Chapter 10 · Explainers

Grounded on your docs: retrieval, not training

Serpio does not train a model on your content. It retrieves passages from your indexed pages and Knowledge Base at write time. What that means for accuracy, privacy, and future articles.

Updated Sep 19, 2026 · 5 min read

On this page

"Grounded on your docs" is a specific claim, and it is worth being precise about what it means. Serpio does not fine-tune or train a model on your content. Instead, each time it writes, it looks up the relevant passages from your indexed pages and Knowledge Base and hands them to the writer as evidence for that one article. The mechanism is retrieval, and the difference matters for accuracy, privacy, and what happens when your facts change.

Two ways to use your content

Training or fine-tuningRetrieval (what Serpio does)
When your content is usedOnce, during a training runEvery time an article is written
What the model keepsPatterns absorbed into its weights, hard to inspect or removeNothing. The passages exist only in that run's evidence bundle
Updating a factRetrainEdit or replace the source; the next article sees it
Tracing a claimNot possibleEach article lists the passages it used

Retrieval is the only one of these where you can open a finished article, see exactly which passage a sentence came from, and change the source for next time.

What is retrieved for an article

When you start a draft, the job card shows four steps: Plan, Research, Write, and Finish. Retrieval happens in Research, described on the card as knowledge and public evidence. Serpio queries three places.

Your indexed pages
The private site index Serpio built by crawling your site. Used for site context and as the only source of internal link targets.
Your Knowledge Base
URLs and uploaded documents you added, retrieved by meaning against the article's topic. Only sources whose permission allows facts are used as facts. Up to two sources set to background context are passed along for understanding only.
Bounded public research
A small number of web searches, capped at four queries with five results each and at most eight fetched pages. Public sources need a public URL to be cited.

The passages retrieved become that run's evidence bundle, recorded once and reused if the job retries. If your Knowledge Base changes while a draft is mid-run, Serpio stops and retries with a consistent version rather than mixing old and new facts. On the finished article, Knowledge used and Public sources show what went in.

Permissions keep private facts out

Every Knowledge Base source has a Permission for writing setting with two values. Facts allowed in articles means relevant facts from the source may appear in article copy. Background context only means the writer may use it to understand your business, but its private facts and quotations are not approved for publication. Retrieval for facts skips context-only sources entirely.

Retrieval also skips anything that should not be live: paused or removed sources, sources with a future effective date, sources whose review date has passed, and versions that have not finished indexing. Documents are never turned into links. Their filenames, storage paths, and download URLs are stripped from what the writer sees, and the article's evidence panel shows them simply as Knowledge document.

What happens to your data

Serpio runs its language models through OpenRouter. On every request, Serpio denies model providers permission to collect or train on your data. Uploaded files stay private in Serpio's own storage and index. The writer receives only the context authorised for the current run, and retrieved text is treated as evidence to cite, not as instructions to follow.

Adding knowledge changes future articles only

When you add a document or URL, the change applies to jobs that start after it finishes indexing. Articles already written keep their original evidence and their quality reports. If an existing article needs the new fact, edit it or generate a fresh one. Knowledge refreshes do not use article credits and do not queue anything for publishing.

This is the practical upside of retrieval over training. Your Knowledge Base is a living set of sources, not a model you have to rebuild. Keep it accurate and the next article inherits the accuracy. For how the retrieved evidence is then checked, read No invented links or facts.

Common questions

  • Does Serpio learn my voice over time?

    Not by training. Voice comes from your Brand Book and voice cues, which are applied to every draft at write time. Change them and the next draft changes. Nothing about your voice is baked into a model.

  • If I delete a document, do old articles change?

    No. Each article keeps the evidence it was written from. Deleting or pausing a source only removes it from future retrieval.

  • Can another customer's articles draw on my documents?

    No. Retrieval is scoped to your site. Only sources you added to that site's Knowledge Base, plus its own indexed pages, are candidates.

Your next article starts here.

Turn a topic that is trending right now into your first article.

Write my first article

3 free articles a month. No credit card.