Forkarium

What It Does

When a supporting actor in a film catches your interest, the usual move is to search, read an answer, and close the tab. On Forkarium, exploring starts by leaving that curiosity as a short post. The people and works named in the post connect to wiki documents, and the links on those documents take you on to the next item: a filmography, a voice actor, the entry before or after in a series. The slogan, “Connect curiosity. Build knowledge.”, sums up those two moves in a single line.

The Gap Between Wikis and Feeds

A wiki is good for reading organized knowledge, but it keeps no trace of who arrived at a page or what question brought them there. A social feed is the opposite: full of reactions and questions, yet pushed out of view within a day. Forkarium ties the two into one service. A post becomes the way into a document, and a document becomes material for the next post, so curiosity that used to drift past has somewhere to settle.

Gathering relationships automatically brings a trust problem with it. Once relations picked out by a machine are mixed with links people made on their own judgment, a reader cannot tell what to rely on. That is why every link records where it came from, from the very start.

Who It Is For

• People who read along works and people: Viewers of films, dramas, and games who go on to look up the cast, the voice actors, and the works that follow
• People who add what they know: Contributors who put content and sources into documents and write down why two items are connected
• People who keep the path they read: Readers who save the order they explored in, to pick it up later or share it with someone else

The Path Through a Visit

• Start with a post: Write a short reaction or question, and the names of people, works, and organizations in the text attach to their documents automatically; the author can confirm or remove each one
• Open a document: If an item has not been enriched yet, its relations are read from public sources the moment it opens, and each link found rises into view with a line drawn to it
• Follow the links: Move on to the next item along each relation, while every place you stop piles up in your exploration trail without signing in
• Save and branch off: Save the trail and share it, or branch off partway through someone else’s trail and carry on along a path of your own

Key Features

• Wiki documents anyone can edit: Edit the body and add sources. Every revision is kept, and a revert does not delete earlier versions; it stacks a new one on top
• Three kinds of source: Facts taken from web sources (W), links people made with a reason (@), and unverified machine candidates (AI), told apart by label and color
• Document links and backlinks: Search for another document from the editor to insert a link, then follow previews and backlinks in reverse. Links inside body text are not counted as factual relations
• Trail editing and sharing: Reorder or remove the stops in a saved trail and add a one-line note to each. Sharing an item carries the path that led there, which unfolds one step at a time on the recipient’s screen
• Start exploring and Serendipity: Open discovery cards built on two confirmed relations without typing a search, and the feed’s Serendipity tab reshuffles every time it opens
• Item universe: A page that floats connected items as a 3D constellation, for a look across the whole landscape
• Reading list: Keep documents to read later in the browser you are using, no sign-in needed (the page states plainly that it does not sync across devices)
• Three editor bots: @scout watches subjects whose interest is climbing fast, @context ties what is happening now to the past, and @question asks one question a day that people answer differently; each writes posts meant as starting points for exploration
• Five languages: Korean, English, Japanese, Simplified Chinese, and Traditional Chinese

Design Principles

• Sort sources first: The language model only restructures relations found in encyclopedia summaries; only results confirmed as real items become links, and even those are marked as unverified candidates. They never share a column with links people made
• No crawling ahead: Nothing is collected in bulk beforehand; an item is enriched only as far as needed when someone opens it
• Bots say what they are: Bot posts carry an “Automated” badge and a numbered source list, and links a bot makes are marked as machine candidates too. Bots never comment on or react to people’s posts, and sensitive topics go up only after review
• No invented numbers: Most viewed, following, and by-field lists appear only when there is real data, and the section is dropped when empty. A bot draft containing a number its sources do not state goes to review instead of being published
• Thin pages stay out of the index: Only substantial documents, posts above a minimum length, and titled trails go into the sitemap; the rest are noindex
• One canvas, one page: Relations inside a document read as a single vertical route in HTML, and the graph canvas appears only on the item universe

Tech Stack

• Frontend: Next.js 15 (App Router), server rendering with streaming
• Hosting: Firebase App Hosting (a main push is the deploy)
• Database: Cloud SQL for PostgreSQL (the forkarium database on a shared instance). The browser never connects to the database; every write goes through route handlers on the Next server
• Auth: Firebase Auth (Google sign-in), with the ID token verified on each route
• Analytics: Google Analytics 4
• Media: Firebase Storage (external images are copied with their original source recorded; user photos are resized and re-encoded to webp in the browser)
• Knowledge sources: Wikidata structured data and Wikipedia summaries
• Automation: A GitHub Actions cron runs the editor bots every three hours