Semantic Chunking
Also called Meaning-based Chunking
Splitting text where the topic shifts, not at fixed lengths, so each chunk is one coherent idea.
Think of it like
Cutting a movie into scenes by where the story turns, not every ten minutes on the clock.
Example
A doc about "install, then configure, then troubleshoot" gets split at each topic change, so retrieval returns just the troubleshooting part.
How it actually works
Semantic chunking embeds consecutive sentences and places a boundary wherever similarity drops — the signal that the topic changed. The payoff is chunks that align to ideas rather than arbitrary token counts, which usually improves retrieval. The cost is more compute up front and sensitivity to the threshold, so it is not always worth it over simpler recursive splitting.
For product teams
Can lift retrieval quality on messy docs; adds preprocessing cost, so measure before adopting.
For engineers
Boundary placement via drops in inter-sentence embedding similarity; threshold-sensitive and compute-heavier than fixed splits.
Related
- Chunk Size — A smarter alternative to fixed chunk size.
- Embeddings — Uses embeddings to detect topic shifts.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome