Wikimedia Enterprise

Group:	Advancement
Start:	2021
Team members:	See list on Meta
Backlog:	#wikimedia_enterprise

Translate this page

The Wikimedia Enterprise API is a new service focused on high-volume commercial reusers of Wikimedia content. It will provide a new funding stream for the Wikimedia movement; greater reliability for commercial reusers; and greater reach for Wikimedia content.

For general information, the relationship to the Wikimedia strategy, operating principles, and FAQ, see Wikimedia Enterprise on Meta. The project was formerly known as "Okapi".

See also our website for up-to-date API documentation. Current development work is tracked on our Phabricator board. Our source code is on Github. For information about Wikimedia community access to this service, please see Access on the project's Meta homepage.

Contact the team if you would like to arrange a conversation about this project with your community.

Updates

This is the most recent months of technical updates. All previous updates can be found at the archive.

2024 - Q2

____________

Machine Readability

Goal: To include structured data into our feeds and to make unstructured Wikimedia content available in pre-parsed formats
Launches:
- Structured Contents snapshots: early beta release of Structured Contents Snapshots endpoint, including pre-parsed articles (abstracts, main images, descriptions, infoboxes, sections) in bulk, and covering several languages. Alongside this release, we’re also making available a Hugging Face dataset of the new beta Structured Contents snapshots and inviting the general public to freely use and provide feedback. All of the information regarding the Hugging Face dataset is posted on our blog here.
- Beta Structured Contents endpoint within On-demand API which gives users access to our team’s latest machine readability features, including the below:
  - Short Description (available in Structured Contents On-demand)
    - A concise explanation of the scope of the page written by Wikipedia and Wikidata editors. This allows rapid clarification and helps with topic disambiguation
  - Pre-parsed infoboxes (available in Structured Contents On-demand)
    - Infoboxes from Wikipedia articles to easily extract the important facts of the topic to enrich your entities.
  - Pre-parsed sections (available in Structured Contents On-demand)
    - Content sections from Wikipedia articles to easily extract and access information hidden deeper in the page.
  - Main Image (available in all Wikimedia Enterprise APIs)
    - The main image is curated by editors to represent a given article’s content. This can be used as a visual representation of the topic.
  - Summaries (aka `abstract`) (available in all Wikimedia Enterprise APIs)
    - Easy to ingest text included with each revision to provide a concise summary of the content without any need to parse HTML or Wikitext.

Content Integrity

Goal: To provide more contextual information alongside each revision to help judge whether or not to trust the revision.
Launches
- Maintenance Tags
  - Key enWiki tags that point to changes in credibility.
  - Small scale POC
- Breaking News Beta [Realtime Streaming v2]
  - A boolean field detecting breaking news events to support prioritization when doing real-time ingestion of new Wikipedia pages
- Liftwing ‘Revertrisk’
  - ORES ‘goodfaith’ and ‘damaging’ scores have been deprecated from our API responses. We are working on the integration of ‘revertrisk’ score to our API response objects.
- No-Index tag per revision

API Usability

Goal: To improve the usability of Wikimedia Enterprise APIs
Launches:
- Snapshots
  - Filtering available snapshots to group snapshots to download
  - Parallel downloading capabilities to optimize ingestion speeds
- On-demand
  - Cross language project entity lookups to connect different language projects for faster knowledge graph ingestion.
  - NDJSON responses to enable data consistency across WME APIs
  - Filtering and customized response payloads
- Realtime Batch
  - Filtering available batch updates to group files to download
  - Parallel downloading capabilities to optimize ingestion speeds
- Realtime Streaming
  - Realtime Streaming reconnection performance improvement
  - Shared credibility signals accuracy results
  - Shared latency distribution for Realtime Streaming events
  - Parallel consumption - enable users to open multiple connections to a stream simultaneously
  - More precise tracking - empower users to reconnect and seamlessly resume message consumption from the exact point where they left off
  - Event filtering by data field/value to narrow down revisions
  - Customized response payloads to control event size
  - Proper ordering of revisions to remove accidental overwrites
  - Lower event latency to ensure faster updates
  - NDJSON responses to enable data consistency across WME APIs

Past updates

For previous months' updates, see the the archive.

Overview

Background

Due to the myriad of sources of information on the internet, compiling public and private data sets together has become a major proprietary asset (seen in customer knowledge graphs) for large tech companies when building their products. It is through this work that a company’s voice assistants and search engines can be more effective than those of their competitors. Wikimedia data is the largest public data source on the internet and is used as the "common knowledge" backbone of knowledge graphs. Not having Wikimedia data in a knowledge graph is detrimental to a product’s value, as we've proven through customer research.

In order for Wikimedia Enterprise API's customers to create effective user experiences, they require two core features from the Wikimedia dataset: completeness and timeliness.

Wikimedia content provides the largest corpus of information freely available on the web. It maps broad topics across hundreds of languages and endows consumer products with a feeling of “all-knowingness” and “completeness” that drives positive user experiences.

Wikimedia content originates from a community that authors content in real time, as history unfolds. Leveraging that community’s work provides customer products with the feeling of being “in-the-know” (i.e., “timeliness”) as events occur, thus generating positive user experiences.

There is currently no way for a data-consuming customer to make one or two API requests to retrieve a complete and recent document that contains all relevant and related information for the topic requested. This has resulted in customers building complex ad-hoc solutions that are difficult to maintain; expensive, due to a large internal investment; error prone, due to inconsistencies in Wikimedia data; and fragile, due to changes in Wikimedia responses.

Research Study, 2020

From June 2020 – October 2020, the Wikimedia Enterprise team conducted a series of interviews with third-party reusers [Users] of Wikimedia data to gain a better understanding of what companies are using our data, how they are using our data, in what products they are using it, and what challenges they face when working with our APIs. Our research showed that:

Users cache our data externally rather than query our APIs for live data
Each user approaches our current stack differently, with unique challenges and requests
The Wikimedia APIs are not viewed as a reliable ingestion mechanism for gathering data and are prone to rate limits, uptime issues, and excessive use to achieve their goals
All users have the same general problems when working with our content, and we have received similar asks from users of all size

The Enterprise API team has identified four pain points that cause large third-party reusers to struggle when using our public suite of APIs for commercial purposes. Note: Many of these concepts overlap with other initiatives currently underway within the Wikimedia movement, for example the API Gateway initiative.

Freshness: Commercial reusers want to be able to ingest our content "off-the-press" so that they can have the most current worldview of common knowledge when presenting information to their users.
System Reliability: Commercial reusers want reliable uptime on critical APIs and file downloads so that they can build using our tools without maintenance or increased risk on their products.
Content Integrity: Commercial reusers inherit the same challenges that Wikimedia projects have in relation to vandalism and evolving stories. Commercial reusers desire more metadata with each revision update in order to inform their judgement calls on whether or not to publish a revision to their products.
Machine Readability: Commercial reusers want a clean and consistent schema for working with data across all of our projects. This is due to the challenges that come from parsing and making sense of the data they get from our current APIs.

For Content Integrity and Machine Readability, the Wikimedia Enterprise team created this list of notably interesting areas to focus our work for third party reusers. This list was created in March 2021 and has thus been refined and prioritized into roadmap features laid out below, however, this serves as an artifact of this research and something that can be used to reference back to some of the problems that reusers are facing.

Theme	Feature	Details
Machine Readability	Parsed Wikipedia Content	Break out the HTML and Wikitext content into clear sections that customers can use when processing our content into their external data structures
	Optimized Wikidata Ontology	Wikidata entries mapped into a commercially consistent ontology
	Wikimedia-Wide Schema	Combine Wikimedia project data together to create “single-view” for multiple projects around topics.
	Topic Specific Exports	Segment corpus into distinct groupings for more targeted consumption.
Content Integrity	Anomaly Signals	Update schema with information guiding customers to understand the context of an edit. Examples: page view / edit data
	Credibility Signals	Packaged data from the community useful to detect larger industry trends in disinfo, misinfo, or bad actors
	Improved Wikimedia Commons license access	More machine readable licensing on Commons media
	Content Quality Scoring (Vandalism detection, “best last revision”)	Packaged data used to understand the editorial decision-making of how communities catch vandalism.

Product Roadmap

The Wikimedia Enterprise APIs are designed to help external content reusers seamlessly and reliably mirror Wikimedia content in real time on their systems. However, even with this system in place, reusers still have many struggles with the Content Integrity and the Machine Readability of Wikimedia content when they try to make it actionable on the other end. This section will lay out all of the work we are actively working on to help alleviate some of the struggles. To reference our previous research work:

Wikimedia Enterprise "Future Roadmap" from March 2021 (annotated with current focus points in bold/italic)
Theme	Feature	Details
Machine Readability	Parsed Wikipedia Content	Break out the HTML and Wikitext content into clear sections that customers can use when processing our content into their external data structures
	Optimized Wikidata Ontology	Wikidata entries mapped into a commercially consistent ontology
	Wikimedia-Wide Schema	Combine Wikimedia project data together to create “single-view” for multiple projects around topics.
	Topic Specific Exports	Segment corpus into distinct groupings for more targeted consumption.
Content Integrity	Anomaly Signals	Update schema with information guiding customers to understand the context of an edit. Examples: page view / edit data
	Credibility Signals	Packaged data from the community useful to detect larger industry trends in disinfo, misinfo, or bad actors
	Improved Wikimedia Commons license access	More machine readable licensing on Commons media
	Content Quality Scoring (Vandalism detection, “best last revision”)	Packaged data used to understand the editorial decision-making of how communities catch vandalism.

In Flight Work

New Functionality

Content Integrity: For external reusers that choose to work with Wikimedia data in real-time or even with a slight delay increase their exposure to the most fluid components of the projects and increase risk of propagating vandalism, dis/mis-disinformation, unstable article content, etc. Our goal is not to prescribe content with a decision as to its credibility, but rather to increase the contextual data "signals" around a revision to allow Wikimedia Enterprise reusers to have a better picture of what this revision is doing and how they might want to handle it on their end. This will manifest in new fields in our responses in the Realtime, Snapshot, and On-demand APIs. We are focused on two main categories of signals:
- Credibility Signals: "Context" of a revision. This looks like diving into "what changed", editor reputation, and general article level flagging. The goal initially is to lean on the information that is publicly used by editors and translate those concepts to the reusers that are otherwise unfamiliar. Track this work here.
- Anomaly Signals: "Activity" around a revision. This looks like temporal edit, page views, or talk page activity. The goal initially is to compile quantitative signals to unpack popularity that can be used to help reusers prioritize updates as well as calibrate around our trends and what that might mean for the reliability of the content.

General Improvements

Accessibility: In order to increase the availability of access to Wikimedia Enterprise APIs, we are developing a new self signup tier for folks to get started working with our APIs. Track this work here.
Reliability: Continuous improvement on our system's health in order to comfortably scale, with more context as to the problems that we'll need to continually solve for. We are building what will become a v2 architecture of Wikimedia Enterprise APIs. Track this work for the Snapshots and Realtime APIs. View our status page.
Freshness: We are working with Wikimedia Foundation teams (Platform and Data Engineering) to better understand and flag where we may have revisions missing in the feeds as to improve performance for our systems and the public systems.

Wikimedia Enterprise (Version 1.0)

See also: Up to date API documentation and more information about the general value offerings on our commercial website.

Name	Compare To	What is it?	What’s New?
Enterprise Realtime API	EventStream HTTP API	A stable, push HTTP stream of real-time activity across "text-based" Wikimedia Enterprise Projects	Push changes to client with stable connection Be notified of suspected vandalism in real time Hourly batch update files Machine readable and consistent JSON schema Guaranteed uptime, no rate-limiting
Enterprise On-demand API	Restbase APIs	Current article content in Wikimedia Enterprise JSON format. Structured Contents beta endpoint with experimental parsing.	Machine Readable and Consistent JSON schema Guaranteed uptime Beta features endpoint
Enterprise Snapshot API	Wikimedia Dumps	Recent, compressed Wikimedia data exports for bulk content ingestion.	Machine Readable and Consistent JSON schema Monthly & Daily “Entire Corpus” snapshots Guaranteed delivery Historical Downloads

On-demand API

High-volume reusers that use an infrastructure reliant on the EventStream platform depend on services like RESTBase to pull HTML from page titles and current revisions to update their products. High-volume reusers have requested a reliable means to gather this data, as well as structures other than HTML when incorporating our content into their KGs and products.

Wikimedia Enterprise On-demand API contains:

A commercial schema
SLA
Beta Structured Contents endpoint (not SLA)

Realtime API

High-volume reusers currently rely heavily on the changes that are pushed from our community to update their products in real time, using EventStream APIs to access such changes.High-volume reusers are interested in a service that will allow them to filter the changes they receive to limit their processing, guarantee stable HTTP connections to ensure no data loss, and supply a more useful schema to limit the number of api calls they need to make per event.

Enterprise Realtime API contains:

Update streams that provides real-time events of changes across supported projects
Batch processing files updated hourly with each day's project changes (formerly classified as part of the Snapshot API)
Commercially useful schema similar* to those that we are building in our On-demand API and Snapshot API
SLA

*We are still in the process of mapping out the technical specifications to determine the limitations of schema in event platforms and will post here when we have finalized our design.

Snapshot API

For high volume reusers that currently rely on the Wikimedia Dumps to access our information, we have created a solution to ingest Wikimedia content in near real time without excessive API calls (On-demand API) or maintaining hooks into our infrastructure (Realtime API - Streaming).

Enterprise Snapshot API contains:

24-hour JSON*, Wikitext, or HTML compressed dumps of supported Wikimedia project
SLA

*JSON dumps will contain the same schema per page as the On-demand API.

These dumps are available for public use fortnightly on Wikimedia Dumps and daily on WMCS users

Past Development

In response to the initial research study in 2020, the Enterprise team is focused on building tools for commercial reusers that will offer the advantages of a relationship while expanding the usability of the content that we provide.

The roadmap was split into two ordered phases focused on helping large third-party reusers with:

Building a "commercial ingestion pipe" (COMPLETE)
Creating more useful data to feed into the "commercial ingestion pipe" (IN PROGRESS)

Building a "Commercial Ingestion Pipe" aka Version 1.0 (Launched September 2021)

The goal of the first phase was to build infrastructure that ensures the Wikimedia Foundation can reasonably guarantee Service Level Agreements (SLAs) for 3rd-party reusers as well as create a "single product" where commercial reusers can confidently ingest our content in a clear and consistent manner. While the main goal of this is not explicitly to remove the load of the large reusers from Wikimedia Foundation infrastructure, it is a significant benefit, for we do not currently know the total capacity of these large reusers on donor-funded infrastructure. For more information on the APIs that are currently available, please reference the section Version 1.0 above or our public API documentation.

Daily HTML Dumps (Launched December 2020)

Tracked in Phabricator
Task T254275

The Enterprise team's first product was building daily dump files of HTML for every "text-based" Wikimedia project. These dumps will help content re-users use a more familiar data type as they work with Wikimedia content.

Reusers have four immediate needs from a service that supports large-scale content reuse: system reliability, freshness or real-time access, content integrity, and machine readability.

Web Interface

A downloader interface now in design stages allows for users to download a daily dump for each "text-based" project, search and download individual pages, and save their preferences for return visits. Currently the software is in Alpha and still in usage and quality testing. This dashboard is built in React with internal-facing client endpoints built on top of our infrastructure. The downloads are hosted and served through S3.

Rationale behind choosing this as the Enterprise API's first product

Already validated: Before the Enterprise team ran research to discover the needs of high-volume data reusers, this was the most historically requested feature. Large technology partners, researchers, and internal stakeholders within the Wikimedia Foundation have long sought a comprehensive way to access all of the Wikimedia "text-based" wikis in a form outside of Wikitext.
Take pressure off internal Wikimedia infrastructure: While not proven, anecdotally we can conclude there is a significant band of traffic to our APIs by high-volume reusers aiming to get the most up-to-date content cached on their systems for reuse. Building a tool where they can achieve this has been the first step to pulling high-volume reusers away from WMF infrastructure and onto a new service.
Standalone in nature: Of the projects already laid out for consideration by the Enterprise team, this is the most standalone. We can easily understand the specs without working with a specific partner. We were not forced to make technical decisions that would affect a later product or offering. In fact, in many ways, this flexibility forced us to build a data platform that produced many of the APIs that we are offering in the near future.
Strong business development case: This project gave the Enterprise team a lot of room to talk through solutions with reusers and open up business development conversations.
Strong introductory project for contractors: The Enterprise team started with a team of outside contractors. This forced the team to become reusers of Wikimedia in order to build this product. In the process, the team was able to identify and relate to the problems with the APIs that our customer base faces, giving them a broader understanding of the issues at hand.

Design Documents

Application Hosting

The engineering goal of this project is to rapidly prototype and build solutions that could scale to the needs of the Enterprise API's intended customers – high volume, high speed, commercial reusers. To do this, the product has been optimized for quick iteration, infrastructural separation from critical Wikimedia projects, and to utilize downstream Service Level Agreements (SLAs). To achieve these goals in the short term, we have built the Enterprise API upon a third-party cloud provider (specifically Amazon Web Services [AWS]). While there are many advantages of using external cloud for our use case, we acknowledge there are also fundamental tensions – given the culture and principles of how applications are built at the Foundation.

Consequently, the goal with the Enterprise API is to create an application that is "cloud-agnostic" and can be spun up on any provider's platform. We have taken reasonable steps to architect abstraction layers within our application to remove any overt dependencies on our current host, Amazon Web Services. This was also a pragmatic decision, due to the unclear nature of where this project will live long-term.

The following steps were taken to ensure that principle. We have:

Designed and built service interfaces to create abstractions from provider-specific tools. For instance, we have layers that tie to general File Storage capabilities, decoupling us from using exclusively "AWS S3" or creating undo dependency on other potential cloud options
Built the application using Terraform as Infrastructure as Code to manage our cloud services. [The Terraform code will be published in the near future and this documentation will be updated when it is]
Used Docker for containerization throughout the application
Implemented hard drive encryption to ensure that the data is protected (we are working to expand our data encryption and will continually as this project develops)

We have intentionally kept our technical stack as general, libre & open source, and lightweight as possible. There is a temptation to use a number of proprietary services that may provide easy solutions to hard problems (including EMR, DynamoDB, etc). However, we have restricted our reliance on Amazon services to what we can be found in most other cloud providers. Below is a list of services used by the Enterprise API within Amazon and its purpose in our infrastructure:

Amazon EC2 - Compute
Amazon S3 - File Storage
Amazon Relational Database Service (PostgreSQL) - PostGRES Database
Amazon ElastiCache for Redis - Cache

Amazon Elasticsearch Service - Search Engine
Amazon MSK - Apache Kafka Cluster
Amazon ELB - Load Balancer
Amazon VPC - Virtual Private Cloud
Amazon Cognito - Authentication

We are looking to provide Service Level Agreements (SLA) to customers similar to those guaranteed by Amazon's EC2. We don't have equivalent uptime information from the Wikimedia Foundation's existing infrastructure. However, this is something we are exploring with Wikimedia Site Reliability Engineering. Any alternative hosting in the future would require equivalent services or time to allow us to add more staff to our team in order to give us confidence to handle the SLA we are promising.

In the meantime, we are researching alternatives to AWS (and remain open to ideas that might fit our use case) when this project is more established and we are confident in knowing what the infrastructure needs are in reality.

Team

For the most up-to-date list of people involved in the project, see Meta:Wikimedia Enterprise#Team.

API	Availability	URL base	Example
MediaWiki Action API	Included with MediaWiki Enabled on Wikimedia projects	/api.php	https://en.wikipedia.org/w/api.php?action=query&prop=info&titles=Earth
MediaWiki REST API	Included with MediaWiki 1.35+ Enabled on Wikimedia projects	/rest.php	https://en.wikipedia.org/w/rest.php/v1/page/Earth
Wikimedia REST API	Not included with MediaWiki Available for Wikimedia projects only	/api/rest	https://en.wikipedia.org/api/rest_v1/page/title/Earth
For commercial-scale APIs for Wikimedia projects, see Wikimedia Enterprise