Perplexity is a brand-new product and you are standing up the simplest answer engine that could work: a user asks a question, you embed the query, look up the top-k most similar chunks in a vector index, stuff them into the prompt, and let the LLM write a grounded answer. No multi-hop, no eval, no humans yet. I care about the retrieval path: which embedder, how the index works, and how a tuned top-k reaches the model.
Educational exercise. Not affiliated with or endorsed by Perplexity.