Finally with your guys help I have begun training my Talkeetna Guide, you are welcome to play with it and the embeddings file.
At some point soon GoDaddy will update and propagate https://talkeetna.info what was interesting was a vector search of my csv embeddings file was always finding the same row. So a word search is a lot better still working with it
I got the vector search working and it is way better than the word search
The plan moving forward is to train three separate bots with three separate domains
This is going to take a long time getting to know this vector search tool php has.
I need hundreds of rows in each of these bots before they take on the personality of their assigned domain
Talkeetna guide is easy I live here
The other two not sure what I will try to do
Perhaps airline agent and maybe a hotel bot
I am going to have to get creative to take reservations and remember them
A funny story about Bing Codex. I got it in a loop where he only knew how to write calls to davinci. If I asked him for chat endpoint he would try and fail then would revert to davinci so I asked him point blank here is a bash curl and it looks nothing like your calls to the chat endpoint
He thought about it and said if those endpoints work for you here try this my data is old school.
And it did once he saw a working curl I bash he could use it to build a php curl call successfully
So I am bored with training so I started playing with writing a response file
Going to search it for any key words and append this to what I find with the vector search perhaps some deeper context will result
I don’t think I need the user input
I bagged the cabin rental and focused on training talkeetna.info
I found a bug in my word search that improved the system input for questions and started repeating back the previous conversation has given real context and accuracy
More is better for training files, getting rid of stop words like the they is that etc helps to reduce the returns from the embeddings as does required 2 words of input found in a row before being selected
Accuracy is important for the system input and too many rows will reject the entire query
I looked into vector analysis and didn’t find any working free code to do it so I am using a word search of the embeddings
Getting pretty good results using this search tool feeding system input using a two keyword criteria and eliminating stop words like this and the etc
So you don’t “need” a vector search a word search of a text file works well enough to train a bot
I think this post was meant to be on one of your existing threads? Not sure it adds anything to your previous work out on it’s own like this. (I’ve moved this thread into your main one about the bot)
Don’t see anyone else doing it yet except Bing no one here is I know of, oh I get it let’s sink the fox
Printing live links is a BIG deal. Users need them OpenAI doesn’t do them and Bing does being able to link users directly to content is elegant.
A independent open source bot that does will be downloaded wait and see I am going to make this bot do every trick bing does one day with 3.5
I am going to get this web search working next …. Without a plug-in
But you are not interested I get it but I am gaining followers and with the competition going on right now OpenAI needs geeks of every skill level
Your pissed I Make it EASY for the kids to get started that is all