Knowledge Base

The Knowledge Base view (sidebar ▸ Data) is the curation face of the knowledge graph: where Cortex shows you the graph as a map, Knowledge Base shows the same workspaces as document collections — lists you ingest into, search, and manage. Same graph, two doors; a workspace you create in either view appears in both.

The view is tabbed:

Your first workspace

On a fresh install the view opens with a first-run panel, because no curated collections exist yet — the automatic root mirror lives in Cortex, and curated workspaces are yours to create. Create one, give it a name, and pick the build depth (Structure only, Structure + Entities, or Full) — the same three-choice depth model as everywhere else, explained in The knowledge graph.

A practical default: start at structure, ingest, and deepen later. Depth is never a trap — the badge always reflects what the workspace actually holds.

Ingest documents

Ingest into the selected workspace from three sources:

• Files — Markdown, plain text, PDF, Word, PowerPoint, Excel, HTML, reStructuredText, and source code.
• Text — paste content directly.
• URLs — fetch and ingest a page.

At the default depth, ingest registers the documents in the graph — mechanical, immediate, and free; no file content is sent to any model. Entity enrichment stays a separate, deliberate step (run it from Cortex or let your agents do it as they work). Only a full-depth workspace runs the embedding pipeline at ingest time.

One size limit applies, and only on the full-depth path: a file whose extracted text runs past roughly half a million characters is refused with a message telling you to split it. Registering that same file at structure depth has no limit — the graph stores the document, not its text.

Browse and manage — the Content tab

The Content tab lists a workspace's documents with their metadata and enrichment state. From here you inspect a document, replace its content (re-ingesting updated material), or delete it from the collection. Deleting from a curated workspace removes the membership — the file on disk is not touched, and the same document can keep living in other workspaces.

Search — and what depth changes

The Search tab queries the selected workspace, and its behavior follows the workspace's depth honestly. The two arms differ in what they hand back, not just in how well they rank:

• On a full workspace, search is semantic — it retrieves by meaning and returns a written answer, synthesized from the material it retrieved. That is what the vector upgrade buys.
• On a structure or entities workspace, the same search box returns a ranked list of documents — matched over names, descriptions, and summaries, each with its path and extraction state. Useful, but lexical. It degrades; it never errors.

That split is a design decision, not a gap: graph lookup is the free default, and meaning-based retrieval is the opt-in tier. If you find yourself rephrasing queries to hit a jargon-heavy collection, that collection is a candidate for the upgrade.

The semantic arm needs the key every time, not only at build time. Search on a full workspace refuses with a plain message when no Gemini key is set. So a workspace you upgraded and then left without a key is not merely frozen at its current vectors — its search stops answering, while every structure- and entities-depth workspace beside it keeps working. Keep the key in place for as long as you want those collections searchable.

Upgrade a collection to semantic search

The upgrade flow adds vectors to an existing workspace without rebuilding it by hand. It pre-flights your Gemini key before starting (semantic search is the one capability in this cluster that uses an external key — see BYO Claude & keys), then runs the embedding pipeline over the collection.

Three things to know before you click:

• It spends tokens. Embedding is the only knowledge-graph operation with an external model bill, and it is proportional to the collection's size.
• It is one-way. A full workspace cannot be quietly downgraded in place; reverting means deleting and recreating the workspace. Upgrade collections that earn it, not everything.
• The key stays required. Searching a full workspace calls the embedding service on every query, so the collection depends on that key from then on — see the note above.

Stats and Graph

The Stats tab is the workspace's dashboard — document and entity counts, derived depth, and vector coverage — and the quickest way to confirm what an ingest or enrichment actually produced. The Graph tab renders the same workspace on the graph canvas, so you can flip from list to map without switching views.

Where to go next

• Cortex — the graph surface: enrichment, lint, analytics, export
• The knowledge graph — the depth model and when each depth is worth it
• Knowledge graph infrastructure — the operating reference