The intended way for using Annif in production is by making the REST API or Web UI available for end-users, either publicly to everyone on internet or in a more private network. For this purpose one minimally needs to have a server and to configure it to expose Annif API and/or Web UI to the outside network.
Also, when training large Annif models, and especially optimizing them by searching for optimal hyperparameters,
it is best to have a separate computer for that with adequote resources.
From that training computer the ready-made models (the data/ directory and the projects configuration file)
can be transferred to the server where an Annif instance is run for the API and/or Web UI.
For a reference value, the memory consumption of the current (in 2022) Annif models run for Finto AI is ~25 GB (there are 12 projects). Note that the memory consumption of a model of a given backend/algorithm depends on the hyperparameters and the data the model is trained on. Usually there is a trade-off between the quality of the model's suggestions and its memory consumption.
Annif is based on Flask, which is a WSGI (Web Server Gateway Interface) application framework well suitable for development and debugging, but in production systems it is recommended to use a WSGI server in front.
Gunicorn is a good choice for the WSGI server.
To start Annif behind Gunicorn run e.g.
gunicorn "annif:create_app()" --bind 0.0.0.0:8000 --timeout 600 --worker-class uvicorn.workers.UvicornWorker.
This exposes port 8000, sets the timeout to 600 s to allow the Annif models to load in case they are large, and uses Uvicorn workers (necessitated by Annif v1.1).
In addition to a WSGI server, in front of all it is recommended to put a dedicated HTTP server to act as a reverse proxy. Gunicorn documentation advices to use NGINX. In addition to increasing performance the proxying HTTP server can serve static files for a web page (in the case you don't want to use the default Annif Web UI), perform TLS/SLL handling, redirections etc.
The platform to run the above components can be a traditional server, but nowdays a more common approach is to employ containers. Docker image of Annif is available on quay.io/natlibfi/annif repository. The image includes Gunicorn as well as all optional dependencies of Annif (see how to customize the Annif image for example to reduce the size of the image). NGINX is better to be run on a separete container.
There are many platforms where the containers can be run. The National Library of Finland is using Annif to power the Finto AI service, which is run on an OpenShift Container Platform environment. The Finto AI's GitHub repository contains code for the service.
See also:
- The video on YouTube demonstrating the use of Finto AI and its integration to the submission system of the publications archive of University of Vaasa.
- The article on Annif and Finto AI published on JLIS.it in 2022.
- The presentation at LIBER 2022 and the paper and presentation at IFLA 2022 by Anna Kasprzik about automating subject indexing process at ZBW.