임베딩 유사도를 활용한 의미 기반 청킹

Semantic chunking is a technique that splits text at points where the topic changes significantly, rather than at fixed character counts. Instead of asking 'have we hit 500 tokens?', it asks 'does the next sentence belong to the same topic as the current chunk?' — using embedding similarity to answer that question.

The algorithm works by embedding each sentence (or small group of sentences) and computing the cosine similarity between consecutive sentence embeddings. When the similarity drops sharply — meaning the topic has shifted — the algorithm inserts a chunk boundary. Sentences that discuss the same concept stay together in the same chunk.

The first step is to split the document into individual sentences using a sentence tokenizer, then embed each sentence with a fast embedding model. You need sentence-level embeddings — not document-level — so you can detect local topic changes as you move through the text.

Once you have sentence embeddings, compute the cosine similarity between each consecutive pair: sentence i and sentence i+1. The result is a list of similarity scores, one per sentence boundary. Low scores indicate that the adjacent sentences cover different topics — these are your candidate split points.

A breakpoint is a sentence boundary where the similarity drops below a threshold. You can use a fixed threshold (e.g., 0.6) or a percentile-based threshold that adapts to the document — for example, split whenever similarity falls below the 25th percentile of all similarity scores in that document.

With breakpoints identified, you can now assemble chunks by joining consecutive sentences between each breakpoint. Each resulting chunk contains a coherent sequence of sentences about the same topic. The chunk boundaries align with natural topic transitions in the original document.

Putting it all together into a single function: embed sentences, compute adjacent similarities, find breakpoints, assemble chunks. The output is a list of semantically coherent text segments ready to be embedded as whole chunks and stored in your vector database.

The percentile threshold controls chunk granularity. A low percentile (e.g., 10th) means you only split at major topic shifts — resulting in fewer, longer chunks. A high percentile (e.g., 40th) splits more aggressively — resulting in many small, highly focused chunks. Tune this against your retrieval hit rate evaluation set.

LangChain provides a built-in SemanticChunker that implements this algorithm. It accepts an embedding model and a breakpoint threshold type. This saves you from implementing the algorithm from scratch and integrates directly with LangChain document loaders and vector stores.

Semantic chunking produces higher-quality chunks with better topic coherence, but it is more expensive: every sentence must be embedded just to determine chunk boundaries — before the chunks are even indexed. For a 100-page document, this means thousands of embedding calls just for chunking. Use semantic chunking when retrieval quality matters more than indexing speed.

Semantic chunking excels on long-form narrative content — blog posts, research papers, legal documents, and books — where topics shift organically. It is less necessary for highly structured documents like product catalogs, FAQ lists, or code files, which are better handled with document-aware splitters that respect the structure directly.

Test your understanding of semantic chunking from this lesson.

In this lesson you learned: semantic chunking uses embedding similarity to detect topic shifts, the percentile threshold controls granularity, and LangChain's SemanticChunker implements this out of the box. Next up we explore parent-child chunking, which combines small precise chunks with large context-rich parent passages.