Knowledge base

Create a dataset

A dataset is the container documents live in. Get the metadata and provider choice right up front — they shape what the agent can retrieve and how fast.

What you'll learn
  • What each dataset field controls
  • Who can see a dataset
  • Which OpenAI key indexes the dataset, and who chooses it
  • How to change a dataset after creating it
  • When to split content into multiple datasets

Open the dataset builder

Go to Knowledge Base in the sidebar, then click Create Dataset. Three steps: Basic Information, Search Engine, Review & Create. Name and search engine are required; the description is optional. If the workspace has no OpenAI key for knowledge bases yet, you are asked to choose one first (a workspace admin does this; anyone else is told which admin to ask).

Fill the dataset fields

  1. 1

    Name

    Short and descriptive, up to 150 characters. Examples: "support-faqs", "hr-policies-2026". Names are unique within a workspace, ignoring case. The name appears in the agent builder's knowledge-base step.
  2. 2

    Description

    One sentence on what's inside and who maintains it. This is what teammates see when they attach the dataset to an agent.
  3. 3

    Who can see it

    Everything in a workspace is shared with that team: every dataset can be seen, searched and attached to agents by everyone in the workspace, according to their role. Use a personal workspace for private material. Anyone who can chat with an agent it is attached to gets answers from it.
  4. 4

    OpenAI key

    Documents are indexed and searched with an OpenAI key from your workspace, never with Dezifi's. The form shows the key it will use: the workspace's knowledge-base key, which a workspace admin chooses with the "OpenAI key" button on the Knowledge Base page. Changing that key re-indexes nothing. Whoever may change the dataset (the person who created it, or a workspace admin) can give it its own key later; a personal key is allowed only on a dataset in your own personal workspace.
  5. 5

    Search engine

    Pick the search engine (stored in pgvector, embedded with an OpenAI model). The default fits most teams; see Embeddings & providers.

Create the dataset

Click Create. You land on the dataset detail page, empty, ready for uploads. The page shows which key it indexes with ("Indexing with: …").

Change it later

Open the dataset and click Edit to change its name or description. Only the people allowed to change the dataset see Edit: the person who created it, or a workspace admin or owner. Viewers can read it but not change it. A dataset declared in a DAK stack is changed in the stack, not in the dashboard.

When to split into multiple datasets

  1. 1

    Different topics

    Support FAQs and HR policies don't belong together — keep retrieval focused.
  2. 2

    Different freshness

    A daily-rotated dataset and an annual one should be separate so re-indexing one doesn't touch the other.
  3. 3

    Different access

    If only some agents should see certain content, isolate it in its own dataset and control attachment.

Frequently asked questions

Can I rename a dataset later?
Yes: Edit on the dataset page changes the name and description. Agents attach by the dataset's id, so renaming does not detach it. The new name must not be used by another dataset in the workspace.
Can I change the RAG provider after creating the dataset?
No. The search engine is fixed at creation because it determines how documents are embedded. To switch, create a new dataset and upload the files again.
How many datasets can a workspace have?
There is no hard limit. Most teams end up with five to twenty — one per topic or content owner. Splitting is cheap; over-merging hurts retrieval quality.