Providers

FixtureForge is provider-agnostic. It auto-detects your provider from environment variables and falls back gracefully when no key is set.

---

Auto-Detection Order

ANTHROPIC_API_KEY  ->  Claude  (claude-haiku-4-5-20251001)
OPENAI_API_KEY     ->  GPT     (gpt-4o-mini)
GOOGLE_API_KEY     ->  Gemini  (gemini-2.0-flash)
GROQ_API_KEY       ->  Groq    (llama-3.3-70b-versatile)
(none found)       ->  Ollama  (localhost:11434)
(Ollama down)      ->  Deterministic-only (no AI)

---

Explicit Configuration

# Anthropic Claude
forge = Forge(provider_name="anthropic", model="claude-sonnet-4-6")

# OpenAI
forge = Forge(provider_name="openai", model="gpt-4o")

# Groq
forge = Forge(provider_name="groq", model="llama-3.3-70b-versatile")

# Local Ollama
forge = Forge(provider_name="ollama", model="llama3.2")

# No AI
forge = Forge(use_ai=False)

---

Recommended: Groq Free Tier

Groq offers a generous free tier that works well for development:

• 14,400 requests/day on llama-3.3-70b-versatile
• Fast inference (~300 tokens/sec)
• No credit card required

pip install "fixtureforge[groq]"
export GROQ_API_KEY=gsk_...

forge = Forge()  # auto-detects Groq from GROQ_API_KEY

---

Response Cache

AI responses are cached locally for 7 days. Identical requests cost nothing after the first call.

forge = Forge(use_cache=True)   # default — cache at ~/.fixtureforge/cache/
forge = Forge(use_cache=False)  # disable caching

This means running your test suite twice with the same models and contexts costs half as many API calls.

---

Cost Optimization

FixtureForge uses several strategies to minimize API costs:

Strategy	Saving
Field routing (only semantic fields hit AI)	~70-90%
Batch prompting (all records in one call)	~95% latency
Response cache (7-day TTL)	100% on repeat runs
SYSTEM_PROMPT_DYNAMIC_BOUNDARY (static prompt cached once)	~40%
DataSwarms shared cache	~90% per additional model