Studio

Knowledge Bank

One library for the whole workspace — and a switch per character deciding who may read what.

Knowledge Bank page with the upload dropzone and the list of files showing size, chunk count and status
The workspace library. Status moves from Processing to Ready as each file is read and indexed.

The Knowledge Bank is how a character stops improvising and starts answering from your actual material — the price list, the policy document, the product manual. Files live here at workspace level; each character is then granted access to specific ones on its own Knowledge Base tab.

What you can upload

FormatNotes
PDFThe most common source. Text is extracted page by page.
Word (.docx)Handled directly — no need to export to PDF first.
Text and MarkdownIdeal, because you control exactly how it is written.
CSVGood for price lists and tables of specifications.
Images — JPG, PNG, WebPThe text inside them is read automatically, so a photographed menu or a screenshot of a table works.

Maximum 50 MB per file. Drag several at once and each is processed independently.

What happens after upload

  1. Processing — the file is read, split into passages, and indexed so it can be searched by meaning rather than keyword.
  2. Ready — the file can now be used. The list refreshes itself, so you can watch it change.
  3. Failed — the file could not be read. Usually a scanned PDF with no text layer, or a file whose extension does not match its actual contents.
Uploading here does not give any character access. A file added on this page starts switched off everywhere. Open each character's Knowledge Base tab and turn it on. If you would rather do both in one step, upload from inside the character instead — files added there are enabled for that character immediately.

Deleting

The trash icon removes a file from the workspace, its indexed passages, and every character's access to it, immediately and without a confirmation prompt. There is no undo — re-upload if you delete the wrong one.

Writing documents that retrieve well

Retrieval finds passages, not documents. How you write has more effect on answer quality than which model you pay for.

DoInstead of
One file per subject — "Returns policy", "Delivery times", "Warranty"A single 200-page company handbook
Headings that read like customer questions — "How long does delivery take?""Section 4.2.1 — Logistics"
The answer immediately under the headingAn answer spread across three pages
Spelling out the subject in each section"As mentioned above, it is 14 days"
Plain sentencesTables of abbreviations with no context
That last one matters more than it looks. Passages are retrieved on their own, without the pages around them. A line that reads "14 days" is useless out of context; "Returns are accepted within 14 days of delivery" answers the question by itself.

Checking it works

Every character's Knowledge Base tab has a Search Test box. Type the question a real customer would ask — not the keyword you think is in the document — and confirm the returned passages actually contain the answer. If they do not, the fix is in the document, not in the character.

Do this before blaming the AI. Nine times out of ten a character "making things up" is a character that retrieved nothing useful, because the answer was buried in a table or phrased in language no customer would use.

Where to go next

Check your storage

Plan & quota →