A long document cannot be handed to the model whole, so it is split into consecutive chunks. A naive cut at a fixed length can land in the middle of a sentence, leaving one chunk with a fragment that no longer carries its meaning and the next chunk with the rest. Overlapping consecutive chunks — repeating a little of the previous chunk at the start of the next — keeps any sentence or idea whole inside at least one chunk.
How AI Answers Questions About Documents It Never Learned
Chunking and Embedding: Turning Documents Into Something Searchable
Cutting a Document Into Chunks
1 / 4
The pipeline starts with a document that is far too long to hand to the model in one piece. So it gets cut into chunks, consecutive slices of text stacked one after another.
0:00 / 0:00