Document Extraction

Turn unstructured documents into structured, reviewable data.

Document Extraction runs uploaded files through OCR, layout analysis, and an LLM extraction pipeline, then presents the result for human review and correction. It draws on the platform's full document type registry — financial statements, tax forms, bank statements, credit reports, real estate documents, ID documents, and more. See Document Types for the complete list.

Use Extract your first document for a guided first run. Use Extract data from documents for submission and batch-processing steps.

In the UI, Document Extraction lives under Capabilities > Extraction, with quick-start entry points from the Home dashboard and the sidebar's + New Document button.

Extraction Upload — drag and drop files on the left, with Document Type, Model and the processing parameters on the right

Supported Document Types

Document Extraction draws from the platform's document-type registry — the same registry used by Document Classification. For the complete, authoritative, categorized list, see Document Types.

Additional document types can be registered dynamically without a code change — new types are picked up automatically by classification and extraction routing, by adding a YAML config file. See Document Types for how the registry works.

Capability	Description	Limit
Document upload	Upload a single file or a batch (ZIP) for extraction	500 MB max per ZIP, 100 MB max per file within a ZIP
Supported file types	PDF, images, Office formats, and structured text	PDF, PNG, JPG, JPEG, GIF, WEBP, SVG, BMP, TIFF, XLS, XLSX, DOC, DOCX, TXT, CSV, JSON, XML, HTML
List filters & search	Filter the extractions list by status, a curated shortlist of common types, or processor (single/batch); search by name	Type quick-filters are a curated subset of the full 115-type registry used when choosing a document type for extraction
Extraction result review	Review extracted fields against the source document, by document type; the list also shows processor and an overall confidence score per extraction	115+ registered document types — see Document Types
Chart of Accounts (COA) review	Review and adjust rollup calculations for balance sheet, P&L, and property operating statements	Form 8825: full extraction for the first 50 properties, basic info only beyond that
Human corrections	Submit a corrected value for a field; recomputes confidence for that field	—
Signed document URLs	Generate a time-limited link to the source document	1–168 hours
Preview URLs	Generate a time-limited preview link	1–24 hours

Extraction results over 15 MB are stored separately rather than inline. This is handled automatically and does not change how you retrieve the result.

Workflow

1. Upload — add a document or a ZIP batch from the Extractions list, choosing a document type from the 115-type registry and, optionally, a model (GPT-4.1, GPT-5.1, Gemini 3, or the account default). Advanced parameters let you toggle caching, document analysis, layout inclusion, LLM-judge validation, quick-extract mode, and raw-extraction-only mode, or supply industry context, a webhook URL, a custom extraction prompt, agent instructions, or metadata.
2. Processing — the extraction moves from Pending to Processing. Large PDFs (over 40 pages) use a combined native-PDF-plus-markdown extraction path; documents between 10 and 40 pages use an enhanced extraction pipeline.
3. Classification — for multi-document PDFs, the system samples at least 20 pages to detect document type boundaries before extraction.
4. Result — status moves to Processed or Failed. The result includes the extracted fields, layout data, and, for V2 processing, a document categorization pass.
5. Review — open the extraction to review fields against the source document, using the reviewer for that document type. For example, the Balance Sheet reviewer or Bank Statements reviewer.
6. Correct — submit a correction for any field. This recomputes the confidence score for that field and feeds the correction into the error-pattern feedback loop used to improve future extractions.
7. Chart of Accounts review (balance sheet, P&L, and property operating statement types only) — review and adjust account rollups, including Form 8825 property-level detail.

Pending → Processing → Processed
→ Failed

Everyone in the account can use the extractions list and the upload flow, whatever their role.

Related pages

• Extract your first document provides a guided first run.
• Extract data from documents covers submission, options, and batches.
• Document Classification helps route unknown or mixed files before extraction starts.
• Document Types defines the supported types and field schemas this flow uses.
• Test Dashboard benchmarks extraction quality, and Webhooks & Integrations pushes results to downstream systems.

---