Will fine-tuning be useful to reduce RAG query processing time

Good day to all! There are tasks that are solved via RAG, I am thinking about fine-tuning. A database of slightly more than 50 questions and answers has been compiled on our data.
When substituted as a database into the RAG system, this set worsens the system’s responses. The database by topics produces a higher-quality result and is more flexible. As a result, there are questions

  1. Does this mean that fine-tuning on this database can also fail?
  2. Does LLM trained on individual data lead to faster information processing or can it only improve context understanding?
  3. What actions can help reduce query processing time