.env.example

# ── Cloud providers ───────────────────────────────────────────────────────────
ANTHROPIC_API_KEY=your-anthropic-key
OPENAI_API_KEY=your-openai-key
GROQ_API_KEY=your-groq-key

# ── LLM selection ─────────────────────────────────────────────────────────────
# Set LLM_PROVIDER to one of: anthropic, openai, ollama, groq
LLM_PROVIDER=anthropic
LLM_MODEL=claude-sonnet-4-20250514

# ── Ollama (local LLM — no API key required) ──────────────────────────────────
# To use Ollama: set LLM_PROVIDER=ollama and pull the model first:
#   ollama pull llama3.1:8b
# llama3.1:8b is recommended — supports tool calling and reliable JSON output.
# OLLAMA_BASE_URL=http://localhost:11434
# LLM_PROVIDER=ollama
# LLM_MODEL=llama3.1:8b

# ── Groq (fast cloud inference — no GPU required) ────────────────────────────
# To use Groq: set LLM_PROVIDER=groq and add your API key above.
# llama-3.3-70b-versatile is recommended for best JSON reliability.
# Note: free tier has a 100k tokens/day limit — Dev Tier recommended for full runs.
# LLM_PROVIDER=groq
# LLM_MODEL=llama-3.3-70b-versatile

# ── LLM-as-judge ──────────────────────────────────────────────────────────────
# Optional: use a lighter model for answer quality scoring to save cost/speed.
# Empty (default) means the judge uses the same model as LLM_MODEL.
# Example with Ollama: JUDGE_MODEL=llama3.2:3b
JUDGE_MODEL=

# ── Data ──────────────────────────────────────────────────────────────────────
DATABASE_URL=sqlite:///data/db/benchmark.db
CHROMA_PERSIST_DIR=data/vectordb
LOG_LEVEL=INFO