Hang Yu. I have not compared the embeddings or done any similarity check on the vectors. The fact is Knowledge Embeddings can be complimentary and add context is what is being experimented. I ihave trued this on multiple input data sets and have comparison of the results. The behavior of the RAG varies and depends upon the particular task and the attributes of the embeddings. When I tried with generic information context, it didnt make a difference in the accuracy of the response, but for business / domain specific implementation, this largely applied. I am following up with an other article on the responses and performance metrics. I will be abstracting the real data i used. please await my update. Thanks for your note!