// Layer 2: MCP Protocol — AI uses tools like a human uses apps
const tools = [{
  name: 'check_inventory',
  description: 'Check real-time stock level and price for a product SKU',
  inputSchema: {
    type: 'object',
    properties: {
      sku: { type: 'string' },
      size: { type: 'string', optional: true }
    }
  }
}];

// GPT-4o (Layer 0) decides to call the tool
const response = await openai.chat.completions.create({
  model: 'gpt-4o',
  messages: [{ role: 'user', content: customerQuery }],
  tools: tools,        // Layer 2: MCP tools exposed to model
  tool_choice: 'auto'
});

// If model chose to call tool, execute it and return result
if (response.choices[0].message.tool_calls) {
  const toolResult = await executeInventoryCheck(toolCall.arguments);
  // Feed result back — model generates final customer-facing response
}

# Layer 0: Llama 4 running on-prem (no PHI leaves firewall)
from transformers import pipeline
import requests

llm = pipeline('text-generation', model='meta-llama/Llama-4-Scout-17B',
               device='cuda')   # on-prem GPU server

# Structure the transcript into DAP format
structured = llm(f"""You are a clinical documentation specialist.
Convert this transcript to DAP format with ICD-10 codes.
Return structured JSON only.

Transcript: {transcript}""")[0]['generated_text']

# Layer 4: Write to Epic via FHIR R4 API
fhir_note = {
    "resourceType": "DocumentReference",
    "status": "current",
    "content": [{"attachment": {"data": encode_b64(structured)}}],
    "context": {"encounter": [{"reference": f"Encounter/{encounter_id}"}]}
}
requests.post(f'{epic_fhir_base}/DocumentReference',
              json=fhir_note, headers={'Authorization': f'Bearer {token}'})

// Supabase Edge Function (Layer 1 infra) calls Claude Haiku (Layer 0 model)
// This runs server-side — API key never exposed to Flutter app
import Anthropic from '@anthropic-ai/sdk'

const anthropic = new Anthropic({ apiKey: Deno.env.get('ANTHROPIC_API_KEY') })

Deno.serve(async (req) => {
  const { ingredients, dietary_prefs } = await req.json()
  
  // Claude Haiku: fast, cheap, perfect for high-volume inference
  const msg = await anthropic.messages.create({
    model: 'claude-haiku-4-5-20251001',   // Layer 0: cheapest Claude tier
    max_tokens: 512,
    messages: [{
      role: 'user',
      content: `Suggest 3 dinner ideas using: ${ingredients}.
Dietary preferences: ${dietary_prefs}.
Format: JSON array with name, time_minutes, difficulty.`
    }]
  })
  
  return new Response(JSON.stringify({ ideas: msg.content[0].text }))
})