Knowledge base
The source documents and structured facts Content draws on, so drafts cite your own evidence instead of general knowledge.
On this page
The knowledge base is the Knowledge tab inside Content settings, at Content → Settings → Knowledge. It holds two things: whole documents the writer can quote passages from, and single facts you have typed in and checked yourself.
Fill it before you generate your first article. With an empty knowledge base a draft can only reach for your brand description and whatever the model already knows, which is the same material every rival's draft reaches for. With your own survey numbers, named experts, and customer results in it, the draft contains something nobody else can publish.
Content is a paid surface, so the knowledge base needs Growth or Scale and editor access to the brand. Client portal users cannot open it.
Sources and facts
The tab has two views, and they behave differently.
| View | What it holds | How generation uses it |
|---|---|---|
| Sources | Whole documents: decks, briefs, research, internal notes, pages from your site | Split into passages, then searched by meaning at draft time. The writer sees only the passages that match the question. |
| Facts | Single items you type in: a statistic, a quote, a framework, a customer result, a testimonial | Matched by topic, then handed to the writer whole as citable assets. |
Both are scoped to one brand. Nothing you add to Nike is visible when you write for Adidas.
Adding a source
The Sources view gives you three buttons.
- File accepts PDF, DOCX, TXT, and Markdown, up to 10MB.
- URL fetches the page once and keeps the extracted text. It is a copy, not a live link, so a page that changes later does not update here.
- Paste text takes a name and a body, for anything that is not already a file or a published page.
Two limits are worth knowing before you start. One source caps at roughly 100,000 tokens, about 75,000 words, so split a long annual report rather than trimming it. And identical content is rejected: Trakkr hashes the extracted text per brand, so re-uploading the same PDF returns an error instead of creating a second copy. Pasted text and URLs also need roughly 50 tokens of content, so a one-line note belongs in Facts.
What processing does
Adding a source creates a row with the status pending and hands the work to a background processor. The row shows its type, its count, and its status as it moves through pending, processing, and ready. A source that fails shows error and a Retry link.
Processing splits the text into passages of about 200 tokens with 50 tokens of overlap, so a sentence is never cut cleanly in half between two of them, then embeds each passage into a vector index kept per brand. A source caps at 500 chunks.
The five fact types
Switch to Facts and a segmented control offers five types. Each is a short form, and each has a minimum before it will save.
| Type | What it records | Must have |
|---|---|---|
| Data points | One statement, its source, and the date it comes from | A statement of at least 12 characters, plus a source |
| Expert voices | A named person, their title and company, and one quote | A name, plus a title or a company |
| Methodologies | A named framework, what it is, and the evidence behind it | A name, plus a description of at least 20 characters |
| Case studies | A customer, the challenge, the approach, and one result metric | A customer name, plus a challenge or a solution |
| Testimonials | A quote with the customer and company behind it | A quote of at least 12 characters, plus a customer or a company |
Every type also takes a Topics field. Topics decide which facts reach which article, so treat them as the most important field on the form, not an afterthought.
How knowledge reaches a draft
You never attach knowledge to an article by hand. Generation assembles it:
- Sources are searched against the article's question. Trakkr embeds the question, also writes a hypothetical answer and embeds that, then keeps up to 10 passages scoring 0.65 or higher across both.
- Facts are matched by plain word overlap between your Topics and the words in the question. Up to 5 data points, 3 expert quotes, and 2 methodologies come through.
- After the draft exists, a verification pass re-reads it against the retrieved passages, flags claims that contradict them or cannot be checked, and rewrites the high-severity contradictions automatically.
This is a check against the retrieved material, not a complete fact check. A source can be wrong or out of date, and an unsupported claim can escape detection. Review the final draft, quotes, figures and source URLs before publishing. Articles explains the limits of the visible checks.
That topic matching is literal, and it is the thing people trip over. A data point tagged attribution reaches an article about "attribution" and nothing else. Tag with the words a buyer would type, not internal shorthand, and tag generously.
Common questions
How much should I add before writing my first article?
Two or three real sources and a handful of facts is enough to change the output. One customer survey, one methodology you can name, and three data points beats twenty marketing PDFs, because retrieval picks passages by relevance and a thin, generic source rarely wins that contest.
Does the writer have to use what I add?
No. Sources and facts are offered to the model as context, not injected as required text. If nothing in the knowledge base is relevant to the question, the draft is written without it, and the verification pass is skipped because there is nothing to check against.
Why did my upload fail as a duplicate?
Trakkr hashes the extracted text of every source per brand. If the same text already exists, even under a different filename, the upload is rejected. Delete the older copy first if you are trying to replace it.
Can I see which knowledge went into an article?
Not from the Content surface today. The passages and facts used are recorded with the article, but there is no panel that lists them back to you, so treat the knowledge base as the audit trail.
What happens if I delete a source?
Its passages are removed from the brand's vector index, so it stops appearing in future drafts. Articles already generated keep their text unchanged.
Should a fact also live as a source?
It can, and the duplication is harmless. A fact is precise and always arrives whole when its topics match. A passage is fuzzy and arrives only when the question is close enough. Numbers you want quoted exactly belong in Facts.