I’ve noticed an inconsistency in how the LLM handles image files:
Working scenario:
- When I directly attach an image or reference a file path explicitly, the LLM can properly tokenize and analyze the image content
Not working scenario:
-
When I ask the LLM to search for and open an image file (e.g., “search for and describe the dashboard screenshot”), it successfully:
-
Lists the files
-
Finds the correct image file
-
But then cannot actually open/see/tokenize the image content
Is there a way to make the LLM actually open and tokenize image files it finds through search operations? Or is this a known limitation where images must be explicitly attached/referenced to be processed?
Thanks!