Your RAG Pipeline Is Wrong About Tables, and the Fix Is grep

READ (arXiv 2608.06305, submitted August 6) took a 780-page government financial report and asked 51 verified questions. Top-k embedding retrieval answered 15.7% of them correctly.

The same agent loop, given three deterministic tools over MCP instead of a vector index, answered 58.8%.

The mechanism is not subtle once you see it. 86.8% of content lines in that document are table rows. A figure inherits its unit from a header a median of 13 lines above it. Your chunker splits between the header and the number, and now your retrieved context says "4,382" without saying whether that's lakh or crore. That's a two-order-of-magnitude error delivered with full confidence and a citation.

The replacement tools are almost insultingly simple: normalized lexical search, structural navigation, and bounded span reads. Search, walk, read. It's grep with a table of contents.

The control is what makes this a finding rather than an anecdote. The authors gave the same agent loop a top-k retrieval tool instead of the three deterministic ones, and it reached only 27.5%. So the gain isn't "agents beat RAG" or "iteration beats single-shot." It's the retrieval interface specifically. Statistically clean too: p_Holm = 2e-5, and 23.5 points ahead of a tuned dense baseline.

I've shipped pgvector RAG in production. Rayni does document intelligence and it chunks-and-embeds like everything else, and this paper describes a failure mode I've watched happen without having the vocabulary for it. Financial tables and structured reports were always the queries where users came back and said "that number's wrong" and I couldn't reproduce it reliably. It wasn't retrieval quality. It was that chunk boundaries destroy the relationship between a value and its unit.

This connects to two other findings today. Leaked Accenture audio, surfaced by Simon Willison via 404 Media, has client group lead Stuart Henderson naming PDF-to-image-to-markdown conversion as one of the company's biggest token expenses, confirmed against internal data by agentic AI strategy lead Justice Kwak. And PaDoc (arXiv 2608.06146) is the engineering answer, decoding layout and content in parallel rather than sequentially for 67 to 118% higher throughput and 39 to 55% lower P95 latency, with code released.

So: PDFs are eating enterprise token budgets, chunking them destroys the information you're paying to extract, and the fix is either better parsing or abandoning chunk-and-embed for search-and-navigate. Willison's take is that the real answer is upstream, PDFs being a terrible medium for communicating information in the first place, and he's right, and it doesn't help any of us shipping this week.