Simple Flask REST service for English question answering. Leverages the transformer API and a RoBERTa model with question answering-head pretrained by deepset on the SQuAD2.0 dataset (learn more here). Runs on CPU only.
This service can be either set up in a Python 3.8 environment or build as Docker image.
Install all necessary requirements in your Python environment via:
pip install -r requirements.txt
To start the server run:
python server.py -p <PORT>
With the optional parameter -p the port can be changed from the default port 5000.
To build the docker image run the following command:
docker build -t qaservice:final .
To start a container with the previously built image run the following command or use your favorite framework for deployment.
docker run -d -p <PORT>:5000 --name qaservice qaservice:final
Other than in the [Python Environment](#Python Environment) the parameter -p is mandatory since the
internal port has to be mapped to a port of the host.
The service exposes a single REST endpoint: POST /inference which expects a JSON object of the following
format
{
impossible: boolean [optional, default: True]
top_k: integer [optional, default: 1]
data: [
{
questions: [
'question1',
'question2',
...
]
context: 'context'
},
...
]
}
The parameter impossible can be set to False if all questions can be answered with the given
context and the model should be forced to return a valid answer. The parameter top_k can be set to an
arbitrary value k >= 1 to get the best k answers for each question.
The questions and contexts are provided under the parameter data in a nested form in which all questions
are grouped by the corresponding context.
Example requests can be found in the folder requests and can be sent with python requests.py -p
. The parameter -p has to be set if the port on the host differs from the default port 5000.
Responses will be logged in an automatically created build folder.