Sections 1–5 explain how Atlas selects and configures services. Sections 6–8 describe shared Backend APIs for application developers.

Each configurable service has a SOURCE variable. It selects one mode: run in Docker, connect to a localhost instance, disable, or a service-specific mode. SOURCE Configuration lists the values for each service.

Tracks select the subset of services needed for a workflow and force-disable out-of-track services. An explicit --<svc>-source flag or a SOURCE value in a consumer manifest's env.values overrides the track for that service. Tracks lists the services in each track.

Each manifest owns service metadata, env vars, source options, dependencies, runtime slices, and data-flow calls.

Kong provides the main local entrypoint and generated aliases. Direct ports remain available for services that expose their own UI or API.

Atlas starts from .env.example, writes or preserves the active .env, then merges user-owned overlays before backfilling missing keys and applying CLI flags. The sibling .env.user file is useful for local checkout-owned values. ATLAS_ENV_USER_FILE points at a parent-owned overlay outside the Atlas checkout and is the preferred submodule-consumer pattern.

Overlay precedence is .env.example baseline, generated or existing .env, sibling .env.user, ATLAS_ENV_USER_FILE, consumer manifest env values, then explicit flags such as --project and --<svc>-source. Both overlays are merged on every start, including --cold. Relative ATLAS_ENV_USER_FILE values resolve against the directory that invoked start.sh.

The backend's provider-neutral media API is POST /media/generate, GET /media/operations/{operation_id} and POST /media/operations/{operation_id}/cancel. Requests choose provider, modality and model:

• provider=fal: modality=image and modality=image_to_3d (verified TRELLIS, Hunyuan3D, Tripo and Rodin endpoints).
• provider=comfyui: modality=image, on the managed or local ComfyUI host.

Provider API keys stay in the backend. Responses normalize status, artifacts, cost, license and provenance. The artifact_url depends on the provider. For fal it is absolute (a hosted CDN URL). For comfyui it is gateway-relative: /media/operations/{operation_id}/artifacts/{index}, an owner-checked backend path. Consumers MUST resolve a relative artifact_url (one that starts with / and has no http(s):// scheme) against their gateway or backend base URL.

Cancellation keeps reserved spend until provider polling proves a terminal outcome. If FAL times out before it returns a request id, the operation becomes submission_unknown with its reservation held. After checking provider billing, an operator with BACKEND_INTERNAL_API_TOKEN calls POST /media/operations/{operation_id}/reconcile with outcome=commit|release. This call is safe to retry after a transient ledger failure.

MEDIA_BUDGET_STORE also backs recovery when the operation record could not be written, even with budget enforcement off. Keep the default postgres, because memory is lost when the process restarts.

The backend exposes POST /api/chunk as the shared Chonkie-powered text-splitting surface for RAG ingestion clients. The endpoint supports token, recursive and semantic strategies. It returns stable character offsets and strategy metadata, so n8n workflows, notebooks and future ingestion services share one chunking contract.

JupyterHub also installs Chonkie for exploratory notebook work, including 13_chonkie_chunking.ipynb. Production workflows should still call the Backend endpoint instead of each service adding its own Chonkie dependency.

The backend exposes POST /api/rag/evaluate as the shared Ragas-powered quality-evaluation surface for supplied RAG question, answer, context, and optional reference records. The endpoint supports faithfulness, answer relevancy, context precision, and context recall metrics while routing evaluator calls through Atlas LiteLLM configuration.

JupyterHub also installs Ragas for exploratory evaluation work, including 14_ragas_evaluation.ipynb. Production workflows should call the Backend endpoint. Then n8n, notebooks and future ingestion jobs share one metric contract, and no service carries its own evaluator package.