Document Intelligence
Enterprise intelligence lives in messy documents. Document Intelligence is an agentic system that extracts the facts, on your own infrastructure with no per-page fees.
schema = {
"total_revenue": "number",
"cloud_revenue": "number",
"growth_yoy": "number"
}
facts = ai_extract("acme_annual_report.pdf", schema)
# {"total_revenue": 48200000, "cloud_revenue": 31400000, "growth_yoy": 0.12}What it's for.
Built for the long, dense documents finance and investment teams read every day.
Financial reporting
Annual reports and financial statements
Revenue, segments, margins and guidance from annual and quarterly reports hundreds of pages long, including the notes.
total_revenuesegment_revenueoperating_marginFunds
Fund documents
Fees, terms, liquidity and strategy from prospectuses, offering memoranda and factsheets.
management_feelock_up_periodredemption_frequencyCredit
Credit and loan agreements
Parties, rates, maturities and covenants from credit agreements and their amendments.
interest_ratematurity_dateleverage_covenantCompliance
Regulatory filings
The disclosures and figures your compliance and risk teams need from regulatory filings.
filing_periodcapital_ratiorisk_factorsAny other document type works the same way: you define the schema, Document Intelligence fills it.
How it works.
Parse
The document is converted into a reasoning-friendly format that keeps its sections, tables, notes and page references.
Reason
The agent moves through the document like an analyst, following cross-references wherever they lead.
Extract
The facts come back structured to your schema, ready for your systems.
Revenue rose 12% year over year (see Note 7).
Note 7. Revenue by segment
Why Document Intelligence.
Reads like an analyst
It follows cross-references across hundreds of pages, from a summary on page 12 to the note on page 200, instead of reading one chunk at a time.
Your schema, your fields
You decide which facts to extract. The agentic harness is customized to your documents and returns exactly the fields you ask for.
ai_extract(document_pdf, schema)On-premise, no per-page fees
It runs on your own infrastructure, so your documents never leave it and the cost does not grow with every page you process.
Built for volume
If you process thousands of documents a day, we can help you.
How it performs.
86%
Internal benchmark
Accuracy on complex financial documents running to hundreds of pages, with a 32B open model and harness engineering alone.
Model
Qwen 32B
Fine-tuning
None
Method
Harness engineering
Public results on FinanceBench coming soon.
Document Intelligence is in private beta.
Request early access and tell us about the documents you process.