Should I self-host a model or call a hosted API?

Call an API until the arithmetic stops working. Hosted inference wins on time to launch, model choice and not operating GPUs. Self-hosting wins when your volume is steady and high enough that per-token pricing exceeds the cost of dedicated hardware, or when data residency and privacy rules make sending the data out impossible. Steady, predictable, high volume is the condition that flips it.

Do I need a vector database?

Less often than the category's prominence implies. Under roughly a hundred thousand documents, the vector extension for the database you already run is usually enough and saves you an entire system to operate. Dedicated vector databases earn their place at large scale, with heavy metadata filtering, or when you need frequent live updates to the index.

What's the real cost driver in an AI application?

Input tokens, usually — not output. Retrieval-augmented systems that stuff a long context into every request spend far more on the prompt than the answer. Prompt caching, tighter retrieval, and routing easy requests to a smaller model typically cut a bill more than negotiating rates. Instrument token usage per request before you optimise anything, because the expensive path is rarely the one you would guess.