Veil

Veil is a framework for building and running Named Entity Recognition & masking pipelines.

Why Veil?

Named Entity Recognition (NER) is the process of identifying and classifying named entities in text (names, organizations, locations, etc.). Although public models and techniques perform well, a production-ready NER system can be comprised of many components, not just a standalone model. Veil provides a framework to build and run production-tested NER pipelines with many different components. It offers primitives for defining a pipeline that may include:

• Multiple entity detectors that identify entities in the text, extracting the most out of many different NER techniques and models
• Entity resolvers that resolve entities to a single canonical form
• Overlap resolvers that resolve overlaps between entities
• Maskers that mask the entities in the text for privacy-focused use cases
• Evaluators that exhaustively measure the pipeline, both in quality of detection and system performance

Veil can be deployed in online mode as an API server, or in offline mode as a batch processor.

Getting Started

First, clone the repository:

git clone https://github.com/chus-chus/veil.git
cd veil

Environment setup (supports all included entity detectors)

First, install make and mamba if you don't have them already. Then, build the environment:

make build

activate the environment with:

mamba activate ./env

Note that you will need CUDA for GPU model execution. If not available, Veil will fall back to CPU execution.

Development packages

You may also need development requirements (build documentation, run tests, etc.). Inside the environment:

python -m uv pip install -r requirements_dev.txt

Documentation

You can extract the most out of Veil when you bring your own entity detectors. To learn how, you can read the documentation on Read the Docs.

You can also build the documentation by running:

make docs/html

and serve it locally with:

make docs/serve

which will start a local server at http://localhost:5500.

Run from the CLI

Veil is highly configurable. All configuration classes, defined in veil/config, have a 1–1 mapping with CLI parameters. You can see the available options with:

python -m veil --help

Run from a file

For example, create a configuration file like run_configs/example_offline.yml:

mode: offline
dataloader:
  path: data/input/example.jsonl
entity_detectors:
  - type: regex
    min_confidence: 0.3

And run:

python -m veil --pipeline-config-from-file run_configs/example_offline.yml

Input data must contain at least an input field with the text to process.

See docs/architecture.md for more details.

Docker

Run the Veil API with Docker

We provide a Docker image for a reproducible API deployment. See the configuration used in the image at run_configs/online_multi_detector.yml. You can also build the image yourself:

make docker/build

or pull it from Docker Hub:

docker pull docker-username/veil:gpu-latest

Then run:

docker run --gpus all -t -e HUGGINGFACE_HUB_TOKEN=hf_your_token -p 8000:8000 docker-username/veil:gpu-latest

This will start the API server on port 8000. See API details in veil/api_server.py.