Knowledge Base

The Knowledge Base is the set of sources your assistant grounds its answers in. It answers from what is here and nothing else, so what you put in is the whole of what it knows.

Open it from the dashboard's Knowledge Base page. The header shows how much of your plan's budget you have used, in pages.

Adding a source

The Add source button opens a form with four tabs.

Paste

For anything you can copy: a policy, a price list, an internal note nobody has published. Give it a title, paste the text, and it is searchable within moments.

Upload

Text-based PDF, Markdown, or plain text, up to 2 MB per file. The title defaults to the filename.

Scanned PDFs do not work. A scan is a picture of a page, and Sondo reads text rather than running optical character recognition, so a PDF with no extractable text is refused rather than silently added as an empty source. Export a text PDF from the original document instead.

URL

Fetches one page and extracts its main content. Login-walled pages do not work, since Sondo visits the page as an anonymous reader.

A URL source keeps its address, so you can refetch it later when the page changes.

Site

Connect a blog and every post becomes a source, kept up to date every night on its own.

Give it the root address of the site. Sondo reads your llms.txt file and your sitemap together and combines what they list, since each one knows something the other does not: the llms.txt carries the titles and descriptions you wrote, and the sitemap carries the dates that tell Sondo what has changed. An RSS or Atom feed is tried only if neither of those turns up a post. If none of them does, Sondo asks you for the address of a sitemap or feed directly, and a pasted address is used on its own.

You can restrict a connected site to one path prefix, for example /blog/, so only the pages under it are synced. Leave it empty to take every post found.

Each sync reads up to fifty posts per site, spaced out so your server is not hammered, and skips anything under about a hundred and twenty characters as too short to be useful. Sondo identifies itself as SondoBot/1.0 in every request.

The Sync now button on a site's card runs a sync immediately, in the background. Posts that come from a connected site are read only in Sondo: to change what your assistant says about one, edit the post on your own site, and the next sync picks it up.

Keeping a source current

Open any source from the Knowledge Base list.

• Edit a document in place. Change the title or the text and save. Sondo re-slices and re-embeds it, and swaps the new version in atomically, so the assistant is never mid-way between two versions of the same source. A shrinking edit always fits, because only the growth counts against your budget.
• Refetch a URL source, or point it at a different address, from the same page. A refetch replaces any hand edits you made to that source's text, which is the trade for keeping it in sync with the page.
• Re-index rebuilds the searchable slices without changing the text. It is the thing to try when a source looks present but never turns up in an answer.
• Delete removes it from the Knowledge Base entirely.

The detail page also has a Chunks tab, which shows the slices the assistant actually searches, and whether each one is searchable yet.

Limits

• One source holds up to 200,000 characters of extracted text, which is about a hundred printed pages. This applies whether you pasted it, uploaded it, or fetched it.
• Your workspace's total is your plan's page budget. See API keys and plans for each plan's number, and for what a page is.
• Re-indexing and deleting are never blocked by the budget. Only adding and growing count, so a full Knowledge Base can always be repaired or trimmed.
• A connected site stops at the budget rather than adding some posts and not others. Its card says so, and the sync resumes once you have room.

Adding anything also needs a working model key on the workspace, because a new source has to be embedded before it can be searched, and that is a paid call on your own account.