Ironically I had implemented this myself with the GPT 3.5 API: Wolfram is a good match because it already supports fuzzy searches, has a well formed API, and t

Ironically I had implemented this myself with the GPT 3.5 API: Wolfram is a good match because it already supports fuzzy searches, has a well formed API, and the API responses provide information in a separate objects sorted by importance. That last bit is important because if you just open a random website, there's almost no chance it fits in the context window, and if you just truncate the contents there's no guarantee you capture important info. I ended up messing around with using non-LLM NLP to summarize webpages, but it's still a pain. - As far as knowing, ChatGPT doesn't need to know it's wrong, it just needs to know what it tends to be wrong at. For example, I told my instance: "Any time you feel like the specificity of your answer would be improved, you may use !q <query>", "Any time a question would benefit from more recent information you may use !q <query>" and "As a Large Language Model your math cannot be trusted, use !q <calculator operations>". That was enough to get it using Wolfram for any math or factual searches