[AGE-370] Improve reproducibility of AI critique outputs

[mmabrouk]

AI critique provides different results from run to run.
The goal of this issue is to determine and implement the best practices / parameters for running AI critic and improving its reliability.

The first step is to determine the best practices in other oss libraries / literature

_{From SyncLinear.com | AGE-370}

[mmabrouk]

I looked into how it is done in Ragas.
In the default mode, they set the temperature to 1e-8
To increase the reproducibility (for example in CI), they increase the temperature to 0.3 and run each call three times.