TsgcAIDocumentExtractor

TsgcAIDocumentExtractor reads PDF documents and images (invoices, receipts, identity documents or any document you describe with a JSON Schema) and returns the data they contain as a JSON document. It sends the document to an AI model through a linked TsgcAIChat, so every provider of TsgcAIChat can be used, and it checks the answer locally before returning it.

 

How It Works

• The data to extract is described by Schema, or by a built in Preset: dpInvoice, dpReceipt or dpIDCard.
• The model is asked for a structured answer (rfJSONSchema) with a data member that follows your schema and, optionally, a confidence value per field.
• The data is validated against the schema. When it does not validate, the model is asked again with the list of errors (ValidationRetries).
• The result is returned as JSON and is also available in LastResult, with the validation errors and the confidence of every field.

 

Supported Documents

PDF, PNG, JPEG, GIF and WEBP files. PDF input depends on the provider selected in the linked TsgcAIChat:

 

Provider	PDF	Images
aicpOpenAI	Yes	Yes
aicpAnthropic	Yes	Yes
aicpGemini	Yes	Yes
aicpDeepSeek	No	Yes
aicpGrok	No	Yes
aicpMistral	No	Yes
aicpOllama	No	Yes

 

Schema Rules

With OpenAI the schema is sent in strict mode: every object must declare "additionalProperties": false, every property must be listed in required, and optional values use a nullable type such as ["string","null"]. The presets already follow these rules. The local validation checks type, properties, required, items and enum, recursing into nested objects and arrays.

 

Local OCR

Link a TsgcAIOCRTesseract or a TsgcAIOCRWindows to OCR and set OCRMode to recognise the text of the images locally, either as a fallback when the vision request fails (domFallback) or always (domAlways), which lets a text only model extract the data. The local OCR accepts images only.

Two local OCR engines are available. TsgcAIOCRTesseract runs the Tesseract command line tool on Windows, Linux and macOS, and needs Tesseract installed. TsgcAIOCRWindows uses the OCR engine built into Windows 10 and later, so nothing has to be installed or deployed; it reads the OCR language packs installed in Windows.

 

Quick Start

oChat.Provider := aicpOpenAI;
oChat.ChatOptions.ApiKey := 'API_KEY';
oChat.ChatOptions.Model := 'gpt-4o-mini';

oExtractor.AIChat := oChat;
oExtractor.Preset := dpInvoice;
vJSON := oExtractor.ExtractFile('invoice.pdf');

 

The step by step guide Extract data from an invoice shows a complete example, with confidence checks and the OCR fallback.