CLI Reference¶

Director-AI ships a command-line interface for scoring, serving, benchmarking, and project scaffolding.

pip install director-ai
director-ai --help

Commands¶

Scoring¶

# Score a single prompt/response pair
director-ai review "What is the capital of France?" "The capital is Berlin."

# Process with agent (generate + score)
director-ai process "What is the refund policy?"

# Batch score from JSONL
director-ai batch input.jsonl --output results.jsonl

Server¶

# Start REST server (default transport: http)
director-ai serve --port 8080 --workers 4

# Start gRPC server
director-ai serve --transport grpc --port 50051 --workers 4

# Health check (via curl, no dedicated CLI command)
curl http://localhost:8080/v1/health

Configuration¶

# Show current config
director-ai config

# Check runtime dependencies and model revision pins
director-ai doctor

# Show a named profile
director-ai config --profile medical

# Generate YAML config
director-ai config --export config.yaml

Project Scaffolding¶

# Create a new project with config, facts, and guard script
director-ai quickstart --profile medical
cd director_guard/
python guard.py

# Create and validate an authenticated production scaffold
director-ai quickstart --profile production
director-ai production-check --path director_guard
director-ai production-check --path director_guard --require-secrets

Benchmarking¶

# Run latency benchmark
director-ai bench

# Run with specific dataset
director-ai bench --dataset e2e

# Run regression suite
python -m benchmarks.regression_suite

Model Export¶

# Export to ONNX
director-ai export --format onnx --output ./models/onnx/

# Export to TensorRT
director-ai export --format tensorrt --output ./models/trt/

Guardrail Forensics¶

# Explain reviewed misses from tenant-safe eval records
director-ai forensics --input eval_records.json --format markdown

The input is either a JSON array of eval records or an object with a records array. It may include director.eval.* attributes from the eval-trace layer plus reviewer labels such as label: "hallucination" or label: "grounded".

Fine-Tuning¶

# Fine-tune NLI model on custom data
director-ai finetune train.jsonl --output ./models/custom/

Managed Training¶

Managed training submissions use one CLI contract across local, portable, and Vertex execution lanes. local runs on the current machine. portable emits a provider-neutral container job request for AWS, Azure, Slurm, Kubernetes, or other customer-owned orchestrators. vertex submits directly to Vertex AI when the managed-training extra and cloud credentials are installed.

# Local dry run
director-ai train submit \
  --backend local \
  --dataset-uri ./train.jsonl \
  --output-uri ./artifacts/customer-run-001 \
  --dry-run

# Portable external-orchestrator contract
director-ai train submit \
  --backend portable \
  --dataset-uri s3://customer-data/train.jsonl \
  --eval-uri azure://customer-data/eval.jsonl \
  --output-uri file:///mnt/customer-artifacts/director-ai/run-001 \
  --image registry.example.com/director-ai/train:2026-05 \
  --dry-run

# Vertex managed submission
director-ai train submit \
  --backend vertex \
  --dataset-uri gs://customer-data/train.jsonl \
  --eval-uri gs://customer-data/eval.jsonl \
  --output-uri gs://customer-artifacts/director-ai/run-001 \
  --project customer-project \
  --region europe-west4 \
  --image europe-west4-docker.pkg.dev/customer-project/director/train:2026-05

The portable backend is dry-run only by design. It redacts secret-looking environment variables in the emitted request and leaves live job lifecycle control to the customer's external orchestrator.

Threshold Tuning¶

# Adaptive threshold calibration on your dataset
director-ai tune eval_data.jsonl

Version¶

director-ai version
# director-ai 3.16.0

Global Options¶

Flag	Description
`--config PATH`	YAML config file
`--profile NAME`	Named profile (fast, thorough, medical, etc.)
`--verbose`	Enable debug logging
`--json`	JSON output format