Choosing a chunking strategy for the content type
Key Takeaway
Each chunking strategy fits a different kind of content. For example, section-aware chunking follows a document's titles and headings, and conversation-based chunking is for transcribed audio and video.
Chunking is how Data 360 splits content into smaller pieces, called chunks, before it turns each chunk into a vector in a search index. Retrievers search these chunks to ground prompts and agents. So the way content is split decides what the LLM can find and how much context it gets.
How Data 360 chooses a chunking strategy
Data 360 automatically chooses the optimal chunking strategy based on the content type. Even so, you still plan three things:
- Which fields carry the most relevant context and meaning
- How to break those fields up
- How the retrieved results will be used