10. Semantic Chunking

Semantic chunking places boundaries where the topic changes, not at fixed character counts — so each chunk is one coherent idea.

How it works

 1. split doc into sentences
 2. embed every sentence
 3. walk neighbours; measure similarity between consecutive sentences
 4. start a NEW chunk wherever similarity drops below a threshold (topic shift)

 sent: A  A  A | B  B | C  C  C
 sim:  hi hi  ↓lo  hi ↓lo  hi hi
              break    break
 → chunks: [A A A] [B B] [C C C]

Code

# pip install langchain-experimental
from langchain_experimental.text_splitter import SemanticChunker
from langchain_huggingface import HuggingFaceEmbeddings

splitter = SemanticChunker(
    embeddings=HuggingFaceEmbeddings(model_name="all-MiniLM-L6-v2"),
    breakpoint_threshold_type="percentile",   # break at biggest similarity drops
)
chunks = splitter.create_documents([text])

The trade-off

It produces noticeably more coherent chunks, but it embeds every sentence just to decide the splits — so for a long document you generate hundreds of extra embeddings at indexing time. That's more compute/cost and slower ingestion.

Reach for semantic chunking when recursive plateaus on your eval set and the indexing-time cost is acceptable. Otherwise recursive is the better default.

Next: Agentic Chunking →