Introducing s-1: Sievly’s new document reading model
All posts
Engineering
Layout-first chunking: why we split by structure, not tokens
Fixed-size chunks cut tables in half and separate headers from their rows. Here is how Sievly decides where one idea ends and the next begins.
Rafael Ström
·
Machine Learning Engineer
sievly
Layout-first chunking
Retrieval quality depends on what goes into each chunk. Most pipelines split text every few hundred tokens, which is simple and wrong for real documents.
The problem with fixed windows
A fixed window does not know that a table is one unit, that a caption belongs to its figure, or that a heading sets the context for the paragraphs below it. Splitting through those boundaries creates chunks that look fine alone and mislead a model in context.
Reading structure first
Sievly builds a layout tree for every page: sections, headings, tables, figures, lists and running text. Chunks follow that tree, so a table travels with its header row and a figure with its caption.
What it means for retrieval
Answers come back with the context that makes them true. We publish the chunking options in the API so you can tune size and overlap for your own retrieval setup.
More from the blog
sievly
Introducing s-1
Announcements
Introducing s-1, our new document reading model
s-1 reads PDFs, scans and spreadsheets the way a careful analyst does, and returns structured data with a citation for every value. It is available in preview today.
sievly
Reading order
Research
Research note: reading order in complex layouts
Multi-column pages, sidebars and callouts confuse simple reading order. We share what we learned teaching s-1 to follow the page the way a person would.
sievly
Quarryline × Sievly
Case Studies
How Quarryline turns research notes into a searchable knowledge base
Quarryline’s data team uses Parse and Extract to turn thousands of analyst notes, filings and spreadsheets into clean, cited records their AI assistant can trust.