Like Laurie Young mentioned, you need bigger models of Llama. I took thought it was Langgraph that was inconsistent. Rewrote my application without Langgraph in python. No use. Llama 1B is the issue here. It can either use LLM or it can use tools. It cannot combine both. The fix is to use a better model.

It took a few days of breaking my head and testing with python/Pydantic and extensive online search and testing with ChatGPT to realize that Llama is the issue. I wish Meta had better documentation on this. Its appalling that they did not mention this anywhere on their documentation. What a waste of my time! I have decided to give up on Llama and stick to ChatGPT just because how unhelpful the documentation is. ChatGPT saves a lot of time, community is bigger and their models are just better. The only downside is the amount of space required. But nobody can put a price on time wasted on a model which is so far behind ChatGPT.