import time
import uuid
import logging
from functools import wraps
from typing import Callable, Any
from fastapi import Request, HTTPException
from redis.asyncio import Redis
from ai_providers import LLMClient

logger = logging.getLogger(__name__)

class MonetizationMiddleware:
    def __init__(self, redis: Redis, pricing_engine: PricingEngine, event_bus: EventBus):
        self.redis = redis
        self.pricing_engine = pricing_engine
        self.event_bus = event_bus

    def metered(self, cost_center: str, model: str):
        def decorator(func: Callable):
            @wraps(func)
            async def wrapper(*args, **kwargs):
                request: Request = kwargs.get('request')
                user_id = request.state.user_id
                org_id = request.state.org_id
                req_id = str(uuid.uuid4())

                # 1. Pre-flight Quota Check (Synchronous)
                quota_key = f"quota:{org_id}:{cost_center}"
                current_usage = await self.redis.get(quota_key)
                limit = await self.pricing_engine.get_limit(org_id, cost_center)
                
                if current_usage and int(current_usage) >= limit:
                    raise HTTPException(status_code=429, detail="Quota exceeded")

            # 2. Execute Inference
            start_time = time.perf_counter()
            try:
                # Assume func returns a response object with token counts
                response = await func(*args, **kwargs)
                latency = (time.perf_counter() - start_time) * 1000
                
                input_tokens = response.usage.prompt_tokens
                output_tokens = response.usage.completion_tokens
                
                # 3. Calculate Cost
                cost = self.pricing_engine.calculate(
                    model=model, 
                    input_tokens=input_tokens, 
                    output_tokens=output_tokens
                )

                # 4. Update Quota & Emit Event (Async/Fire-and-forget for billing)
                await self.redis.incrby(quota_key, 1)
                await self.redis.expire(quota_key, 86400) # Reset daily

                await self.event_bus.publish({
                    "event": "ai_usage",
                    "request_id": req_id,
                    "user_id": user_id,
                    "org_id": org_id,
                    "cost_center": cost_center,
                    "model": model,
                    "input_tokens": input_tokens,
                    "output_tokens": output_tokens,
                    "cost_usd": cost,
                    "timestamp": time.time()
                })
                
                return response

            except Exception as e:
                logger.error(f"Monetization failure for {req_id}: {e}")
                # Fail open or closed based on policy? 
                # Recommendation: Fail closed for cost protection.
                raise HTTPException(status_code=500, detail="Service unavailable")
        return wrapper
    return decorator

#### 4. Pricing Engine Design
Hardcoding prices is an anti-pattern. Model providers change rates frequently. The pricing engine should load configurations from a versioned store.

```python
class PricingEngine:
    def __init__(self, config_store: ConfigStore):
        self.config = config_store.load_pricing_config()

    def calculate(self, model: str, input_tokens: int, output_tokens: int) -> float:
        pricing = self.config.models.get(model)
        if not pricing:
            raise ValueError(f"Unknown model: {model}")
        
        # Support for tiered pricing (e.g., first 1M tokens @ $X, next @ $Y)
        input_cost = self._apply_tiers(input_tokens, pricing.input_tiers)
        output_cost = self._apply_tiers(output_tokens, pricing.output_tiers)
        
        return input_cost + output_cost

    def _apply_tiers(self, tokens: int, tiers: list) -> float:
        cost = 0.0
        remaining = tokens
        for tier in tiers:
            if remaining <= 0:
                break
            volume = min(remaining, tier.limit)
            cost += (volume / 1000) * tier.rate
            remaining -= volume
        return cost

---

Pitfall Guide

1. Ignoring Input vs. Output Token Asymmetry

Most models charge significantly more for output tokens. If your AI generates long responses (e.g., code completion, summaries), costs scale non-linearly.

• Fix: Implement output token caps and stream responses to allow early termination if costs exceed thresholds.

2. The "Prompt Injection" Cost Attack

Malicious users can craft prompts designed to maximize output tokens or trigger infinite loops in agentic workflows.

• Fix: Enforce strict max_tokens limits. Monitor for anomalous output lengths per request. Implement circuit breakers for cost-per-minute spikes.

3. Race Conditions in Quota Deduction

Using standard read-modify-write patterns for quotas without atomic operations leads to over-consumption.

• Fix: Use Redis INCR or Lua scripts for atomic quota checks. Never read, calculate, and write separately in distributed environments.

4. Forgetting Non-Inference Costs

Monetization often focuses only on LLM tokens. Embeddings, vector database storage, and GPU compute for fine-tuning also incur costs.

• Fix: Expand the cost_center taxonomy. Track embedding tokens and storage GBs. Include these in the total cost of ownership calculation.

5. Hardcoded Pricing Models

API providers update pricing monthly. Hardcoded values result in margin erosion or overcharging.

• Fix: Externalize pricing to a configuration service or database. Implement a pricing update webhook that refreshes the engine cache without redeployment.

6. Lack of User Transparency

Users churn when they receive a bill they cannot explain.

• Fix: Provide a real-time usage dashboard. Include a "Cost Estimator" before execution where possible. Send usage alerts at 50%, 80%, and 100% of quotas.

7. Model Drift Impacting Costs

As models improve, prompt effectiveness may change, altering token usage patterns. A prompt that worked efficiently on v3 might waste tokens on v4 due to different behavior.

• Fix: Monitor "Cost per Successful Outcome" metrics, not just token volume. A/B test prompts for cost efficiency alongside quality.

---

Production Bundle

Action Checklist

• Define Cost Centers: Map every AI feature to a cost_center (e.g., search, summarize, code_assist).
• Deploy Monetization Middleware: Implement the middleware pattern to intercept all AI calls.
• Configure Atomic Quotas: Set up Redis-backed quota enforcement with TTLs.
• Externalize Pricing: Build a pricing engine that loads from a dynamic config source.
• Implement Idempotent Billing: Ensure usage events include request_id for deduplication in your billing ledger.
• Add Cost Alerts: Configure alerts for cost spikes, quota breaches, and anomalous usage patterns.
• Build User Dashboard: Expose usage metrics and cost breakdown to end-users.
• Stress Test Billing: Simulate high-concurrency scenarios to verify quota atomicity and event bus throughput.

Decision Matrix: Billing Integration

Feature	Stripe Metered Billing	Custom Ledger + Stripe	Hybrid (LemonSqueezy/Razorpay)
Flexibility	Medium	High	Medium
Dev Effort	Low	High	Low
Auditability	Medium	High	Medium
Real-time Limits	Requires Webhooks	Native	Medium
Best For	Standard SaaS	Complex AI Usage	Global/Creator Economy
Recommendation	Use for simple per-seat + flat overage.	Recommended for AI. Full control over token logic, complex tiers, and custom cost attribution.	Good for rapid validation, limited by API constraints.

Configuration Template

Save this as pricing_config.yaml to drive your pricing engine. This structure supports multi-model, tiered pricing.

version: "2024-05-15"
models:
  gpt-4o:
    input_tiers:
      - limit: 1000000
        rate_per_1k: 0.005
      - limit: 10000000
        rate_per_1k: 0.004
    output_tiers:
      - limit: 1000000
        rate_per_1k: 0.015
      - limit: 10000000
        rate_per_1k: 0.012
  text-embedding-3-large:
    input_tiers:
      - limit: 100000000
        rate_per_1k: 0.00013
    output_tiers: [] # Embeddings usually input-only

quotas:
  free_tier:
    daily_limit: 100000 # tokens
    monthly_limit: 2000000
  pro_tier:
    daily_limit: 5000000
    monthly_limit: 100000000

Quick Start Guide

1. Initialize Middleware:

   pip install redis fastapi ai-monetizer
   # Configure Redis connection and PricingEngine in your app entry point
   

2. Wrap Endpoints: Apply the @metered decorator to all routes invoking AI models. Ensure request.state contains user/org context.
3. Configure Pricing: Create pricing_config.yaml based on current provider rates. Load into your PricingEngine.
4. Deploy & Monitor: Deploy the service. Monitor the ai_usage event stream. Verify that Redis quotas are decrementing and events are firing.
5. Iterate: After 7 days, analyze cost attribution data. Adjust pricing tiers and quotas based on actual usage distribution.

---

Codcompass Insight: Monetization is not just a revenue function; it is a system design constraint. By engineering monetization into the architecture from day one, you gain real-time visibility into unit economics, protect margins against usage volatility, and build a product that scales profitably. Treat tokens as currency, and your middleware as the central bank.