I really liked this. The economics of it are similar to what I've seen at work on a project to apply LLMs to RAG to summarise some complex technical information

I really liked this. The economics of it are similar to what I've seen at work on a project to apply LLMs to RAG to summarise some complex technical information. Basically, it works fairly OK 90% of the time, but the cost is ~$1 per thing we need it to do. That's not a sustainable price when someone might be hitting this a lot. If costs don't rapidly come down then it's not worth introducing it to customers (currently only being used for internal tests).