AI Config | DirectoryLaunch

OpenAI or Anthropic for auto-generating descriptions — provider choice, model selection, cost ceilings and caching strategy.

Why this matters

This is the smallest config file in the project, but it controls the most expensive runtime behaviour: every LLM call costs real money, and a misconfigured retry loop can drain $100/day before you notice. The good news is that the right defaults — gpt-4o-mini or claude-haiku-4-5, capped tokens, modest temperature — keep AI generation under $0.05 per project description even at production scale.

This page documents the toggles, recommends models for different niches, and lays out the cost-control patterns that prevent surprises. If you're enabling AI for the first time, read Common pitfalls before deploying — three of the four issues are budget killers.

File: config/ai.config.ts

When AI is on, users get "Generate with AI" buttons in the submit flow — the LLM writes descriptions, suggests categories, and fills SEO metadata from a URL.

Two switches to flip

You need both ai: true in features.config.ts and a valid AI_API_KEY. Either missing → AI silently skips.

Supported providers

Provider	Where to get a key
OpenAI	platform.openai.com → API keys
Anthropic	console.anthropic.com → API keys

Setup

1. Turn on the feature flag

config/features.config.ts

export const featuresConfig = {
  // ...
  ai: true,
};

2. Add env vars

.env.local

AI_PROVIDER=openai          # or 'anthropic'
AI_MODEL=gpt-4o-mini        # or 'claude-haiku-4-5' etc.
AI_API_KEY=sk-...

Restart the dev server.

3. Verify

Go to /submit → fill in a project URL → click "Generate description with AI". If nothing happens, check the dev console / network tab for errors.

The config file

config/ai.config.ts

export const aiConfig = {
  provider: (process.env.AI_PROVIDER as 'openai' | 'anthropic') || 'openai',
  model: process.env.AI_MODEL || 'gpt-4o-mini',
  apiKey: process.env.AI_API_KEY || '',
  maxTokens: 1000,
  temperature: 0.7,
};

Field	Default	What it does
`provider`	`openai`	Provider. Must match your key.
`model`	`gpt-4o-mini`	Model ID. See provider's docs for available models.
`apiKey`	—	Pulled from `AI_API_KEY`.
`maxTokens`	`1000`	Max response length.
`temperature`	`0.7`	0 = deterministic, 1 = creative. Good middle ground for descriptions.

Recommended models

For directory descriptions (short, factual), cheap fast models win:

Provider	Model	Input / Output cost
OpenAI	`gpt-4o-mini`	cheapest GPT-4-class
Anthropic	`claude-haiku-4-5`	cheap and fast
OpenAI	`gpt-4o`	better quality, more expensive
Anthropic	`claude-sonnet-4-5`	best quality Anthropic

Start with the cheap one. Only upgrade if outputs are noticeably worse.

Cost control

Each description generation is ~300-500 tokens → fractions of a cent per request.
Rate-limit AI endpoints in lib/rate-limit.ts to prevent abuse.
Consider caching generated descriptions by URL so re-submissions don't re-pay.

A real-world walkthrough: choosing a model for a 200-listing AI directory

You're operating an AI-tools directory and expect to publish 30–50 new listings per week. Each listing needs a description (~80 words) and category suggestions. Math:

30 listings/week × 4 weeks = ~120 generations/month
Each generation: ~600 input tokens (URL fetched), ~200 output tokens
gpt-4o-mini at current pricing: roughly $0.04 per generation → $5/month
claude-haiku-4-5: similar
gpt-4o (premium): roughly $0.40 per generation → $50/month

Start with gpt-4o-mini. Generate 10 sample descriptions, manually review for tone and accuracy. If quality is acceptable (it usually is for short marketing copy), stay there. If you find consistent issues — tone too generic, missing nuance for a specific niche — try claude-haiku-4-5 next. Only escalate to gpt-4o or claude-sonnet-4-6 if both small models struggle.

Set a $50 monthly cap in your provider dashboard regardless. That's a 10× safety margin against runaway loops, and it triggers an email alert before any meaningful damage.

Common pitfalls

Setting temperature: 1.0 to "make it creative". For directory descriptions, you want consistent voice across listings. 0.6–0.8 is the right range. 1.0+ produces variability that reads as inconsistency at scale.
Forgetting the provider's spend cap. Both OpenAI and Anthropic let you set hard monthly spend limits in the dashboard. Set one. A misconfigured loop in your generation code can run unbounded calls until the limit kicks in.
Hardcoding the API key in the config file. Always pull from env. The config file gets committed; the env file shouldn't. If you accidentally commit a key, rotate it immediately at the provider — it's already scraped by GitHub bots.
Mixing providers in one deploy. AI_PROVIDER=openai with AI_MODEL=claude-haiku-4-5 silently fails — the OpenAI client doesn't know about Anthropic models. Make sure both env vars are set together, and test locally before pushing.
Caching too aggressively. Caching by URL alone means a project that updates its tagline keeps showing the stale AI description for months. Cache TTL of 7 days is a reasonable default; expire on user-triggered regenerate.

FAQ

Can I run a local model (Ollama, vLLM) for free generations?

Yes — Ollama exposes an OpenAI-compatible API on :11434/v1. Set AI_PROVIDER=openai and AI_BASE_URL=http://localhost:11434/v1, plus a model like llama3.1:8b. Quality of small open-source models is noticeably worse for short marketing copy than the hosted commercial models — usable for prototyping, not great for production.

How do I switch from OpenAI to Anthropic without code changes?

Update three env vars: AI_PROVIDER=anthropic, AI_MODEL=claude-haiku-4-5, AI_API_KEY=<anthropic-key>. Restart the server. The internal lib/ai.ts adapter handles both providers transparently — no other code changes needed.

Should I let users see the AI was used?

Optional, but transparency is increasingly the norm. Add a small "Generated with AI — review before submitting" notice next to the Generate button. This sets expectations and reduces support tickets when output isn't perfect.