# Financial document OCR ## Financial document OCR Keep track of the changes and updates for the Financial document OCR API. ## Version 1 #### ⚡️ Features and Changes (April 28th, 2025) * :sparkles: :zap: Expand the granularity of the **`category`** and **`sub_category`** output. * :sparkles: New categories: **`energy`, `shopping` , `software`** * :sparkles: New sub-categories: * **`food`** * **`delivery`** (Deliveroo, UberEats...) * **`transport`** * \*\*`public` \*\*(tram, rer, metro, underground, Bus...) * **`car_rental`** * **`micro_mobility`** (Short-term rentals, like scooters and bikes.. For further information, you can consult ) * **`shopping`** * **`office_supplies`** (notebooks, pens, scissors, ...) * **`electronics`** (computer, printer, iPhone, cables, ...) * **`cultural`** (books, movies, music, ...) * **`groceries`** (food, cleaning, ...) * **`other`**
* :zap: **Significant Performance Improvement for Supplier Name Extraction on receipts:**\ \ We've also significantly boosted the accuracy of `supplier_name` extraction in the Receipt API. We've achieved a 30-to-50% error rate reduction by developing and training a novel, domain-specific NLP model. #### ⚡️ Features and Changes (April 7th, 2025) * ✨ New field **`document_type_extended`**.\ This field provides more granular document classifications compared to **`document_type`**. The possible output values are: * **`CREDIT NOTE`**: Reduces the amount a buyer owes. * **`INVOICE`**: Requests payment for goods or services. * **`PAYSLIP`**: Details employee earnings and deductions. * **`PURCHASE ORDER`**: Buyer's official request to purchase. * **`QUOTE`**: Seller's estimated cost for goods or services. * **`RECEIPT`**: Acknowledges payment. * **`STATEMENT`**: Summary of financial transactions over a period. * **`OTHER FINANCIAL`**: Miscellaneous financial documents. * **`OTHER`**: Documents not fitting other financial categories. * 🔥 Strong improvement on **`currency`** field for invoices * ✨ Address decomposition into its’ sub-component: `street_number` , `street_name`, `po_box`, `address_complement`, `city`, `postal_code`, `state`, `country` for invoices.
#### ⚡️ Features and Changes (October 10th, 2024) * :bug: Fix `category` / `subcategory` consistency for receipts. * 🐛 Fix French Guyana company IDs not correctly detected for invoices. * 🐛 Fix text recognition errors for emails and websites for native PDFs for invoices. * ✨ New field \*\*`po_number` \*\*: The unique identifier which is issued by a buyer to a seller to authorize the purchase of goods or services. * ✨ New field \*\*`payment_date` \*\*: The date by which the payment is expected or was made. * ✨ New sub-field **`is_computed`** in **`due_date`**: It is set to **`True`** if the **`due_date`** is calculated based on payment terms or natural language, and **`False`** if it is directly specified as a date in the document.
#### ⚡️ Features and Changes (July 17th, 2024) We added the LiLT for **`line_items`** reconstruction for invoices on the the Financial document API. * 🔥 **Strong improvement on `line_items` for invoices.**\ Reduction in errors in terms of perfectly reconstructed lines of about **30 to 40%.** * ⚡️ **Strong reduction of processing time for invoices with many line items** * ✨ **New field `unit_of_measure` in `line_items` that represents the unit of measurement for the item, such as kilograms, liters, units, etc.**
#### ⚡️ Features and Changes (July 11th, 2024) We added the LiLT for **`total_amount`, `total_tax` and `taxes`** for invoices on the the Financial document API. * 🔥 **Strong improvement on `total_amount`, `total_tax` and `taxes` .** We have observed a reduction in error rates as follows: * 36% for **`total_amount`** * 25% for **`total_tax` (+7% on precision)** * 15% for **`taxes`**
#### ⚡️ Features and Changes (June 12th, 2024) * 🚀 **Extended latin alphabet support for receipt**\ We released new models for our generic text detection and recognition pipeline. This release has increased the overall performances on all fields and supports extended latin alphabet characters: ``` {'`', '¡', '¥', '¿', 'Á', 'Ã', 'Ä', 'Å', 'Æ', 'Ì', 'Í', 'Ð', 'Ñ', 'Ò', 'Ó', 'Õ', 'Ö', 'Ø', 'Ú', 'Ü', 'Ý', 'Þ', 'ß', 'á', 'ã', 'ä', 'å', 'æ', 'ì', 'í', 'ð', 'ñ', 'ò', 'ó', 'õ', 'ö', 'ø', 'ú', 'ü', 'ý', 'þ', 'Ā', 'ā', 'Ă', 'ă', 'Ą', 'ą', 'Ć', 'ć', 'Č', 'č', 'Ď', 'ď', 'Đ', 'đ', 'Ē', 'ē', 'Ė', 'ė', 'Ę', 'ę', 'Ě', 'ě', 'Ğ', 'ğ', 'Ģ', 'ģ', 'Ī', 'ī', 'Į', 'į', 'İ', 'ı', 'Ķ', 'ķ', 'Ĺ', 'ĺ', 'Ļ', 'ļ', 'Ľ', 'ľ', 'Ł', 'ł', 'Ń', 'ń', 'Ņ', 'ņ', 'Ň', 'ň', 'Ō', 'ō', 'Ő', 'ő', 'Ŕ', 'ŕ', 'Ŗ', 'ŗ', 'Ř', 'ř', 'Ś', 'ś', 'Ş', 'ş', 'Š', 'š', 'Ť', 'ť', 'Ū', 'ū', 'Ů', 'ů', 'Ű', 'ű', 'Ų', 'ų', 'Ź', 'ź', 'Ż', 'ż', 'Ž', 'ž', 'Ș', 'ș', 'Ț', 'ț', 'ẞ', '₿'} ``` * 🔥 **Strong improvement on `taxes` for multi-taxes extraction for receipt**\ The main focus of this release was to improve drastically the multi-taxes extraction.\ We measured a decrease in error rates of **56%** on `taxes` for multi-taxes extraction.
#### ⚡️ Features and Changes (May 16th, 2024) * 🔥 **Strong improvement on `supplier_name` and `supplier_address` for receipts**\ We measured a decrease in error rates for receipts of: * 27% for `supplier_name` * 14% for `supplier_address` * ✨ **New field: `receipt_number` and `document_number`**\ The API is now extracting the `receipt_number` returned as a string. It also extracts the `document_number` as a string, which is the document unique identifier. * :bug: **Fix date parsing for spanish/italian invoices** * :bug: **Fix reading errors for invoice number** * 🔥 **Strong improvement on `document_type` for credit note / invoice classification** * :zap: **Taxes from line items** `taxes` field outputs taxes from the line items when no tax summary is present on the document. #### ⚡️ Features and Changes (April 25th, 2024) * 🚀 **Extended latin alphabet support for invoices**\ We released new models for our generic text detection and recognition pipeline. This release has increased the overall performances on all fields and supports extended latin alphabet characters: ``` {'`', '¡', '¥', '¿', 'Á', 'Ã', 'Ä', 'Å', 'Æ', 'Ì', 'Í', 'Ð', 'Ñ', 'Ò', 'Ó', 'Õ', 'Ö', 'Ø', 'Ú', 'Ü', 'Ý', 'Þ', 'ß', 'á', 'ã', 'ä', 'å', 'æ', 'ì', 'í', 'ð', 'ñ', 'ò', 'ó', 'õ', 'ö', 'ø', 'ú', 'ü', 'ý', 'þ', 'Ā', 'ā', 'Ă', 'ă', 'Ą', 'ą', 'Ć', 'ć', 'Č', 'č', 'Ď', 'ď', 'Đ', 'đ', 'Ē', 'ē', 'Ė', 'ė', 'Ę', 'ę', 'Ě', 'ě', 'Ğ', 'ğ', 'Ģ', 'ģ', 'Ī', 'ī', 'Į', 'į', 'İ', 'ı', 'Ķ', 'ķ', 'Ĺ', 'ĺ', 'Ļ', 'ļ', 'Ľ', 'ľ', 'Ł', 'ł', 'Ń', 'ń', 'Ņ', 'ņ', 'Ň', 'ň', 'Ō', 'ō', 'Ő', 'ő', 'Ŕ', 'ŕ', 'Ŗ', 'ŗ', 'Ř', 'ř', 'Ś', 'ś', 'Ş', 'ş', 'Š', 'š', 'Ť', 'ť', 'Ū', 'ū', 'Ů', 'ů', 'Ű', 'ű', 'Ų', 'ų', 'Ź', 'ź', 'Ż', 'ż', 'Ž', 'ž', 'Ș', 'ș', 'Ț', 'ț', 'ẞ', '₿'} ``` * 🔥 **Strong improvement on `due_date`, `line_items` for invoices**\ We have observed a reduction in error rates as follows: * 20% for `due_date` * 25% for `line_items` * ✨ **New fields for invoices**:\ The API is now extracting the following fields:\ `customer_id`: The identifier of the customer in the supplier’s referential. It can also refer to the client ID, client / customer account number…\ `supplier_phone_number`: The phone number of the supplier\ `supplier_email`: The supplier email address\ `supplier_website`: The supplier website URL * 🔥 **General accuracy improvement**\ Thanks to the improvement done on our generic text detection and recognition algorithms, we measured a reduction in error rates on all fields, especially for supplier and customer information. * ✨ **New field for receipts**:\ The `locale` field now contains the following subfields when the document sent to the endpoint is a receipt:\ `country`: country code of the country where the receipt was issued (ex: US)\ `value`: concatenation of language and country codes in ISO format (ex: en-US) #### ⚡️ Features and Changes (March 11th, 2024) * 🚀 **Integration of company ID & logo database for invoices**\ We have integrated a company ID database and a vector database featuring millions of logos. This enhancement enables our R\&D team to efficiently rectify any issues with non-functional supplier names. * 🔥 **Strong improvement on invoices for `supplier_name`, `customer_name`, and `invoice_number`**\ We have observed a reduction in error rates as follows: * 20% for `customer_name` * 15% for `supplier_name` * 10% for `invoice_number` The improvement in `supplier_name` was achieved by incorporating information from the databases. The `customer_name` algorithm now mirrors the `supplier_name` one. `invoice_number` now employs an NLP modality to boost its precision. #### ⚡️ Features and Changes (January 30th, 2024) * 🚀 **Integration of a proprietary language model in the algorithm pipeline: LiLT**\ LILT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding.\ LiLT's design combines textual content with layout structure. This means it doesn't just read the text but also understands how the text is organized within the document. For instance, it recognizes headings, paragraphs, tables, and other structural elements, which is a crucial aspect of context awareness in document processing.\ The integration of this new language model in our pipeline helps us achieve better accuracy, and more flexibility when adding new supported fields. * 🔥 **Strong improvement on `supplier_name`, `supplier_address`, and `supplier_company_registrations`on invoices**\ The main focus of this release was to improve drastically the supplier information extraction.\ We measured a decrease in error rates of: * 42% for `**supplier_name**` * 10% for `**supplier_address**` * 10% for `**supplier_company_registrations**`\ Moreover, the integration of the LILT offers more robustness in terms of languages thanks to its language-independent component and will help us improve all other fields in the next releases. * ✨ **New field: `total_tax`on invoices**\ The API is now extracting the total tax information, returned as a number. It corresponds to the total tax explicitly written in the document. * 🔥 **General improvement for all fields on invoices**\ More training data was added to our training set, including different geographies and more variability. We’ve measured an improvement in accuracy for all extracted fields. #### ⚡️ Features and Changes (September 1st, 2023) * New feature: Raw Value available for both Supplier Name and Customer Name. The Raw Value extracts the name without post processing nor formatting. It can thus be different from the Value.
#### ⚡️ Features and Changes (May 23rd, 2023) New extracted field: * supplier\_phone\_number Updated fields for receipts: * supplier\_address is now available for receipts * supplier\_company\_registrations is now available for receipts * line items is now available for receipts but limited to the following features: description, unit\_price quantity, total\_amount #### ⚡️Feature: First Release (January 17th, 2023) Extracted fields: * total\_amount * total\_net * taxes * supplier\_address * supplier\_name * payment\_details ( Null for receipt) * orientation * locale (currency, language) * invoice\_number (Null for receipt) * reference\_numbers ( Empty list for receipt) * due\_date * document\_type * date * customer\_company\_registration (Null for receipt) * customer\_address (Null for receipt) * customer\_name (Null for receipt) * supplier\_company\_registration (Null for receipt) * category * Subcategory * time (Null for Invoices) * tip (Null for Invoices) * total\_tax (sum of taxes for Invoices) * line\_items (empty list for Receipt) :`product_code`, `description`, `quantity`, `unit_price`, `total_amount`, `tax_amount`, `tax_rate`