How nvidia-nemo/nemo-helix works
NeMo Helix is a modular control plane for evaluating, improving, and hardening production agents. Users enter through the CLI, Python SDK/REST APIs, or NeMo Studio; the CLI supports setup and operations, while platform services and plugins provide agent execution, data generation, evaluation, optimization, and safety capabilities. Shared foundations include authentication, files, jobs, and inference. The README documents the product surfaces and capability breadth; sampled source confirms selected client, controller, and storage behavior, but not most runtime wiring, so unverified links are dashed.
Main entry points: plan_storage.py Evaluation workflows · main.py Platform services · inference_gateway/client.py Inference gateway · jobs/client.py Jobs and orchestration
Components 20
User surfaces
- NeMo CLI
packages/nemo_helix_ext/src/nemo_helix_ext/cli/app.py - Python and REST
packages/nemo_helix_plugin/src/nemo_helix_plugin/client/client.py - NeMo Studio
web/packages/studio
Platform core
- Platform services
packages/nhx_platform/src/nhx/platform/main.py - Plugin framework
packages/nemo_helix_plugin/src/nemo_helix_plugin/controller.py
Agent operations
- Agent execution
plugins/nemo-agents/src/nemo_agents_plugin/runner/controller.py - Traces and insights
plugins/nemo-compass/src/nemo_compass_plugin/analysis/pipeline.py - Safety and guardrails
services/guardrails/src/nhx/guardrails/app/services/rails/engine.py
Agent improvement
- Evaluation workflows
plugins/nemo-evals/src/nemo_evals/taskset_runs/plan_storage.py - Scaled evaluations
plugins/_temporary-scaled-evals/src/nemo_scaled_evals_plugin/controller.py - Data Designer
plugins/nemo-data-designer/src/nemo_data_designer_plugin/sdk/_engine_pass.py - Optimization
plugins/nemo-optimization/src/nemo_optimization/router.py - Model fine-tuning
plugins/nemo-customizer/src/nemo_customizer/router.py - Data anonymization
plugins/nemo-anonymizer/src/nemo_anonymizer_plugin/functions/_preview_worker.py
Shared foundations
- Inference gateway
packages/nemo_helix_plugin/src/nemo_helix_plugin/inference_gateway/client.py - Authentication and policy
services/core/auth/src/nhx/core/auth/app/embedded_pdp/engine.py - Files and artifacts
services/core/files/src/nhx/core/files/app/log_storage.py - Jobs and orchestration
packages/nemo_helix_plugin/src/nemo_helix_plugin/jobs/client.py - Trace ingestion
services/intake/src/nhx/intake/spans/storage.py
Other components
- Helix user
Connections 23
- Helix user → NeMo CLI: operates
README.md - Helix user → Python and REST: integrates through
README.md - Helix user → NeMo Studio: uses
README.md - NeMo CLI → Platform services: manages services
- Python and REST → Platform services: calls APIs
- NeMo Studio → Platform services: accesses capabilities
- Platform services → Plugin framework: hosts plugins
- Plugin framework → Agent execution: runs controller
packages/nemo_helix_plugin/src/nemo_helix_plugin/controller.py - Agent execution → Jobs and orchestration: manages deployments
- Agent execution → Inference gateway: runs agent models
- Trace ingestion → Traces and insights: supplies traces
- Traces and insights → Evaluation workflows: informs evaluations
- Evaluation workflows → Jobs and orchestration: schedules runs
- Evaluation workflows → Files and artifacts: stores plans
plugins/nemo-evals/src/nemo_evals/taskset_runs/plan_storage.py - Scaled evaluations → Jobs and orchestration: reconciles jobs
plugins/_temporary-scaled-evals/src/nemo_scaled_evals_plugin/controller.py - Data Designer → Inference gateway: resolves providers
- Data Designer → Files and artifacts: produces datasets
- Optimization → Evaluation workflows: compares candidates
- Model fine-tuning → Jobs and orchestration: launches tuning
- Safety and guardrails → Inference gateway: guards model calls
- Data anonymization → Files and artifacts: protects data
- Python and REST → Authentication and policy: carries identity
- Evaluation workflows → Inference gateway: scores candidates