Upstage document classification

Upstage Document Classification

Classify documents into user-defined categories with confidence scores. Also supports Document Split for separating multi-document PDFs into individual documents.

Quick Start

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["UPSTAGE_API_KEY"],
    base_url="https://api.upstage.ai/v1/document-classification"
)

response = client.chat.completions.create(
    model="document-classify",
    messages=[{
        "role": "user",
        "content": [{"type": "image_url", "image_url": {"url": "https://example.com/document.pdf"}}]
    }],
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "document-classify",
            "schema": {
                "type": "string",
                "oneOf": [
                    {"const": "invoice", "description": "Commercial invoice with itemized charges"},
                    {"const": "receipt", "description": "Payment receipt"},
                    {"const": "contract", "description": "Legal contract or agreement"},
                    {"const": "resume", "description": "Personal resume or CV"}
                ]
            }
        }
    }
)
print(response.choices[0].message.content)

API Key: Always use os.environ["UPSTAGE_API_KEY"]. Get your key at console.upstage.ai.

Endpoint

POST https://api.upstage.ai/v1/document-classification

OpenAI SDK compatible — set base_url to https://api.upstage.ai/v1/document-classification.

Parameters

Parameter	Type	Required	Description
model	string	Yes	document-classify or document-classify-nightly
messages	array	Yes	Single user message with image_url
response_format	object	Yes	JSON Schema defining classification categories
split	boolean	No	Enable multi-document splitting (default: false)
split_criteria	array	No	Additional splitting criteria (used with split=true)

Schema Definition (Important)

Categories are defined using oneOf with const values. The root schema type must be "string", not "object".

{
  "type": "json_schema",
  "json_schema": {
    "name": "document-classify",
    "schema": {
      "type": "string",
      "oneOf": [
        {"const": "invoice", "description": "Commercial invoice"},
        {"const": "receipt", "description": "Payment receipt"},
        {"const": "other", "description": "Other document type"}
      ]
    }
  }
}

Important: Using enum or object-based schemas will return a 400 error. The Classification API requires oneOf with const/description pairs.

Response Structure

{
  "choices": [{
    "message": {
      "content": "invoice",
      "tool_calls": [{
        "function": {
          "arguments": {
            "document_type": {"_value": "invoice", "confidence_score": 0.99},
            "pages": [1, 2]
          }
        }
      }]
    }
  }]
}

• content: classified category name
• tool_calls.function.arguments.document_type._value: classified value
• tool_calls.function.arguments.document_type.confidence_score: 0.0–1.0
• tool_calls.function.arguments.pages: page range (most useful with split mode)

Document Split (Multi-Document PDFs)

For PDFs containing multiple document types, set extra_body={"split": True} to separate them into groups. Each group is returned as a separate choices entry. See references/document-split.md for the full split workflow with optional split_criteria.

Output Files

• Default (classify only): <system-temp>/<input-stem>.classified.json (e.g., /tmp/contract.classified.json).
• Default (split mode): directory <system-temp>/<input-stem>.split/ with one file per detected document (e.g., page-001.invoice.pdf).
• Override: if the user specifies an output path, use it.
• Always print the resolved absolute path(s) in your response so the user can locate the file(s).

Tips

• Include "other" in your categories to handle unclassified documents.
• split is useful as the first step in a document processing pipeline for scanned mixed-document files.
• A common pattern: classify first → apply category-specific schemas with upstage-information-extraction.
• Use confidence_score to flag low-confidence documents for manual review.

Detailed References

File	Content
references/document-split.md	Document split (basic + with criteria), curl example