Document types
Pass document_type to select the extraction schema.
Use auto to let the pipeline detect the type, or combined for multi-page PDFs.
Supported types
| document_type | Description | Key fields |
|---|---|---|
invoice |
Carrier or supplier invoice | invoice_number, invoice_date, due_date, total_amount, currency, vat_amount, supplier_name, customer_name, shipment_reference |
cmr |
CMR consignment note | cmr_number, consignor, consignee, carrier, place_of_loading, place_of_delivery, goods_description, gross_weight, packages_count |
pod |
Proof of delivery | delivery_date, recipient_name, signature_present, reference_number, notes |
delivery_note |
Delivery note / packing list | delivery_note_number, shipper, recipient, delivery_date, line_items |
customs |
Customs declaration | declaration_type, declaration_number, hs_codes, total_value, currency, country_of_origin |
bill_of_lading |
Bill of lading | bl_number, vessel_name, port_of_loading, port_of_discharge, shipper, consignee, container_numbers |
rail_waybill |
Rail waybill | waybill_number, wagon_number, departure_station, destination_station, cargo_description |
combined |
Multi-document PDF | Per-page classification + aggregated fields |
auto |
Auto-detect (default) | Pipeline selects best matching schema |
Digital vs scanned PDFs
- Digital PDFs — text layer extracted locally via Poppler on our servers. Fast (~100–500 ms). No third-party AI.
- Scanned PDFs — when OCR is enabled, PDF content may be sent to Mistral AI over TLS. Slower (~2–5 s cold; cached instant for repeats on your account).
- Structured / vision stages — optional Mistral or OpenAI passes for complex fields. See Data processing.
Language support
Optimised for freight documents in:
- English (EN)
- German (DE)
- Estonian (ET)
- Lithuanian (LT)
- Latvian (LV)
- Russian (RU)
Language hints are detected automatically from document text.
Confidence scores
Each field includes a confidence score. Tune sensitivity with
options[confidence_threshold] (default 0.7).
Low-confidence fields may appear in warnings.
Business rules
After extraction, a business rules engine validates cross-field consistency
(e.g. VAT calculation, date ordering). Violations appear in
business_rule_violations with severity levels.