SplitBench Wants to Put RAG Chunking Under Test

Field notes

The proposed workflow accepts document sets or repository sources, runs multiple chunking strategies, and compares their boundaries and retrieval results with a baseline. Its mockup shows a 1,280-token document split into three topic-labelled chunks, alongside Recall @ 5, context preservation, and table-boundary metrics. The initial beta is described as targeting Markdown and code documentation in GitHub Actions, with compatibility plans for several vector databases and embedding providers. The page identifies itself as a design mockup, not a working product.