Hello everyone, I suppose this has probably been posted here before, sorry for the repeat.
I’m having stability issues with the REAL TIME API in a telephony tool. It’s sometimes unstable and struggles to understand certain sentences, especially in French. I’ve added some NLP logic, including an example with multiple possible expected responses so it can grasp nuances, like interpreting “pourquoi pas” (why not) as a “yes” for it.
Things have improved a bit, but it’s still not perfect. If you have any tips for this, I’d love to hear them.
On another note, I was wondering if attaching a RAG (Retrieval-Augmented Generation) could fully help it? The idea of being able to pull real-time data if needed to keep it contextual, in addition to the prompt—has anyone already tested this and gotten better results?
Is it difficult to implement a RAG with the Real-Time API? How much time should I budget for this with 4-5 senior developers?
Thanks so much in advance for all your answers, and a general thank you for your kindness on this forum.
Cheers