intextum-worker

The Intextum processing worker: an HTTP-polling worker that pulls tasks from an Intextum API instance and runs the Docling / FFmpeg document, image and audio pipeline (OCR, ASR, chunking, classification, content enrichment, embeddings).

The worker is always-remote: it downloads source files from and uploads results to the API over HTTP, so it does not need a shared volume and can run anywhere — including on a host with a GPU while the rest of the stack runs in Docker.

Install

Pick the bundle that matches your accelerator. The macOS (Apple MPS) wheels are on PyPI, so it installs with no extra flags:

pip install 'intextum-worker[mps]'

Linux CPU and NVIDIA CUDA pull their Torch build from the PyTorch index, so add the matching --extra-index-url:

# Linux, CPU only
pip install 'intextum-worker[cpu]'  --extra-index-url https://download.pytorch.org/whl/cpu

# Linux, NVIDIA CUDA 12.6
pip install 'intextum-worker[cuda]' --extra-index-url https://download.pytorch.org/whl/cu126

Available extras: mps, cpu, cuda, cpu-document (document/image only), plus the granular document, asr, enrichment stacks.

Run

export API_URL="https://your-intextum-host"   # the API to poll
export WORKER_TOKEN="<token from the Add Worker dialog>"
intextum-worker --capabilities document,video,image

intextum-worker --help lists all flags. Every flag also has an environment variable (API_URL, WORKER_TOKEN, WORK_DIR, CAPABILITIES, POLL_INTERVAL, CLASSIFICATION_DEVICE, DOCLING_OCR_ENGINE, …); CLI flags take precedence.

Development

This package uses a src/ layout. The repo-root VERSION file is the single source of truth for the version; it is staged into worker/VERSION at build time (worker/VERSION is gitignored).

cp ../VERSION VERSION           # stage the version for an editable install
pip install -e '.[mps,test]'    # or [cpu,test] / [cuda,test]
pytest

export API_URL="http://localhost:8000"
export WORKER_TOKEN="<token from the Add Worker dialog>"
export CLASSIFICATION_DEVICE="mps"   # macOS Apple Silicon; use cpu/cuda elsewhere
intextum-worker --capabilities document,video,image,training