Document OCR API
Extract text from scanned documents with Upstage OCRβhighly accurate for Korean and complex layouts, even at low resolution.
Schema generation
Generate an extraction schema automatically from up to three sample documents. Covers the request body (model, messages), the chat.completion response carrying the schema, and an example request and response.
AuthorizationIn: header
application/jsonRequiredmodelA string value indicating the model used for inference. We recommend using the alias information-extract, which always points to the latest stable model. You may also specify a particular model version, but please note that models are frequently updated, and services tied to specific versions may be affected when those versions are deprecated. Available snapshot models are listed in the capabilities document.
messagesA list of messages
response_formatA JSON object containing the target schema.
modeExtraction mode. "standard" is the default mode. "enhanced" mode provides better robustness for complex tables, poor quality scans, and handwritten text. The "enhanced" mode incurs additional costs. See the pricing page for details.
"standard"Value in: "standard" | "enhanced"locationWhen set to true, returns the location of extracted values in the original document, including page number and bounding box coordinates. The bounding box is defined by 4 coordinates (x,y), where x and y values are normalized between 0 and 1.
falselocation_granularitySpecifies the granularity of the location coordinates. "element" returns the coordinates of the entire HTML element containing the extracted value. "word" returns the coordinates of the specific words matching the value.
"element"Value in: "element" | "word" | "all"splitAutomatically split and process multiple documents within a single file. When enabled, each document will be processed separately and return individual results.
falseconfidenceWhen set to true, returns confidence levels ("high" or "low") for each extracted value. Our confidence scoring prioritizes high recall (>95%) to ensure potentially incorrect extractions are reliably flagged as "low" confidence, with approximately half of "low" confidence results being actually incorrect extractions (precision 50%).
falsechunkingParameter for chunking documents into smaller parts for processing. Recommended for documents with 30+ pages or tables with 50+ rows to improve extraction performance.