Confusion on LLM inference capability

Hi, i’m a new learner of AI and NLP related technoloy, i built a RAG program on LLM, but it seems the inference capability of this program is weak when i’m asking question to it.
this is the LLM and tokenizr i used for this program,

this RAG program is designed to load and analysize a tabular data, and answer question pertaining to it.

the answer to the simple question is fine;

however, with the increasing complexity of the question, the answer is getting unacceptale,

can anyone help to explain what is the problem on the program,is it problem of the LLM or the tokenizer?
Thanks in advance.

is the excel sheet your data or from the huggingface?

not from huggingface, I constructed the tabular data as test case.