Review, search and upkeep
The review queue, reading and searching, and keeping a collection healthy.
Review#
Knowledge > Review is the collection's queue: everything that waits for a person. The sidebar badge counts what waits for you across your collections.
| Item | What you do |
|---|---|
| Contradiction: new evidence directly contradicts a statement on the existing page | The review screen shows the published page beside the page as it would become, the new statements with their quotes and the statement being contradicted. Approve (the old statement is superseded), reject (the change is discarded), or leave it. |
| Probable duplicate: a new topic very similar to an existing one (Similarity at which a new topic is held, default 0.92) | Approve, reject or merge it into the existing topic. |
| Source page with objections | Approve as it is, or have it written again. |
| Suggested link to the data model | Accept or reject in place. |
| Document that could not be read | Retry or remove. |
| Vocabulary proposal | The collection's manager accepts or rejects (an editor sees it as waiting for the manager). |
A held change is written but not published: its statements stay pending, readers see the old page and agents see nothing of it. Two held changes on one topic never build on each other's unapproved work. Everyone who can write in the collection sees the same queue, so whoever is free picks it up; readers do not see it because held changes are unpublished. Interviews are different: the conversation belongs to the person who held it.
Read and search#
- Topics (Knowledge > Topics): the article with source chips and links to related topics, and beside it every claim with its quote, source, source wiki version and the wiki passage it sits in. Aliases, related topics (with their relation), the data-model links and the version history are on the page.
- Search (Knowledge > Sources): a search box for the selected collection combines meaning and words; each result shows its source, place, section and similarity and carries the collection's usage note. Smallest similarity (Settings > Library) can hide weak matches; it is off by default because models differ. Search is always inside one collection: the database itself refuses a passage that names a source of another collection.
- The search box in the top bar finds topics, documents and collections across everything you may read, each result showing its collection and usage tag.
- Graph (Knowledge > Graph): the topics of one collection and their links (colour = type); tick Also show the data-model entities these topics are about to add the entities. The model's own graph can overlay the topics of the collections you can read.
Keep it healthy#
- Merging: any topic can be merged into another (topic page > Merge into another topic). Its statements are renumbered into the target (a statement the target already makes keeps one copy with both pieces of evidence), its title and aliases become aliases, its links and sources move, the target's page is written again and the old address keeps working as a redirect.
- Revising (the lint pass): when a source is deleted, the claims only it supported are superseded (their quotes are kept) and the topics that relied on it are marked needs revision. A background job runs the surgical rewrite: the writer checks each withdrawn claim against the remaining sources' wikis and removes or softens only what nothing else supports. A topic with no source left is archived (not listed, not found; its history stays).
- Vocabulary (Knowledge > Vocabulary): the collection's manager adds kinds, edits their meaning and (for topic types) the page template the writer follows, accepts or rejects what the analyzer proposed, deletes a kind nothing uses, or merges one kind into another (every topic or link of the first becomes the second).
Topicandrelated_toare the fallbacks and cannot be removed. When nothing fits, the analyzer proposes a new kind and the closest existing kind is used meanwhile. Nothing one collection adds reaches another, and the model's ontology is never touched. - Health (Knowledge > Health): a check without any AI model: held changes, topics needing revision, topics with no source evidence or no published page, active claims citing no source, source pages with objections, failed sources, approved sources not yet turned into topics, isolated topics, links to archived topics, statements whose quote is no longer on their source's page, and pairs of very similar topics that may be one subject. The last 20 runs are kept.
- Export (Collections > Export, managers; audited and rate limited): the whole collection as one zip. It holds the original files as uploaded, what is known about each source, every source page with all versions and checks, every topic as Markdown with front matter plus a JSON with every claim, its exact quote, the source and the source wiki version and passage it comes from, the links, the vocabulary, the collection's settings and a manifest that lists anything that could not be included. Held changes and interviews are left out. It is an archive and handover format; there is no import yet.