PDF Annotations with Labels and Structure is software that makes it easy to collect a series of annotations associated with a PDF document. It was written specifically for annotating academic papers within the Semantic Scholar corpus, but can be used with any collection of PDF documents.
Quick start will download some pre-processed PDFs and get the UI set up so that you can see them. If you want to pre-process your own PDFs, keep reading! If it's your first time working with PAWLS, we recommend you try the quick start first though.
First, we need to download some processed PDFs to view in the UI. PAWLS uses the PDFs themselves to render in the browser, as well as using a JSON file of extracted token bounding boxes per page, called pdf_structure.json. The PAWLS CLI can be used to do this pre-processing, but for the quick start, we have done it for you. Download them from the provided AWS S3 Bucket like so:
aws s3 sync s3://ai2-s2-pawls-public/example-data ./skiff_files/apps/pawls/papers/ --no-sign-requestConfiguration in PAWLS is controlled by a JSON file, located in the api/config directory. The location that we downloaded the PDFs to above corresponds to the location in the config file, where it is mounted in using docker-compose.yaml. So, when PAWLS starts up, the API knows where to look to serve the PDFs we want.
Next, we can start the services required to use PAWLS using docker-compose:
~ docker-compose up --build
This process launches 4 services:
- the
ui, which renders the user interface that PAWLS uses - the
api, which serves PDFs and saves/recieves annotations - a
proxyresponsible for forwarding traffic to the appropriate services. - A
grobidservice, running a fork of Grobid. This is not actually necessary for the application, but is useful for the CLI.
You'll see output from each.
Once all of these have come up, navigate to localhost:8080 in your browser and you should see the PAWLS UI! Happy annotating.
In order to run a local environment, you'll need to use the PAWLS CLI to preprocess and assign the PDFs you want to serve. When using PDFs from semantic scholar, the CLI is also used to download the PDFs. The PDFs have to be put in a directory structure within skiff_files/apps/pawls (see PAWLS CLI usage for details).
For instance, you can run the following commands to download, preprocess, and assign PDFs:
# Fetches PDFs from semantic scholar's S3 buckets.
python scripts/ai2-internal/fetch_pdfs.py skiff_files/apps/pawls/papers 34f25a8704614163c4095b3ee2fc969b60de4698 3febb2bed8865945e7fddc99efd791887bb7e14f 553c58a05e25f794d24e8db8c2b8fdb9603e6a29
# ensure that the papers are pre-processed with grobid so that they have token information.
pawls preprocess grobid skiff_files/apps/pawls/papers
# Assign the development user to all the papers we've downloaded.
pawls assign skiff_files/apps/pawls/papers [email protected] --all --name-file skiff_files/apps/pawls/papers/name_mapping.jsonand then open up the UI locally by running docker-compose up.
Authentication is simply checking that users are who they say they are. Whether or not these users' requests are allowed (e.g., to view a PDFs) is considerd authorization. See more about this distinction at Skiff Login.
All requests must be authenticated.
- The production deployment of PAWLS uses Skiff Login to authenticate requests. New users are bounced to a Google login workflow, and redirected back to the site if they authenticate with Google. Authenticated requests carry an HTTP header that identifies the user.
- For local development, there is no login workflow. Instead, all requests are
supplemented with a hard-coded authentication header in
proxy/local.conf specifying that the user is
[email protected].
Look at the function get_user_from_header in main.py for
details.
Authorization is enforced by the PAWLS app. A file of allowed user email addresses is consulted on every request.
- In production, this file is sourced from the secret named
"users" in Marina, which is
projected to
/users/allowed.txtin the container. - For local development, this file is sourced from
allowed_users_local_development.txt,
and also projected to
/users/allowed.txtin the Docker container.
The format of the file is simply a list of allowed email addresses.
There's a special case when an allowed email address in this file starts with "@", meaning all users in that domain are allowed. That is, an entry "@allenai.org" will grant access to all AI2 people.
Look at the function user_is_allowed in main.py for details.
The Python service and Python cli are formatted using black and flake8. Currently this is run in a local environment
using the app's requirements.txt. To run the linters:
black api/
flake8 api/
Make sure that you have the latest version of Docker 🐳 installed on your local machine.
To start a version of the application locally for development purposes, run this command:
~ docker-compose up --build
This process launches 3 services, the ui, api and a proxy responsible
for forwarding traffic to the appropriate services. You'll see output
from each.
It might take a minute or two for the application to start, particularly if it's the first time you've executed this command. Be patience and wait for a clear message indicating that all of the required services have started up.
As you make changes the running application will be automatically updated. Simply refresh your browser to see them.
Sometimes one portion of your application will crash due to errors in the code.
When this occurs resolve the related issue and re-run docker-compose up --build
to start things back up.
The skiff template contains some features which are ideal for a robust web application, but might be un-intuitive for researchers. Below are some small technical points that might help you if you are making substantial changes to the skiff template.
-
Skiff uses
sonarto check that all parts of the application (frontend, backend) are up and running before serving requests. To do this, it checks that your api returns 2XX codes from its root url - if you change the server, you'll need to make sure to add code which returns a 2XX response from your server. -
To ease development/deployment differences, skiff uses a proxy to route different urls to different containers in your application. The TL;DR of this is the following:
| External URL | Internal URL | Container |
|---|---|---|
localhost:8080/* |
localhost:3000/* |
ui |
localhost:8080/api/* |
localhost:8000/* |
api |
So, in your web application, you would make a request, e.g axios.get("/api/route", data), which the server recieves at localhost:8000/route.
This makes it easy to develop without worrying about where apis will be hosted in production vs development, and also allows for things like
rate limiting. The configuration for the proxy lives here for development and here for production.
For example, if you wanted to expose the docs route localhost:8000/docs from your api container to users of your app in production, you would add this to prod.conf:
location /docs/ {
limit_req zone=api;
proxy_pass http://api:8000/;
}
- Add the package and version reqs in the
resolutionsfield from thepackage.jsonfile; - Run
yarn installto update theyarn.lockfile - Start the docker and test whether the UI still works
docker-compose up --build
The application was developed for Linux, and might fail to start on Windows because of line-ending differences.
To fix this, run this command from the root of the repository:
~ (cd ./ui && yarn && yarn lint:fix) # with parenthesis, to stay in same directoryIf you find PAWLS helpful for your research, please consider cite PAWLS.
@misc{neumann2021pawls,
title={PAWLS: PDF Annotation With Labels and Structure},
author={Mark Neumann and Zejiang Shen and Sam Skjonsberg},
year={2021},
eprint={2101.10281},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
PAWLS is an open-source project developed by the Allen Institute for Artificial Intelligence (AI2). AI2 is a non-profit institute with the mission to contribute to humanity through high-impact AI research and engineering.
![dependabot[bot] avatar](https://avatars.githubusercontent.com/in/29110?v=4 "dependabot[bot]")
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent