langfuse

name	langfuse
description	LLM observability for EstimateProof's Claude API calls. Traces every call with cost, latency, tokens. Protects the $100/mo company cap. Self-hosted on Render. Use when checking LLM spend, debugging prompt issues, or setting cost alerts.

Langfuse (LLM Observability)

Status: Deploy ready — Render service can be created via API

Why Langfuse

EstimateProof calls Claude API in report-engine/ai/synthesize.mjs with a $100/mo company cap. Without observability:

• One edge-case VIN (salvage title, 47 recalls) could generate a $5 prompt
• No visibility into cost per report
• No alerts before hitting cap

Integration Point

File: companies/estimateproof/report-engine/ai/synthesize.mjs

import { Langfuse } from 'langfuse';

const langfuse = new Langfuse({
  secretKey: process.env.LANGFUSE_SECRET_KEY,
  publicKey: process.env.LANGFUSE_PUBLIC_KEY,
  baseUrl: process.env.LANGFUSE_BASEURL,
});

// Wrap each ask() call:
const trace = langfuse.trace({ name: 'report-generation', metadata: { vin } });
const span = trace.span({ name: 'synthesize', input: { prompt } });
// ... existing Claude call ...
span.end({ output: result, metadata: { tokens, cost } });

Deploy Checklist

1. Create Render Docker service: image langfuse/langfuse, starter plan (~$7/mo)
2. Connect to Supabase Postgres (direct connection string from dashboard)
3. First login → create project → get keys
4. Set env vars in root .env:
   • LANGFUSE_SECRET_KEY
   • LANGFUSE_PUBLIC_KEY
   • LANGFUSE_BASEURL
5. Add langfuse npm dependency to report-engine
6. Instrument synthesize.mjs

Cost Cap Guard

// Before each LLM call, check monthly spend
const usage = await langfuse.fetchUsage({ period: 'month' });
if (usage.totalCost >= 100) {
  throw new Error('CostCapExceeded: $100/mo limit reached');
}

Cost

~$7/mo on Render starter plan