Skip to main content

​
What This Example Is

icrl-ts/web-example is a Next.js app that demonstrates feedback-driven training with Convex-backed persistence and configurable LLM providers.

​
1. Install Dependencies

From icrl-ts/web-example/: 
pnpm install

​
2. Set Up Convex

npx convex dev
This generates convex/_generated/* and a deployment URL.

​
3. Configure .env.local

Required: 
NEXT_PUBLIC_CONVEX_URL=https://<your-deployment>.convex.cloud
Provider options: 
# Option A: Anthropic standard API
ANTHROPIC_API_KEY=...

# Optional explicit provider selection: anthropic | gemini-vertex
LLM_PROVIDER=anthropic

# Option B: Gemini Vertex
GOOGLE_CREDENTIALS_JSON={...full service account json...}
GOOGLE_VERTEX_LOCATION=global
Notes:
  • If LLM_PROVIDER is unset, selection order is: Anthropic first, then Gemini Vertex.
  • Unsupported LLM_PROVIDER values fail configuration checks.

​
4. Start Dev Server

pnpm dev
Open http://localhost:3000.

​
5. Data Model (Convex)

The web demo uses these tables:
  • databases
  • trajectories
  • curationMetadata
  • embeddings
  • examples
  • suggestionCache
Defined in icrl-ts/web-example/convex/schema.ts.

​
6. Production Deploy

  • Deploy Next app (for example Vercel).
  • Configure same env vars in deployment settings.
  • Keep NEXT_PUBLIC_CONVEX_URL aligned with production Convex deployment.