Methodology Chapter 04 of 08
How Mosaic knows these articles belong together.
Dozens of outlets cover the same event. Mosaic News gathers their articles into one Story rather than showing you the same news over and over. It does that by comparing what each article means, not the words it uses. Related Stories then group into a Big Picture.
Watch separate articles become one Story.
Separate articles about the same event gather into Stories, and those Stories build the US-Iran Escalation Big Picture.
Turning meaning into numbers
An embedding turns a piece of text into a long list of numbers. Think of that list as a position: text that means something similar ends up in a similar position, even when the two share no words.
Mosaic News uses a model from OpenAI called text-embedding-3-small. For an article it reads the headline and the short preview text, with the headline counted twice so the main event weighs more than the details. Each article is measured once and the result is kept.
“electric vehicle production cuts”
“automaker reduces EV output”
How a Story forms and stays on topic
No editor built the Story above. It formed on its own as the articles arrived, following three rules: what joins a Story that already exists, what starts a new one, and what keeps a Story from wandering off the event it began with.
Join an existing Story
Every Story has an average position, the middle of all the articles already in it. A new article is measured against the middle of each Story still running. If it lands close enough to one, and both are filed under the same broad subject such as World or Politics, it joins that Story.
Start a new Story
Articles that did not join any Story are then compared with each other. When several land close together and come from at least two different newsrooms, they start a new Story. That two-newsroom rule keeps a single outlet from creating a Story by itself.
Stay on the original event
The first articles in a Story count double when that middle is calculated. Without that, weeks of new coverage could slowly pull a Story away from the event that started it.
The same measurements power your topics
Every topic Mosaic holds is measured the same way an article is, so any article, Story, or Big Picture can be compared against all of them at once. That comparison is what makes topic pages, pinned sections, and My Topics possible.
Closeness is not sufficient to file an article under a topic, though. Two subjects can register as similar without either being what the article is about, and separating the two takes more than distance. How an article is labeled is the Article labels chapter.
The boundary
What embeddings do not decide
- They do not judge whether an article is true.
- They do not decide a publisher's political lean or credibility.
- They do not write any summary, title, or framing note you read.
- They do not let one outlet count as several independent sources.
- They do not guess: an article that could not be measured is left out rather than filed somewhere close enough.