Marble Name Entity Recognition API

This repository powers Marble's Name Entity Recognition (NER) used within the Sanctions Check feature.

Build and run

$ docker build -t marble/marble-ner:latest -f Dockerfile.gpu . # Or Dockerfile.gpu
$ docker run --rm -p 9000:9000 -e GLINER_LABELS=Person,Company,Country marble/marble-ner:latest

You can instruct the model to try and label specific names from your input by specifying a comma-separated list of labels in the GLINER_LABELS environment variable. By default, the used labels are Person and Company.

Query

$ curl -XPOST 127.0.0.1:9000/detect -H content-type:application/json -d \
  '{"text":"flying to canada to have dinner at acme\'s headquarters with joe"}'
[
  {
    "type": "Country",
    "text": "canada"
  },
  {
    "type": "Company",
    "text": "acme"
  },
  {
    "type": "Person",
    "text": "joe"
  }
]

Develop

Before pulling the dependencies, you need to rename/symlink the proper pyproject.toml and poetry.lock for your platform, mostly whether or not you are running on a GPU. A Makefile is provided to help with this.

Note that the CPU built wheel of pytorch is not provided for macOS. If running that platform, you need to be using the gpu build. The Docker build requires a TARGET=<cpu|gpu> build argument to target the proper build (defaults to CPU).

$ make prepare TARGET=cpu # Or TARGET=gpu
$ poetry install
# Either run it directly with uvicorn
$ NER_API_KEY=apikey make run

In order to build the Docker image from macOS, you either have to use --build-arg TARGET=gpu or --platform linux/x86_64.

Configure

You can configure which labels the Name Entity Recognition will try and detect from the input string. By default, it is set to Person and Company, but you can adjust it by setting a comma-separated list of labels in GLINER_LABELS. For example:

$ NER_API_KEY=apikey GLINER_LABELS='Person,Company,Country,Illegal Activity' make run
$ http POST :9000/detect text='Dinner with Martin Finnigan at Moneycorp, flying to Germany for some light money laundering and have a meeting with Bob Gloss' --bearer apikey
[
    {
        "type": "Person",
        "text": "Martin Finnigan"
    },
    {
        "type": "Company",
        "text": "Moneycorp"
    },
    {
        "type": "Country",
        "text": "Germany"
    },
    {
        "type": "Illegal Activity",
        "text": "money laundering"
    },
    {
        "type": "Person",
        "text": "Bob Gloss"
    }
]