Internet Archive Reference Inventory IARI
This document is under development
This API is capable of fetching, extracting, transforming and storing reference information from Wikipedia articles as structured data.
The endpoints make it possible to get structured data about the references from any Wikipedia article in any language version.
IARI was developed as part of the Turn All References Blue project of the Wayback Machine, an integral part of the Internet Archive.
Mark Graham is the director of the project. The primary developer is currently Chris Lombardi (under the nom de silico "mojomonger"). Initial development was done by Dennis Priskorn.
There are at least 200 million references in at least 40 million articles and together with the article text in the Wikipedias one of the most valuable collections of knowledge ever made by humans, see size comparison.
Wikimedia movement currently does not have good and effective tools to help editors keep up the quality of references over time. The references are stored in templates that differ between language versions of Wikipedia which makes it hard for tool developers to develop good tools that work well across different language versions.
The endpoint providing a detailed analysis of a Wikipedia article and it's references enable wikipedians to get an overview of the state of the references and using the API it is possible for the Wikimedia tech-community to build tools that help make it easier to curate and improve the references.
This is part of a wider initiative help raise the quality of references in Wikipedia to enable everyone in the world to make decisions based on trustworthy knowledge that is derived from trustworthy sources.
This project is a part of the Wikicite initiative.
On the longer term Turn All References Blue project is planning on populating a database based on the data we extract. This part of the effort is led by James Hare.
The end goal is a large database with all references from all Wikipedias. We call it the Wikipedia Citations Database (WCD).
IARI features a number of endpoints that help patrons get structured data about references in a Wikipedia article:
- an article endpoint which analyzes a given article and returns basic statistics about it
- a references endpoint which gives back all ids of references found
- a reference endpoint which gives back all details about a reference including templates and wikitext
- a check-url endpoint which looks up the URL and gives back standardized information about its status
- a check-doi endpoint which looks up the DOI and gives back standardized information about it from FatCat, OpenAlex and Wikidata including abstract, retracted status, and more.
- a pdf endpoint which extracts links both from annotations and free text from PDFs.
- a xhtml endpoint which extracts links both from any XHTML-page.
See known limitations under each endpoint below.
Currently we support a handful of the 200+ language versions of Wikipedia but we plan on extending the support to all Wikipedia language versions and you can help us by submitting sections to search for references in issues and pull requests.
We also would like to support non-Wikimedia wikis using MediaWiki in the future and perhaps also any webpage on the internet with outlinks (e.g. news articles).
English Wikipedia for example has hundreds of special reference templates in use and a handful of widely used generic templates. WARI exposes them all when found in a reference.
We support detecting the following types. A reference cannot have multiple types. We distinguish between two main types of references:
- named reference - type of reference with only a name and no content e.g. '' (Unsupported beyond counting because we have not decided if they contain any value)
- content reference - reference with any type of content
- general reference - subtype of content reference which is outside a and usually found in
sections called "Further reading" or "Bibliography"
(unsupported but we want to support it, see
https://github.com/internetarchive/wcdimportbot/labels/general%20reference)
- footnote reference - subtype of content reference which is inside a (supported partially, see below)
- general reference - subtype of content reference which is outside a and usually found in
sections called "Further reading" or "Bibliography"
(unsupported but we want to support it, see
https://github.com/internetarchive/wcdimportbot/labels/general%20reference)
Example of a URL-template reference:
<ref>Mueller Report, p12 {{url|http://example.com}} {{bare url inline}}</ref>
This is a content reference -> a footnote reference -> a mixed reference with a URL template.
Example of an plain text reference:
<ref>Muller Report, p12</ref>
This is a footnote reference -> content reference -> Short citation reference aka naked named footnote.
This endpoint requires an authkey for the testdeadlink API from Internet Archive to work. It looks for the environment variable TESTDEADLINK_KEY.
the check-url endpoint accepts the following parameters:
- url (string, mandatory)
- refresh (boolean, optional)
- testing (boolean, optional)
- timeout (int, optional)
- debug (boolean, optional)
On error it returns 400.
It returns json similar to
{
"first_level_domain": "github.com",
"fld_is_ip": false,
"url": "https://github.com/internetarchive/iari/issues",
"scheme": "https",
"netloc": "github.com",
"tld": "com",
"malformed_url": false,
"malformed_url_details": null,
"archived_url": "",
"wayback_machine_timestamp": "",
"is_valid": true,
"request_error": false,
"request_error_details": "",
"dns_record_found": true,
"dns_no_answer": false,
"dns_error": false,
"status_code": 200,
"testdeadlink_status_code": 200,
"timeout": 2,
"dns_error_details": "",
"response_headers": {
"Server": "GitHub.com",
"Date": "Mon, 12 Jun 2023 12:00:28 GMT",
"Content-Type": "text/html; charset=utf-8",
"Vary": "X-PJAX, X-PJAX-Container, Turbo-Visit, Turbo-Frame, Accept-Encoding, Accept, X-Requested-With",
"ETag": "W/\"03d14b06291209737bb700da5adf5968\"",
"Cache-Control": "max-age=0, private, must-revalidate",
"Strict-Transport-Security": "max-age=31536000; includeSubdomains; preload",
"X-Frame-Options": "deny",
"X-Content-Type-Options": "nosniff",
"X-XSS-Protection": "0",
"Referrer-Policy": "no-referrer-when-downgrade",
"Content-Security-Policy": "default-src 'none'; base-uri 'self'; block-all-mixed-content; child-src github.com/assets-cdn/worker/ gist.github.com/assets-cdn/worker/; connect-src 'self' uploads.github.com objects-origin.githubusercontent.com www.githubstatus.com collector.github.com raw.githubusercontent.com api.github.com github-cloud.s3.amazonaws.com github-production-repository-file-5c1aeb.s3.amazonaws.com github-production-upload-manifest-file-7fdce7.s3.amazonaws.com github-production-user-asset-6210df.s3.amazonaws.com cdn.optimizely.com logx.optimizely.com/v1/events *.actions.githubusercontent.com productionresultssa0.blob.core.windows.net/ productionresultssa1.blob.core.windows.net/ productionresultssa2.blob.core.windows.net/ productionresultssa3.blob.core.windows.net/ productionresultssa4.blob.core.windows.net/ wss://*.actions.githubusercontent.com github-production-repository-image-32fea6.s3.amazonaws.com github-production-release-asset-2e65be.s3.amazonaws.com insights.github.com wss://alive.github.com; font-src github.githubassets.com; form-action 'self' github.com gist.github.com objects-origin.githubusercontent.com; frame-ancestors 'none'; frame-src viewscreen.githubusercontent.com notebooks.githubusercontent.com; img-src 'self' data: github.githubassets.com media.githubusercontent.com camo.githubusercontent.com identicons.github.com avatars.githubusercontent.com github-cloud.s3.amazonaws.com objects.githubusercontent.com objects-origin.githubusercontent.com secured-user-images.githubusercontent.com/ user-images.githubusercontent.com/ private-user-images.githubusercontent.com opengraph.githubassets.com github-production-user-asset-6210df.s3.amazonaws.com customer-stories-feed.github.com spotlights-feed.github.com *.githubusercontent.com; manifest-src 'self'; media-src github.com user-images.githubusercontent.com/ secured-user-images.githubusercontent.com/ private-user-images.githubusercontent.com; script-src github.githubassets.com; style-src 'unsafe-inline' github.githubassets.com; worker-src github.com/assets-cdn/worker/ gist.github.com/assets-cdn/worker/",
"Content-Encoding": "gzip",
"Set-Cookie": "_gh_sess=Kt%2FbTAsPUlbQMLzfnhA6DaZ%2FrXJio0ITQdf7cs83MUTpiSwa0ALcco5FkiN3Gne6MBf%2Fzf001SNsFAAt1IXJVU9kFP5EOV1P1UZsps9JI2dyn9lAG0Zt%2FKocfikDwnYU03uuJAmDpI82mOt6L0ULDzxEm1hc0vCMRy72T8saF%2Ba8qoSFPJR296%2FjZVoqrg9cHzAnYr5uo0cj9dQL6pGfoDreWMMV41kG7S%2FbwvU5%2FTWkJnmYTK8XoioOtrVjnvQ%2Fw%2FNuMVh4dkEEeprnmfe8%2Bw%3D%3D--CpOcS3YATJ%2FQZh%2Fk--zuEd5QQlr0A8Pif7t41ywQ%3D%3D; Path=/; HttpOnly; Secure; SameSite=Lax, _octo=GH1.1.1345489649.1686571228; Path=/; Domain=github.com; Expires=Wed, 12 Jun 2024 12:00:28 GMT; Secure; SameSite=Lax, logged_in=no; Path=/; Domain=github.com; Expires=Wed, 12 Jun 2024 12:00:28 GMT; HttpOnly; Secure; SameSite=Lax",
"Accept-Ranges": "bytes",
"Transfer-Encoding": "chunked",
"X-GitHub-Request-Id": "4C01:6326:116FFBB:1199F4E:648708DB"
},
"detected_language": "en",
"detected_language_error": false,
"detected_language_error_details": "",
"timestamp": 1686564033,
"isodate": "2023-06-12T12:00:33.941494",
"id": "7e8230ce"
}
When the debug parameter is true the text from the resource is returned. This text is the basis for the language detection.
You are very welcome to suggest improvements by opening an issue or sending a pull request. :)
The status_code attribute is not as reliable as the testdeadlink_status_code.
Sometimes we get back a 403 because an intermediary like Cloudflare detected that we are not a person behind a browser doing the request. We don't have any ways to detect these soft200s.
Also sometimes a server returns status code 200 but the content is an error page or any other content than what the person asking for the information wants. These are called soft404s and we currently do not make any effort to detect them.
We need at least 200 characters to be able to reliably detect the language.
the statistics/article endpoint accepts the following parameters:
- url (mandatory)
- refresh (bool, optional)
- testing (bool, optional)
- revision (int, optional, defaults to the most recent)
- dehydrate (bool, optional, defaults to True)
- regex (string, mandatory)
On error it returns 400. On timeout it returns 504 or 502 (this is a bug and should be reported).
Regex is a string of section headers separated by "|" to analyze for general references. For enwiki a string like "bibliography|further%20reading|works%20cited|sources|external%20links" catches most references outside -tags.
Dehydrate means it returns a limited reference object. Currently the keys "templates" and "url_objects" are being removed to make sure we don't return a huge blob of json by default.
If you set dehydrate to False the whole reference will be returned. For an article like https://en.wikipedia.org/wiki/United_States with >600 references 1,3 MB of JSON is returned.
It will return json like this (by default with dehydrated references) and like this with dehydrate set to false.
- the general references parsing relies on 2 things:
- a manually supplied list of sections to search using the 'regex' to the article and all endpoints. The list is case insensitive and should be delimited by the '|' character.
- that every line with a general reference begins with a star character (*)
- URLs in comments are not currently supported. Comments begin with .
Any line that doesn't begin with a star is ignored when parsing lines in the sections specified by the regex.
the statistics/references endpoint accepts the following parameters:
- wari_id (mandatory, str) (this is obtained from the article endpoint)
- all (optional, boolean) (default: false)
- offset (optional, int) (default: 0)
- chunk_size (optional, int) (default: 10)
On error it returns 400. If data is not found it returns 404.
It will return json similar to:
{
"total": 31,
"references": [
{
"id": "cfa8b438",
"template_names": [],
"wikitext": "<ref name = \"Hartmann1\" />",
"type": "footnote",
"footnote_subtype": "named",
"flds": [],
"urls": [],
"templates": [],
"titles": [],
"section": "History",
"served_from_cache": true
},
]
}
None
the statistics/reference/id endpoint accepts the following parameters:
- id (mandatory) (this is unique for each reference and is obtained from the article or references endpoint)
On error it returns 400. If data is not found it returns 404.
It will return json similar to:
{
"id": "cfa8b438",
"template_names": [],
"wikitext": "<ref name = \"Hartmann1\" />",
"type": "footnote",
"footnote_subtype": "named",
"flds": [],
"urls": [],
"templates": [],
"titles": [],
"section": "History",
"served_from_cache": true
}
None
the statistics/pdf endpoint accepts the following parameters:
- url (mandatory)
- refresh (bool, optional, default false)
- testing (bool, optional, default false)
- timeout (int, optional, default 2 (seconds))
- debug (bool, optional) Note: you have to set refresh=true to be sure to get debug output.
- html (bool, optional, default false)
- xml (bool, optional, default false)
- json_ (bool, optional, default false)
- blocks (bool, optional, default false)
On error it returns 404 or 415. The first is when we could not find/fetch the url and the second is when it is not a valid PDF.
If not given debug=true it will return json similar this
{
"words_mean": 660,
"words_max": 836,
"words_min": 474,
"annotation_links": [
{
"url": "http://arxiv.org/abs/2210.02667v1",
"page": 0
},...
],
"links_from_original_text": [
{
"url": "https://www.un.org/en/about-us/universal-declaration-of-human-rights",
"page": 1
},
],
"links_from_text_without_linebreaks": [
{
"url": "https://www.un.org/en/about-us/universal-declaration-of-human-rights3https://www.ohchr.org/en/issues/pages/whatarehumanrights.aspx4Legal",
"page": 1
},
],
"links_from_text_without_spaces": [
{
"url": "https://www.un.org/en/about-us/universal-declaration-of-human-rights",
"page": 1
},
],
"url": "https://arxiv.org/pdf/2210.02667.pdf",
"timeout": 2,
"pages_total": 17,
"detected_language": "en",
"detected_language_error": false,
"detected_language_error_details": "",
"characters": 74196,
"timestamp": 1686055925,
"isodate": "2023-06-06T14:52:05.228754",
"id": "639dc1df"
}
Using the debug and refresh parameter, all the text before and after cleaning is exposed as well as the link-annotations before url extraction. Using the 4 additional parameters you can get the corresponding output from PyMuPDF for further debugging.
Example debug expressions:
To show block debug output:
https://archive.org/services/context/iari/v2/statistics/pdf?url=https://ia601600.us.archive.org/31/items/Book_URLs/DeSantis.pdf&refresh=true&debug=true&blocks=true
To show html debug output:
https://archive.org/services/context/iari/v2/statistics/pdf?url=https://ia601600.us.archive.org/31/items/Book_URLs/DeSantis.pdf&refresh=true&debug=true&html=true
Note: Setting debug=true parameter without refresh=true will often not yield any debug output since we don't have it stored in the cache.
Warning: The debug outputs can generate very large output up to hundreds of MB so use with care or from the command line to avoid crashing your browser
This output permits the data consumer to count number of links per page, which links or domains appear most, number of characters in the pdf and using the debug outputs you can investigate why the API could not find the links correctly.
The extraction of URLs from unstructured text in any random PDF is a difficult thing to do reliably. This is because PDF is not a structured data format with clear boundaries between different type of information.
We currently use PyMuPDF to extract all the text as it appears on the page. We clean the text in various ways to get the links out via a regex. This approach results in some links containing information that was part of the text after the link ended but we have yet to find a way to reliably determine the end boundary of links.
We are currently investigating if using the html output of PyMuPdf or using advanced Machine Learning could improve the output.
You are very welcome to suggest improvements by opening an issue or sending a pull request. :)
the statistics/pdf endpoint accepts the following parameters:
- url (mandatory)
- refresh (optional)
- testing (optional)
On error it returns 400.
It will return json similar to:
{
"links": [
{
"context": "<a accesskey=\"t\" href=\"http://www.hut.fi/u/hsivonen/test/xhtml-suite/\" title=\"The main page of this test suite\">This is a link to the <acronym title=\"Extensible HyperText Markup Language\">XHTML</acronym> test suite table of contents.</a>",
"parent": "<p><a accesskey=\"t\" href=\"http://www.hut.fi/u/hsivonen/test/xhtml-suite/\" title=\"The main page of this test suite\">This is a link to the <acronym title=\"Extensible HyperText Markup Language\">XHTML</acronym> test suite table of contents.</a> The link contain inline markup (<code><acronym></code>), has a title and an accesskey \u2018t\u2019.</p>",
"title": "The main page of this test suite",
"href": "http://www.hut.fi/u/hsivonen/test/xhtml-suite/"
},
{
"context": "<a href=\"base-target\">This is a relative link.</a>",
"parent": "<p><a href=\"base-target\">This is a relative link.</a> If the link points to http://www.hut.fi/u/hsivonen/test/base-test/base-target, the user agent supports <code><base/></code>. if it points to http://www.hut.fi/u/hsivonen/test/xhtml-suite/base-target, the user agent has failed the test.</p>",
"title": "",
"href": "base-target"
}
],
"links_total": 2,
"timestamp": 1682497512,
"isodate": "2023-04-26T10:25:12.798840",
"id": "fc5aa88d",
"refreshed_now": false
}
None
Clone the git repo:
$ git clone https://github.com/internetarchive/iari.git && cd iari
We recommend checking out the latest release before proceeding.
- python pip
- python gunicorn
- python poetry
We use pip and poetry to set everything up.
$ pip install poetry gunicorn && poetry install
Setup:
$ python3 -m venv venv
Enter in (in a Linux/Mac terminal):
$ source venv/bin/activate
Lastly setup the directories for the json cache files
$ ./setup_json_directories.sh
Run these commands in different shells or in GNU screen.
Before running them please make sure you have activated the virtual environment first, see above.
Start GNU screen (if you want to have a persisting session)
$ screen -D -RR
Run it with
$ export TESTDEADLINK_KEY="your key here" && ./run-debug-api.sh
Test it in another Screen window or local terminal with
$ curl -i "localhost:5000/v2/statistics/article?regex=external%20links&url=https://en.wikipedia.org/wiki/Test"
Run it with
$ export TESTDEADLINK_KEY="your key here" && ./run-api.sh
Test it in another Screen window or local terminal with
$ curl -i "localhost:8000/v2/statistics/article?regex=external%20links&url=https://en.wikipedia.org/wiki/Test"
Make sure the following directories are created at the root of the iari install (they are not in the git repo)
json/articles/
json/references/
json/dois/
json/urls/
json/urls/archives
json/xhtmls/
json/pdfs/
# (for new V2 development)
json/articlesV2/
json/referencesV2/
Make an ".env" file at same level as docker-compose.yml file with the contents:
PORT=<port you'd like the IARI Server to run on>
# some suggested values:
# 5042: for iari-test
# 5088: for iari-stage
# 5000: for iari-prod
TESTDEADLINK_KEY=<api key for iabot's testdeadlink method>
docker compose up
Make sure this setting is checked. https://www.google.com/search?client=firefox-b-d&q=make+pycharm+automatically+enter+the+venv
Note for mac users: This might not work, so you have to manually enter the venv in all terminals.
See KNOWN_DEPLOYED_INSTANCES.md
- version 1.0.0 a proof of concept import tool based on WikidataIntegrator (by James Hare)
- version 2.0.0+ a scalable ETL-framework with an API and capability of reading EventStreams (by Dennis Priskorn)
- version 3.0.0+ WARI, a host of API endpoints that returns statistics about a Wikipedia article and its references. (by Dennis Priskorn)
- version 4.0.0+ IARI, a host of API endpoints that returns statistics about both Wikipedia articles and its references and can extract links from any PDF or XHTML page. (by Dennis Priskor
This project is licensed under GPLv3+. Copyright Dennis Priskorn 2022 The diagram PNG files are CC0.
See the development notes