A single-pass pipeline reports 78% to 86% cheaper chunking for RAG ingestion

single source· 1 articles · confidence: medium · first seen 2026-09-20 20:00 UTC

What this means for you

If you pay a frontier model to chunk enterprise documents, this is a number to watch, not to act on: no code, licence or run date is given, and the comparison is against the authors' own agentic baseline on one 236-document corpus. Nothing to install today; worth re-testing on your own PDFs if a release appears.

A preprint posted to arXiv on 20 September describes a pipeline that normalises any renderable file format to PDF, uses one multimodal model pass to turn pages into Markdown — tables rewritten as prose — then chunks the result deterministically over identifiers rather than text, never regenerating source wording. On a 236-document, 795-page PDF subset of the RAG-Multi-Corpus benchmark, the authors report the corpus converted and chunked in 72 minutes with no errors, 1,748 chunks, 95.7% fewer chunking-stage output tokens than agentic chunking with frontier models, cost down 77.8–85.6% depending on the pricing used, and 75% less time. No run date, code or peer review is given.

Key facts

  • ·The paper reports converting and chunking a 236-document, 795-page PDF subset of the RAG-Multi-Corpus benchmark in 72 minutes with zero errors, producing 1,748 chunks. source
  • ·It reports 95.7% fewer chunking-stage output tokens than agentic chunking with frontier LLMs. source
  • ·Chunking cost falls 77.8% at GPT-4.1 pricing and 85.6% at Gemini 2.5 Pro pricing, with chunking time down 75%. source
  • ·The method normalises any renderable input format to PDF, then uses one multimodal LLM pass to render pages as retrieval-optimised Markdown, rewriting tables as prose statements. source
  • ·Chunking is deterministic parsing into ID-addressable units with lightweight LLM chunk planning over identifiers, so source text is not regenerated. source
  • ·Posted as arXiv 2609.24220 on 20 September 2026; the input gives no code release, no benchmark run date and no peer review. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire