annie sets it up for you
Our AI expert configures the pipeline in seconds for you, no manual setup required.
I want to extract info from invoices
Document processing
your systems can trust
Heavy lifting on your hardest documents. Parse, extract, classify, split and validate them into clean, reliable data. Live in minutes, with EU data residency, ISO 27001 and self-hosting.
Trusted by teams automating document-heavy operations
Document operations
composable into one workflow
Parse any document
into clean, structured text
Turn any document into clean markdown your pipeline, RAG or agents can read.
Structured markdownClean markdown with text, tables and figures that any downstream step, RAG pipeline or agent can read.
Layout detectionDetects headings, columns, reading order and tables, so document structure survives even on dense, multi-column pages.
Agentic routingAgentic parse mode routes tricky pages and elements for higher accuracy on messy, real-world documents.
Any format, any languageReads PDFs, images and scans in any language, with no templates and no per-format setup.
Parse
Parse any document into clean, structured text
Turn any document into clean markdown your pipeline, RAG or agents can read.
Structured markdownClean markdown with text, tables and figures that any downstream step, RAG pipeline or agent can read.
Layout detectionDetects headings, columns, reading order and tables, so document structure survives even on dense, multi-column pages.
Agentic routingAgentic parse mode routes tricky pages and elements for higher accuracy on messy, real-world documents.
Any format, any languageReads PDFs, images and scans in any language, with no templates and no per-format setup.
Extract
Extract exactly the data your systems need
Turn parsed documents into typed, structured fields you can use anywhere.
Hundreds of pages, one shotExtract across hundreds of pages in a single pass, no chunking or manual stitching.
Visual citationsEvery value comes back cited to its exact position on the page, with click-through navigation to the source.
Confidence scoringA per-field confidence score flags what to trust and what to send to human review.
Complex tablesCapture long, multipage and nested tables, like line items spanning dozens of pages, in one pass.
Classify
Sort every document into the right type
Label each document so a single workflow can handle many document types.
Categories you defineSet up your own labels, each with a short description of what belongs in it.
No training dataClassifies from your plain-language descriptions alone, no labeled examples needed.
Built-in routingSend each category down its own path, so one workflow handles every document type.
Self-learningCorrections feed back in, so classification keeps getting sharper on your documents over time.
Split
Break bundled files into clean documents
Separate multi-document files so each piece is processed on its own.
Split by categoryDefine split rules in plain language, and each segment is processed on its own.
Repeated documentsSeparate several documents of the same type, like four invoices stapled into one PDF, into independent runs.
Long documentsBreak a 50-page bundle into its sections, so each one is handled separately.
Nothing falls throughAnything that doesn't match a rule lands in a catch-all group, so no document is ever silently dropped.
Validate
Catch bad data before it ships
Check every result against rules you write in plain language.
Rules in plain languageWrite checks like “the IBAN is valid” or “the document is not expired”, no code required.
Automatic flaggingEvery document is checked, and anything that fails surfaces for review instead of slipping through.
Cross-field checksCompare values across fields, like the name on an ID against the name on the contract.
Human-in-the-loopAutomated rules pair with human sign-off, placed exactly where you want it, at scale.
Orchestrate
Run the whole pipeline in a single call
Wire every step into one workflow that runs end to end.
Composable graphConnect Parse, Extract, Classify, Split and Validate into a typed graph for any document.
Branching and conditionsConditional nodes send each document type or segment down its own path.
One call, end to endRun an entire pipeline on a document in a single API call.
Run in bulkApply a published workflow to large batches via API, the UI, or cloud-storage sync.
Benchmark
Real documents. Real numbers.
Frontier-level accuracy,
with confidence and visual citations
Measuring document parsing quality is complex. It spans multiple dimensions, including content extraction, tables, and document structure, each requiring its own evaluation metrics. Our Parse Score brings them together into a single benchmark, combining state-of-the-art metrics across 1,000+ real-world documents against a carefully curated ground truth. anyformat outperforms every dedicated OCR engine on the market, and is the only solution that also provides calibrated confidence scores and visual citations, making every extracted result both accurate and verifiable.
Accurate on page one.
Still accurate on page fifty.
Parsing long documents is where most systems break down. As tables span dozens of pages, models gradually lose coverage and silently skip rows, even when the document itself is perfectly readable. We benchmarked real documents of up to 50 pages (≈2,400 rows) and measured how many rows were correctly recovered. anyformat stays consistently accurate from beginning to end, while leading frontier models and default OCR pipelines degrade rapidly as document length increases.
Native understanding
of complex layouts
We measure the percentage of documents extracted perfectly: every line item correct and every total reconciled. Across nearly 100 real, complex, mixed-format invoices, grouped by page count, the gap widens as documents grow. On documents of 16+ pages, anyformat keeps 80% of invoices perfect, while the next-best model reaches just 27%.
Calibrated confidence
at field-level extraction
Every value carries a calibrated confidence, not a non-calibrated signal, so you can set a threshold and auto-approve the confident majority. It isn’t one global number either: anyformat calibrates each operator on its own, so the score fits the work that produced it.
99.1%
calibration accuracy · Adaptive ECE 0.009
Why anyformatBuilt for production,
Built for production,
not demos
Raw OCR can't read structure. A raw LLM hallucinates and can't be audited. anyformat combines both with deterministic rules, confidence scoring and validation, so the output is structured, traceable and ready for production.
Traditional OCR
Raw LLM
anyformatReads layout, tables and structure
–
~
Long documents, hundreds of pages
~
–
Complex, multipage tables
–
~
Poor scans and low-quality inputs
~
–
Any format and language, no templates
–
Visual citations: every value cited to its exact position on the page
–
–
Confidence scoring and human-in-the-loop
–
~
EU data residency, ISO 27001, zero retention
–
–
Built for the agentic eraBy engineers and agents
By engineers and agents
for engineers and agents
Call anyformat workflows straight from your codebase with the SDK or CLI, or drop in the agent skill and let your coding agent build and run them for you in minutes, not months.
Terminal
Claude Code
Opus 4.8
SDKs
Install the SDK to call anyformat workflows straight from your codebase.
npm install @anyformat/sdkCLI
Use the Python CLI to run and manage your workflows from the terminal.
uv tool install --python 3.13 anyformat && afx --helpAgent Skill
Add this skill to teach your agent the anyformat workflow API.
npx @anyformat/skillBeyond the model
Everything around the model, built in.
Advanced OCR and parsing, enhanced by the capabilities required to build and run reliable workflows, fast and at scale.
Evaluate your workflows with your own dataset
Upload your dataset and expected outputs to run evaluations, and test every change before you ship.
Monitor your workflow
Track confidence scores, accuracy, and throughput rate in real time, so you always know exactly how your pipeline is performing.
Optimize your workflows automatically
Ask annie to improve your workflow, she finds the fix and applies it automatically.
Accuracy by iteration
#1
#2
#3
#4
#5
#6
1st Optimization
5 fields · 6 iterations
81.0%85.5%+4.5%
Security and certifications
Enterprise-grade security
anyformat meets enterprise security and compliance requirements out of the box, so teams can run document workflows in production with confidence.
GDPRPersonal data handled to GDPR standards, with EU data residency.
ISO 27001Independently certified to ISO 27001 for information security management.
Self-hosted deploymentRun anyformat in our cloud, in a dedicated VPC, or fully self-hosted in your own infrastructure. Your data stays exactly where your policy requires.
Use casesBuilt for industries
Built for industries
that can't afford mistakes
Invoices, bills of lading, claims, contracts… anyformat handles the high-stakes documents every industry depends on, and goes deep where accuracy matters most.
Accounts payable
Turn every invoice, credit note and statement your vendors send into structured data for your ERP.
anyformat reads them across vendors, layouts and languages, validates them against POs and pushes clean data into SAP, Oracle or NetSuite.
“We moved from intensive manual data entry to agile validation, with every figure traceable back to the source.”
AP Operations LeadGlobal manufacturer
See it in actionInvoice
Frequently Asked Questions
Answers to common questions about anyformat and its features. If you have any other questions, please don't hesitate to contact us.
anyformat is a document intelligence platform that extracts, structures, and routes data from any document, automatically. It combines AI models with deterministic rules and no-code workflow orchestration, built for enterprise teams that need production-grade accuracy.
