Try Deep Extract for Free

New customers can get up to $1,000 in free credits to try Deep Extract on your hardest extraction use cases.

Deep Extract ranked first overall in micro1's independent document extraction benchmark — 100% completion and 99.6% precision and recall across 225 real-world documents. Full results below.

Trusted by companies worldwide

Vanta logoScale AI logoHarvey logoToast logo
Air quote icon

"With Reducto, we tripled our processing speed and built a pipeline that accurately processes millions of documents per month, all while reducing engineering time spent on chunking by 90%."

— Co-founder, Legora

Document grid

What is Deep Extract?

Deep Extract is the highest-accuracy extraction mode on Reducto's agentic document platform. Where single-pass extraction reads a document once and hopes for the best, Deep Extract plans, extracts, and verifies its own output across the full document — built for invoices, financial statements, and claims where every field matters.

Agentic verification

Instead of a single pass, Deep Extract runs an agentic loop that re-reads the document and verifies its own output, so hundreds of line items come back complete — no truncated arrays, no skipped pages.

Citations on every value

Every field is returned as typed JSON with a citation and bounding box back to the exact page it came from, so every value in your pipeline is auditable.

No templates or retraining

Define a schema and go. Deep Extract handles complex tables, scans, handwriting, and 30+ file types out of the box — no per-document templates to maintain.

Independently benchmarked, first overall

We commissioned micro1, an independent data lab, to build and publish the first benchmark for complex structured extraction: 225 real documents — invoices, financial statements, and claims, averaging 88,700+ fields each — with a human-validated answer key. micro1 sourced every document and published the results. Six systems were tested: Reducto Deep Extract, Extend MAX, LlamaExtract-Agentic, GPT-5.5, Claude Opus 4.8, and Gemini 3.1 Pro.

100%

Coverage

225 of 225 documents completed

99.6%

Precision

Returned values that match ground truth

99.6%

Recall

Required fields actually found

99.3%

Leaf accuracy

Individual cell values correct

What the benchmark measures

Coverage

Whether a system can complete a full extraction for every document, or fails and returns nothing.

Precision

The fraction of what a system returned that is actually real and exists in the ground truth.

Recall

The fraction of what should have been returned that the system actually found.

Leaf accuracy

The fraction of individual cell values that match the answer key on the rows a system got right.

Bar chart of benchmark coverage: Reducto Deep Extract completed all 225 documents while every other system failed on some
Deep Extract completed every document in the benchmark. Every other system failed on at least some.
Bar chart of benchmark recall across systems, with Reducto Deep Extract at 99.6%
Recall: Deep Extract found 99.6% of required fields. GPT-5.5 returned 52.7%; Gemini 3.1 Pro returned 48.6%.
Bar chart of leaf accuracy across systems, with Reducto Deep Extract at 99.3%
Leaf accuracy: 99.3% of individual cell values matched the human-validated answer key.
Full benchmark results table comparing Reducto Deep Extract, Extend MAX, LlamaExtract-Agentic, GPT-5.5, Claude Opus 4.8, and Gemini 3.1 Pro
Full results across all six systems tested by micro1.

See the same results on your documents

Get up to $1,000 in free credits and run Deep Extract on your hardest extraction use cases.

Search Reducto