A vector database stores embeddings and retrieves the entries most similar to a query embedding. In semantic hotel search, it replaces a FAISS index as collections grow across cities and languages, returning relevant reviews together with their attached data. section 6.2
Similarity alone can return hotels from every city. Metadata filtering narrows retrieval before results reach the answer generator: a query for a centrally located hotel can be limited to Istanbul through a locality field. This reduces irrelevant context and supports more personalized recommendations. section 6.2 section 6.0
City-filtered hotel retrieval
For “Amazing hotel close to everything,” filter the search to Istanbul, retrieve up to 25 matching reviews, and pass hotel details, locality, review text, and similarity scores into a concise generated response. The resulting answer cites hotels located near attractions and amenities in Istanbul.
In a RAG pipeline, retrieved reviews ground an LLM response in current domain information rather than relying only on generated text. The database therefore connects embedding search to a conversational recommendation answer. section 6.0 section 6.2

































