Skip to main content
After partitioning, you can have a vision language model (VLM) optimize the fidelity of text blocks that Unstructured initially processed during its partitioning phase. Here are a few examples of Unstructured’s output of text blocks that were initially processed, and the more accurate version of these text blocks that were optimized by using Claude Sonnet 4. Irrelevant lines of output have been omitted here for brevity. Example 1: Vertical watermarked text Vertical watermarked text example Before (vertical watermarked text, represented incorrectly):
After (vertical watermarked text, now represented correctly from the original content):
Example 2: Hyperlink Hyperlinked text example Before (hyperlink, represented incorrectly):
After (hyperlink, now represented correctly from the original content):
Example 3: Chinese characters Chinese characters example Before (Chinese characters, represented incorrectly):
After (Chinese characters, now represented correctly from the original content, expressed as Unicode):

Improve text fidelity with generative OCR

To have Unstructured perform generative OCR optimization, do the following:

Customize the generative OCR prompt

Unstructured API users can replace the default prompt that the model follows for generative OCR. Set prompt_interface_overrides.prompt.user in the Generative OCR task settings. The workflow designer does not support custom prompts for generative OCR.