Knowledge base

In the portal, under Environments, click Development, and there Open. Flow opens in a new tab and you are signed in straight away.

Click ＋ next to General and choose New knowledge base. Under Settings, give it a name and a description. The description is not just for you: an agent reads it to decide when to search this knowledge base.

Under Documents, click Add document and paste a text. Dragging in a file works too: PDF, Word, Excel, HTML or CSV. Flow splits the document into chunks — for this text, by paragraph — and makes them searchable.

The Notebook is where you try things out: search, ask questions, take notes, as in Jupyter. Add a Search cell, type what you are looking for and run the cell. You see the best matching chunks, with their score and the document they come from.

A Question cell first finds the best matching chunks and lets an AI model answer the question with them, with numbered sources. Choose the model; without one you get the best passages back. The notebook is kept: everyone who may read the knowledge base sees the same cells.

A Flow table can be a source too: every row becomes one chunk. Under Documents, choose Add document again, then Flow table, and choose the table. With many chunks at once — a big table or upload, or a sync with many new rows — Flow makes the first series of vectors straight away; the rest waits. Until they are there, search finds those chunks by words only, not by meaning. The knowledge base says so on every tab, with how many.

Run the search cell in the Notebook once more. Below the results it says how many chunks were searched by words only. An agent or workflow that searches this knowledge base gets the same warning; you see it in the agent's trace.

Choose Re-index in the notice. Flow makes the missing vectors, series after series, and counts down. Then the notice disappears: every chunk can be found by meaning again. When the embedder stops halfway, Flow says how many chunks are still waiting.