Difference in outcome ollama native and ellmer

I try to analyse images in png format. Using Ollama I get useful results:

Using the following code I get no useful results:
I am using the same local models.
Why are they so different

type_line <-  type_object(
        set = type_string(description = "set"),
        kuerzel = type_string(description = "Kuerzel"),
        name = type_string(description = "Name")
    )
type_ops <- type_array(type_line)

#| label: chat-ollama
# Initialise chat with gemma3 model
chat <- chat_ollama(
model = "Keyvan/german-ocr"
#system_prompt = system_prompt
)
png_files <- list.files(here("data"), include.dirs = TRUE, full.names = TRUE)
file <- png_files[1] 

x <- chat$chat_structured(
  content_image_file(png_files[1]),
  type = type_ops
)

Result:
1 \nBlutungen\n1-591,04 abh.\n\nofte chirurgische Operation am Gefäßknotenepithel und an den Mesentern. Entnahme eines Gefäßperitoneoplasmaeine Kniegelenk.\n1-592,03,\n\nAtherosklerotische Operationen an der Gefäßknotenepithel-Va.\n1-631,01,\n\nAtherosklerotische Operationen am Gefäßknobelsystem und an den Mesentern. Entnahme eines Gefäßperitoneoplasmaeine Kniegelenk.\n1-631,01,\n\nAtherosklerotische Operation am Gefäßknobelsystem und an den Mesentern. Subchordale Kniegelenkchirurgie. [...]\n(4b) nach Prufba, Microinfrastruktur. Abzesseinflussfrequenz(eins.) Kniegelenk.\n