Collecting the sources
It all starts with what the platforms, and the companies that own them, put out themselves. No marketing blogs, no agency theories, no “I tested it and here’s what happened” threads. For Instagram that’s 1,951 sources so far, everything from Help Center pages and Mosseri’s videos to Meta’s patents and earnings calls.
Plenty of what they publish has nothing to do with the feed, so a model reads each source first and drops what doesn’t belong. Earnings calls are long and mostly about money, so only the parts about Instagram are kept, and only if they match the transcript word for word. Everything keeps the date it came out, because advice from 2019 and advice from last month aren’t the same thing.
Figure 1
What goes into the library
Sources of five kinds are gathered; a model leaves out personal posts and off-topic patents; videos are transcribed with timestamps; every source keeps its date.
Step 1 of 4: It starts with what Instagram and Meta publish: Help Center pages, Mosseri’s posts and videos, announcements, system cards, engineering and research posts, patents and earnings calls.
Cutting it into passages
You can’t usefully search a twenty-minute video as one thing, so everything gets cut into smaller passages. For writing that’s about 270 words each, with a little overlap so nothing gets chopped off mid-thought. For video it’s 20 to 60 seconds, cut where the shot changes.
Patents and research papers are hard going, so they also get a plain-English summary of what they’d mean for creators, and that summary can be found like any other passage. Then every passage is filed twice, once by the words in it and once by what it means, and tagged with whichever of the 80 topics it’s about.
Figure 2
Two ways to find a passage
A source is cut into overlapping passages. Each goes into a keyword index and onto a map of meaning, where similar passages sit close together, and is tagged with checked topics.
Step 1 of 5: Take one source, say a help page or a video transcript. As one long block, it’s too much to search well.
Finding the passages for your question
Searching by words is good at exact terms, like “Trial Reels”. Searching by meaning catches the rest, like when you ask why your posts flop and the platform talks about reach. Neither is enough on its own, so the site does both, for your question and a few rewordings of it.
Everything it finds is merged into one ranked list for each kind of source. That’s a good shortlist, but it’s only a shortlist.
Figure 3
From one question to a shortlist
The question is reworded three or four ways; each version is searched by words and by meaning; the results for each kind of source are merged into one ranked shortlist.
Step 1 of 4: You ask in your own words. Typos get fixed first, against the words the library actually uses.
Rereading the shortlist
Both kinds of search are quick because they never really read the passage against your question. They match words, or measure how close two meanings sit on the map. That’s how a passage can rank high just for using the word “hashtags” a lot, without saying anything about whether they help.
So the shortlist gets reread by a reranker, a model built to do one job: take your question and a passage together and score how well the passage answers it. It’s much slower than searching, which is why it only sees the top 100 of each kind of source, but it’s much better at telling a real answer from a near miss. The best 48, by its scores, are what the answer gets written from.
Figure 4
How the best 48 are chosen
The top of each kind's shortlist is read together with the question by a reranking model, each passage gets a score, and the best 48 across all kinds go to the writer.
Step 1 of 4: The merge leaves one list for each kind of source. The top 100 of each, up to 300 passages, go on to be reread. Anything further down can’t make it into the answer.
Adding what’s already known
Search isn’t the only place the writer’s material comes from. If your question is about something the encyclopedia already has an article on, like Reels or hashtags, the site adds that article’s most relevant sentences, which have already been checked, along with the passages they cite. A broad question like “how does the algorithm work?” gets the article’s outline instead, so the answer can cover every part.
On Instagram it also adds the topic’s system card, Meta’s own list of what it predicts about each post and what each prediction looks at. And if you’re asking about your own numbers, like why your views dropped, it adds what each Insights number means.
This happens by fixed rules, not by a model deciding it needs to read more. The writer gets one round of material and has to work with what’s there, which is why it tells you when the sources don’t cover something.
Writing and checking the answer
Before anything gets written, the site checks your question is actually about how the platforms rank and recommend things. If even the closest passages are a long way off, it tells you that instead of making something up.
The writer is also told what a careful reader would know. A patent doesn’t prove Instagram uses something today. Plans mentioned on an earnings call aren’t promises. And when newer advice has replaced older advice, the old version shouldn’t be quoted as if it still holds.
Figure 5
Writing from the passages, then checking
The writer gets numbered passages and nothing else; it gives a verdict and cites a passage on every factual sentence; its own reasoning is marked Our read; a second pass rewrites or removes any sentence its passages don't support.
Step 1 of 4: The writer, an OpenAI model, gets only the numbered passages it’s been handed. It isn’t allowed to use anything else it knows.
Popular questions are answered ahead of time in exactly the same way, so they open straight away.
Topics, changes and the glossary
Topic articles come from the same passages. Each one is outlined, then written, then checked line by line against its sources, just like an answer.
Changes come from lining up everything said about a topic by date, and having a model point out where something newer updates or contradicts something older. It’s found 400 so far.
The glossary defines each term from the passages that use it. A second model then cuts every definition back to what those passages actually say.
Keeping it honest
There’s a test set of 118 questions, many with a list of the facts a good answer should include. Any change to how searching or writing works gets scored against it, including whether each claim is really backed by the passage it cites.
It has limits, too. It can’t read text inside images or on screen in a video, only what’s said and written. Help Center pages don’t carry a date, so they’re treated as current. And it only knows what the platforms have chosen to make public, which isn’t everything.