Vision with the OpenAI API

Use image understanding for a first analysis of screenshots, documents or mixed visual/text material, with appropriate rule or human verification. It is not error-free OCR or a final financial approval mechanism.

#Responses example

pythonCopy
import os
from openai import OpenAI

client = OpenAI(timeout=30.0)
response = client.responses.create(
    model=os.environ["OPENAI_VISION_MODEL"],
    input=[{
        "role": "user",
        "content": [
            {"type": "input_text", "text": "Describe the visible layout and uncertainties."},
            {"type": "input_image", "image_url": os.environ["IMAGE_URL"],
             "detail": "auto"},
        ],
    }],
)
print(response.output_text)

Use an authorized image and a model that supports the chosen input. Supported workflows can use an image URL, a base64 data URL or an uploaded file reference; check the current endpoint requirements.

#Detail and cost

Where supported, low/high/auto detail controls trade input handling and detail inspection. Start with the detail the task needs rather than assuming the largest image always yields the best result.

#Review boundaries

Check tiny text, ambiguous fields, crop/orientation, missing context and any important extracted figures against the original. Preserve uncertainty instead of turning a plausible observation into a confirmed fact.

#Reference

• Images and vision↗

#Related pages

• Quickstart
• OpenAI API overview