RCRAG Converter

Why not just send the PDF to the model?

Sending a PDF straight to a vision model works for a short document you will read once; it is not a RAG system, because cost scales with every page, citations have nowhere to land, and nothing is reusable across questions.

The demo and the system are different jobs

Unstructured's own FAQ opens with this question, because it is the one every team asks after their first Claude or GPT vision demo. A model that can see a page can answer a question about that page. That is a one-shot read, not retrieval. The moment you have a folder, a second question, or a file longer than the context window, you are paying to re-send every page you already paid to send, and you still cannot point at the passage that justified the answer.

Cost grows with pages, not with questions

Vision tokens are charged on the pixels, whether or not the page contained anything useful. A 80-page filing asked three times is billed three times. A converter does the structural work once: parse, chunk, embed. Every later question is a cheap vector lookup plus a handful of text passages. That is why hosted parsers exist as a product at all - LlamaParse, Reducto and Unstructured all sell the same bargain: pay for understanding up front so you stop paying for raw pages on every turn.

A page the model saw is not a citation

A vision answer can be right and still be useless in production, because you cannot show a reviewer which cell or which paragraph it used. Conversion produces an id, a source file, a page when the format has one, and a passage you can print. That is the difference between 'the model said so' and a grounded answer. If you need the former, send the PDF. If you need the latter, convert first.

Scans, tables and anything past the window still break

Vision helps on a clean digital page and still fails the cases conversion was invented for. A scanned PDF is a stack of photographs; the model is doing OCR under another name, with no layout you can audit. A wide table is often cropped or summarised. A 200-page manual does not fit. Convert those files into passages you own, then let the model read the few that matter. RAG Converter does that conversion in the tab for text-layer documents, so the file never has to go to a model provider to become searchable.

More on converting for RAG

  • rusty_ragdocument ingest, chunking and retrieval
  • rusty_xmlWord, PowerPoint and Excel, without LibreOffice
  • rusty_jpeg / rusty_pngimage decode
  • rusty_flacaudio decode
  • rusty_zstdcompression
  • rusty_allocthe allocator, in wasm too
  • rusty_timeNTP timestamps on every chunk
  • spacedbthe storage seam, CRDT per entry
  • FFAIthe AI toolkit the three engines below come from
  • CarmentaOCR - reads scans and screenshots, in your browser
  • Mercuryspeech recognition - turns recordings into transcripts
  • Argusvision - describes images and video frames

MATA Corporate Holdings © 2026 — Made in Scottsdale, Arizona

Remade with Rust