A knowledge base for legal documents, powered by LlamaIndex Index v2 (the LlamaParse Platform).

Sign in, create a project, upload files, and chat with an agent that can run hybrid semantic search, find/read files, and grep across your documents. Each project is mirrored as a managed LlamaCloud index — uploads are parsed and indexed automatically, and the chat agent queries that index live.

• Auth: WorkOS (AuthKit) — sign-in flow only; user ids drive ownership
• Storage: PostgreSQL via Prisma. User API keys are encrypted at rest (AES-256-GCM)
• Indexing: LlamaCloud Index v2 — one index per project, files uploaded + synced via the Index v2 API
• Chat: Vercel AI SDK 6 (useChat) + TanStack Start server route, streaming UIMessages. Markdown rendered with Streamdown
• Persistence: chat messages stored as full UIMessage JSON, grouped into conversations (multiple threads per project)
• Models: pick OpenAI or Anthropic per turn; bring your own keys via the profile page

TanStack Start · React 19 · Tailwind v4 · Prisma 7 · Vite · Bun · AI SDK 6 · @llamaindex/llama-cloud · WorkOS AuthKit

• Bun (1.3+)
• A PostgreSQL database (local or hosted — Neon, Supabase, Render Postgres all work)
• A WorkOS account with an AuthKit application
• A LlamaCloud account (users add their keys via the profile page at runtime, so this isn't strictly needed at boot — but you'll want one to actually use the app)
• An OpenAI and/or Anthropic API key (also added via the profile page)

bun install
cp .env.example .env.local   # then fill it in (see below)
bun run db:push              # apply Prisma schema to your DB
bun run dev                  # http://localhost:5173

All vars live in .env.local. Required:

Optional:

User-facing API keys (LlamaCloud, OpenAI, Anthropic) are not in env. Each user enters them on /user-profile, where they're encrypted with ENCRYPTION_KEY and persisted per-user.

Schema lives in prisma/schema.prisma. Useful scripts:

bun run db:push       # push schema (no migration files) — good for dev
bun run db:migrate    # create a named migration
bun run db:studio     # open Prisma Studio
bun run db:generate   # regenerate Prisma client after editing the schema

Core tables:

• User — mirrors WorkOS users (lazily upserted)
• ApiKeys — encrypted LlamaCloud / OpenAI / Anthropic credentials per user
• Project — one row per project; holds the LlamaCloud indexId
• ProjectFile — files uploaded to a project's index, with sync status
• Conversation — chat threads scoped to (userId, projectId)
• Message — full UIMessage JSON, ordered within a conversation

1. Sign in via WorkOS.
2. Profile (/user-profile): paste your LlamaCloud API key + project id, plus your OpenAI and/or Anthropic key. Saved keys are encrypted server-side.
3. Projects (/projects): create a project. This creates a new LlamaCloud Index v2 mirror. Upload files (drag/drop or picker). Files sync to the index in the background — status polls until ready.
4. Chat (/chat): pick a project, start (or pick) a conversation, pick a model, and ask away. The agent has these tools available:
   • hybrid search over the project's index
   • find files by name/glob
   • read file contents
   • grep across the project

bun run build
node dist/server/index.mjs

The output is a self-contained Node server (via Nitro). Any Node-compatible host works — Render, Fly.io, your own VPS, etc. For host-specific presets (Vercel, Netlify, Cloudflare, AWS Lambda), see nitro.build/deploy.

Make sure DATABASE_URL, ENCRYPTION_KEY, and the VITE_WORKOS_* vars are set in your production environment. VITE_* vars are baked in at build time, the rest are read at runtime.

src/
  routes/
    index.tsx           landing page
    chat.tsx            chat UI with conversation sidebar
    projects.tsx        project CRUD + file upload/sync UI
    user-profile.tsx    encrypted API key management
    api/chat.ts         streaming chat endpoint (TanStack Start server route)
  lib/
    agent.ts            AI SDK agent definition + tools
    files.ts            LlamaCloud Index v2 file ops
    projects.ts         project CRUD server fns
    messages.ts         conversation + message persistence
    apiKeys.ts          encrypted credential storage
    crypto.ts           AES-256-GCM helpers
  components/           Header, Footer, etc.
prisma/schema.prisma    DB schema

bun run dev               # vite dev server on :5173
bun run build             # production build
bun run preview           # preview the production build
bun run test              # vitest
bun run lint              # eslint
bun run format            # prettier + eslint --fix
bun run check             # prettier --check
bun run generate-routes   # regenerate routeTree.gen.ts

• "configure your LlamaCloud credentials" — go to /user-profile and add them.
• Chat stream cut off mid-response — set NITRO_BUN_IDLE_TIMEOUT=255 in .env.local.
• ENCRYPTION_KEY errors at boot — must be exactly 64 hex chars (32 bytes). Regenerate with openssl rand -hex 32.
• Files stuck "syncing" — the projects page polls the index status; if a file errors, LlamaCloud's status will surface the reason. Detach + re-upload usually fixes transient parse failures.

TBD.