Can you guys suggest a way to add browsing feature to my AI writer?

I’m building an AI writer that is targeting a niche that focuses on Fact-checking and content accuracy.

Is there any way to add a workflow before the Openai API calls that would allow it to write content that is factually correct based on the given list of URLs?

Thanks

Yes. You need to scrape the URLs first and then feed it to the openAI via the prompt

I actually tried an hour ago, It spend like 5000 in a single request.

The horror!!!

It is based on how OpenAI charges for both inputs and outputs. It you give less URLs, so a lower amount of input, it would cost different, but Microsoft does need a way to recoup some of their $10 Billion investment

I’ll have to wait till they figure out the browsing capabilities through API I guess. What I made spent around 40000 tokens :sweat_smile:. I just tested it with 1 URL, the final application must work with 5-10 URLs, Imagine the token it’s going to use.

The purpose it’s gonna be used for isn’t worth this many tokens.

I:

Scrape a URL
Vectorise URL
Get relevant text
Insert in prompt

You can view a demo at the bottom of the page here: Demos | Not Quite Unicorns

Only insert the relevant text. You don’t need to pass all of every URL.

Here it returns the relevant news headline from BBC news based on the input (as opposed to the one that appears first on the page)

2 Likes

I see the possibility. Can you tell how you vectorize the URL and extract the relevant text?

I wrote a custom API hosted in Google Cloud platform that does those four functions. You can (I believe) get a website’s text in Bubble using a plugin, then split into chunks, then schedule a recursive workflow that embeds and upserts each chunk, and when that’s done queries the Pinecone index, but it’s obscenely slow so I went custom.

1 Like