MDKeyChunker: What Does One LLM Call per Chunk Buy for Markdown Retrieval?
MDKeyChunker improves RAG for Markdown by replacing multiple extraction passes with one structure-aware LLM call per chunk. It first splits documents based on structural elements like headers and code blocks. Then, it enriches each chunk with metadata using a single call while maintaining document context via a rolling key dictionary. Finally, it merges semantically related chunks into larger units, significantly reducing total chunks and improving retrieval accuracy.