Search and result flows use extractive or generative models to get users answers.
Extractive models are fast and cheap to use. They help with simple fact extraction over well formed text.
They use natural language to extract facts from a piece of text and return it with some surrounding context. They don't create new or novel text unlike generative models. They will always return "something" along with a score. It's essential users verify the facts.
Generative models are large language models that generate text and other media. We use fine-tuned models that are good at conversational dialogue. Sometimes chat chat or instruction models.
We pass the model some of the returned results and use them as context for the conversation. This is called Open Book engagement.
Azure Open AI GPT 3.5 and 4
Falcon AI
Mistral
Qwen
Llama3 (and Llama2)
