I’m working on developing a GPT with a client as a resource to search their network of websites for insights and information.
We’ve built a GPT with ChatGPT which lead us down a path of looking for a direct web integration as OpenAI only offers this with the assistants framework afaik.
Pickaxe is now the solution - however when training Pickaxe vs GPT.. there seems to be some limitations and the output and outcomes aren’t quite the same or similar for that matter.
Any ideas how to feed the GPT to closely resemble the one we’ve trained on ChatGPT directly?
NOTE: I’m new here and not connected to Pickaxe or an expert in it in anyway, just interested in your topic…
When you say “training pixkaxe” I’m assuming you’re referring to knowledge retrieval, and are you talking about fine tuning the output with say lots of examples of how you expect the output to be, stored in your system prompt?
Or are you referring to something completely different?
Pickaxe should behave pretty similarly to a custom GPT when using the same model.
This is a bit complicated because OpenAI custom GPTs will scale their model based on the subscriber’s level. Similarly it will change to adapt to new models as they are released.
Pickaxe allows you to select the specific model and maintain it, but I still believe the models themselves are continuously being tuned by the AI companies based on what I’ve seen.
There’s one other element here which is that OpenAI’s web interface essentially is running its own pre-script. That script seems to vary based on context — when on a phone app it directs the system to keep answers succinct, and it has a similar set of specific instructions when using OpenAi’s voice tools.
I haven’t explored adding that script form OpenAI onto the top of my prompt, but I suspect doing so would make the results between the two almost indistinguishable.