Results Of Using The Assistant API

Great project.

As someone who has implemented an advisor for a very large Heating company, where you had a mix of PDF’s, website info and Technical specification, here is how it was implemented:

  • We tested with OpenAI and then later migrated to Microsoft Azure, as it provided more security and reliability (with SLA’s)
  • With OpenAI, we used PineCone as our RAG db, and did some intelligent chunking → we later migrated this to Azure AI Search
  • with images that you have on these technical diagrams, you can use GPT-4V, or Azure AI Search automatically does this for you (at a premium)
  • all the data went into PineCone or Azure AI Search with different settings (we found both equally good)
  • we used a top 20 results most of the time
  • the actual bot was a mix of models (GPT3.5 Turbo to create the first response) and then GPT4 if it required a more granular response (again you can play around with how the response is presented back

Hope this helps…

2 Likes

This will not work in my case as final idea is that users of that Assistant should get those figures on their request, they are not supposed to upload anything. Maybe I am fundamentally on the wrong way to solve that task…

Assistants are essentially a combination of the underlying base model + some type of black box memory. You can essentially create that yourself using (PineCone or Azure Storage/Search).

At the end of the day, there are different ways to achieve the same thing. :slight_smile:

2 Likes

@mouimet
i would like to talk to u
is there any way that i can reach u

Nice investigative discoveries…