Reducto

The industry lacks a unified benchmark for document extraction evaluation, and customers often need to build their own evaluation sets; Reducto says it will release a benchmark in the near future.

Omar mentioned having seen a schema containing 156 fields, with nesting limited to four or five levels.

In addition to Reducto's extraction model, there is a dedicated citation model responsible for locating the source of each piece of data in the original text.

Over the past three years, Reducto has raised $108 million from institutions including ACC and Benchmark.

Reducto has processed more than 2 billion documents.

Reducto's customers include Harvey, Scale, and one of the world's top five hedge funds.

Reducto adopts a rasterization-based technical approach, treating documents as images and not relying on information embedded in the PDF's underlying layer.

The Reducto platform provides five core capabilities: parse, split, extract, edit, and classify.

Reducto was responsible for parsing more than 3 million pages of redacted PDFs, scanned documents, and government files in the J-Mail project.

After parsing PDF metadata, Reducto found that the redaction of some files was incomplete, and the underlying information could still be read.

Reducto's schema generation offers two modes: the standard mode generates a JSON schema directly with a single LLM call; the enhanced mode first parses the document content and then customizes the schema description.

Reducto is a US document parsing API company founded around 2023, has raised a cumulative $108 million, and its extraction is Schema-driven and zero-shot, but it has no UI.

Agentic OCR has the ability to self-audit and correct failures; when recognition results are suspicious, the system can go back to re-examine and reprocess, forming a closed-loop error correction mechanism.