You have two routes. TigZig built a reference app for each, both connected to a real Postgres database (a static snapshot of 1.5M cricket records, though the pattern works for any domain: finance, retail, anything). Both apps were retired in October 2026. The trade-off below is what they showed.
Build your own: a realtime model over WebRTC with function calling, plus a FastAPI layer for database connectivity. You get full UI control and a much lower per-minute cost, but it is not plug-and-play. You will go deep into event flows, WebRTC and the API - expect to write and debug a lot of code. Drop in a vendor widget (TigZig's version used ElevenLabs' conversational widget): plug-and-play, excellent performance, running in minutes - but limited UI customization unless you adopt their SDK, and it cost several times more per minute.
Two caveats. Pricing and models move fast in realtime voice, so check current rates rather than trusting any figure you read (including older ones on this site). And secure it properly - the reference apps used basic domain whitelisting, which is fine for a demo and too weak for production. Because it uses function calling, the agent can also take actions on the database.
Write-up: https://www.tigzig.com/post/realtime-voice-ai-openai-webrtc-implementation-live-app-open-source. Related: where to get a free Postgres database https://www.tigzig.com/agents-faq/where-to-get-a-free-postgres-database and querying a database in plain English https://www.tigzig.com/agents-faq/how-to-query-a-database-in-plain-english.
Building something like this? How I work covers the rates, the availability and what I take on.