What if u embed `askLLM("5 thins this Diff could fix" + chunk)` instead of `chunk`? That should be closer in the latent space.