Author notes

An author note is a short text you attach to a document to guide how kb2b reads its content when extracting facts. The note doesn’t become a fact and doesn’t appear in the POT’s graph — it lives pinned to the document and only influences the next extraction. It’s the tool for things the LLM can’t infer on its own: that the document is internal, that it’s third-party marketing, that a section is a legal annex and not a team statement. Before, this was a compromise — you hid the context inside the document and polluted the facts. Now it lives separately.

​

When to add a note

Notes help when the document, on its own, is ambiguous or misleading for a literal extraction: If the document is clear and needs no framing, don’t add a note — the LLM extracts well without help.

​

How to add a note when uploading

1. Navigate to Documents and drag the file (/dashboard/documents).
2. Before pressing Extract, expand Attach context under the document in the staging area.
3. Type a short Name (e.g. “Scope: internal document”) and pick a mode:
   • Paste text — type the note directly (up to 16 KB).
   • Upload file — attach a .txt or .md with the note (up to 80 KB).
4. Press Create attachment. An amber chip appears under the document.
5. Repeat for more notes (up to 5 per document), or press Extract to start extraction.

Every note is optional. A document with no notes extracts normally, the same as before.

​

Automatic suggestions

When you upload a document, kb2b reads it with a fast model (Claude Haiku) and proposes draft notes if it detects useful signals — for example, that the file is JSON with an external $schema, or that the text has a clearly promotional tone. Suggestions appear as cards in an amber panel headed Suggestions for extraction:

• Each card has a short title (the angle the model detected) and a draft text for the note.
• Press Use this note and the editor opens pre-filled. You can edit the text before pressing Create attachment.
• If none of the suggestions fits, ignore them and write your own — the manual flow stays right below.
• Close the panel with the X if you’d rather not use them.

Suggestions are advisory. The model never invents content from the document — it only proposes framings. You decide what to attach.

​

Edit or delete a note

Before extracting, each amber chip has a trash icon to delete. Deleting a chip removes the note immediately. To change a note, delete it and create another — there’s no in-place edit. That’s deliberate: a note is a curatorial decision, not a draft. After extraction, the notes that were in scope stay recorded on each fact (see Fact context). If you delete the note later, the facts already extracted keep a visual trace (“Attachment deleted”) so the team sees the historical record.

​

Limits

If you need to pass much more context, the right material is probably to upload it as a separate document, not as a note. Notes are short by design — they’re interpretation anchors, not a source of information.

​

What notes do NOT do

• They aren’t extracted as facts. Extraction rejects, by design, any fact whose evidence is only the note. The note guides how the main document is read; facts come from the document.
• They don’t appear in chat. When the team talks with the POT, notes aren’t cited or mentioned — they’re extraction metadata.
• They have no POT Score. A note isn’t domain knowledge; it’s an instruction to the pipeline.
• They don’t carry across documents. Every note lives pinned to one specific document. If the same directive applies to 20 documents, write it 20 times (or use the POT Constitution for something that applies POT-wide).

​

Best practices

• One note = one reason. If you need to say three things, write three short notes. The LLM reads them better than one long note with three ideas mixed in.
• Name the note like a headline, not like a long sentence. “Scope: external” works better than “This note explains that the document is from an external source and therefore…”
• Look at suggestions before writing. The model catches things you might miss (a $schema on line 12 of the JSON, a disclaimer at the end of the PDF).
• If a note changes, start it over. Delete the old chip, create a new one, re-extract. Cleaner than mutating.