Add lazy initialization and FAISS caching for vector store

[Sasyamerugu]

This change improves the initialization of the RAG vector store by introducing
lazy loading and FAISS index caching.

Previously, the vector store was built during server startup, which caused
slower startup times and recomputation of embeddings on every restart.

Changes:

• Initialize the retriever lazily when the /ask endpoint is first used
• Persist the FAISS index locally
• Load the existing index if available instead of rebuilding embeddings

Benefits:

• Faster server startup
• Avoids recomputing embeddings unnecessarily
• Improves performance when working with larger document sets