Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MDKeyChunker: Single-Call LLM Enrichment with Rolling Keys and Key-Based Restructuring for High-Accuracy RAG

Bhavik Mangla

Abstract

RAG pipelines typically rely on fixed-size chunking, which ignores document structure, fragments semantic units across boundaries, and requires multiple LLM calls per chunk for metadata extraction. We present MDKeyChunker, a three-stage pipeline for Markdown documents that (1) performs structure-aware chunking treating headers, code blocks, tables, and lists as atomic units; (2) enriches each chunk via a single LLM call extracting title, summary, keywords, typed entities, hypothetical questions, and a semantic key, while propagating a rolling key dictionary to maintain document-level context; and (3) restructures chunks by merging those sharing the same semantic key via bin-packing, co-locating related content for retrieval. The single-call design extracts all seven metadata fields in one LLM invocation, eliminating the need for separate per-field extraction passes. Rolling key propagation replaces hand-tuned scoring with LLM-native semantic matching. An empirical evaluation on 30 queries over an 18-document Markdown corpus shows Config D (BM25 over structural chunks) achieves Recall@5=1.000 and MRR=0.911, while dense retrieval over the full pipeline (Config C) reaches Recall@5=0.867. MDKeyChunker is implemented in Python with four dependencies and supports any OpenAI-compatible endpoint.

MDKeyChunker: Single-Call LLM Enrichment with Rolling Keys and Key-Based Restructuring for High-Accuracy RAG

Abstract

RAG pipelines typically rely on fixed-size chunking, which ignores document structure, fragments semantic units across boundaries, and requires multiple LLM calls per chunk for metadata extraction. We present MDKeyChunker, a three-stage pipeline for Markdown documents that (1) performs structure-aware chunking treating headers, code blocks, tables, and lists as atomic units; (2) enriches each chunk via a single LLM call extracting title, summary, keywords, typed entities, hypothetical questions, and a semantic key, while propagating a rolling key dictionary to maintain document-level context; and (3) restructures chunks by merging those sharing the same semantic key via bin-packing, co-locating related content for retrieval. The single-call design extracts all seven metadata fields in one LLM invocation, eliminating the need for separate per-field extraction passes. Rolling key propagation replaces hand-tuned scoring with LLM-native semantic matching. An empirical evaluation on 30 queries over an 18-document Markdown corpus shows Config D (BM25 over structural chunks) achieves Recall@5=1.000 and MRR=0.911, while dense retrieval over the full pipeline (Config C) reaches Recall@5=0.867. MDKeyChunker is implemented in Python with four dependencies and supports any OpenAI-compatible endpoint.
Paper Structure (42 sections, 2 equations, 4 figures, 7 tables, 2 algorithms)

This paper contains 42 sections, 2 equations, 4 figures, 7 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Chunk boundary fragmentation.
  3. Metadata extraction cost.
  4. Contextual isolation.
  5. Contributions.
  6. Related Work
  7. Chunking Strategies for RAG
  8. Semantic Enrichment
  9. Graph-Based and Structured RAG
  10. Comparison of Approaches
  11. Methodology
  12. Stage 1: Markdown Structural Splitting
  13. Block parsing.
  14. Atomicity constraints.
  15. Size management.
  16. ...and 27 more sections

Figures (4)

  • Figure 1: MDKeyChunker three-stage pipeline: structural Markdown splitting (Stage 1), single-call LLM enrichment with rolling key propagation (Stage 2), and key-based bin-packing restructuring (Stage 3).
  • Figure 2: Rolling key propagation: each chunk enrichment call receives the accumulated key dictionary $K$, enabling the LLM to reuse prior keys (e.g. "admissions process" introduced at $c_2$ is reused at $c_5$) instead of coining synonyms.
  • Figure 3: Chunk-size distributions. Config A (fixed-size) concentrates 99% of chunks at exactly 512 chars. Configs B and C (structural splitting) produce variable-length chunks; Config D shares Config B's chunk set and is labeled on that panel.
  • Figure 4: Retrieval performance across four configurations. Config D (BM25 over structural chunks) achieves perfect Recall@5 and Recall@10; Config A (dense, fixed-size) leads among dense configs at all cut-offs.