Image Extraction and Searchable Descriptions
Enhanced sync can extract images through Mistral OCR and append generated descriptions to document chunks. Those descriptions are embedded as text, so supported text embedding providers can retrieve passages that describe visual content.
The current shared sync pipeline generates text embeddings. A model's Multimodal badge describes a provider capability; it does not mean the app sends extracted images into embedding requests. Native image-vector similarity search is not part of this workflow.
Set up image extraction
- Configure your embedding provider and project vector store.
- Configure the LLM used for Enhanced sync and the Mistral API key required for image extraction.
- Run Enhanced sync on a supported document containing images.
- Review the sync results. When extraction and annotation succeed, image descriptions are included in relevant chunks and can be found through text search.
Extraction depends on the document and OCR result; not every image or format produces an annotation. Without Mistral configuration, Enhanced sync can process document text but skips image extraction.
Image extraction accepts up to 50 PNG, JPEG, or WebP images per document, with a 2 MiB limit per image and 10 MiB combined. If the OCR result exceeds these limits, text sync continues without storing its extracted images.
What descriptions contain
Generated annotations may include an image type, short description, detailed summary, and extracted data points. Review important details against the original document because generated descriptions can be incomplete or inaccurate.
Storage and access
Extracted images are stored privately with organization ownership. Both uploaded files and extracted images count toward storage billing at 200 credits per GiB per month. Mistral, embedding, and LLM provider charges are separate from app credits.
The app can supply temporary signed image links with a ten-minute expiry. A link is not a permanent public asset URL. External applications should use their configured vector store and their own authorized image-access integration; app session endpoints are not a documented customer API.
See Embedding Model Configuration, Sync Settings, and Credits & Usage.