# Copy to .env and fill in your values. # Model-endpoint triplets. Every triplet has three fields: # _ENDPOINT (URL) / _API_KEY / _MODEL # Any field left unset falls back to the matching DEFAULT_MODELS_ value below, so a # single DEFAULT_MODELS_ triplet can supply the shared endpoint/key/model for all of # them. Only set the per-triplet fields that need to differ from the default. # When DEFAULT_MODELS_* is also unset, built-in defaults are chosen from the API # key prefix (sk- → inference-api.nvidia.com, nvapi- → integrate.api.nvidia.com). # --- Default triplet (fallback for every field of every triplet below) --- DEFAULT_MODELS_API_KEY=your-api-key-here # DEFAULT_MODELS_ENDPOINT=https://inference-api.nvidia.com/v1 # DEFAULT_MODELS_MODEL=azure/anthropic/claude-opus-5-5 # --- Agent LLM (the Auto Ontology agent: the benchmark's ontology arm) --- # REASONING_API_KEY falls back to DEFAULT_MODELS_API_KEY (then legacy NVIDIA_API_KEY). # REASONING_API_KEY=your-api-key-here # REASONING_ENDPOINT=https://inference-api.nvidia.com/v1 # REASONING_MODEL=azure/anthropic/claude-opus-5-5 # --- Control-arm model (the benchmark's schema-only arm, run by NeMo Gym) --- # Authenticates with DEFAULT_MODELS_API_KEY (or GYM_API_KEY, which then also # replaces it for the agent). Keep it the same model as REASONING_MODEL so the # delta measures the ontology rather than a model difference. # GYM_MODEL=azure/anthropic/claude-opus-5-5 # GYM_MODEL_URL=https://inference-api.nvidia.com/v1 # --- Embeddings (used by ingestion + the benchmark's ontology arm) --- # MUST be identical for ingest and query, otherwise stored vectors and query # vectors live in different spaces. # EMBED_API_KEY=your-api-key-here # EMBED_ENDPOINT=https://inference-api.nvidia.com/v1 # EMBED_MODEL=nvidia/nvidia/nemotron-3.5-embed-1b-nvfp4 # --- Non-reasoning model (e.g. entity extraction) --- # Has its own key/endpoint/model so it can point at a different endpoint than the # main reasoning model. ENDPOINT/MODEL fall back to the REASONING_* values. # NON_REASONING_API_KEY=your-api-key-here # NON_REASONING_ENDPOINT=https://inference-api.nvidia.com/v1 # NON_REASONING_MODEL=azure/anthropic/claude-opus-5-5 # --- Postgres store: Auto Ontology's catalog + pgvector collections (from its # docker-compose.yml) --- # POSTGRES_HOST=postgres — when running inside Docker (service name as host) POSTGRES_HOST=localhost POSTGRES_PORT=5432 POSTGRES_USER=postgres POSTGRES_PASSWORD=change-me POSTGRES_DATABASE=auto_ontology # --- Source database that ontology_sql_eval/ingestion/ingest.py extracts schema from --- # If you seeded a local DB via scripts/seed/seed_wwi.py, point this at it, e.g.: # CONNECTION_STRINGS=postgresql://postgres:password@localhost:5432/wideworldimporters # For BIRD Mini-Dev (SQLite), list the databases comma-separated: # CONNECTION_STRINGS=sqlite:////absolute/path/datasets/bird/dev/california_schools/california_schools.sqlite,sqlite:////absolute/path/datasets/bird/dev/card_games/card_games.sqlite # Examples: # CONNECTION_STRINGS=postgresql://user:password@host:5432/dbname CONNECTION_STRINGS=postgresql://user:password@host:5432/wideworldimporters # --- BIRD-Interact master switch --- # Set BIRD_INTERACT=true to turn on every flag this deployment has approved # for BIRD-Interact runs (INTERACTIVE, DB_PROBE_JSONB_PATH_CHECK, # DB_PROBE_JOIN_PATH_CHECK, DB_PROBE_PROACTIVE, DETECT_VACUOUS_GROUP_BY, # RELEVANCE_FILTER_INCLUDE_COLUMNS, HUB_SIBLING_EXPANSION_ENABLED, # HUB_SIBLING_CAP=6, TABLE_BRIDGE_RECONCILIATION_ENABLED) in one place, # instead of setting each individually. Applied by # ontology_sql_eval/env.py:load_env() via os.environ.setdefault(), so any of # these already set explicitly in either ontology-sql-eval or Auto Ontology # still wins over this master flag's defaults. ASK_OUTPUT_TYPE, # INJECT_CONDITIONAL_OUTPUT_HINT, and CLARIFY_MAX_DISTANCE are deliberately # not part of this bundle -- they already default to INTERACTIVE's value # when unset. # BIRD_INTERACT=true