I’ve worked on something similar, and for open-source options, I’d recommend checking out Rasa for NLU—it’s super flexible and works well for conversational bots. For voice, Coqui TTS is great for text-to-speech, and Vosk works nicely for speech-to-text. If you’re dealing with PDFs, you can use LangChain to pull data and connect it to an open-source LLM from Hugging Face—some of their models are surprisingly good and don’t cost a thing if you have the hardware to host them. I’d avoid commercial options at first since the costs can add up fast, but Google’s APIs do have free tiers you can experiment with. Start small, keep it simple, and see how users respond before scaling up.