Architecture

The interactive diagrams live on the Architecture page. Each one is drawn from the current code, and each has views that highlight one part of it.

Group	Diagram	What it shows
System	System overview	Every compose service and profile: clients, api, Triton, OpenSearch, the shared job volumes, the multi-project curation workers, segmenter, VLM endpoints, trainer, evaluator, MLflow and monitoring.
System	Backend modules	src/main.py, the core and curation routers (project-scoped), the FastAPI-free services under src/services/ (projects, config store, dataset import, reprocess, combine), the clients, src/config/, and the workers that import the same services.
System	Data model	The per-project curation indexes (images, items with their region_boxes, classes, UMAP state and the configs store), the global project and config indexes, the project's files, and the separate visual-search indexes.
Workflows	Core inference	/detect, /faces, /embed and /ocr: CPU preprocessing in the API, TensorRT models in Triton, and the ocr_pipeline BLS model.
Workflows	Curation pipeline	Create a project, ingest or import, embed, cluster, VLM suggest, human review, reprocess, export, train, bake-off, promote, and probe predictions back into review.
Workflows	Region cascade	The detection worker's legs (region detector, segmenter, OCR text hint) over a per-project profile, candidate selection up to max_regions_per_item, one numbered VLM verify per item, per-box verdicts and text, and the terminal statuses.
Sequences	Training and promote	The job-file protocol between the API and the trainer for one project, the GPU arbiter, and the gated promote into Triton.

The curation parts of these diagrams share three ideas:

• A project owns its indexes, files and settings. Requests bind one project through the route; workers visit every active project in turn. See Projects.
• The config store holds each project's prompt packs, region profiles, active VLM endpoint, keymap and settings, with revisions. Workers re-read it at quiesce points. The VLM endpoint registry is the one global store. See Settings and config store.
• A region is a list of boxes on an item (region_boxes), each with its own state, text, cluster and embedding. See Multi-box regions.

Related pages

• Introduction: the services and what each capability is backed by.
• Compose profiles: every service, which profile starts it, and how a frontend joins the network.
• Model export: the Triton models and how they are built.
• Workers and heartbeats: the curation workers and their shared job state.
• Training, bake-off and promote: the file protocol between the API, trainer and evaluator.

Regenerating the diagrams

The sources are hand-authored Archify specs (MIT) in docs-site/architecture-diagrams/specs/. After changing one, run scripts/generate-architecture-diagrams.sh from the repo root; it validates each spec at showcase quality and writes the HTML to docs-site/static/architecture/.