Understanding Knowledge Base

A knowledge base is one named container

A knowledge base is a single named container that belongs to your organization, optionally scoped to one workspace. It holds two things: the documents you uploaded, and the indexed form of their text that the platform builds from them. Each document is its own record, with its own name and its own status, but they all live inside the one container you named when you created it.

Indexing happens once, at upload. When you add a file, the platform reads it and extracts its text right away. It does not wait for something else to ask for that content first. That single pass is why the container carries a status of its own:

• Processing: its files are being read
• Completed: its files are usable
• Failed: a file’s extraction didn’t work; the reason shows on that file’s row

Three surfaces read it

• Synthetic data generation (generates dataset rows from a schema you define) points at a knowledge base by name so the rows it produces echo your domain instead of reading like generic text
• Agent-type evaluations (evals authored as the Agent Evaluator type, which can reason over multiple turns and use tools) attach a knowledge base to give the eval a reference to check the agent’s output against
• A Simulation agent definition (the config that governs how a simulated agent behaves) can attach one knowledge base of its own, giving the simulated agent something to draw on when it answers

flowchart TD
accTitle: How a knowledge base sits between your documents and the surfaces that read it
accDescr: A knowledge base holds the documents you uploaded and their indexed text. Synthetic data generation, agent-type evaluations, and a Simulation agent definition each read from that knowledge base.
KB["Knowledge base"] -->|"holds"| DOCS["Documents you uploaded"]
KB -->|"holds"| IDX["Indexed text"]
SDG["Synthetic data generation"] -->|"reads"| KB
EVAL["Agent-type evaluation"] -->|"reads"| KB
SIM["Simulation agent definition"] -->|"reads"| KB

Not your agent’s retrieval store

Read it as reference material, not as your agent’s live retrieval path. When your production agent answers a real request, it is not opening this container and searching it. A knowledge base is a document store that the three surfaces above consult when they run. There is no versioning either: a knowledge base holds whatever documents are in it right now, not snapshots of what it held before. If you’re looking for a live index your agent queries during retrieval, that’s the dataset’s Retrieval column, which connects to Pinecone, Qdrant, or Weaviate, not to a knowledge base.

Why it matters

Without a knowledge base behind it, a generator or an evaluator has nothing to ground itself in: ask for synthetic data on a topic you never described and it produces plausible wording that isn’t yours. Point that generation, or an agent-type eval, at a knowledge base and it draws on your actual wording and your actual steps instead.

Limits and supported files

• Supported file types: PDF, DOCX, TXT, and RTF only, anything else yields no extracted text
• Size cap: 1 GB per knowledge base, across all its documents combined

Keep exploring