|

Reducto Releases r-1: A Single Pass Document Parsing Model That Cuts Errors 20% at 1 Cent Per Page

ByRicardo

Last week, Reducto introduced r-1. It is the primary mannequin in a brand new parsing household constructed on a rewritten structure, and it replaces the corporate’s multi stage agentic OCR with one full web page go. Reducto says r-1 is extra correct than its strongest legacy agentic fashions, sooner, and as much as 6x cheaper.

Is it deployable at present? Yes, in preview. r-1 runs by way of Reducto’s hosted Parse API on V3 and is switched on with a configuration flag. There are not any open weights and no native checkpoint to self host. Reducto’s platform individually helps multi tenant cloud, buyer VPC, on premises, and air gapped installs, with SOC 2 Type II attestation and HIPAA processing on larger tiers, per its security policies.

One go as an alternative of a pipeline

Legacy Parse runs OCR, format detection, and put up processing as separate levels, with non-obligatory agentic imaginative and prescient language passes layered on high. Each additional mannequin name provides latency. r-1 folds textual content, tables, figures, format, studying order, formatting, and grounding right into a single full web page go. Every block returns with web page relative bounding packing containers that tie content material again to its place on the web page.

The consolidation is the true product declare. Teams engaged on monetary statements, insurance coverage claims, or contracts usually route information throughout a number of suppliers and bolt on put up processing to succeed in usable accuracy. r-1 targets that orchestration price, not solely uncooked character accuracy.

The numbers

Reducto studies a 20% discount in error charge for the early r-1 preview in opposition to its personal legacy agentic pipelines. It additionally states that r-1 outperformed generally used hyperscaler merchandise and huge LLMs on complicated paperwork in inner evaluations. Amazon Textract and Azure Document Intelligence are named within the launch because the incumbent baseline class.

On value, legacy agentic fashions ran between 3 and 6 cents per web page relying on workload. r-1 is 1 cent per web page, all in, with no characteristic multipliers or credit score prices layered on to succeed in excessive accuracy. Reducto frames this as a part of a wider transfer to flat product charge playing cards.

Please be aware that the 20% error discount is measured relative to Reducto’s personal prior pipeline, to not a 3rd occasion baseline. And the face to face in opposition to hyperscalers and LLMs is vendor run, with no public eval harness or dataset launched alongside the announcement.

What r-1 resolves on the web page

Per the documentation, r-1 handles the next natively within the full web page go: digital textual content, scans, and handwriting; desk construction learn with surrounding web page context, together with merged cells and nested headers; columns, headers, footers, sidebars, and studying order resolved collectively; determine detection with a brief generated description; formatting that carries that means, together with headings, lists, daring, underlines, and strikethroughs; and grounding by way of web page relative bounding packing containers.

The lengthy tail instances Reducto calls out are dense tables, uncommon layouts, low high quality scans, watermarked content material, and paperwork that observe no predictable template. A dropped strikethrough can invert a contract clause, and a misinterpret desk can hand an agent the flawed determine, which is why these edge instances carry weight in regulated pipelines.

(*1*)

Migration path and what stays on the previous stack

r-1 requires the V3 API. A Parse request that omits settings.mannequin nonetheless runs on legacy Parse, so nothing breaks silently. New pipelines created in Studio default to r-1.

Copy Code
consequence = consumer.parse.run(
    enter=add.file_id,
    settings={"mannequin": "r-1"}
)

Agentic processing isn’t gone. Workflows needing customized prompts or superior chart extraction nonetheless route these pages by way of the agentic pipeline, which Reducto augments onto the r-1 consequence, and which provides latency. Teams migrating an present configuration ought to test the r-1 configuration compatibility web page first, since some legacy settings are ignored or unsupported.

Reducto has additionally signalled the following two items: r-1 mini for pace and value delicate workloads, and computerized per web page routing that picks the fitting mannequin for every web page. Organisations shifting off one other parser can request up to $5,000 in credits for a facet by facet comparability.

Key Takeaways

  • r-1 collapses OCR, format, tables, formatting, and grounding into one full web page go.
  • Reducto studies a 20% error discount in opposition to its personal legacy agentic pipelines.
  • Pricing drops from 3 to six cents per web page to a flat 1 cent, all in.
  • Available in preview through the V3 Parse API utilizing settings.mannequin: "r-1", no open weights.

Check out the Reducto blog, Parse r-1 docs and announcement on X. Also, be happy to observe us on Twitter and don’t overlook to affix our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

Need to companion with us for selling your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar and so forth.? Connect with us

The put up Reducto Releases r-1: A Single Pass Document Parsing Model That Cuts Errors 20% at 1 Cent Per Page appeared first on MarkTechPost.

Similar Posts