Comparison
ChunkingvsRow group
Chunking
you split the document into pieces before you store it, because embedding a whole 40-page PDF as one blob retrieves nothing useful.
Splitting documents into pieces before embedding, because embeddings represent one 'thought' well and a 40-page document poorly, and because retrieved chunks must fit a context budget. Chunking quality caps retrieval quality; garbage chunks, garbage answers.
Full entry →Row group
the file is split into blocks of a hundred and twenty-eight megabytes, each with its own min and max per column.
The unit of horizontal partitioning inside a columnar file, holding a range of rows across all columns along with per-column statistics. It is the unit of skipping and the unit of parallelism, so its size sets both how precisely a reader can avoid data and how evenly work spreads. Very small row groups defeat the statistics and add overhead; very large ones mean reading far more than the query needed.
Full entry →