existing panel behavior is unchanged (380px wide, anchored bottom-right, max-height 480px)",
        "Panel open/close animation still respects prefers-reduced-motion",
        "Safe area insets applied via env(safe-area-inset-*) for notched devices",
        "Input area stays pinned to bottom of screen on mobile",
        "Typecheck passes",
        "Verify in browser using dev-browser skill"
      ],
      "priority": 11,
      "passes": true,
      "notes": "The current ChatWidget already has some mobile handling (bottom-sheet style at <640px). This story changes the breakpoint to 768px (md) and makes it truly full-screen instead of 85vh. Use 100dvh (dynamic viewport height) to account for mobile browser chrome. The floating button visibility can be controlled by combining isOpen state with a CSS media query or a useMediaQuery hook. The <style> block with data-chat-panel attribute is the place to update responsive rules."
    },
    {
      "id": "US-012",
      "title": "Welcome message with suggested question chips",
      "description": "As a visitor opening the chat, I see a friendly welcome message and clickable suggested questions so I know what to ask.",
      "acceptanceCriteria": [
        "When chat panel is open and conversation is empty, display welcome text: 'Hey! I'm here to help you learn more about Andy. What would you like to know?'",
        "Welcome text is styled as an AI message bubble (left-aligned, light background, same styling as assistant messages)",
        "Below the welcome bubble, show 2-3 clickable pill/chip buttons with suggested questions",
        "Suggested questions: 'What's his NHS experience?', 'Tell me about his data skills', 'What projects has he built?'",
        "Chips styled with: teal accent border, rounded-full, font-ui 12-13px, hover state (teal background tint)",
        "Clicking a chip sends that question as a user message (same codepath as typing + Enter)",
        "Welcome message and chips always visible when conversation is empty (persist across panel open/close)",
        "Once any message is sent, the welcome/chips area is replaced by the conversation messages",
        "Typecheck passes",
        "Verify in browser using dev-browser skill"
      ],
      "priority": 12,
      "passes": true,
      "notes": "Replace the current empty-state text ('Ask me anything about Andy's experience, skills, or projects.') with the new welcome bubble + chips. The chips should call handleSubmit (or equivalent) with the chip text pre-filled — simplest approach is setInputValue(chipText) then immediately trigger submit. Check that the welcome state reappears if the user hasn't sent a message (messages.length === 0). The suggested questions could live in a const array at the top of ChatWidget for easy future editing."
    },
    {
      "id": "US-013",
      "title": "Self-host ONNX embedding model",
      "description": "As a developer, I want the ONNX model files served from the same host as the site to eliminate dependency on Hugging Face CDN.",
      "acceptanceCriteria": [
        "Model files for Xenova/all-MiniLM-L6-v2 downloaded and placed in public/models/all-MiniLM-L6-v2/onnx/ (matching HF repo structure)",
        "Required files present: model_quantized.onnx, tokenizer.json, tokenizer_config.json, config.json, and any other files the pipeline expects",
        "src/lib/embedding-model.ts updated: configure @xenova/transformers env to use local model path (e.g., env.localModelPath or custom model URL pointing to /models/)",
        "scripts/generate-embeddings.ts also updated to use the same local model path for consistency",
        "Model files are NOT in .gitignore — they are committed as static assets",
        "No network requests to huggingface.co in the browser network tab when semantic search is used",
        "Semantic search still works correctly in the command palette after the change",
        "Typecheck passes"
      ],
      "priority": 13,
      "passes": true,
      "notes": "Transformers.js uses env.localModelPath or env.remoteHost to control where models are fetched from. Setting env.localModelPath = '/models/' should make it look for files at /models/Xenova/all-MiniLM-L6-v2/onnx/model_quantized.onnx etc. The Vite public/ directory serves files at the root — so public/models/ becomes /models/ at runtime. For the build script (Node.js), use a file:// path or the local filesystem path instead. Download model files from https://huggingface.co/Xenova/all-MiniLM-L6-v2/tree/main — the quantized ONNX model is ~23MB. Check what files the pipeline actually requests by watching network tab before making this change."
    },
    {
      "id": "US-014",
      "title": "Update to Gemini 3 Flash Preview with model indicator",
      "description": "As a developer, I want to use the latest free Gemini model, and as a visitor, I want to see what model powers the chat.",
      "acceptanceCriteria": [
        "Extract model name to a single constant (e.g., GEMINI_MODEL = 'gemini-3-flash-preview') used for both the API URL and display",
        "GEMINI_API_BASE URL updated to use the new model constant",
        "Review and tighten the system prompt — ensure it's well-structured, concise, and clear for the new model",
        "Review the [ITEMS: ...] suffix instruction — ensure new model follows the format reliably",
        "Small model indicator in chat panel header: 'Gemini 3 Flash' in font-geist, 11px, var(--text-tertiary)",
        "Model indicator positioned right-aligned in the header bar or as a subtle line below the header",
        "Streaming SSE parsing still works correctly with the new model endpoint",
        "Typecheck passes",
        "Verify in browser using dev-browser skill"
      ],
      "priority": 14,
      "passes": true,
      "notes": "The current API base is 'https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash'. Change the model segment to 'gemini-3-flash-preview'. The API path structure (v1beta/models/{model}:streamGenerateContent) should be the same. Verify that gemini-3-flash-preview is the correct model ID — check Google AI Studio or the API docs. For the display name, use a human-friendly string like 'Gemini 3 Flash' (not the full model ID). The constant should be defined at the top of gemini.ts and exported for use in ChatWidget."
    }
  ]
 }
@@ -0,0 +1,315 @@
 # Progress Log — Semantic Search & AI Chat
 # Branch: ralph/semantic-search
 # Started: 2026-02-15
 ## Codebase Patterns
 - `@xenova/transformers` pipeline with `pooling: 'mean'` and `normalize: true` returns a Tensor; use `Array.from(output.data as Float32Array)` to extract the 384-d vector
 - Scripts live in `scripts/` and run via `npx tsx` (tsx is not a project dep, npx fetches it)
 - tsconfig `include` only covers `src/` — scripts are type-checked by tsx at runtime, not by `tsc --noEmit`
 - Project uses `"type": "module"` in package.json
 - Palette item IDs: `exp-{consultation.id}`, `skill-{skill.id}`, `proj-{investigation.id}`, `ach-{0-3}`, `edu-{0-3}`, `action-{0-3}`
 - `buildEmbeddingTexts()` in `src/lib/search.ts` returns `Array<{ id: string, text: string }>` with IDs matching PaletteItem IDs — use this for both embedding generation and chat context
 - `src/data/embeddings.json` is an array of `{ id: string, embedding: number[] }` — 42 items, 384-d vectors, IDs match PaletteItem IDs. Vite imports JSON natively.
 - `src/lib/embedding-model.ts` exports `initModel()`, `embedQuery(text)`, `isModelReady()` — check `isModelReady()` before calling `embedQuery()`
 - `initModel()` is called fire-and-forget in `App.tsx` on mount — model loads during boot/ECG/login phases
 - ONNX model files self-hosted in `public/models/Xenova/all-MiniLM-L6-v2/` — `env.localModelPath = '/models/'`, `env.allowRemoteModels = false`, `env.useBrowserCache = false` eliminates HF CDN dependency
 - `src/lib/semantic-search.ts` exports `semanticSearch(queryEmbedding, embeddings, threshold?)` and `loadEmbeddings()` — embeddings are normalized so cosine similarity is dot(a,b)/(mag(a)*mag(b))
 - CommandPalette uses `semanticResults` state + debounced `useEffect` for async semantic search, falling back to Fuse.js when `isModelReady()` returns false or on any error
 - `loadEmbeddings()` and `paletteMap` (Map<id, PaletteItem>) are precomputed via `useMemo` — no re-computation on each search
 - ChatWidget is mounted in DashboardLayout alongside CommandPalette and DetailPanel — z-index 90 (below command palette z-1000)
 - `prefersReducedMotion` pattern: read `window.matchMedia` at module level, use in framer-motion variants to skip animation
 - ChatWidget stores messages as `Array<{ role: 'user' | 'assistant', content: string }>` — same shape as LLM message format, ready for Gemini integration
 - ChatWidget `isOpen` state controls both panel visibility and button icon (MessageCircle ↔ X) — panel rendering handled by AnimatePresence
 - `src/lib/gemini.ts` exports `sendChatMessage(messages)` (async generator), `isGeminiAvailable()`, `parseItemIds(text)`, `stripItemsSuffix(text)` — ChatMessage type is `{ role: 'user' | 'assistant', content: string }`
 - Gemini API uses SSE streaming: POST to `:streamGenerateContent?alt=sse&key=KEY`, parse `data:` lines as JSON, extract `candidates[0].content.parts[0].text`
 - System prompt built from `buildEmbeddingTexts()` — instructs model to end responses with `[ITEMS: id1, id2, id3]` for portfolio item linking
 - `isGeminiAvailable()` checks `import.meta.env.VITE_GEMINI_API_KEY` — when missing, chat panel shows "unavailable" message but button remains visible
 - Assistant messages store item IDs as `<!--ITEMS:id1,id2-->` HTML comment suffix for US-010 to parse — `getDisplayText()` strips this before rendering
 - Conversation history capped at 10 messages (`MAX_HISTORY`), metadata stripped before sending to API
 - Icon/color mappings (`iconByType`, `iconColorStyles`) live in `src/lib/palette-icons.ts` — shared between CommandPalette and ChatWidget
 - ChatWidget accepts optional `onAction?: (action: PaletteAction) => void` prop — same pattern as CommandPalette's `onAction`
 - `DashboardLayout` passes `handlePaletteAction` to both CommandPalette and ChatWidget for unified action routing
 - TopBar is `z-index: 100` (fixed), nav is `z-index: 99` (sticky) — mobile full-screen overlays need `z-index > 100` to appear above them
 - Inline `style={{ display: 'flex' }}` overrides Tailwind's `hidden` class — use `!important` modifier (`max-md:!hidden`) or move display to Tailwind classes to allow responsive hiding
 - ChatWidget mobile breakpoint is `md` (768px) — below this, panel is full-screen; above, it's 380px anchored bottom-right
 - `handleSubmit(overrideText?)` accepts optional text param — use this when programmatically sending messages (e.g., suggested question chips) to avoid stale `inputValue` state
 - `SUGGESTED_QUESTIONS` const array at top of ChatWidget — edit here to change welcome screen chip text
 - `GEMINI_MODEL` and `GEMINI_DISPLAY_NAME` exported from `src/lib/gemini.ts` — single source of truth for model ID and display name; update both when changing models
 ---
 ## 2026-02-15 - US-001
 - Installed `@xenova/transformers` (^2.17.2)
 - Created `scripts/generate-embeddings.ts` with main() that loads `Xenova/all-MiniLM-L6-v2` and embeds a test string
 - Added `"generate-embeddings"` npm script
 - Verified: outputs vector length 384 and exits cleanly
 - Typecheck passes
 - Files changed: `package.json`, `package-lock.json`, `scripts/generate-embeddings.ts`
 - **Learnings for future iterations:**
  - `pipeline('feature-extraction', 'Xenova/all-MiniLM-L6-v2')` auto-downloads and caches the ONNX model (~23MB)
  - First run takes a few seconds for model download; subsequent runs are near-instant from cache
  - The pipeline's `pooling: 'mean'` and `normalize: true` options handle mean-pooling and L2 normalization in one step — no manual tensor manipulation needed
  - `output.data` is a `Float32Array`; wrap in `Array.from()` for a plain number array
 ---
 ## 2026-02-15 - US-002
 - Added `buildEmbeddingTexts()` function to `src/lib/search.ts`
 - Imports all raw data files (consultations, skills, kpis, investigations, documents)
 - Generates natural-language paragraphs for each palette item type:
  - Consultations: role, org, duration, history narrative, examination bullets, coded entry descriptions
  - Skills: name, category, frequency, proficiency %, years of experience
  - Achievements: title, subtitle, full KPI explanation + story context + outcomes
  - Investigations: name, methodology, tech stack, results
  - Education: title, type, institution, duration, classification, research detail, notes (from documents.ts)
  - Quick Actions: title + subtitle
 - IDs match PaletteItem IDs (e.g. `exp-{id}`, `skill-{id}`, `ach-{i}`, `proj-{id}`, `edu-{i}`, `action-{i}`)
 - Typecheck and lint pass
 - Files changed: `src/lib/search.ts`
 - **Learnings for future iterations:**
  - Education items in `buildPaletteData()` are hardcoded arrays (not iterated from `documents`), with ids `edu-0` through `edu-3`. The mapping to `documents.ts` entries is: edu-0→doc-mary-seacole, edu-1→doc-mpharm, edu-2→doc-alevels, edu-3→doc-gphc
  - Achievement items are similarly hardcoded with ids `ach-0` through `ach-3`, each linked to a KPI id
  - Quick action items are `action-0` through `action-3`
  - `documents.ts` is imported but wasn't previously used in `search.ts` — now used for education embedding text
 ---
 ## 2026-02-15 - US-003
 - Updated `scripts/generate-embeddings.ts` to import `buildEmbeddingTexts()` and generate full embeddings
 - Script embeds all 42 palette items sequentially using `Xenova/all-MiniLM-L6-v2`
 - Outputs `src/data/embeddings.json` as `Array<{ id: string, embedding: number[] }>`
 - Each embedding is a 384-dimensional float array
 - File is ~453KB (42 items × 384 floats with pretty-printed JSON)
 - `npm run generate-embeddings` regenerates the file successfully
 - Typecheck and lint pass
 - Files changed: `scripts/generate-embeddings.ts`, `src/data/embeddings.json`
 - **Learnings for future iterations:**
  - `import.meta.dirname` works in tsx/Node ESM scripts — use it instead of `__dirname` (which isn't available in ESM)
  - `@/` path alias works in `npx tsx` scripts because tsx resolves tsconfig paths automatically
  - The embeddings file is ~450KB with pretty-print; could be reduced with compact JSON but readability is preferred for now
  - Processing 42 items takes ~10-15 seconds on first run (model cached after first download)
 ---
 ## 2026-02-15 - US-004
 - Created `src/lib/embedding-model.ts` with three exports: `initModel()`, `embedQuery()`, `isModelReady()`
 - Module-level `let extractor` pattern avoids React re-render issues
 - `initModel()` uses `loading` guard to prevent duplicate pipeline loads
 - `embedQuery()` uses same `pooling: 'mean'` and `normalize: true` as the build script
 - `initModel()` called fire-and-forget in `App.tsx` `useEffect([], [])` — runs during boot phase
 - Silent failure: try/catch swallows errors, `isModelReady()` stays false
 - Typecheck, lint, and build all pass
 - Files changed: `src/lib/embedding-model.ts` (new), `src/App.tsx`
 - **Learnings for future iterations:**
  - `FeatureExtractionPipeline` type is exported from `@xenova/transformers` and can be used for the module-level variable
  - The `loading` boolean guard prevents race conditions if `initModel()` is called multiple times (e.g., React strict mode double-mount)
  - `initModel()` is intentionally not awaited — it's fire-and-forget so it doesn't block the boot animation
  - Consumers should check `isModelReady()` before calling `embedQuery()` — it throws if model isn't loaded
 ---
 ## 2026-02-15 - US-005
 - Created `src/lib/semantic-search.ts` with cosine similarity search and embeddings loader
 - `semanticSearch()` computes cosine similarity, filters by threshold (default 0.3), returns sorted by score descending
 - `loadEmbeddings()` imports `embeddings.json` via Vite's native JSON import and returns typed array
 - Typecheck and lint pass (0 new warnings)
 - Files changed: `src/lib/semantic-search.ts` (new)
 - **Learnings for future iterations:**
  - Vite handles JSON imports natively — `import data from '@/data/embeddings.json'` just works, no dynamic import needed
  - Since embeddings are already L2-normalized (from pipeline's `normalize: true`), cosine similarity simplifies to just the dot product. However, the full formula is kept for correctness in case non-normalized vectors are ever used
  - With only ~42 items and 384-d vectors, brute-force cosine similarity is fast enough — no need for approximate nearest neighbor libraries
 ---
 ## 2026-02-15 - US-006
 - Integrated semantic search into CommandPalette with Fuse.js fallback
 - When `isModelReady()` is true: debounces query by 200ms, calls `embedQuery()`, runs `semanticSearch()` against preloaded embeddings, maps result IDs back to PaletteItems via O(1) Map lookup
 - When model is NOT ready: uses existing Fuse.js search (behavior preserved exactly)
 - Results maintain `groupBySection()` grouping and section ordering
 - Existing keyboard navigation, action routing, and UI unchanged
 - Semantic results state is cleared when palette opens/closes and when query is empty
 - Error handling: any failure in embedQuery/semanticSearch silently falls back to Fuse.js
 - Typecheck, lint, and build all pass
 - Browser verified: Fuse.js fallback works correctly; ONNX model loads asynchronously during boot and activates semantic search when ready
 - Files changed: `src/components/CommandPalette.tsx`
 - **Learnings for future iterations:**
  - Semantic search is async so it can't live in a `useMemo` — use `useState` + debounced `useEffect` pattern instead
  - The `useRef + setTimeout` debounce pattern works well here: set `debounceRef.current = setTimeout(...)`, clear it in the cleanup function, and in early-return paths
  - `isModelReady()` is a synchronous check — call it before setting up the debounce timeout to avoid unnecessary delays when model isn't loaded
  - The ONNX model takes several seconds to load in the browser (downloads ~23MB first time, then cached in IndexedDB), so initial searches will always use Fuse.js fallback
  - `loadEmbeddings()` is cheap (just returns the already-imported JSON) — safe to call in `useMemo` without performance concern
 ---
 ## 2026-02-15 - US-007
 - Created `src/components/ChatWidget.tsx` — floating chat button with toggle state
 - 48px circular button (40px on mobile <640px), fixed bottom-right, teal accent background, white MessageCircle icon
 - Entrance animation: fade + translateY(8px→0), 1s delay after mount, via framer-motion variants
 - Respects `prefers-reduced-motion` — skips animation, shows immediately
 - Hover: shadow-md → shadow-lg + scale(1.05), 150ms transition
 - z-index 90 (below command palette z-1000)
 - onClick toggles `isOpen` state, swaps icon between MessageCircle and X
 - Mounted in `DashboardLayout.tsx` alongside CommandPalette and DetailPanel
 - Typecheck, lint (0 errors), and build all pass
 - Browser verified: button visible at bottom-right, toggle works (Open chat ↔ Close chat)
 - Files changed: `src/components/ChatWidget.tsx` (new), `src/components/DashboardLayout.tsx`
 - **Learnings for future iterations:**
  - Responsive sizing via Tailwind classes (`h-10 w-10 sm:h-12 sm:w-12`) works well with inline style for non-Tailwind properties (boxShadow, border-radius)
  - `AnimatePresence` is already imported and ready for the panel animation in US-008
  - The `isOpen` state lives in ChatWidget — US-008 will add the panel UI inside the same component
  - Hover effects use `onMouseEnter/Leave` with direct style mutation (same pattern as other dashboard components)
 ---
 ## 2026-02-15 - US-008
 - Built chat panel UI inside `ChatWidget.tsx` with header, message area, and input
 - Panel opens above the floating button with scale+opacity entrance/exit animation via framer-motion `AnimatePresence`
 - Messages stored as `Array<{ role: 'user' | 'assistant', content: string }>` in component state
 - User messages right-aligned in teal-tinted bubbles (`var(--accent-light)` bg, `var(--accent-border)` border)
 - Assistant messages left-aligned in light gray bubbles (`var(--bg-dashboard)` bg, `var(--border-light)` border)
 - Message corner radii differ: user bubbles have small bottom-right radius, assistant bubbles small bottom-left (conversational feel)
 - Input area: textarea with Enter to submit, Shift+Enter for newline. Send button enabled/disabled based on input content
 - Empty state shows placeholder text when no messages yet
 - Auto-scrolls to latest message via `useRef` + `scrollIntoView`
 - Auto-focuses input when panel opens (200ms delay for animation)
 - Responsive: on mobile (<640px), panel is full-width bottom sheet with rounded top corners; on desktop, 380px wide positioned above the button
 - Panel entrance: scale(0.95)+opacity(0) → scale(1)+opacity(1), 200ms. Exit: reverse, 150ms
 - Respects `prefers-reduced-motion` — skips all animation
 - Close button in header triggers `setIsOpen(false)` (same as floating button toggle)
 - Submitting appends both user message and placeholder assistant response to state
 - Typecheck, lint (0 errors), and build all pass
 - Browser verified: panel opens/closes correctly, messages display, input works, Enter submits, close button works
 - Files changed: `src/components/ChatWidget.tsx`
 - **Learnings for future iterations:**
  - `AnimatePresence` with `key` prop on the panel div is needed for exit animations to work
  - Panel uses `transformOrigin: 'bottom right'` for natural scale animation from the button corner
  - CSS-in-JS `<style>` tag with `data-chat-panel` attribute handles responsive width/height (Tailwind can't express max-height conditionally based on viewport width easily)
  - `textarea` with `rows={1}` and `maxHeight: 80px` gives auto-growing feel; `resize: none` prevents manual resize
  - The `ChatMessage` interface (`{ role, content }`) is ready to be extended for US-009 Gemini integration — same shape as typical LLM message format
  - `onFocus/onBlur` border color transitions on the textarea give a polished input interaction
 ---
 ## 2026-02-15 - US-009
 - Created `src/lib/gemini.ts` — Gemini Flash streaming integration module
  - `sendChatMessage(messages)` async generator that streams SSE tokens from Gemini 2.0 Flash
  - `isGeminiAvailable()` checks for `VITE_GEMINI_API_KEY` env var
  - `parseItemIds(text)` extracts `[ITEMS: id1, id2]` from response text
  - `stripItemsSuffix(text)` removes the `[ITEMS: ...]` line for clean display
  - System prompt built from `buildEmbeddingTexts()` output — full CV context (~42 items)
  - Model instructed to answer concisely and append relevant palette item IDs
 - Rewired `ChatWidget.tsx` to use real Gemini API instead of placeholder responses
  - Streaming: tokens progressively appear in assistant message bubble
  - Typing indicator (Loader2 spinner + "Thinking...") shown while waiting for first token
  - Input disabled during streaming, send button grayed out
  - Error handling: API failures show "Sorry, I couldn't process that. Please try again."
  - Missing API key: panel shows "Chat is currently unavailable", input area hidden
  - Conversation history capped at 10 messages before sending to API
  - Assistant messages store parsed item IDs as `<!--ITEMS:id1,id2-->` HTML comment (for US-010)
  - Messages sent to API have metadata stripped to keep context clean
 - Typecheck, lint (0 errors), and build all pass
 - Files changed: `src/lib/gemini.ts` (new), `src/components/ChatWidget.tsx`
 - **Learnings for future iterations:**
  - Gemini SSE format: `data:` prefix per line, JSON body with `candidates[0].content.parts[0].text`
  - `system_instruction` field in Gemini request body sets the system prompt (not a message in `contents`)
  - Gemini role mapping: `'assistant'` → `'model'` in the API's `contents` array
  - Buffer-based SSE parsing handles chunk boundaries: split on `\n`, keep last incomplete line in buffer
  - `buildEmbeddingTexts()` is a great source for structured CV context — natural language paragraphs per item
  - The `<!--ITEMS:-->` HTML comment pattern is invisible when rendered but parseable by US-010 for item card display
  - `useCallback` on `handleSubmit` with `[inputValue, isStreaming, messages]` deps is needed because it reads all three
 ---
 ## 2026-02-15 - US-010
 - Extracted `iconByType` and `iconColorStyles` from `CommandPalette.tsx` into shared `src/lib/palette-icons.ts`
 - Updated `CommandPalette.tsx` to import from the shared module (no behavioral change)
 - Added `onAction?: (action: PaletteAction) => void` prop to `ChatWidget` — same pattern as `CommandPalette`
 - `DashboardLayout.tsx` passes `handlePaletteAction` to `ChatWidget` (same handler used by CommandPalette)
 - ChatWidget builds a `paletteMap` (Map<id, PaletteItem>) via `useMemo` for O(1) item lookups
 - Added `getMessageItemIds()` to parse `<!--ITEMS:id1,id2-->` HTML comments from message content
 - Added `getMessageItems()` to resolve parsed IDs to PaletteItem objects via the map
 - Assistant message bubbles now render compact clickable item cards below text when items are referenced:
  - Cards use same icon/color scheme from CommandPalette (22px icon + title + subtitle)
  - Cards have hover highlight (`var(--accent-light)`) and trigger `onAction(item.action)` on click
  - Cards only appear after streaming completes (when `<!--ITEMS:-->` metadata is in final content)
  - If no items referenced or IDs don't match, no cards shown — just text
 - Typecheck, lint (0 errors), and build all pass
 - Files changed: `src/lib/palette-icons.ts` (new), `src/components/ChatWidget.tsx`, `src/components/CommandPalette.tsx`, `src/components/DashboardLayout.tsx`
 - **Learnings for future iterations:**
  - Extracting shared constants to `src/lib/` is the right pattern — both `CommandPalette` and `ChatWidget` now use the same icon mappings without duplication
  - `buildPaletteData()` is pure (no side effects) and idempotent — safe to call in `useMemo` with empty deps
  - The `<!--ITEMS:-->` HTML comment regex `<!--ITEMS:([^>]*)-->` works reliably; `[^>]*` captures everything between the colons and closing
  - Item card buttons use `fontFamily: 'inherit'` to pick up the panel's `font-ui` — without this, browser defaults apply
  - The `overflow: 'hidden'` on the message bubble container is needed so the item cards section (with its own border-top) stays visually contained within the bubble's border-radius
 ---
 ## 2026-02-15 - US-011
 - Updated ChatWidget mobile breakpoint from `sm` (640px) to `md` (768px)
 - Changed mobile panel from 85vh bottom-sheet to full-screen overlay using `position: fixed; inset: 0` with `100dvh` height
 - Panel z-index on mobile bumped to 101 (`max-md:z-[101]`) to render above TopBar (z-100) and nav (z-99)
 - Floating chat button hidden on mobile when panel is open via `max-md:!hidden` Tailwind class
  - Fixed specificity issue: inline `style={{ display: 'flex' }}` was overriding Tailwind's `hidden` — moved flex/centering to Tailwind classes (`flex items-center justify-center`)
 - Safe area insets applied via `env(safe-area-inset-*)` CSS on the `[data-chat-panel]` element for notched devices
 - Input area stays pinned to bottom via existing flex layout (flex-col container + flex-1 message area + flex-shrink-0 input)
 - Desktop behavior unchanged: 380px wide, anchored bottom-right, max-height 480px, floating button visible
 - Panel open/close animations still respect `prefers-reduced-motion`
 - Typecheck, lint (0 errors), and build all pass
 - Browser verified at 375×812 (mobile) and 1280×800 (desktop): full-screen overlay works, button hides/shows correctly, close button works
 - Files changed: `src/components/ChatWidget.tsx`
 - **Learnings for future iterations:**
  - Inline `style` properties always override CSS classes — to allow Tailwind responsive utilities (like `max-md:hidden`) to work, move conflicting properties (especially `display`) to Tailwind classes instead
  - Use `!important` modifier (`max-md:!hidden`) when competing with framer-motion's inline styles that can't be easily removed
  - TopBar (`z-100`) and nav (`z-99`) sit above the chat panel's default `z-90` — mobile full-screen panels need `z-101+` to overlay properly
  - `100dvh` (dynamic viewport height) is essential for mobile full-screen panels — it accounts for browser chrome (address bar, toolbar) unlike `100vh`
  - The `[data-chat-panel]` CSS selector in the `<style>` block is the right place for responsive size rules since Tailwind can't conditionally set max-height based on viewport width
 ---
 ## 2026-02-15 - US-012
 - Replaced empty-state centered text with welcome bubble + suggested question chips
 - Welcome bubble styled as assistant message (left-aligned, `var(--bg-dashboard)` bg, `var(--border-light)` border)
 - Added `SUGGESTED_QUESTIONS` const array at module top for easy future editing
 - Three chips: "What's his NHS experience?", "Tell me about his data skills", "What projects has he built?"
 - Chips styled: rounded-full, teal accent border, teal hover tint, `font-ui` 12.5px
 - Clicking a chip calls `handleSubmit(questionText)` — same codepath as typing + Enter
 - Refactored `handleSubmit` to accept optional `overrideText` parameter (avoids stale state issue with `setInputValue` + immediate submit)
 - Wrapped send button `onClick` in arrow function to prevent passing MouseEvent as text argument
 - Welcome/chips visible when `messages.length === 0`, replaced by conversation once any message is sent
 - Typecheck passes (0 errors), lint passes (0 new errors/warnings)
 - Browser verified: welcome bubble displays correctly, chips render, clicking chip sends message and replaces welcome state
 - Files changed: `src/components/ChatWidget.tsx`
 - **Learnings for future iterations:**
  - When refactoring a callback to accept optional parameters, wrap `onClick={handler}` as `onClick={() => handler()}` to prevent React from passing the SyntheticEvent as the first argument
  - `SUGGESTED_QUESTIONS` as a module-level const is the simplest approach — easily editable, no data file needed for 3 items
  - The `handleSubmit(overrideText?)` pattern avoids the stale-state problem: `setInputValue(text)` followed by immediate `handleSubmit()` would read the old `inputValue` since React batches state updates
 ---
 ## 2026-02-15 - US-013
 - Downloaded all-MiniLM-L6-v2 model files to `public/models/Xenova/all-MiniLM-L6-v2/`:
  - `config.json`, `tokenizer.json`, `tokenizer_config.json`, `onnx/model_quantized.onnx` (~22MB)
 - Updated `src/lib/embedding-model.ts`:
  - `env.localModelPath = '/models/'` — Vite serves `public/` at root
  - `env.allowRemoteModels = false` — prevents any HF CDN fallback
  - `env.useBrowserCache = false` — prevents stale Cache API entries from interfering
 - Updated `scripts/generate-embeddings.ts`:
  - `env.localModelPath = resolve(import.meta.dirname, '..', 'public', 'models')` — absolute path for Node.js
  - `env.allowRemoteModels = false`
 - Model files committed as static assets (not in .gitignore)
 - Browser verified: all 4 model files fetched from `localhost:5173/models/` with 200 OK, zero `huggingface.co` requests
 - Semantic search verified working: "data analysis" returns multi-category results (Core Skills, Active Projects, Achievements)
 - Build script (`npm run generate-embeddings`) still works with local model files
 - Typecheck passes (0 errors), lint passes (0 new errors/warnings)
 - Files changed: `src/lib/embedding-model.ts`, `scripts/generate-embeddings.ts`, `public/models/Xenova/all-MiniLM-L6-v2/` (new directory with 4 files)
 - **Learnings for future iterations:**
  - `@xenova/transformers` env configuration: `env.localModelPath` sets the base path, `env.allowRemoteModels = false` prevents CDN fallback, `env.useBrowserCache = false` bypasses Browser Cache API
  - The library constructs paths as `{localModelPath}/{modelId}/{filename}` — so `/models/` + `Xenova/all-MiniLM-L6-v2` + `/onnx/model_quantized.onnx`
  - Browser Cache API can retain stale entries from previous HF CDN loads — setting `useBrowserCache = false` forces fresh fetches from the configured local path
  - For Node.js scripts, use an absolute filesystem path for `localModelPath` (not a URL)
  - The quantized ONNX model (`model_quantized.onnx`) is ~22MB — acceptable for a static asset since it's cached after first load
 ---
 ## 2026-02-15 - US-014
 - Extracted `GEMINI_MODEL` and `GEMINI_DISPLAY_NAME` constants in `src/lib/gemini.ts`
 - Updated `GEMINI_API_BASE` to use template literal with `GEMINI_MODEL` constant (`gemini-3-flash-preview`)
 - Tightened system prompt: restructured with markdown headers, more concise instructions, clearer `[ITEMS: ...]` format specification
 - Added model indicator to ChatWidget header: "Gemini 3 Flash" in `font-geist`, 11px, `var(--text-tertiary)`, right-aligned next to title
 - Imported `GEMINI_DISPLAY_NAME` in ChatWidget for the indicator text
 - Typecheck passes (0 errors), lint passes (0 new errors/warnings), build succeeds
 - Files changed: `src/lib/gemini.ts`, `src/components/ChatWidget.tsx`
 - **Learnings for future iterations:**
  - `gemini-3-flash-preview` is the correct model ID for Gemini 3 Flash (confirmed via Google AI docs); Gemini 2.0 Flash deprecated, shutdown scheduled for March 31 2026
  - The API path structure (`v1beta/models/{model}:streamGenerateContent?alt=sse&key=KEY`) is unchanged between Gemini 2 and 3
  - Extracting both `GEMINI_MODEL` (for API URL) and `GEMINI_DISPLAY_NAME` (for UI) as separate constants keeps the API ID decoupled from the human-readable name
  - System prompt with markdown headers (##) gives the model clearer section boundaries — improves instruction following for structured output like `[ITEMS: ...]`
  - Pre-existing uncommitted change in `src/App.tsx` (boot→login phase skip) was excluded from the commit — always check `git diff --stat` and stage specific files
 ---
@@ -195,6 +195,82 @@
      "priority": 10,
      "passes": true,
      "notes": "The action routing needs to flow from ChatWidget up to DashboardLayout. Add an onAction prop to ChatWidget (same pattern as CommandPalette). DashboardLayout passes handlePaletteAction to ChatWidget. Export iconByType and iconColorStyles from CommandPalette (or extract to a shared module) so ChatWidget can reuse them."
    },
    {
      "id": "US-011",
      "title": "Mobile full-screen chat panel",
      "description": "As a mobile visitor, I want the chat panel to be a full-screen overlay so it's easy to use on small screens.",
      "acceptanceCriteria": [
        "Below md breakpoint (768px), chat panel renders as full-screen overlay using position: fixed; inset: 0 with 100dvh height",
        "Full-screen mode has the existing header with close button (no visual change needed, just full-width)",
        "Floating chat button is hidden (display: none or opacity: 0) while panel is open on mobile (<768px)",
        "Above 768px, existing panel behavior is unchanged (380px wide, anchored bottom-right, max-height 480px)",
        "Panel open/close animation still respects prefers-reduced-motion",
        "Safe area insets applied via env(safe-area-inset-*) for notched devices",
        "Input area stays pinned to bottom of screen on mobile",
        "Typecheck passes",
        "Verify in browser using dev-browser skill"
      ],
      "priority": 11,
      "passes": true,
      "notes": "The current ChatWidget already has some mobile handling (bottom-sheet style at <640px). This story changes the breakpoint to 768px (md) and makes it truly full-screen instead of 85vh. Use 100dvh (dynamic viewport height) to account for mobile browser chrome. The floating button visibility can be controlled by combining isOpen state with a CSS media query or a useMediaQuery hook. The <style> block with data-chat-panel attribute is the place to update responsive rules."
    },
    {
      "id": "US-012",
      "title": "Welcome message with suggested question chips",
      "description": "As a visitor opening the chat, I see a friendly welcome message and clickable suggested questions so I know what to ask.",
      "acceptanceCriteria": [
        "When chat panel is open and conversation is empty, display welcome text: 'Hey! I'm here to help you learn more about Andy. What would you like to know?'",
        "Welcome text is styled as an AI message bubble (left-aligned, light background, same styling as assistant messages)",
        "Below the welcome bubble, show 2-3 clickable pill/chip buttons with suggested questions",
        "Suggested questions: 'What's his NHS experience?', 'Tell me about his data skills', 'What projects has he built?'",
        "Chips styled with: teal accent border, rounded-full, font-ui 12-13px, hover state (teal background tint)",
        "Clicking a chip sends that question as a user message (same codepath as typing + Enter)",
        "Welcome message and chips always visible when conversation is empty (persist across panel open/close)",
        "Once any message is sent, the welcome/chips area is replaced by the conversation messages",
        "Typecheck passes",
        "Verify in browser using dev-browser skill"
      ],
      "priority": 12,
      "passes": true,
      "notes": "Replace the current empty-state text ('Ask me anything about Andy's experience, skills, or projects.') with the new welcome bubble + chips. The chips should call handleSubmit (or equivalent) with the chip text pre-filled — simplest approach is setInputValue(chipText) then immediately trigger submit. Check that the welcome state reappears if the user hasn't sent a message (messages.length === 0). The suggested questions could live in a const array at the top of ChatWidget for easy future editing."
    },
    {
      "id": "US-013",
      "title": "Self-host ONNX embedding model",
      "description": "As a developer, I want the ONNX model files served from the same host as the site to eliminate dependency on Hugging Face CDN.",
      "acceptanceCriteria": [
        "Model files for Xenova/all-MiniLM-L6-v2 downloaded and placed in public/models/all-MiniLM-L6-v2/onnx/ (matching HF repo structure)",
        "Required files present: model_quantized.onnx, tokenizer.json, tokenizer_config.json, config.json, and any other files the pipeline expects",
        "src/lib/embedding-model.ts updated: configure @xenova/transformers env to use local model path (e.g., env.localModelPath or custom model URL pointing to /models/)",
        "scripts/generate-embeddings.ts also updated to use the same local model path for consistency",
        "Model files are NOT in .gitignore — they are committed as static assets",
        "No network requests to huggingface.co in the browser network tab when semantic search is used",
        "Semantic search still works correctly in the command palette after the change",
        "Typecheck passes"
      ],
      "priority": 13,
      "passes": true,
      "notes": "Transformers.js uses env.localModelPath or env.remoteHost to control where models are fetched from. Setting env.localModelPath = '/models/' should make it look for files at /models/Xenova/all-MiniLM-L6-v2/onnx/model_quantized.onnx etc. The Vite public/ directory serves files at the root — so public/models/ becomes /models/ at runtime. For the build script (Node.js), use a file:// path or the local filesystem path instead. Download model files from https://huggingface.co/Xenova/all-MiniLM-L6-v2/tree/main — the quantized ONNX model is ~23MB. Check what files the pipeline actually requests by watching network tab before making this change."
    },
    {
      "id": "US-014",
      "title": "Update to Gemini 3 Flash Preview with model indicator",
      "description": "As a developer, I want to use the latest free Gemini model, and as a visitor, I want to see what model powers the chat.",
      "acceptanceCriteria": [
        "Extract model name to a single constant (e.g., GEMINI_MODEL = 'gemini-3-flash-preview') used for both the API URL and display",
        "GEMINI_API_BASE URL updated to use the new model constant",
        "Review and tighten the system prompt — ensure it's well-structured, concise, and clear for the new model",
        "Review the [ITEMS: ...] suffix instruction — ensure new model follows the format reliably",
        "Small model indicator in chat panel header: 'Gemini 3 Flash' in font-geist, 11px, var(--text-tertiary)",
        "Model indicator positioned right-aligned in the header bar or as a subtle line below the header",
        "Streaming SSE parsing still works correctly with the new model endpoint",
        "Typecheck passes",
        "Verify in browser using dev-browser skill"
      ],
      "priority": 14,
      "passes": false,
      "notes": "The current API base is 'https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash'. Change the model segment to 'gemini-3-flash-preview'. The API path structure (v1beta/models/{model}:streamGenerateContent) should be the same. Verify that gemini-3-flash-preview is the correct model ID — check Google AI Studio or the API docs. For the display name, use a human-friendly string like 'Gemini 3 Flash' (not the full model ID). The constant should be defined at the top of gemini.ts and exported for use in ChatWidget."
    }
  ]
 }
@@ -12,6 +12,7 @@
 - `src/data/embeddings.json` is an array of `{ id: string, embedding: number[] }` — 42 items, 384-d vectors, IDs match PaletteItem IDs. Vite imports JSON natively.
 - `src/lib/embedding-model.ts` exports `initModel()`, `embedQuery(text)`, `isModelReady()` — check `isModelReady()` before calling `embedQuery()`
 - `initModel()` is called fire-and-forget in `App.tsx` on mount — model loads during boot/ECG/login phases
 - ONNX model files self-hosted in `public/models/Xenova/all-MiniLM-L6-v2/` — `env.localModelPath = '/models/'`, `env.allowRemoteModels = false`, `env.useBrowserCache = false` eliminates HF CDN dependency
 - `src/lib/semantic-search.ts` exports `semanticSearch(queryEmbedding, embeddings, threshold?)` and `loadEmbeddings()` — embeddings are normalized so cosine similarity is dot(a,b)/(mag(a)*mag(b))
 - CommandPalette uses `semanticResults` state + debounced `useEffect` for async semantic search, falling back to Fuse.js when `isModelReady()` returns false or on any error
 - `loadEmbeddings()` and `paletteMap` (Map<id, PaletteItem>) are precomputed via `useMemo` — no re-computation on each search
@@ -28,6 +29,11 @@
 - Icon/color mappings (`iconByType`, `iconColorStyles`) live in `src/lib/palette-icons.ts` — shared between CommandPalette and ChatWidget
 - ChatWidget accepts optional `onAction?: (action: PaletteAction) => void` prop — same pattern as CommandPalette's `onAction`
 - `DashboardLayout` passes `handlePaletteAction` to both CommandPalette and ChatWidget for unified action routing
 - TopBar is `z-index: 100` (fixed), nav is `z-index: 99` (sticky) — mobile full-screen overlays need `z-index > 100` to appear above them
 - Inline `style={{ display: 'flex' }}` overrides Tailwind's `hidden` class — use `!important` modifier (`max-md:!hidden`) or move display to Tailwind classes to allow responsive hiding
 - ChatWidget mobile breakpoint is `md` (768px) — below this, panel is full-screen; above, it's 380px anchored bottom-right
 - `handleSubmit(overrideText?)` accepts optional text param — use this when programmatically sending messages (e.g., suggested question chips) to avoid stale `inputValue` state
 - `SUGGESTED_QUESTIONS` const array at top of ChatWidget — edit here to change welcome screen chip text
 ---
@@ -226,3 +232,67 @@
  - Item card buttons use `fontFamily: 'inherit'` to pick up the panel's `font-ui` — without this, browser defaults apply
  - The `overflow: 'hidden'` on the message bubble container is needed so the item cards section (with its own border-top) stays visually contained within the bubble's border-radius
 ---
 ## 2026-02-15 - US-011
 - Updated ChatWidget mobile breakpoint from `sm` (640px) to `md` (768px)
 - Changed mobile panel from 85vh bottom-sheet to full-screen overlay using `position: fixed; inset: 0` with `100dvh` height
 - Panel z-index on mobile bumped to 101 (`max-md:z-[101]`) to render above TopBar (z-100) and nav (z-99)
 - Floating chat button hidden on mobile when panel is open via `max-md:!hidden` Tailwind class
  - Fixed specificity issue: inline `style={{ display: 'flex' }}` was overriding Tailwind's `hidden` — moved flex/centering to Tailwind classes (`flex items-center justify-center`)
 - Safe area insets applied via `env(safe-area-inset-*)` CSS on the `[data-chat-panel]` element for notched devices
 - Input area stays pinned to bottom via existing flex layout (flex-col container + flex-1 message area + flex-shrink-0 input)
 - Desktop behavior unchanged: 380px wide, anchored bottom-right, max-height 480px, floating button visible
 - Panel open/close animations still respect `prefers-reduced-motion`
 - Typecheck, lint (0 errors), and build all pass
 - Browser verified at 375×812 (mobile) and 1280×800 (desktop): full-screen overlay works, button hides/shows correctly, close button works
 - Files changed: `src/components/ChatWidget.tsx`
 - **Learnings for future iterations:**
  - Inline `style` properties always override CSS classes — to allow Tailwind responsive utilities (like `max-md:hidden`) to work, move conflicting properties (especially `display`) to Tailwind classes instead
  - Use `!important` modifier (`max-md:!hidden`) when competing with framer-motion's inline styles that can't be easily removed
  - TopBar (`z-100`) and nav (`z-99`) sit above the chat panel's default `z-90` — mobile full-screen panels need `z-101+` to overlay properly
  - `100dvh` (dynamic viewport height) is essential for mobile full-screen panels — it accounts for browser chrome (address bar, toolbar) unlike `100vh`
  - The `[data-chat-panel]` CSS selector in the `<style>` block is the right place for responsive size rules since Tailwind can't conditionally set max-height based on viewport width
 ---
 ## 2026-02-15 - US-012
 - Replaced empty-state centered text with welcome bubble + suggested question chips
 - Welcome bubble styled as assistant message (left-aligned, `var(--bg-dashboard)` bg, `var(--border-light)` border)
 - Added `SUGGESTED_QUESTIONS` const array at module top for easy future editing
 - Three chips: "What's his NHS experience?", "Tell me about his data skills", "What projects has he built?"
 - Chips styled: rounded-full, teal accent border, teal hover tint, `font-ui` 12.5px
 - Clicking a chip calls `handleSubmit(questionText)` — same codepath as typing + Enter
 - Refactored `handleSubmit` to accept optional `overrideText` parameter (avoids stale state issue with `setInputValue` + immediate submit)
 - Wrapped send button `onClick` in arrow function to prevent passing MouseEvent as text argument
 - Welcome/chips visible when `messages.length === 0`, replaced by conversation once any message is sent
 - Typecheck passes (0 errors), lint passes (0 new errors/warnings)
 - Browser verified: welcome bubble displays correctly, chips render, clicking chip sends message and replaces welcome state
 - Files changed: `src/components/ChatWidget.tsx`
 - **Learnings for future iterations:**
  - When refactoring a callback to accept optional parameters, wrap `onClick={handler}` as `onClick={() => handler()}` to prevent React from passing the SyntheticEvent as the first argument
  - `SUGGESTED_QUESTIONS` as a module-level const is the simplest approach — easily editable, no data file needed for 3 items
  - The `handleSubmit(overrideText?)` pattern avoids the stale-state problem: `setInputValue(text)` followed by immediate `handleSubmit()` would read the old `inputValue` since React batches state updates
 ---
 ## 2026-02-15 - US-013
 - Downloaded all-MiniLM-L6-v2 model files to `public/models/Xenova/all-MiniLM-L6-v2/`:
  - `config.json`, `tokenizer.json`, `tokenizer_config.json`, `onnx/model_quantized.onnx` (~22MB)
 - Updated `src/lib/embedding-model.ts`:
  - `env.localModelPath = '/models/'` — Vite serves `public/` at root
  - `env.allowRemoteModels = false` — prevents any HF CDN fallback
  - `env.useBrowserCache = false` — prevents stale Cache API entries from interfering
 - Updated `scripts/generate-embeddings.ts`:
  - `env.localModelPath = resolve(import.meta.dirname, '..', 'public', 'models')` — absolute path for Node.js
  - `env.allowRemoteModels = false`
 - Model files committed as static assets (not in .gitignore)
 - Browser verified: all 4 model files fetched from `localhost:5173/models/` with 200 OK, zero `huggingface.co` requests
 - Semantic search verified working: "data analysis" returns multi-category results (Core Skills, Active Projects, Achievements)
 - Build script (`npm run generate-embeddings`) still works with local model files
 - Typecheck passes (0 errors), lint passes (0 new errors/warnings)
 - Files changed: `src/lib/embedding-model.ts`, `scripts/generate-embeddings.ts`, `public/models/Xenova/all-MiniLM-L6-v2/` (new directory with 4 files)
 - **Learnings for future iterations:**
  - `@xenova/transformers` env configuration: `env.localModelPath` sets the base path, `env.allowRemoteModels = false` prevents CDN fallback, `env.useBrowserCache = false` bypasses Browser Cache API
  - The library constructs paths as `{localModelPath}/{modelId}/{filename}` — so `/models/` + `Xenova/all-MiniLM-L6-v2` + `/onnx/model_quantized.onnx`
  - Browser Cache API can retain stale entries from previous HF CDN loads — setting `useBrowserCache = false` forces fresh fetches from the configured local path
  - For Node.js scripts, use an absolute filesystem path for `localModelPath` (not a URL)
  - The quantized ONNX model (`model_quantized.onnx`) is ~22MB — acceptable for a static asset since it's cached after first load
 ---
@@ -0,0 +1,111 @@
 # Landing Page Polish Plan
 ## KPI Cards (Make Evidence Drawer Obvious)
 ### Core copy change
 - Update subsection header to: `Latest Results (click to view full reference range)`.
 ### Recommended interaction and affordance updates
 1. Add explicit CTA text on every KPI card:
 - `Click to view evidence` or `Open case summary`.
 - Keep this always visible (do not hide behind hover).
 2. Add a visible action affordance icon:
 - Use a chevron, plus, or document icon in the card corner.
 - Keep icon visible at all times to signal clickability.
 3. Strengthen hover and focus states:
 - On hover/focus: slightly lift card, increase border contrast, subtle shadow/glow.
 - Ensure clear keyboard focus ring for accessibility.
 - Keep `cursor: pointer` on full card.
 4. Add a one-time coachmark:
 - Pulse a single KPI card on first visit.
 - Message: `Open any metric to see evidence`.
 - Dismiss permanently after first KPI click.
 5. Add a section-level helper hint above KPI grid:
 - `Select a metric to inspect methodology, impact, and outcomes`.
 6. Keep interaction labels persistent for mobile:
 - Do not rely on hover-only affordances.
 - Ensure all cues are visible on touch devices.
 7. Add click/tap micro-feedback:
 - Subtle pressed-state animation on card tap.
 - Immediate drawer motion to confirm action.
 ### Priority (low effort -> high gain)
 1. Header copy update
 2. Persistent CTA text + action icon
 3. Strong hover/focus states
 4. One-time coachmark
 5. Micro-animation polish
 ---
 ## Network Graph (Career Constellation) Improvements
 ## Key issues identified in current implementation
 1. Keyboard accessibility overlay is incorrect:
 - Hidden focus buttons are all centered rather than mapped to real node coordinates.
 2. Simulation starts from poor initial state:
 - Nodes initialize from `(0,0)`, causing visual jumpiness and unstable first impression.
 3. Label readability and collision handling are weak:
 - Dense regions become hard to scan quickly.
 4. Interaction is hover-first:
 - Mobile/touch and keyboard parity is limited.
 5. Timeline logic is invisible:
 - Layout uses years but lacks visual timeline scaffolding (ticks/axis/era cues).
 ## Direction agreed
 - Desktop: pivot to a two-column workspace.
 - Left column: graph (sticky).
 - Right column: chronological clinical record stream (work + education).
 - Mobile/tablet: keep stacked layout (graph above timeline).
 ## Important implementation note
 - Do **not** visually rotate the SVG with CSS transforms.
 - Instead, remap the graph layout so time runs vertically:
 - Roles aligned by year from top (oldest) to bottom (newest).
 - Skills positioned around their linked roles.
 ## Recommended graph changes
 1. Add timeline guides:
 - Year ticks/markers and subtle era separators.
 - Small legend for node/link semantics.
 2. Seed deterministic initial positions:
 - Pre-place role nodes on year track.
 - Pre-place skill nodes near connected role clusters.
 - Then run constrained simulation for gentle settling, not dramatic motion.
 3. Fix keyboard/touch interaction model:
 - Map focusable hit targets to actual node positions.
 - Add tap-to-pin highlight mode for mobile.
 - Keep Enter/Space behavior equivalent for keyboard users.
 4. Improve label system:
 - Smarter truncation, optional reveal-on-hover/focus, and collision avoidance.
 - Increase contrast and spacing where clusters are dense.
 5. Preserve and enhance relationship highlighting:
 - Keep connected-node/link emphasis behavior.
 - Improve selected state persistence (not just hover transient state).
 ## Priority (low effort -> high gain)
 1. Timeline guides + legend
 2. Deterministic initial positions
 3. Correct keyboard hit-target mapping
 4. Tap-to-pin for mobile
 5. Label collision/declutter strategy
 ---
 ## Layout Note for Chronology Column
 - Use a single chronological stream with type badges (`Role`, `Education`).
 - This preserves the same current visual order while staying future-proof if entries interleave later.
@@ -9,7 +9,8 @@
    "lint": "eslint .",
    "typecheck": "tsc --noEmit",
    "preview": "vite preview",
-    "generate-embeddings": "npx tsx scripts/generate-embeddings.ts"
+    "generate-embeddings": "npx tsx scripts/generate-embeddings.ts",
    "benchmark": "npx tsx scripts/benchmark.ts"
  },
  "dependencies": {
    "@types/d3": "^7.4.3",
@@ -0,0 +1,25 @@
 {
  "_name_or_path": "sentence-transformers/all-MiniLM-L6-v2",
  "architectures": [
    "BertModel"
  ],
  "attention_probs_dropout_prob": 0.1,
  "classifier_dropout": null,
  "gradient_checkpointing": false,
  "hidden_act": "gelu",
  "hidden_dropout_prob": 0.1,
  "hidden_size": 384,
  "initializer_range": 0.02,
  "intermediate_size": 1536,
  "layer_norm_eps": 1e-12,
  "max_position_embeddings": 512,
  "model_type": "bert",
  "num_attention_heads": 12,
  "num_hidden_layers": 6,
  "pad_token_id": 0,
  "position_embedding_type": "absolute",
  "transformers_version": "4.29.2",
  "type_vocab_size": 2,
  "use_cache": true,
  "vocab_size": 30522
 }
@@ -0,0 +1,15 @@
 {
  "clean_up_tokenization_spaces": true,
  "cls_token": "[CLS]",
  "do_basic_tokenize": true,
  "do_lower_case": true,
  "mask_token": "[MASK]",
  "model_max_length": 512,
  "never_split": null,
  "pad_token": "[PAD]",
  "sep_token": "[SEP]",
  "strip_accents": null,
  "tokenize_chinese_chars": true,
  "tokenizer_class": "BertTokenizer",
  "unk_token": "[UNK]"
 }
@@ -0,0 +1,120 @@
 {
  "passThreshold": 18,
  "maxScore": 20,
  "questions": [
    {
      "id": "Q01",
      "question": "How many years has Andy been employed by the NHS?",
      "expectedAnswer": "Approximately 3-4 years. Andy's NHS employment started in May 2022 when he joined NHS Norfolk and Waveney ICB. His previous role at Tesco PLC was in the private sector, not the NHS.",
      "keyFacts": [
        "NHS employment started May 2022",
        "Tesco was private employer",
        "approximately 3-4 years NHS employment"
      ]
    },
    {
      "id": "Q02",
      "question": "What was Andy's involvement with tirzepatide?",
      "expectedAnswer": "Andy supported commissioning of NICE TA1026 (tirzepatide). He authored the initial executive paper advocating a primary care delivery model over specialist provider, which drove a system shift to GP-led model.",
      "keyFacts": [
        "NICE TA1026",
        "authored executive paper",
        "primary care model",
        "GP-led delivery"
      ]
    },
    {
      "id": "Q03",
      "question": "What specific tools and software has Andy built?",
      "expectedAnswer": "Andy has built 5 notable projects: a patient switching algorithm (Python, 14000 patients, £2.6M savings), a Blueteq generator for high-cost drug forms, a controlled drugs monitoring system, a Sankey chart tool for visualising patient flows, and PharMetrics — a Power BI analytics dashboard.",
      "keyFacts": [
        "patient switching algorithm",
        "Blueteq generator",
        "CD monitoring system",
        "Sankey chart tool",
        "PharMetrics dashboard"
      ]
    },
    {
      "id": "Q04",
      "question": "What were Andy's A-level subjects and grades?",
      "expectedAnswer": "Andy achieved Mathematics A*, Chemistry B, and Politics C at Highworth Grammar School between 2009-2011.",
      "keyFacts": [
        "Mathematics A*",
        "Chemistry B",
        "Politics C",
        "Highworth Grammar School"
      ]
    },
    {
      "id": "Q05",
      "question": "Was Andy's Tesco role part of the NHS?",
      "expectedAnswer": "No. Andy's role at Tesco PLC was in the private sector as a community pharmacist. Tesco PLC is a private employer. He was an LPC representative during this time.",
      "keyFacts": [
        "Tesco PLC is private/not NHS",
        "community pharmacy",
        "LPC representative"
      ]
    },
    {
      "id": "Q06",
      "question": "How did the patient switching algorithm work?",
      "expectedAnswer": "It was Python-based and used real-world GP prescribing data to auto-identify patients eligible for cost-effective medication alternatives. It compressed months of manual work into 3 days, covered 14,000 patients, and identified £2.6M in savings.",
      "keyFacts": [
        "Python",
        "GP prescribing data",
        "14000 patients",
        "£2.6M savings",
        "compressed months to 3 days"
      ]
    },
    {
      "id": "Q07",
      "question": "What clinical specialties has Andy worked across?",
      "expectedAnswer": "Andy has worked across rheumatology, ophthalmology (wet AMD, DMO, RVO), dermatology, gastroenterology, neurology, and migraine through his high-cost drugs role.",
      "keyFacts": [
        "rheumatology",
        "ophthalmology",
        "dermatology",
        "gastroenterology",
        "neurology",
        "migraine"
      ]
    },
    {
      "id": "Q08",
      "question": "What is Andy's experience with the dm+d?",
      "expectedAnswer": "Andy created a comprehensive medicines data table integrating all dm+d products with standardised strengths, morphine equivalents, and Anticholinergic Burden scoring, serving as a single source of truth.",
      "keyFacts": [
        "dm+d integration",
        "standardised strengths",
        "morphine equivalents",
        "Anticholinergic Burden",
        "single source of truth"
      ]
    },
    {
      "id": "Q09",
      "question": "What budget does Andy manage and how?",
      "expectedAnswer": "Andy manages a £220M prescribing budget using forecasting models, variance analysis, and financial reporting to the executive team, enabling proactive financial planning.",
      "keyFacts": [
        "£220M",
        "forecasting models",
        "variance analysis",
        "proactive financial planning"
      ]
    },
    {
      "id": "Q10",
      "question": "What leadership training does Andy have?",
      "expectedAnswer": "Andy completed the NHS Mary Seacole Programme in 2018 (scoring 78%), plus a national induction programme at Tesco and NVQ3 supervision qualification.",
      "keyFacts": [
        "Mary Seacole Programme",
        "2018",
        "78%",
        "national induction training at Tesco",
        "NVQ3 supervision"
      ]
    }
  ]
 }
@@ -0,0 +1,382 @@
 import { readFileSync, writeFileSync, readdirSync, mkdirSync, existsSync } from 'node:fs'
 import { resolve } from 'node:path'
 import { buildEmbeddingTexts } from '@/lib/search'
 // Load .env file manually (avoid adding dotenv dependency)
 function loadEnvFile(): void {
  const envPath = resolve(import.meta.dirname, '..', '.env')
  if (!existsSync(envPath)) return
  const content = readFileSync(envPath, 'utf-8')
  for (const line of content.split('\n')) {
    const trimmed = line.trim()
    if (!trimmed || trimmed.startsWith('#')) continue
    const eqIndex = trimmed.indexOf('=')
    if (eqIndex === -1) continue
    const key = trimmed.slice(0, eqIndex)
    const value = trimmed.slice(eqIndex + 1)
    if (!process.env[key]) {
      process.env[key] = value
    }
  }
 }
 loadEnvFile()
 // --- Types ---
 interface BenchmarkQuestion {
  id: string
  question: string
  expectedAnswer: string
  keyFacts: string[]
 }
 interface BenchmarkConfig {
  passThreshold: number
  maxScore: number
  questions: BenchmarkQuestion[]
 }
 interface ScoringResult {
  score: 0 | 1 | 2
  justification: string
 }
 interface QuestionResult {
  id: string
  question: string
  expectedAnswer: string
  actualAnswer: string
  score: number
  justification: string
 }
 interface BenchmarkResults {
  iteration: number
  timestamp: string
  model: string
  totalScore: number
  maxPossibleScore: number
  passThreshold: number
  passed: boolean
  hasZeros: boolean
  results: QuestionResult[]
 }
 // --- Gemini API ---
 const GEMINI_MODEL = 'gemini-3-flash-preview'
 const GEMINI_API_BASE = `https://generativelanguage.googleapis.com/v1beta/models/${GEMINI_MODEL}`
 function getApiKey(): string {
  const key = process.env.VITE_GEMINI_API_KEY
  if (!key) {
    throw new Error('VITE_GEMINI_API_KEY not set. Ensure .env file exists with this key.')
  }
  return key
 }
 function buildSystemPrompt(): string {
  const texts = buildEmbeddingTexts()
  const cvContent = texts.map((t) => `- ${t.text}`).join('\n')
  return `You are an AI assistant on Andy Charlwood's portfolio website. Answer questions about his experience, skills, projects, and qualifications.
 ## Andy's Professional Profile
 ${cvContent}
 ## Rules
 1. Use ONLY the profile above. Never invent roles, dates, or achievements.
 2. Be concise (2-4 sentences). Be professional but friendly.
 3. If the information isn't in the profile, say so.
 ## Item References
 After your answer, on a NEW line, list relevant portfolio item IDs:
 [ITEMS: id1, id2, id3]
 - IDs match the profile entries above (exp-*, skill-*, proj-*, ach-*, edu-*, action-*).
 - Only include IDs directly relevant to your answer.
 - If no items are relevant, omit the [ITEMS: ...] line entirely.`
 }
 function sleep(ms: number): Promise<void> {
  return new Promise((resolve) => setTimeout(resolve, ms))
 }
 async function callGemini(
  systemPrompt: string,
  userMessage: string,
  temperature = 0.7,
  maxOutputTokens = 512,
 ): Promise<string> {
  const apiKey = getApiKey()
  const maxRetries = 5
  for (let attempt = 0; attempt < maxRetries; attempt++) {
    const response = await fetch(
      `${GEMINI_API_BASE}:generateContent?key=${apiKey}`,
      {
        method: 'POST',
        headers: { 'Content-Type': 'application/json' },
        body: JSON.stringify({
          system_instruction: {
            parts: [{ text: systemPrompt }],
          },
          contents: [
            { role: 'user', parts: [{ text: userMessage }] },
          ],
          generationConfig: {
            temperature,
            maxOutputTokens,
          },
        }),
      },
    )
    if (response.status === 429 || response.status === 503) {
      const errorBody = await response.text()
      const retryMatch = errorBody.match(/retry in ([\d.]+)s/)
      const waitSeconds = retryMatch ? Math.ceil(parseFloat(retryMatch[1])) + 2 : (attempt + 1) * 15
      const reason = response.status === 429 ? 'Rate limited' : 'Service unavailable'
      console.log(`  ${reason}. Waiting ${waitSeconds}s (attempt ${attempt + 1}/${maxRetries})...`)
      await sleep(waitSeconds * 1000)
      continue
    }
    if (!response.ok) {
      const errorBody = await response.text()
      throw new Error(`Gemini API error ${response.status}: ${errorBody}`)
    }
    const data = await response.json()
    const text = data?.candidates?.[0]?.content?.parts?.[0]?.text
    if (!text) {
      throw new Error(`No text in Gemini response: ${JSON.stringify(data)}`)
    }
    return text
  }
  throw new Error('Max retries exceeded for rate limiting')
 }
 // --- Scoring ---
 function extractJson(text: string): string | null {
  // Try parsing directly first
  try {
    JSON.parse(text)
    return text
  } catch { /* not direct JSON, continue extraction */ }
  // Strip markdown code fences
  const fenceMatch = text.match(/```(?:json)?\s*([\s\S]*?)```/)
  if (fenceMatch) {
    return fenceMatch[1].trim()
  }
  // Find first { ... } block
  const braceStart = text.indexOf('{')
  if (braceStart === -1) return null
  // Find matching closing brace
  let depth = 0
  let inString = false
  let escaped = false
  for (let i = braceStart; i < text.length; i++) {
    const ch = text[i]
    if (escaped) { escaped = false; continue }
    if (ch === '\\') { escaped = true; continue }
    if (ch === '"') { inString = !inString; continue }
    if (inString) continue
    if (ch === '{') depth++
    if (ch === '}') { depth--; if (depth === 0) return text.slice(braceStart, i + 1) }
  }
  return null
 }
 async function scoreAnswer(
  question: string,
  expectedAnswer: string,
  keyFacts: string[],
  actualAnswer: string,
 ): Promise<ScoringResult> {
  const scoringPrompt = `You are a strict evaluator. Compare an ACTUAL answer to an EXPECTED answer about a person's CV.
 Rubric:
 - 2 = ACCURATE: Covers key facts correctly. Minor omissions OK if no errors.
 - 1 = PARTIAL: Some key facts right but misses important details or is vague.
 - 0 = INCORRECT: Contains factual errors, contradicts expected answer, or misses the point.
 Key facts for score 2:
 ${keyFacts.map((f) => `- ${f}`).join('\n')}
 IMPORTANT: Respond with ONLY a single-line JSON object. No markdown, no code fences, no extra text.
 Example: {"score":2,"justification":"Covers all key facts accurately"}
 Keep justification under 30 words.`
  const userMessage = `QUESTION: ${question}
 EXPECTED ANSWER: ${expectedAnswer}
 ACTUAL ANSWER: ${actualAnswer}`
  const rawResponse = await callGemini(scoringPrompt, userMessage, 0, 512)
  // Extract JSON — handle code fences, preamble text, multiline responses
  const extracted = extractJson(rawResponse)
  if (!extracted) {
    console.warn(`  Warning: Could not extract JSON from scoring response: ${rawResponse.slice(0, 200)}`)
    return { score: 0, justification: `Failed to parse scoring response` }
  }
  try {
    const parsed = JSON.parse(extracted) as ScoringResult
    if (![0, 1, 2].includes(parsed.score)) {
      console.warn(`  Warning: Invalid score value: ${parsed.score}`)
      return { score: 0, justification: `Invalid score value: ${parsed.score}` }
    }
    return parsed
  } catch {
    console.warn(`  Warning: Invalid JSON: ${extracted.slice(0, 150)}`)
    return { score: 0, justification: `Invalid JSON in response` }
  }
 }
 // --- Iteration Management ---
 function getNextIteration(resultsDir: string): number {
  if (!existsSync(resultsDir)) return 0
  const files = readdirSync(resultsDir).filter((f) => f.startsWith('iteration-') && f.endsWith('.json'))
  if (files.length === 0) return 0
  const iterations = files.map((f) => {
    const match = f.match(/iteration-(\d+)\.json/)
    return match ? parseInt(match[1], 10) : -1
  })
  return Math.max(...iterations) + 1
 }
 // --- Console Output ---
 function printSummary(results: BenchmarkResults): void {
  console.log('\n' + '='.repeat(80))
  console.log(`BENCHMARK RESULTS — Iteration ${results.iteration}`)
  console.log(`Model: ${results.model} | ${results.timestamp}`)
  console.log('='.repeat(80))
  // Table header
  console.log(
    'ID'.padEnd(6) +
    'Score'.padEnd(8) +
    'Question'.padEnd(50) +
    'Justification'
  )
  console.log('-'.repeat(80))
  for (const r of results.results) {
    const scoreLabel = r.score === 2 ? '2 ✓' : r.score === 1 ? '1 ~' : '0 ✗'
    const questionTruncated = r.question.length > 47 ? r.question.slice(0, 44) + '...' : r.question
    const justTruncated = r.justification.length > 60 ? r.justification.slice(0, 57) + '...' : r.justification
    console.log(
      r.id.padEnd(6) +
      scoreLabel.padEnd(8) +
      questionTruncated.padEnd(50) +
      justTruncated
    )
  }
  console.log('-'.repeat(80))
  console.log(
    `TOTAL: ${results.totalScore}/${results.maxPossibleScore}` +
    ` | Threshold: ${results.passThreshold}/${results.maxPossibleScore}` +
    ` | Has zeros: ${results.hasZeros ? 'YES' : 'No'}` +
    ` | ${results.passed ? 'PASSED ✓' : 'FAILED ✗'}`
  )
  console.log('='.repeat(80))
 }
 // --- Main ---
 async function main() {
  const scriptDir = import.meta.dirname
  const configPath = resolve(scriptDir, 'benchmark-config.json')
  const resultsDir = resolve(scriptDir, 'benchmark-results')
  // Load config
  const config: BenchmarkConfig = JSON.parse(readFileSync(configPath, 'utf-8'))
  console.log(`Loaded ${config.questions.length} benchmark questions.`)
  // Determine iteration number
  const iteration = getNextIteration(resultsDir)
  console.log(`Running iteration ${iteration}...`)
  // Build system prompt (same as production)
  const systemPrompt = buildSystemPrompt()
  console.log(`System prompt built (${systemPrompt.length} chars).`)
  // Run each question
  const questionResults: QuestionResult[] = []
  for (const q of config.questions) {
    console.log(`\n[${q.id}] ${q.question}`)
    // Get answer from Gemini
    console.log('  Getting answer...')
    const actualAnswer = await callGemini(systemPrompt, q.question)
    console.log(`  Answer: ${actualAnswer.slice(0, 100)}...`)
    // Score the answer
    console.log('  Scoring...')
    const { score, justification } = await scoreAnswer(
      q.question,
      q.expectedAnswer,
      q.keyFacts,
      actualAnswer,
    )
    console.log(`  Score: ${score}/2 — ${justification}`)
    questionResults.push({
      id: q.id,
      question: q.question,
      expectedAnswer: q.expectedAnswer,
      actualAnswer,
      score,
      justification,
    })
  }
  // Calculate totals
  const totalScore = questionResults.reduce((sum, r) => sum + r.score, 0)
  const hasZeros = questionResults.some((r) => r.score === 0)
  const passed = totalScore >= config.passThreshold && !hasZeros
  const results: BenchmarkResults = {
    iteration,
    timestamp: new Date().toISOString(),
    model: GEMINI_MODEL,
    totalScore,
    maxPossibleScore: config.maxScore,
    passThreshold: config.passThreshold,
    passed,
    hasZeros,
    results: questionResults,
  }
  // Save results
  mkdirSync(resultsDir, { recursive: true })
  const resultsPath = resolve(resultsDir, `iteration-${iteration}.json`)
  writeFileSync(resultsPath, JSON.stringify(results, null, 2))
  console.log(`\nResults saved to ${resultsPath}`)
  // Print summary table
  printSummary(results)
  // Exit with appropriate code
  process.exit(passed ? 0 : 1)
 }
 main().catch((err) => {
  console.error('Benchmark failed:', err)
  process.exit(2)
 })
@@ -1,8 +1,12 @@
 import { writeFileSync } from 'node:fs'
 import { resolve } from 'node:path'
-import { pipeline } from '@xenova/transformers'
+import { env, pipeline } from '@xenova/transformers'
 import { buildEmbeddingTexts } from '@/lib/search'
 // Use local model files from public/models/ (same files the browser uses)
 env.localModelPath = resolve(import.meta.dirname, '..', 'public', 'models')
 env.allowRemoteModels = false
 async function main() {
  const items = buildEmbeddingTexts()
  console.log(`Found ${items.length} items to embed.`)
@@ -6,6 +6,7 @@ import {
  isGeminiAvailable,
  parseItemIds,
  stripItemsSuffix,
  GEMINI_DISPLAY_NAME,
  type ChatMessage,
 } from '@/lib/gemini'
 import { buildPaletteData } from '@/lib/search'
@@ -16,6 +17,12 @@ const prefersReducedMotion = window.matchMedia('(prefers-reduced-motion: reduce)
 const MAX_HISTORY = 10
 const SUGGESTED_QUESTIONS = [
  "What's his NHS experience?",
  'Tell me about his data skills',
  'What projects has he built?',
 ]
 const buttonVariants = {
  hidden: prefersReducedMotion
    ? { opacity: 1, y: 0 }
@@ -79,8 +86,8 @@ export function ChatWidget({ onAction }: ChatWidgetProps) {
    }
  }, [isOpen])
-  const handleSubmit = useCallback(async () => {
+  const handleSubmit = useCallback(async (overrideText?: string) => {
-    const trimmed = inputValue.trim()
+    const trimmed = (overrideText ?? inputValue).trim()
    if (!trimmed || isStreaming) return
    const userMessage: ChatMessage = { role: 'user', content: trimmed }
@@ -192,8 +199,8 @@ export function ChatWidget({ onAction }: ChatWidgetProps) {
            role="dialog"
            aria-label="Chat with AI about Andy"
            className="fixed z-[90] font-ui
-              bottom-0 left-0 right-0 rounded-t-xl
+              inset-0 rounded-none max-md:z-[101]
-              sm:bottom-[88px] sm:right-6 sm:left-auto sm:rounded-xl"
+              md:inset-auto md:bottom-[88px] md:right-6 md:rounded-xl"
            style={{
              width: undefined,
              background: 'var(--surface)',
@@ -206,11 +213,18 @@ export function ChatWidget({ onAction }: ChatWidgetProps) {
            }}
          >
            <style>{`
-              @media (min-width: 640px) {
+              @media (min-width: 768px) {
                [data-chat-panel] { width: 380px; max-height: 480px; }
              }
-              @media (max-width: 639px) {
+              @media (max-width: 767px) {
-                [data-chat-panel] { height: 85vh; max-height: 85vh; }
+                [data-chat-panel] {
                  height: 100dvh;
                  max-height: 100dvh;
                  padding-top: env(safe-area-inset-top, 0px);
                  padding-bottom: env(safe-area-inset-bottom, 0px);
                  padding-left: env(safe-area-inset-left, 0px);
                  padding-right: env(safe-area-inset-right, 0px);
                }
              }
            `}</style>
            <div
@@ -220,7 +234,6 @@ export function ChatWidget({ onAction }: ChatWidgetProps) {
                flexDirection: 'column',
                width: '100%',
                height: '100%',
                maxHeight: '480px',
              }}
            >
              {/* Header */}
@@ -234,15 +247,26 @@ export function ChatWidget({ onAction }: ChatWidgetProps) {
                  flexShrink: 0,
                }}
              >
-                <span
+                <div style={{ display: 'flex', alignItems: 'baseline', gap: '8px' }}>
-                  style={{
+                  <span
-                    fontSize: '14px',
+                    style={{
-                    fontWeight: 600,
+                      fontSize: '14px',
-                    color: 'var(--text-primary)',
+                      fontWeight: 600,
-                  }}
+                      color: 'var(--text-primary)',
-                >
+                    }}
-                  Ask about Andy
+                  >
-                </span>
+                    Ask about Andy
                  </span>
                  <span
                    className="font-geist"
                    style={{
                      fontSize: '11px',
                      color: 'var(--text-tertiary)',
                    }}
                  >
                    {GEMINI_DISPLAY_NAME}
                  </span>
                </div>
                <button
                  onClick={() => setIsOpen(false)}
                  aria-label="Close chat"
@@ -297,16 +321,64 @@ export function ChatWidget({ onAction }: ChatWidgetProps) {
                )}
                {geminiAvailable && messages.length === 0 && (
-                  <div
+                  <div style={{ display: 'flex', flexDirection: 'column', gap: '12px' }}>
-                    style={{
+                    {/* Welcome bubble — styled as assistant message */}
-                      textAlign: 'center',
+                    <div style={{ display: 'flex', justifyContent: 'flex-start' }}>
-                      padding: '32px 16px',
+                      <div
-                      color: 'var(--text-tertiary)',
+                        style={{
-                      fontSize: '13px',
+                          maxWidth: '85%',
-                      lineHeight: 1.5,
+                          padding: '10px 14px',
-                    }}
+                          borderRadius: '12px 12px 12px 4px',
-                  >
+                          fontSize: '13px',
-                    Ask me anything about Andy's experience, skills, or projects.
+                          lineHeight: 1.5,
                          background: 'var(--bg-dashboard)',
                          color: 'var(--text-primary)',
                          border: '1px solid var(--border-light)',
                          whiteSpace: 'pre-wrap',
                        }}
                      >
                        Hey! I'm here to help you learn more about Andy. What would you like to know?
                      </div>
                    </div>
                    {/* Suggested question chips */}
                    <div
                      style={{
                        display: 'flex',
                        flexWrap: 'wrap',
                        gap: '8px',
                        paddingLeft: '4px',
                      }}
                    >
                      {SUGGESTED_QUESTIONS.map((question) => (
                        <button
                          key={question}
                          onClick={() => handleSubmit(question)}
                          style={{
                            padding: '6px 14px',
                            borderRadius: '9999px',
                            border: '1px solid var(--accent-border)',
                            background: 'transparent',
                            color: 'var(--text-secondary)',
                            fontSize: '12.5px',
                            fontFamily: 'inherit',
                            cursor: 'pointer',
                            transition: 'background-color 150ms ease-out, color 150ms ease-out',
                            whiteSpace: 'nowrap',
                          }}
                          onMouseEnter={(e) => {
                            e.currentTarget.style.backgroundColor = 'var(--accent-light)'
                            e.currentTarget.style.color = 'var(--accent)'
                          }}
                          onMouseLeave={(e) => {
                            e.currentTarget.style.backgroundColor = 'transparent'
                            e.currentTarget.style.color = 'var(--text-secondary)'
                          }}
                        >
                          {question}
                        </button>
                      ))}
                    </div>
                  </div>
                )}
@@ -509,7 +581,7 @@ export function ChatWidget({ onAction }: ChatWidgetProps) {
                    }}
                  />
                  <button
-                    onClick={handleSubmit}
+                    onClick={() => handleSubmit()}
                    disabled={!inputValue.trim() || isStreaming}
                    aria-label="Send message"
                    style={{
@@ -536,18 +608,15 @@ export function ChatWidget({ onAction }: ChatWidgetProps) {
        )}
      </AnimatePresence>
-      {/* Floating chat button */}
+      {/* Floating chat button — hidden on mobile when panel is open */}
      <motion.button
        initial="hidden"
        animate="visible"
        variants={buttonVariants}
        onClick={() => setIsOpen((prev) => !prev)}
        aria-label={isOpen ? 'Close chat' : 'Open chat'}
-        className="fixed z-[90] cursor-pointer bottom-4 right-4 h-10 w-10 sm:bottom-6 sm:right-6 sm:h-12 sm:w-12"
+        className={`fixed z-[90] cursor-pointer flex items-center justify-center bottom-4 right-4 h-10 w-10 md:bottom-6 md:right-6 md:h-12 md:w-12${isOpen ? ' max-md:!hidden' : ''}`}
        style={{
          display: 'flex',
          alignItems: 'center',
          justifyContent: 'center',
          borderRadius: '50%',
          border: 'none',
          background: 'var(--accent)',
@@ -1,4 +1,9 @@
-import { pipeline, type FeatureExtractionPipeline } from '@xenova/transformers'
+import { env, pipeline, type FeatureExtractionPipeline } from '@xenova/transformers'
 // Serve model files from /models/ (Vite serves public/ at root)
 env.localModelPath = '/models/'
 env.allowRemoteModels = false
 env.useBrowserCache = false
 let extractor: FeatureExtractionPipeline | null = null
 let loading = false
@@ -5,7 +5,10 @@ export interface ChatMessage {
  content: string
 }
-const GEMINI_API_BASE = 'https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash'
+export const GEMINI_MODEL = 'gemini-3-flash-preview'
 export const GEMINI_DISPLAY_NAME = 'Gemini 3 Flash'
 const GEMINI_API_BASE = `https://generativelanguage.googleapis.com/v1beta/models/${GEMINI_MODEL}`
 function getApiKey(): string | undefined {
  return import.meta.env.VITE_GEMINI_API_KEY as string | undefined
@@ -19,20 +22,23 @@ function buildSystemPrompt(): string {
  const texts = buildEmbeddingTexts()
  const cvContent = texts.map((t) => `- ${t.text}`).join('\n')
-  return `You are an AI assistant embedded in Andy Charlwood's professional portfolio website. Your role is to answer questions about Andy's professional experience, skills, projects, and qualifications accurately and concisely.
+  return `You are an AI assistant on Andy Charlwood's portfolio website. Answer questions about his experience, skills, projects, and qualifications.
-Here is Andy's complete professional profile:
+## Andy's Professional Profile
 ${cvContent}
-Instructions:
+## Rules
- Answer questions based ONLY on the information above. Do not invent roles, dates, or achievements.
+1. Use ONLY the profile above. Never invent roles, dates, or achievements.
- Be concise — 2-4 sentences for most answers.
+2. Be concise (2-4 sentences). Be professional but friendly.
- Be professional but friendly in tone.
+3. If the information isn't in the profile, say so.
- If asked something not covered by the profile data, say you don't have that information.
+
- At the end of your response, on a new line, include relevant portfolio item IDs in this format: [ITEMS: id1, id2, id3]
+## Item References
- Only include item IDs that are directly relevant to your answer. The available IDs are the ones listed above (e.g., exp-*, skill-*, proj-*, ach-*, edu-*, action-*).
+After your answer, on a NEW line, list relevant portfolio item IDs:
- If no items are particularly relevant, omit the [ITEMS: ...] line entirely.`
+[ITEMS: id1, id2, id3]
 - IDs match the profile entries above (exp-*, skill-*, proj-*, ach-*, edu-*, action-*).
 - Only include IDs directly relevant to your answer.
 - If no items are relevant, omit the [ITEMS: ...] line entirely.`
 }
 function buildRequestBody(
@@ -0,0 +1,125 @@
 # PRD: Chat Widget Polish & Model Updates
 ## Introduction
 The semantic search and AI chat features are functionally complete (US-001 through US-010). This PRD covers four polish items: mobile full-screen chat experience, a welcome message with suggested questions, self-hosting the ONNX embedding model, and updating from Gemini 2.0 Flash to Gemini 3 Flash Preview.
 ## Goals
 - Full-screen chat on mobile (<768px) for a better small-screen experience
 - Welcome message with suggested question chips to reduce blank-state friction
 - Self-host the ONNX model (`all-MiniLM-L6-v2`) to eliminate dependency on Hugging Face CDN
 - Update Gemini model to `gemini-3-flash-preview` and show which model powers the chat
 - Refresh system prompt while updating the model
 ## User Stories
 ### US-011: Mobile full-screen chat panel
 **Description:** As a mobile visitor, I want the chat panel to be a full-screen overlay so it's easy to use on small screens.
 **Acceptance Criteria:**
 - [ ] Below `md` breakpoint (768px), chat panel renders as full-screen overlay (100vw x 100vh, or using `dvh` for mobile browser chrome)
 - [ ] Full-screen mode has a visible header with close button
 - [ ] Floating chat button is hidden while panel is open on mobile
 - [ ] Above 768px, existing panel behavior unchanged (380px wide, anchored bottom-right)
 - [ ] Smooth transition between open/closed states respects `prefers-reduced-motion`
 - [ ] Typecheck passes
 - [ ] Verify in browser using dev-browser skill
 ### US-012: Welcome message with suggested questions
 **Description:** As a visitor opening the chat for the first time, I see a friendly welcome and clickable suggested questions so I know what to ask.
 **Acceptance Criteria:**
 - [ ] When chat panel opens and conversation is empty, display welcome message: "Hey! I'm here to help you learn more about Andy. What would you like to know?"
 - [ ] Below the welcome message, show 2-3 clickable pill/chip buttons with suggested questions (e.g., "What's his NHS experience?", "Tell me about his data skills", "What projects has he built?")
 - [ ] Clicking a suggested question sends it as a user message (same as typing and pressing Enter)
 - [ ] Welcome message and chips are always visible when conversation is empty (persist across open/close if no messages sent)
 - [ ] Once a message is sent, the welcome/chips area is replaced by the conversation
 - [ ] Chips use design system tokens (teal accent border, hover state)
 - [ ] Typecheck passes
 - [ ] Verify in browser using dev-browser skill
 ### US-013: Self-host ONNX embedding model
 **Description:** As a developer, I want the ONNX model files served from the same host as the site, so there's no runtime dependency on Hugging Face CDN.
 **Acceptance Criteria:**
 - [ ] Model files for `all-MiniLM-L6-v2` downloaded and placed in `public/models/all-MiniLM-L6-v2/` (or `public/models/onnx/` — whichever is cleaner)
 - [ ] Files include at minimum: `onnx/model_quantized.onnx`, `tokenizer.json`, `tokenizer_config.json`, `config.json`
 - [ ] `src/lib/embedding-model.ts` updated to load from local path instead of Hugging Face CDN
 - [ ] Build-time embedding script (`scripts/generate-embeddings.ts`) also uses local model path
 - [ ] `.gitignore` does NOT ignore the model files — they are committed as static assets
 - [ ] Verify model loads correctly in browser (semantic search still works in command palette)
 - [ ] Typecheck passes
 ### US-014: Update to Gemini 3 Flash Preview + model indicator
 **Description:** As a developer, I want to use the latest free Gemini model, and as a visitor, I want to see what model powers the chat.
 **Acceptance Criteria:**
 - [ ] `GEMINI_API_BASE` in `src/lib/gemini.ts` updated from `gemini-2.0-flash` to `gemini-3-flash-preview`
 - [ ] Review and update the system prompt for clarity (ensure it's well-structured for the new model)
 - [ ] Review and update the response format instructions (the `[ITEMS: ...]` suffix pattern)
 - [ ] Small text indicator in chat panel header or footer showing the model name (e.g., "Gemini 3 Flash" in `font-geist`, 11px, tertiary color)
 - [ ] If the model string needs to change in future, it should be a single constant — not hardcoded in multiple places
 - [ ] Typecheck passes
 - [ ] Verify in browser using dev-browser skill
 ## Functional Requirements
 - FR-1: Chat panel below 768px uses full-screen overlay layout (`position: fixed; inset: 0`)
 - FR-2: Chat button hidden when full-screen panel is open on mobile
 - FR-3: Welcome message and suggested question chips shown when conversation is empty
 - FR-4: Clicking a suggested question chip triggers the same flow as manually typing and sending
 - FR-5: ONNX model files served from `public/models/` as static assets
 - FR-6: `embedding-model.ts` configures Transformers.js to use local model path
 - FR-7: Gemini API calls use `gemini-3-flash-preview` model
 - FR-8: Chat UI displays model name indicator
 ## Non-Goals
 - No changes to the command palette UI or semantic search ranking logic
 - No persistent chat history across page loads
 - No rate limiting or abuse prevention
 - No changes to the boot/ECG/login flow
 - No model fine-tuning or custom training
 ## Design Considerations
 ### Mobile Full-Screen Chat
 - Full viewport with safe area insets (`env(safe-area-inset-*)`) for notched devices
 - Header matches existing panel header style but full-width
 - Input pinned to bottom, messages scroll above
 ### Welcome Message & Chips
 - Welcome text styled as an AI message bubble (left-aligned, light background)
 - Chips: small rounded pills with teal border, teal text on hover, `font-ui` 12-13px
 - 2-3 chips arranged in a flex-wrap row below the welcome bubble
 - Example questions: "What's his NHS experience?", "Tell me about his data skills", "What projects has he built?"
 ### Model Indicator
 - Placed in the chat panel header, right-aligned or below the "Ask about Andy" title
 - `font-geist`, 11px, `var(--text-tertiary)` color
 - Format: "Powered by Gemini 3 Flash" or just "Gemini 3 Flash"
 ## Technical Considerations
 ### Self-Hosting ONNX Model
 - Transformers.js supports a `localURL` or custom `env.localModelPath` configuration to redirect model loading from HF CDN to a local path
 - The quantized model (`model_quantized.onnx`) is ~23MB — acceptable for a static deploy
 - Files must be served with correct MIME types (`.onnx` as `application/octet-stream`)
 - The build-time script and browser runtime must both point to the same model files
 ### Gemini Model Update
 - `gemini-3-flash-preview` may have a different API path structure — verify against the Generative Language API docs
 - The streaming SSE format should be identical across Flash models, but verify the response shape
 ## Success Metrics
 - Mobile chat is comfortable to use on a phone-sized viewport (no overflow, no cropping)
 - Suggested questions reduce "blank screen" hesitation — visitors engage faster
 - ONNX model loads successfully from local path (no HF CDN requests in network tab)
 - Chat responses come through on the new Gemini model with correct item references
 ## Open Questions
 - Should the suggested question chips be configurable from a data file, or hardcoded in the component?
 - Does `gemini-3-flash-preview` require a different API version path (`v1beta` vs `v1`)?
@@ -0,0 +1,197 @@
 # PRD: Improve LLM CV Knowledge Accuracy
 ## Introduction
 The portfolio's AI chat gives inaccurate or shallow answers about Andy's work history. The root cause: the system prompt feeds `buildEmbeddingTexts()` summaries rather than the full CV detail. Questions about specific achievements, methodology, clinical specialties, or cross-role context produce vague or incorrect responses. This PRD defines an iterative improvement process: enrich the LLM's context, measure accuracy against 10 verifiable benchmark questions, and repeat until all pass — while ensuring changes are structural (not question-specific hacks).
 Additionally, the LLM provider is changing from Gemini to **OpenRouter** using the **z-ai/glm-5** model. This requires migrating the API integration, renaming the module, and updating the benchmark harness to use the new provider.
 ## Goals
 - Achieve 10/10 accuracy on benchmark questions with factually correct, detailed, citation-worthy answers
 - Ensure improvements are structural — benefiting all possible queries, not just the 10 benchmarks
 - Maintain the existing architecture (no new APIs beyond OpenRouter, no RAG infrastructure, no backend)
 - Migrate from Gemini to OpenRouter (z-ai/glm-5) for both production chat and benchmark scoring
 - Regenerate embeddings when embedding texts change, keeping search and LLM context in sync
 ## Benchmark Questions
 These 10 questions have verifiable answers from CV_v4.md and the structured data files. Each tests a different knowledge gap.
 | # | Question | Expected Answer (summary) | Tests |
 |---|----------|--------------------------|-------|
 | Q1 | "How many years has Andy been employed by the NHS?" | ~3.5 years (May 2022–present). Tesco was private sector. | NHS vs non-NHS employer distinction |
 | Q2 | "What was Andy's involvement with tirzepatide?" | Supported NICE TA1026 commissioning, authored executive paper advocating primary care model, drove GP-led delivery. | Deep role-specific detail |
 | Q3 | "What specific tools and software has Andy built?" | 5 projects: switching algorithm, Blueteq generator, CD monitoring, Sankey tool, PharMetrics. Each with outcomes. | Cross-role aggregation |
 | Q4 | "What were Andy's A-level subjects and grades?" | Maths A*, Chemistry B, Politics C. Highworth Grammar School, 2009–2011. | Specific education detail |
 | Q5 | "Was Andy's Tesco role part of the NHS?" | No. Tesco PLC is private. Community pharmacy, not NHS employment. LPC representative for Norfolk. | Employer classification |
 | Q6 | "How did the patient switching algorithm work?" | Python, real-world GP data, auto-identified patients for alternatives, 3 days vs months manual, 14,000 patients, £2.6M, novel GP payment system. | Methodology depth |
 | Q7 | "What clinical specialties has Andy worked across?" | Rheumatology, ophthalmology (wet AMD, DMO, RVO), dermatology, gastroenterology, neurology, migraine — from high-cost drugs role. | Narrative detail not in bullet summaries |
 | Q8 | "What is Andy's experience with the dm+d?" | Created comprehensive medicines data table integrating all dm+d products with standardised strengths, morphine equivalents, Anticholinergic Burden scoring — single source of truth. | Technical achievement context |
 | Q9 | "What budget does Andy manage and how?" | £220M prescribing budget. Forecasting models, variance analysis, financial reporting to executive team, interactive expenditure dashboard. | Figure + methodology |
 | Q10 | "What leadership training does Andy have?" | Mary Seacole Programme (2018, 78%). Also national induction programme at Tesco, NVQ3 supervision. | Cross-role synthesis |
 ### Scoring Criteria
 Each question scored 0–2:
 - **0 — Incorrect**: Wrong facts, invented detail, or contradicts CV
 - **1 — Partial**: Correct but missing key detail, or vague where specifics are available
 - **2 — Accurate**: Factually correct, appropriately detailed, cites specific achievements/metrics
 **Pass threshold**: 18/20 (90%), with no question scoring 0.
 ### Anti-Benchmaxing Rules
 - No hardcoded answers or question-specific prompt clauses
 - Every change must be a structural improvement (richer context, better prompt patterns, enriched embeddings)
 - After each iteration, mentally evaluate: "Would this help a question NOT in the benchmark?" — if no, reject the change
 - The system prompt must not reference benchmark questions or their specific phrasings
 ## User Stories
 ### US-001: Migrate production chat from Gemini to OpenRouter
 **Description:** As a developer, I need to replace the Gemini API integration with OpenRouter so the chat uses z-ai/glm-5.
 **Acceptance Criteria:**
 - [ ] Rename `src/lib/gemini.ts` → `src/lib/llm.ts`
 - [ ] Update all imports across the codebase (`ChatWidget.tsx`, `search.ts`, etc.)
 - [ ] Replace Gemini API calls with OpenRouter's OpenAI-compatible API (`https://openrouter.ai/api/v1/chat/completions`)
 - [ ] Model set to `z-ai/glm-5`
 - [ ] API key read from `VITE_OPEN_ROUTER_API_KEY` env var
 - [ ] SSE streaming still works (OpenRouter supports `stream: true`)
 - [ ] System prompt and message format adapted to OpenAI chat completions format (`messages` array with `role`/`content`)
 - [ ] Export updated display name constant (e.g., `LLM_DISPLAY_NAME = 'GLM-5'`) and update model indicator in chat UI
 - [ ] Rename `isGeminiAvailable()` → `isLLMAvailable()` (or similar)
 - [ ] Typecheck passes
 - [ ] **Verify in browser**: chat opens, sends a message, streams a response
 ### US-002: Migrate benchmark script to OpenRouter
 **Description:** As a developer, I need the benchmark harness to use OpenRouter so it tests the same model and prompt path as production.
 **Acceptance Criteria:**
 - [ ] `scripts/benchmark.ts` uses OpenRouter API instead of Gemini
 - [ ] API key read from `VITE_OPEN_ROUTER_API_KEY` (loaded from `.env`)
 - [ ] Request format uses OpenAI chat completions structure
 - [ ] Model identifier set to `z-ai/glm-5`
 - [ ] Rate limit/retry logic updated for OpenRouter's error responses
 - [ ] Scoring calls also use OpenRouter (same provider for all LLM calls)
 - [ ] `npm run benchmark` still works end-to-end
 - [ ] Typecheck passes
 ### US-003: Enrich system prompt with full CV context
 **Description:** As a portfolio visitor, I want the AI to have comprehensive knowledge of Andy's background so it can answer detailed questions accurately.
 **Acceptance Criteria:**
 - [ ] System prompt includes full professional profile narrative (from CV_v4.md profile section)
 - [ ] Each role includes full achievement bullets, not just summaries
 - [ ] Clear distinction between NHS employment (May 2022+) and private sector (Tesco)
 - [ ] Clinical specialties, methodology details, and specific outcomes included
 - [ ] Education includes specific grades, subjects, research topics
 - [ ] Prompt is well-structured with clear sections for easy LLM parsing
 - [ ] No invented or extrapolated content — everything sourced from CV_v4.md and data files
 - [ ] Typecheck passes
 ### US-004: Improve system prompt instructions
 **Description:** As a portfolio visitor, I want the AI to use its knowledge effectively — citing specifics, distinguishing between employers, and aggregating across roles when asked.
 **Acceptance Criteria:**
 - [ ] Prompt instructs LLM to distinguish NHS employment from private sector roles
 - [ ] Prompt instructs LLM to aggregate across roles when asked broad questions (e.g., "what tools has Andy built?")
 - [ ] Prompt instructs LLM to cite specific metrics, dates, and outcomes when available
 - [ ] Temperature and token limits are appropriate for detailed answers (review current 0.7 temp, 512 max tokens)
 - [ ] Typecheck passes
 ### US-005: Enrich embedding texts for semantic search
 **Description:** As a portfolio visitor, I want semantic search to surface relevant results even for nuanced queries so the chat and command palette find the right content.
 **Acceptance Criteria:**
 - [ ] `buildEmbeddingTexts()` generates richer text per item — full achievement narratives, methodology detail, clinical specialties
 - [ ] Role `history` narratives are included (currently only `examination` bullets and `codedEntries`)
 - [ ] Cross-references included where items relate (e.g., CD monitoring links to controlled drugs skill)
 - [ ] Embedding texts remain well-formed natural language (not keyword soup)
 - [ ] Typecheck passes
 ### US-006: Regenerate embeddings
 **Description:** As a developer, I need embeddings regenerated whenever embedding texts change so semantic search results match the enriched content.
 **Acceptance Criteria:**
 - [ ] Embeddings regenerated using the same model (all-MiniLM-L6-v2)
 - [ ] Output written to `src/data/embeddings.json`
 - [ ] Number of embeddings matches number of palette items
 - [ ] Regeneration can be triggered via script (`npm run generate-embeddings` or similar)
 - [ ] Typecheck passes
 ### US-007: Iterative benchmark loop
 **Description:** As a developer, I want to run the benchmark, review scores, make improvements, and repeat until the pass threshold is met.
 **Acceptance Criteria:**
 - [ ] Run benchmark → review scores → identify failing questions → make structural improvements → repeat
 - [ ] Each iteration logged with: changes made, scores before/after, rationale
 - [ ] Minimum 2 iterations, maximum 10
 - [ ] Stop when 18/20 achieved with no question scoring 0
 - [ ] Final iteration results saved as evidence
 - [ ] All changes pass typecheck before benchmarking
 ### US-008: Validate no regression on general queries
 **Description:** As a portfolio visitor, I want the AI to still handle general questions well after the benchmark-focused improvements.
 **Acceptance Criteria:**
 - [ ] Test 5 general questions not in the benchmark (e.g., "Tell me about Andy", "What does Andy do?", "How can I contact Andy?", "What is this website?", "What are Andy's strongest skills?")
 - [ ] All general questions produce sensible, accurate responses
 - [ ] No degradation in response quality for broad queries
 - [ ] System prompt size hasn't grown to a point that degrades response speed noticeably
 ## Functional Requirements
 - FR-1: Production chat must use OpenRouter API with model `z-ai/glm-5`
 - FR-2: API key sourced from `VITE_OPEN_ROUTER_API_KEY` environment variable
 - FR-3: LLM module renamed from `gemini.ts` to `llm.ts` with updated exports
 - FR-4: Chat UI displays "GLM-5" as the model indicator (replacing "Gemini 3 Flash")
 - FR-5: Benchmark harness must use the identical system prompt construction path as production (`buildSystemPrompt()` from `llm.ts`)
 - FR-6: System prompt changes must be made in `llm.ts` and/or `search.ts` — the same files that serve production
 - FR-7: Embedding text changes must be in `buildEmbeddingTexts()` in `search.ts`
 - FR-8: Scoring must be automated via LLM (OpenRouter), not manual review
 - FR-9: All benchmark artifacts (questions, expected answers, results) stored in `scripts/`
 - FR-10: Embedding regeneration must produce deterministic output for the same input texts
 - FR-11: System prompt must remain a single self-contained context block (no external retrieval at runtime)
 ## Non-Goals
 - No RAG infrastructure or vector database
 - No additional API integrations beyond OpenRouter
 - No changes to the chat UI layout, streaming UX, or item linking (beyond model name display)
 - No changes to the command palette search UX
 - No changes to boot sequence, ECG, or login phases
 - No new backend or server-side components
 - Not optimising for adversarial/trick questions — focus is on legitimate CV queries
 - No keeping Gemini as a fallback — this is a full replacement
 ## Technical Considerations
 - **OpenRouter API format**: Uses OpenAI-compatible chat completions endpoint (`POST https://openrouter.ai/api/v1/chat/completions`). Messages use `{ role: 'system' | 'user' | 'assistant', content: string }` format. Streaming uses `stream: true` with SSE `data:` lines containing `choices[0].delta.content`.
 - **Authentication**: `Authorization: Bearer <VITE_OPEN_ROUTER_API_KEY>` header. Include `HTTP-Referer` and `X-Title` headers as recommended by OpenRouter.
 - **Rate limits**: OpenRouter has per-model rate limits. Add retry logic for 429 responses. The benchmark script should include delays between calls.
 - **Embedding regeneration**: Needs Node.js script that loads the ONNX model and processes all texts. Existing `scripts/generate-embeddings` script should be reused.
 - **Temperature**: Current 0.7 may introduce variability in answers. Consider lowering to 0.3–0.5 for more consistent factual responses. Benchmark both.
 - **Max tokens**: Current 512 may truncate detailed answers. Consider increasing to 768 or 1024 for benchmark testing.
 - **Prompt structure**: Well-structured prompts with clear headings/sections parse better for LLMs than flat text. Consider markdown structure in system prompt.
 - **CORS**: OpenRouter supports browser-side calls. The existing client-side fetch pattern should work without changes.
 ## Success Metrics
 - 18/20 or higher on benchmark (90%+ accuracy)
 - No question scores 0 (no factual errors)
 - 5/5 general validation questions pass
 - System prompt remains under 8KB
 - No typecheck or lint regressions
 - Embedding regeneration completes without errors
 - Chat streaming works in-browser with OpenRouter
 ## Resolved Questions
 - **Model provider**: OpenRouter with z-ai/glm-5 (replaces Gemini 3 Flash).
 - **File naming**: `gemini.ts` renamed to `llm.ts` for provider-agnostic naming.
 - **Benchmark provider**: OpenRouter used for both chat answers and scoring (single provider).
 - **Benchmark results are git-tracked.** Each iteration's scores are committed so improvement over time is visible and auditable.
 - **Existing `scripts/generate-embeddings` script exists.** Review and adapt as needed rather than building from scratch.
 - **Benchmark harness is permanent.** Kept as an ongoing regression test (`npm run benchmark`) for validating LLM accuracy after any data or prompt changes. Question set can be expanded over time.
Author	SHA1	Message	Date
admin	0fbbf9e46f	merge	2026-02-15 23:20:24 +00:00
admin	4580ca9c84	feat: US-014 - Update to Gemini 3 Flash Preview with model indicator	2026-02-15 21:02:52 +00:00
admin	667e5b249c	feat: US-013 - Self-host ONNX embedding model Download all-MiniLM-L6-v2 model files to public/models/ and configure @xenova/transformers to load from local path instead of Hugging Face CDN. Eliminates external dependency for semantic search embedding model.	2026-02-15 20:59:03 +00:00
admin	9e9dd1ae4b	feat: US-012 - Welcome message with suggested question chips	2026-02-15 20:48:00 +00:00
admin	ab5444ee94	feat: US-011 - Mobile full-screen chat panel	2026-02-15 20:43:48 +00:00