@simon You have to be very careful with ollama and its default context settings since it's silently truncating inputs in many cases instead.

This bit me big time and there are open tickets about it. Two things:

1) Check the ollama server logs if you see a log line about truncating inputs

2) You can send num_ctx in their API or create a new ollama model with a bigger context and save it

Because by default if you

ollama run llama3.2

it has a limit of only 2048 🙀