OCR

OCR (Optical Character Recognition, optical character recognition) is responsible for converting the text in images into copyable text that AI can read. The following all depend on it:

• Drag a screenshot/scanned document into the chat box and want AI to understand the text inside
• Put image-format invoices and materials into Knowledge baseand hope to be able to search for them later
• agent Open a local image for analysis

OCR is a separate settings page. In 【Settings】→【OCR】, configure the recognition engine once, and all places that use image text recognition will use the same configuration.

OCR settings: ① Use the dropdown in the top-right corner to select the recognition engine (the illustration shows Mistral). Fill in the selected engine's API key and API address below

Select recognition engine

The dropdown in the panel's top-right corner is used to switch OCR engines,the selected engine will be used as the default. Built-in engines:

engine	Integration / operation method	Who it's for
System OCR	Offline, no configuration required	Borrow the system's built-in recognition (macOS Live Text / Windows OCR), out of the box and fastest
PaddleOCR	Fill in API key (PaddlePaddle Star River Community); if self-hosted, point the API address to your service. Optional parsing model	Don't want to use local resources, but want Paddle's recognition results
Local PaddleOCR	Offline, you need to first download the local OCR model (about 140 MB) in 【Settings】→【Local Models】	Good Chinese recognition and runs fully on the local machine, privacy first
Tesseract OCR	Offline, built-in	Classic open-source OCR, supports multiple languages, can be used as a fallback
Mistral	Mistral API key	Recognition powered by multimodal large models, smarter for complex layouts / handwriting, etc.
Intel OV OCR	Runs locally (Intel OpenVINO, NPU acceleration)	Only appears on Windows + Intel Core Ultra (with NPU), and when the OV model has been deployedother devices will not see this item

Not sure which to choose? Start with System OCR—Most screenshots and clear scans can be handled directly, with no configuration required. If the recognition results are unsatisfactory, switch to Local PaddleOCR or Mistral.

When System OCR is selected, the panel will show Detected available macOS Live Text / Windows OCR engine(If the system doesn't support it, this item will not appear in the dropdown.)

• Before selecting 'Local PaddleOCR', please first download the 'Local OCR Model' in 【Settings】→【Local Models】, otherwise it cannot be called.
• Tesseract(and System OCR on Windows) can select the languages to recognize in the panel's 'Language' dropdown.

Difference from document processing

Many people confuse OCR and Document processing document processing,

• OCR: handles images / scans text recognition in images (image → text).
• Document processing: handles PDF / complex-layout documents structured parsing of documents (PDFs with tables and multiple columns → clean text).

The two are independent and configured separately. Pure text PDFs,.md/.txt/.docx text paragraphs in documents don't go through either; just read them directly.

---

Get help and submit feedback

If you encounter any questions, bugs, or have suggestions for feature improvements during configuration or use, please refer to Feedback and Suggestions for the official channels provided.