From document to topic

What happens to a document: split, write, check, approve, build topics, link to the model.

1. Read and split. The text is cut into passages that fit the embedding model (about 75% of EMBEDDING_MAX_TOKENS), at paragraphs, with a little overlap, starting afresh at each page, slide or sheet. Each passage is embedded and indexed for keywords. This needs no AI endpoint, so even a collection that forbids AI is stored and searchable.
2. Write the source page. A writer model writes a standalone page that keeps every stated fact, number, condition and exception and never mentions the document itself. Long documents are written section by section.
3. Check it. A second checker model verifies the page claim by claim against the document, with the complete text when it fits and otherwise with tools (search the document, read a passage with its neighbours, list the outline). A FAIL that names no problem is asked for specifics and, if it still names none, the same version is checked again: the writer never amends blindly. A FAIL is also reviewed once by the checker itself, which drops entries it calls supported. If real problems remain, the writer corrects the page from the evidence and the checker looks again, up to Correction rounds (default 3).
4. Approve or hold. A page that passes is approved. One that still has objections goes to needs review: a person reads the objections next to the page and either approves it as it is or has it written again. Every version and every check is kept and can be opened.
5. Build topics. For an approved source, an analyzer model (with tools to search and read the collection's existing topics) decides which topics the source creates, updates or complements (at most 12). For each topic it extracts atomic claims, each with a quote that must appear word for word in the approved source wiki (a claim whose quote is not there is dropped). The topic writer then writes or amends each page from the full verified source wiki, the candidate and its claims, following the topic type's template; it can search and read related topics and other approved source wikis of the collection, keeps every cited fact the new evidence does not supersede, and cites sources. Each claim records the wiki version and passage its quote comes from. Every page version is chunked and embedded, the topic's identity vector is rebuilt from its claims, and links are embedded with their context.
6. Link to the model. In the same pass the analyzer says which model entities each topic is about (see The bridge).
7. The review gate. Most updates publish at once; contradictions and probable duplicates wait (see Review).

The wiki models see only the one document of the job; the topic models see the collection's topics and approved source wikis, never another collection's. Which model does which job is set per job (Settings > AI); a job without its own model uses the assistant model. If the topic writer fails, a plain page is built from the candidate so nothing is lost. The same claim found again in another source is attached to the existing one (numbers stay stable). Reading a source again supersedes the claims only it supported and states them afresh; a topic it no longer mentions gets a surgical rewrite that checks each withdrawn claim against the remaining sources' wikis and removes only what nothing else supports.

Editing and downloads. With write access and the editor role, Edit the page (or Write the page by hand where AI is off) saves a new version marked edited, with your name. With AI available it is checked once against the document, not rewritten: a pass approves it, otherwise it waits for review. Without AI, saving it is the approval. Anyone who can read the collection can download the original file and the page (Markdown, any version).