Data Extraction (Document module → Extract) converts documents into structured data: the invoice number, the contract's effective date, the purchase order total. Where the Parser reads a whole document, Extraction targets the specific fields your business process needs.
Extraction vs parsing vs knowledge
| Feature | Question it answers | Output |
|---|---|---|
| Parser | What does this document say? | Full structured content |
| Extract | What are these specific values? | Field → value table |
| Knowledge | What should the assistant know? | Index for grounded chat answers |
How it works
- Provide documents — upload files or select from a processed folder.
- Specify what to extract — either a ready-made schema (for common documents like invoices or resumes) or a custom field list you define (name, date, amount, currency, clause references…).
- Run — the platform combines document understanding, OCR and model reasoning to locate each field, including values inside tables.
- Review and correct — each extracted value can be checked against the source page before you accept it. Human review on low-confidence items keeps downstream data trustworthy.
- Use the output — download as structured data (CSV/JSON) or continue into a workflow that files, validates or posts the values onward.
Supported sources
- Text PDFs and scanned PDFs (OCR)
- Images of documents (photos, screenshots, faxes)
- Office documents and forms
- Multilingual documents where the extraction languages are configured
Handwriting is supported on a best-effort basis depending on the engines enabled by your administrator. See Documents (Admin).
Typical use cases
| Use case | Fields to define |
|---|---|
| Invoice processing | Supplier, invoice number, date, line items, totals, tax |
| Purchase orders | PO number, requester, items, approval status |
| Contracts | Parties, effective/expiry dates, renewal terms, liabilities |
| HR forms | Name, role, start date, salary band, signatures present |
| Insurance claims | Policy number, incident date, claim amount, status |
Quality practices
- Start with a pilot set of 10–20 real documents and measure accuracy per field before automating.
- Define narrow fields ("Total amount excluding tax") rather than vague ones ("money details").
- Review low-confidence results — corrections feed better outcomes and protect reporting.
- Keep originals — extraction is a view of the document, not a replacement; link outputs back to source files.
- For high-volume pipelines, chain Extraction in a workflow with Format Check and User Approve nodes. See Workflow Builder.
Related
- Document Parser — structured reading of whole files
- Document Classification — sort by type first, then extract
- Document Processing reference