Knowledge Bases

A knowledge base (KB) is an indexed corpus of documents that can be attached to one or more projects. Semantic search, also known as RAG query, retrieves information from the KBs attached to a project.

What lives in a KB

  • Documents — typically academic papers or books, but any PDF works. Each document has bibliographic metadata: title, authors, year, and a full bibliographic reference.
  • Files — the actual PDF attachments tied to documents. A single document can have multiple PDFs (e.g., multiple chapters in an editorial book).
  • Indexed chunks — what the retrieval system actually searches over. Indexing happens automatically when you upload a file.

KBs are independent of projects

KBs are top-level objects on your account, not nested inside a project. A KB can be attached to many projects, and a project can attach many KBs.

Creating a KB

From the Knowledge Bases view, click New Knowledge Base and give it a name and optional description. The KB is empty until you populate it.

Populating a KB

There are two paths:

  • Upload PDFs directly in the web app — open the KB detail view and drag files in. Files are uploaded, queued, and indexed in the background. See Uploading Documents for the indexing pipeline.
  • Sync from Zotero using the Zotero plugin. This is the recommended path if you already keep references in Zotero — metadata stays in sync and you keep one source of truth.

How the agent uses a KB

When you issue an instruction, the agent generates multiple search queries from different angles, and runs them in parallel against every KB attached to the current project.

See RAG & Retrieval for the full mechanics.

When the KB doesn't have what's needed

The agent transparently reports when retrieval came up short rather than inventing claims. You'll see a KB limitation note in the workflow summary, with suggestions for what to add to the corpus.

What's next