Knowledge Bases

📖 Overview · Last updated: September 17, 2026

<aside> 📚

WHAT IT DOES

A knowledge base is a searchable library your agents answer from. You gather your workspace content into one — or several — and connect it to an agent through a knowledge project.

</aside>

What a knowledge base is

Your content lives in your workspace (in Sources). A knowledge base gathers the content you choose into a single searchable index that an agent can look things up in. A piece of content can belong to more than one knowledge base, so you can organize the same material different ways — and deleting a knowledge base removes its index but keeps your original content safe.

The Knowledge Bases page — each library shows how many entries it holds and whether it uses vector or keyword search.

Creating a knowledge base

Select New knowledge base and fill in the create form:

Creating a knowledge base: give it a name, choose keyword or vector search and a split style, and set the passage size — the defaults suit most content.

Name (required) and Description — what this library contains.
Use vector search — off by default. Turn it on for meaning-based search that matches on intent, not just keywords; pick an embedding model or leave the workspace default.
Split style — how documents are broken into passages: Structural (splits on headings and paragraphs — fastest), Semantic (keeps related sentences together, and needs an embedding model even when vector search is off), or AI (an AI model decides the best split — highest quality, slowest; adds a Splitting model to pick). See Choosing a split style.
Evaluate passage quality — an optional extra AI pass that checks each passage as it's added.
Chunk size and Chunk overlap — how large each passage is and how much neighboring passages share. Overlap must be smaller than the chunk size. The defaults suit most content.
Advanced — an optional language hint and an extraction strategy for how text is pulled from documents.

Name and description can be changed later; the search settings (splitting and embedding) are set when the knowledge base is created.

Setting reference

Setting	Default	Change later?	Notes
Name	—	Yes	Required.
Description	—	Yes	What this library contains.
Use vector search	Off	No	Meaning-based search; needs an embedding model.
Embedding model	Workspace default	No	Used when vector search is on.
Split style	Structural	No	Structural (fastest), Semantic, or AI (best, slowest). See Choosing a split style.
Evaluate passage quality	Off	No	Optional AI check that scores each passage 1–5. Diagnostic only — nothing is discarded.
Chunk size	1024	No	Passage size, 64–8192 tokens.
Chunk overlap	0	No	Text shared between neighboring passages. Must be less than the chunk size.
Language / Extraction strategy	—	No	Optional Advanced hints for how text is read.

<aside> ⭐

TIP

The defaults are well chosen — you rarely need to touch chunking. The one setting worth a thought is vector search: turn it on for natural-language questions.

</aside>

Adding content to a knowledge base

From Sources, select the items you want and choose Add to knowledge base — or create a new one right there. Content is filed into the library without re-uploading. Once content is in, see Managing a knowledge base to reprocess it, test what it returns, connect it to your apps, and set rules for what's included automatically.