.env.example

# ─────────────────────────────────────────────────────────────────────────
# SDK Orchestrator — environment configuration (V2)
#
# Copy to `.env.local` and fill in. NEVER commit the filled-in file.
#
# COST MODEL (V2 decision): generation runs on NVIDIA's HOSTED NIM free tier,
# NOT self-hosted Brev. Self-hosting the 120B model is ~$120/day and burns the
# entire compute budget in days. Brev compute is reserved exclusively for
# ephemeral notebook verification. See docs/V2_ARCHITECTURE.md.
# ─────────────────────────────────────────────────────────────────────────

# ── Generation — hosted NIM (the default, free-tier) ─────────────────────
# Get a key at https://build.nvidia.com — free credits cover dev + demo.
# Leave NIM_BASE_URL unset (or at the hosted default) to use hosted NIM.
NVIDIA_API_KEY=nvapi-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
NIM_BASE_URL=https://integrate.api.nvidia.com/v1
NIM_REASONING=true

# ── Retrieval / grounding — hosted NIM embeddings + rerank ───────────────
# Same NVIDIA_API_KEY signs these. Models are called via hosted NIM.
NIM_EMBED_MODEL=nvidia/nv-embedqa-e5-v5
NIM_RERANK_MODEL=nvidia/nv-rerankqa-mistral-4b-v3

# ── Vector store ─────────────────────────────────────────────────────────
# Default: Milvus Lite (embedded, file-backed, zero infra, free).
# Point MILVUS_URI at a standalone Milvus only when scale demands it.
MILVUS_URI=./data/milvus_lite.db
BLUEPRINT_COLLECTION=nvidia_blueprints

# ── State — Postgres (free tier: Supabase / local docker) ────────────────
POSTGRES_DSN=postgresql://orchestrator:orchestrator@localhost:5432/orchestrator

# ── Session memory — Redis (local docker, free) ──────────────────────────
REDIS_URL=redis://localhost:6379

# ── Observability — Phoenix (local docker, free) ─────────────────────────
PHOENIX_ENDPOINT=http://localhost:6006/v1/traces
PHOENIX_PROJECT=sdk-orchestrator

# ── Verification — Brev (THE compute budget lives here) ──────────────────
# Ephemeral GPU provisioning for notebook execution. Scale-to-zero.
# Tier-A (cheap "does it run"): A10G. Tier-B (real hardware): H100, on demand.
BREV_API_TOKEN=
BREV_DEFAULT_GPU=A10G
BREV_HEAL_MAX_TURNS=30
BREV_HEAL_BUDGET_USD=2.00

# ── Anthropic — verification agent (Tier-2 hard-case fixer only) ─────────
# Deterministic heal loop is Tier-1 (free, Nemotron). Claude Code is reserved
# for cells the deterministic loop can't fix. Optional.
ANTHROPIC_API_KEY=

# ── Fixture replay (dev only — short-circuits all network LLM calls) ──────
# Set to a fixtures dir to run the pipeline offline against saved responses.
# LLM_REPLAY=fixtures/
# LLM_SAVE_FIXTURES=1

# ── Eval mode ────────────────────────────────────────────────────────────
EVAL_MODE=false