added combined langchain_document content metadata extraction to reduce the calls to llm

removed few shot metadata extraction and added sementic check for same keyword

code push