How things are found

There are six ways somebody comes across a post or a profile on Anahita: their home page, a custom feed, search, suggestions of whom to follow, and what is trending. This page is how they fit together, and what each one orders by. The pages it links to have the numbers.

Six ways in

Way	What it shows	In what order	Page
The home page	What the people, groups and events somebody follows posted	Newest first, always	Custom feeds
A custom feed	What a person chose: sources, hashtags, words, locations, languages, a distance	Newest first	Custom feeds
Search	People, groups, events, posts and shared feeds that match the words	Most relevant, or most recent	Search
A hashtag's or a location's posts	The posts that carry it	Most relevant, without the words	Hashtags and locations
Suggestions	People, groups and events to follow	By the reasons for each, scored	Recommendations
Trending	The hashtags and locations the most profiles used lately	By how many profiles, each once	Trending

The home page and custom feeds are by date and nothing else. Only search, a tag's posts and suggestions rank, and every one of them says why.

What they share

Every way in walks the same graph, over the same edges, and reads only what the viewer may see.

Sign	The edges and columns	Used by
Who somebody follows	edge.actor.follow.v1	Every one: the home page is made of it; search puts what you follow higher; suggestions look at whom your follows follow
What people found useful	Likes (edge.base.voteup.v1), saves (edge.base.bookmark.v1, private: never shown to anybody), replies, reposts and quotes	Search's standing; you_engaged in suggestions
What people cite	edge.medium.citation.v1, counted once for each person who cites	Search's standing; cited and cited_by_leaders in suggestions
Who is in a photo or a video together	edge.tag.person.v1, approved tags only	pictured_with_followed in search; pictured_with_you and pictured_with_leaders in suggestions
What a post is about	edge.tag.hashtag.v1, edge.tag.location.v1	Custom feeds' filters; search; shared_interest in suggestions; trending
How lately a profile was active	When it last posted, or was visited	Suggestions put the quiet last, and leave out those quiet longer than the reader chose

A sign is never more than one of several. A logarithm keeps one post liked a thousand times from burying everything else in search, and every reason for a suggestion has a most it can add.

What is never used

What somebody looked at, for how long, what they scrolled past, what they clicked on, where they were when they did: none of it is recorded, so none of it can rank anything. Each sign above is something a person chose to do and could see themselves doing: follow, like, save, reply, cite, tag. There are no hidden signals, and no profile of interests is built from behaviour to rank by (Why Anahita).

Interest fades; usefulness does not

A guide a library wrote years ago is as useful as one from this week. So nothing is pushed down for being old. What may fade is interest: each like, save, reply or follow is counted by when it happened, halved every half_life_days, a year by default. A guide that people saved this month stands as high as a new post saved as often.

A citation does not fade. Citing takes effort and is lasting, and what is cited stays useful. See Old and useful counts as much as new.

Each person's own weights

Each person who is signed in can change how much each sign counts, for them alone, and nobody else's results change:

• In search, six weights: likes, saves, replies, followers, citations, and what they follow (Their own weights), and two choices, Profiles I follow first and Newer results first.
• In suggestions, a weight for each reason, from 0, not used, to 10; whether what they liked and wrote may be used at all (use_activity); and how lately somebody must have been active to be suggested (Each person's own weights).
• Custom feeds are the rest: a person who wants a different feed makes one, and calls it what they like.

The operator's numbers

Every default is the installation's to change:

What	Where
Search: the points for a match, the weights of each sign, half_life_days	graph-grpc-defaults/search_ranking.json, read again when it changes (Tuning it)
Suggestions: the score of each reason and its most	recommendation-service/constants, in the code: a change is a new build
Trending: the window	TRENDING_WINDOW (Configuration)

A newsroom may want interest to fade in a month; a library, a school or a research group in a year or more, or never.