Documentation

Cortex

MetaLoom // Cortex is the media processing engine of the MetaLoom platform. It analyses media files — images, videos, audio, documents — and extracts metadata such as hashes, thumbnails, fingerprints, face embeddings, transcripts, captions and OCR text.

Cortex is a Daemon that Serves Nodes

Cortex runs as a long-running worker daemon, not a one-shot CLI batch tool. On startup it opens a persistent WebSocket to a Loom server, registers itself and the node kinds it can run, and then waits for work. Loom pushes tasks down that connection; Cortex never polls.

Crucially, the pipeline graph lives on Loom, not on Cortex. Loom owns the DAG and dispatches individual tasks — a source task to enumerate media, then one node task per graph node per item — to whichever registered Cortex instance is allowed to run that kind. Cortex only ever sees one node (or affinity segment) at a time and answers with a result. See Pipeline Mechanism and Operation.

This lets you scale processing horizontally: run many Cortex instances, each advertising a subset of node kinds (e.g. a GPU box that only runs facedetect and whisper), and Loom routes tasks accordingly.

Components

  • Pipeline Nodes — The built-in processing nodes and what they extract. See Nodes.

  • Configuration — Environment variables and settings. See Configuration.

  • Monitoring API — Health and readiness endpoints. See Monitoring API.

  • Metrics — The Prometheus scrape endpoint and the cortex_* catalog. See Metrics.

  • Artifacts — Maven coordinates and the container image. See Artifacts.

  • Examples — Custom node, custom daemon (Java) and custom worker (Python). See Examples.