hivemind-ui

A sleek RAG chat UI — index any website, ask questions, get cited answers.

Tech stack

Features

• iMessage-style chat thread — user messages right-aligned in a pink-to-violet gradient bubble; AI responses left-aligned with a Sparkles avatar.
• Cited answers — each AI response links its sources as pill badges that resolve scraped file paths (e.g. scraped_websites/support_squadcast_com/adding_a_schedule.txt) back to the real URL (https://support.squadcast.com/adding-a-schedule).
• Website indexing panel — submit any public URL to trigger a background crawl; 409 (already indexed) is handled gracefully and the URL is accepted as a scope.
• Conversation scope — once a site is crawled, all subsequent queries in that conversation are scoped to that site's namespace.
• Multi-turn conversations — backend conversation_id is persisted per conversation and re-sent on every message for full context continuity.
• Auto-title — the first user message (up to 50 chars) becomes the conversation title automatically.
• FAQ chips — TanStack Query fetches auto-generated FAQs from the backend; up to 5 are shown as clickable chips above the input bar.
• Sidebar FAQ list — all FAQs are also listed in the sidebar as suggestion buttons.
• Health indicator — a live Wifi icon in the header reflects the backend health (healthy / degraded / offline), polled every 60 seconds.
• Offline / waking-up recovery — network errors show a "Server is waking up" banner with a one-click retry that restores the unsent message.
• Not-indexed error — a 404 from the query API surfaces a human-readable banner instead of an error message in the thread.
• Responsive sidebar — collapses to a slide-over drawer on mobile with an overlay backdrop.
• Markdown rendering — AI responses are rendered as structured prose (headings, lists, code blocks, bold, links) with custom prose styles scoped to .ai-bubble.
• Auto-growing textarea — the input expands as the user types, up to 144 px.
• Keyboard shortcut — Enter sends, Shift+Enter inserts a newline.

Getting started

Prerequisites

• Bun >= 1.0
• Node.js >= 20 (used by Bun internally)

Install

bun install

Development

bun run dev

The app is served at http://localhost:5173 by default.

Build

bun run build

Preview production build

bun run preview

Type-check

bun run check

Running tests

Install Vitest if it is not yet present:

bun add -D vitest

Run all tests:

bunx vitest run

Watch mode:

bunx vitest

Project structure

hivemind/
├── src/
│   ├── lib/
│   │   ├── api.ts              # Typed fetch wrappers for all backend endpoints
│   │   ├── chat.svelte.ts      # Svelte 5 rune-based chat store (conversations, messages)
│   │   ├── source-url.ts       # sourceHref / sourceLabel helpers (scraped path to URL)
│   │   ├── api.test.ts         # Vitest unit tests for api.ts
│   │   ├── chat.test.ts        # Vitest unit tests for the chat store
│   │   ├── sourceUrl.test.ts   # Vitest unit tests for source-url.ts
│   │   ├── components/         # shadcn-svelte UI primitives (Button, Badge, Textarea, ...)
│   │   ├── hooks/              # Svelte hooks
│   │   ├── utils.ts            # cn() class-name helper
│   │   └── index.ts            # Barrel re-exports
│   └── routes/
│       ├── +layout.svelte      # QueryClientProvider wrapper
│       ├── +page.svelte        # Main chat page (sidebar + chat thread + input bar)
│       └── layout.css          # Global styles and font imports
├── .doc/
│   └── API_INFO.md             # Backend API reference
├── package.json
├── svelte.config.js
├── vite.config.ts
├── vitest.config.ts
└── tsconfig.json

API integration

The frontend communicates exclusively with the hosted backend at:

https://hivemind-2f75.onrender.com

All API calls are centralised in src/lib/api.ts.

Error handling

All non-2xx responses are thrown as ApiError instances that carry a status code. The UI handles the following cases explicitly:

• 404 from /api/v1/query — the requested URL has not been indexed yet.
• 409 from /api/v1/crawl — the URL is already indexed; the UI accepts this silently and scopes the conversation.
• Network error (TypeError) — server is offline or still waking up from the Render free-tier cold start.

Query scoping

When a user indexes a website, its URL is stored on the active conversation via chat.setScopeUrl. All subsequent sendQuery calls in that conversation include url in the request body, which tells the backend to search only that site's Pinecone namespace.

Conversation continuity

The conversation_id UUID returned by the first sendQuery response is stored on the conversation and sent with every follow-up message, giving the LLM full history access.

Screenshots

Add screenshots here once the UI is deployed.

License

Private — all rights reserved.