Knowledge base

The knowledge base turns your documents into a graph that every other agent can reason against. Drop a PDF, paste a meeting note, or link a webpage — KOBI handles the rest.

How it works

1. Ingest — KOBI accepts PDFs, images, audio, plain text, and URLs. Text is extracted automatically; audio is transcribed.
2. Organize — content is cleaned, broken into searchable sections, and saved as wiki pages keyed to the source document.
3. Index — sections are indexed for fast, meaning-based search across your project.
4. Reason — when any agent answers, it retrieves the relevant section and cites the source page.

Uploading

From the project sidebar:

• Click Files → Upload.
• Drag-and-drop multiple files at once. Up to 30MB per file in beta.
• Watch the status pill: WAITING → PROCESSING → INDEXING → READY.

If something goes wrong you’ll see one of: NO_CONTENT (empty file), WIKI_STALE (re-index needed), or ERROR (with detail).

The wiki view

Every uploaded document gets a synthesized wiki page — a clean, structured version of the content. From the Knowledge agent’s canvas you can:

• Browse the graph view to see how documents link to each other.
• Click a node to read the wiki page in the side panel.
• Edit titles, tags, and summaries — your edits override the synthesized version.

Asking grounded questions

Once a document is READY, just ask any agent. Answers will include citation chips you can click to jump to the source page.

Examples:

• “What did the Q3 board deck say about hiring?”
• “Summarize the Acme Co contract — what are the renewal terms?”
• “Find inconsistencies between the latest pricing spec and last month’s pricing memo.”

Re-indexing

If you upload a new version of a document, KOBI marks the affected pages WIKI_STALE and re-indexes them in the background. You don’t need to do anything.

Privacy

• All documents are stored securely with project-level access control.
• Search indexes are scoped to your project — never shared across workspaces.
• We process content only to deliver KOBI’s features; nothing is used to train external models.