Chunker

SemanticChunker

Class for splitting markdown notes into contextualized chunks using the Anthropic Contextual Retrieval pattern.

Constructor

SemanticChunker(mode: ChunkMode = "headers", chunk_size: int = 512, chunk_overlap: int = 0)

• mode (ChunkMode): The splitting mode, one of:
• "headers" (default): Split by H2 (##) and H3 (###) headers.
• "paragraphs": Split by double newlines.
• "fixed": Split by fixed character counts.
• chunk_size (int): Target character count for "fixed" mode. Default is 512.
• chunk_overlap (int): Overlap character count for "fixed" mode. Default is 0.

Methods

chunk(content: str, title: str = "", description: str = "") -> list[str]

Split raw markdown content into chunks, prepending document-level context to each chunk.

• Parameters:
• content (str): Raw markdown document (frontmatter is automatically stripped).
• title (str): Optional document title.
• description (str): Optional document description.
• Returns: A list of chunk strings, each prefixed with [Document: {title} | Description: {description}]\n.