D-RAC turns PDFs into retrieval Markdown, cutting chunking output tokens 95.7%

Document Retrieval-Aware Chunking (D-RAC): Universal Retrieval-Aware Ingestion of Enterprise Documents via PDF Normalization and Multimodal Markdown Conversion

Uday Allu, Abhivanth Sivaprakash, Pratik Singh, Aman Manocha

AI Research Team Yellow.ai

cs.CV

2026-09-21

D-RAC normalizes enterprise files to PDF, converts pages to retrieval-oriented Markdown in one multimodal pass, then plans chunks over element IDs. On 236 PDFs (795 pages) it builds 1,748 chunks in 72 minutes and cuts chunking output tokens 95.7% versus agentic rewriting.

What problem this solves

Enterprise knowledge bases are mostly PDFs, Word files, slide decks, and scans, not clean HTML. Rule-based extractors scramble reading order, flatten tables, and drop heading hierarchy. Fully agentic chunking over that broken text regenerates the document as output tokens, which is expensive and invites silent rewrites. Yellow.ai's earlier W-RAC framed web chunking as planning over element IDs rather than regenerating prose, but it assumed recoverable markup. PDFs do not provide that markup.

D-RAC asks whether one multimodal conversion pass can turn any renderable file into Markdown that W-RAC can consume unchanged.

Method

Four stages.

• Normalize: office files, HTML, and scans are converted to PDF with standard tooling, then rendered at 200 DPI with a 1,568-pixel cap.
• Convert: pages go to Gemma-3 27B on AWS Bedrock in batches of five. The prompt keeps text verbatim, forbids Markdown tables, writes each table row as a self-contained sentence that names the columns, drops decorative images, rebuilds heading levels, and tags page numbers. A rule-based fallback catches leftover tables.
• Parse and section: Markdown becomes ID-tagged headers and paragraphs. Documents over a 60-element budget split on heading boundaries, carrying ancestor headers as context.
• Plan: the model sees IDs, truncated previews, and hierarchy, and returns ID arrays targeting 3–8 content blocks per chunk. Coverage is checked in code; missing IDs fall into a backup chunk. Final text is concatenated from the converted source and prefixed with breadcrumbs.

Retrieval awareness is pushed into conversion: one embeddable sentence per row, no merged disjunctions such as "16 or 20 years." Charts are dropped so hallucinated captions never enter the index. Persisted Markdown can be re-chunked at ID-planning cost.

Results

Evaluation uses the authors' RAG-Multi-Corpus: 236 PDFs, 795 pages, five fictional enterprise domains. Inputs are already PDF, so Stage 1 is identity. Gemma-3 27B, temperature 0.1, five parallel workers.

Wall-clock conversion plus chunking is 71.7 minutes. Conversion totals 3,758 s (about 4.7 s/page), planning 542 s, with zero conversion or chunking errors and 1,748 chunks. A 503-page prospectus converts in 21.6 minutes on the 27B model; planning 5,060 elements takes 68.7 s.

Retrieval uses Titan Text Embeddings V2 on 762 queries with supporting-fact labels. A chunk counts as relevant if 60% of a fact's content words appear in it.

D-RAC matches or exceeds agentic chunking on all seven overall metrics. Boolean queries are the category where the agentic reference still leads. Chunking-stage output tokens fall from 270,454 to 11,714 (95.7% less). At list prices, GPT-4.1 chunking costs $2.815 versus $0.624 (77.8% less) and Gemini 2.5 Pro $3.112 versus $0.448 (85.6% less). Planning time is 75% below the authors' earlier agentic measurement.

Why it matters

PDF ingestion is the unglamorous bottleneck in enterprise RAG. D-RAC pays for multimodal understanding once, then chunking becomes ID planning, so retrieval-policy changes do not require another OCR pass. Row-level table prose targets the usual dense-retrieval failure on cells stripped of headers.

The dollar figures cover the chunking stage only. Conversion is the visual call and is billed separately. For teams already on W-RAC, this is a format extension, not a new retriever.

Limitations

The headline experiments never ingest DOCX, PPTX, or scans; format-agnosticism is a design claim, not a result on this table. The corpus is the authors' own fictional suite. Agentic reference chunks were built from clean structured sources while D-RAC read rendered pages, which is a harder input, yet transfer to messy real KBs still needs a separate test. There is no standalone OCR or structure metric, only end-to-end retrieval. Figures are discarded, so questions that need chart values lose evidence. Prose rows may hurt aggregation queries that want the whole table. The prompt says "verbatim"; Gemma-3 can still rewrite. "Zero errors" counts API failures, not factual fidelity.

Terms

• W-RAC: Web Retrieval-Aware Chunking, which plans over element IDs instead of regenerating text.
• Retrieval-aware table normalization: One self-contained sentence per row, carrying column headers.
• ID planning: The chunker emits identifier arrays; source text is reconstructed locally.
• RAG-Multi-Corpus: The authors' multi-domain enterprise retrieval benchmark; this paper uses the PDF subset.

Source

• arXiv 2609.24220v1
• PDF
• Hugging Face

Related papers

• Mira-Scene places objects via pixel-aligned CCM, lifting 3D-IoU 39.8% over SAM3D
• Per-document cliff detection compresses multi-vector retrieval 75% at a 0.002 nDCG cost
• VideoGen-Agent learns tool use and lifts VABench from 56.5 to 75.6
• Tencent ARC's 5,000-hour AAA GameHorizon suite finds action planning is the real bottleneck
• WorldCrafter's pose-queryable 3D memory cuts revisit LPIPS from 0.487 to 0.255
• Video diffusion breaks physics because RoPE spatially locks trajectories too early

All paper explainers