Choosing a split style

📋 Reference

<aside> 🧩

WHAT IT DOES

Split style decides how your documents are cut into the short passages an agent searches and quotes. Three styles are available — Structural, Semantic, and AI — trading speed and cost against how cleanly each passage holds one complete idea. You choose it once, when you create a knowledge base.

</aside>

Why splitting matters

An agent never reads a whole document to answer a question. It searches for the most relevant passages and writes its answer from those. So the quality of the split sets the ceiling on answer quality.

A good passage holds one complete thought. A bad one starts halfway through a sentence about park hours and ends in the middle of a paragraph about permits — the agent then quotes half an answer, or misses the passage entirely because it reads as two unrelated topics at once.

The three split styles

Style	How it decides where to cut	Speed & cost	Best for
Structural (default)	Follows the shape of the document — headings, paragraphs, then sentences. Markdown files are split on their heading structure.	Fastest. No AI cost.	Well-formatted material: web pages, FAQ lists, anything with clear headings. Start here.
Semantic	Compares the meaning of neighboring sentences and starts a new passage where the subject changes.	Moderate. Reads every sentence once.	Long, flat documents with few headings — a transcript, meeting notes, a wall-of-text policy.
AI	An AI model reads the document in overlapping sections and picks the points where the topic genuinely turns.	Slowest and most expensive.	High-value, low-volume material: ordinances, contracts, formal policy. Not for bulk content.

Every style respects your Chunk size limit. If a passage would come out too long, it is split further using the Structural method, so no passage ever exceeds the cap.

What each style needs

Picking a style can add a second field to the form:

Structural — nothing extra.
Semantic — needs an embedding model. Leave it on the workspace default unless you have a reason to change it. This is required even when vector search is off, because the style uses sentence meaning to find its cut points.
AI — adds a Splitting model. Leave it on the built-in model unless you want a specific one.

Passage size and overlap

Two more settings shape the result, and the defaults suit most content:

Chunk size (default 1024, range 64–8192) — roughly how much text one passage may hold. Smaller passages are more precise; larger ones carry more surrounding context.
Chunk overlap (default 0) — how much text neighboring passages share, so a sentence at a boundary isn't stranded. It must be smaller than the chunk size. A modest overlap helps when your content runs continuously with few natural breaks.

Evaluate passage quality is an optional extra check. When it's on, an AI model scores every passage from 1 to 5 with a one-line reason. It's purely a diagnostic — nothing is discarded because of a low score — and it adds time and cost to processing.

Checking your split

After your content is processed, open a knowledge base entry and choose Passage preview to see exactly what was produced. Four figures tell you whether the split worked: