A knowledge base gives your workflows searchable memory. You put in large volumes of text, and an agent retrieves just the parts that matter — by meaning, not keywords.
What you will learn
The context problem
You cannot stuff everything into a model's context. A knowledge base lets you search a large corpus and pass back only what is relevant.
What is inside one
Documents and connectors that sync them, automatically indexed into searchable chunks you can query by meaning.
How it grounds an answer
A search returns the most relevant chunks with similarity scores; feeding those into the model produces accurate, cited answers.
The context problem
The naive way to give a model knowledge is to paste everything into its context. That breaks down fast — too much text, too much cost, too much noise. A knowledge base is the answer: store the corpus once, and at run time pull back only the passages that bear on the question.
Here's the shape of it — a workflow that searches a knowledge base and grounds the answer in what comes back:
Inside a knowledge base
A knowledge base is a resource that lives in your workspace — you can have several. Open one and it holds documents, plus connectors that bring in and keep documents in sync from outside sources.
Sim automatically indexes and embeds every document into searchable chunks, so the contents become queryable by meaning — you don't manage any of that yourself.
Searching by meaning
Use a Knowledge block with a query and it returns the most relevant chunks, each with a similarity score — how closely it matches the meaning of the query, not its words. So "refund timelines" can match a passage that says "we process returns within 14 days," even with no shared words.
Grounding an answer
On its own, that's retrieval. The power comes one step later: feed those chunks into an Agent block's context, and the model answers from your actual documents — accurate, up to date, and able to cite where each fact came from. That's grounding plus retrieval.
Knowledge bases power systems that need to give agents accurate, current context — and they're where you store what your systems learn over time.