Un-documented features that I found

Is the way to use the system input.

Seems like the bot listens for a while then quits but that is a feature not a huge bug

Apparently the way it was designed was to listen to the system input for the one user input but if it just repeats it ignores it

So you cannot make a huge document and feed it every call it won’t work and will stop using it at all after just a few calls

This way the bot is always using fresh information for every input feeding its response back to it helps in context and the contiguous conversation between inputs

But giving the system relevant information for the user input every time sparks a personality being driven by what you feed it

Using a simply text file with numerous rows to search on to feed personality information relevant to user input is the key :key:

I don’t think vector searches would improve it and not worth the money :moneybag: or the time to implement

Text files I am having enormous success over just a large document fed each input but having relevant info to the input seems essential to creating a bot with the personality sought

This seems a major shift from the older training process in older engines
And why it was never released for 3.5 I suspect

The amount of information and it’s relevance to the input can’t be stressed enough too much diverse unrelated info will make the bot mute :shushing_face:

Not enough and the bot will resort to its own training

Just short stories that have at least two key words in them from the user input works best

I’m incredibly unclear what idea you are trying to communicate here.

Did you just discover that the models have a context limit, after which they no longer have access to information which is further back in the conversation?


No, not at all. I am keeping a text file with messages that I search for keywords to feed each input keywords are found.

What I am saying is this guidance I read some where talked about feeding the same large prompt response document into the system port for factual relevant information to be used to reply to the users questions

This only works for a while before the bot simply refuses to do it anymore

By instead searching a text file with rows of specific information found by two keywords limited in diverse rows being returned

The bot starts to get consistent

I can now ask my bot it’s. Name and it always knows it now and I can tag rows with “todays news “ that I can keep updated so it is informative and accurate

I am saying using the system input is a valid way to train a 3.5 bot that works consistently if your rows you feed are short to the point of the question

Not more but less and more accurate relevance to the input

I am very happy with where my bot is this five minutes hoping soon to see more users join into the training who know so much more about talkeetna.info

Having the text file public opens up to the faster my bot gets fully trained mr. fox is only about a week old

But can already remember names and do tricks I can see a future in this bot :robot:


This is a known issue with 3.5 Turbo. It doesn’t listen to the System prompt as well after a few turns.

They are planning to patch this.

In the meantime, try GPT-4 if you have access, it will listen to the System much better.

1 Like

Well that is the best news I have heard in some time thank you.

But thinking about it perhaps it really is not a bug but a feature

Our minds have to flush a lot of information to listen to something new

I am just glad I found the sweet spot where 3.5 listens consistently if I obey his rules

Like just a row or two with key words more than two in the row versus the question

Not sure how the replay of last conversation plays into it exactly but context of that seems vital

Now I am playing with saving names in the text.txt file then seeing how I can replay it back

But when I get access it will be fascinating to see how my bot does

Tags seem consistently returning if there is only the one row for them like “todays talkeetna news”.

Name same way i could never get the bot to remember his name long now seems like has not ignored the system input

Consistently seems really dependent upon the keywords and relevance of the system input to the users questions if it it isn’t the bot will not answer or go off into his own training

Your intuition is spot-on. One “hack” is to simply list the System message at the end of the array, instead of the beginning,

1 Like

You mean as it is concatenated with the previous response ? I do that but not sure which is first I only imagine it like $assistsnt_content . $result (of the text search)

When using the “chat” endpoint, the messages contains a single array of dictionaries (maps) and the last element is the system.

Mostly python user here, not php, but should be the same to the API.

The API is stateless, so System should always the last thing you send if you want to try to fix this. So try sending System after you send the “User” / “Assistant” stuff.

This might mitigate the 3.5 Turbo issues pre-patch. Otherwise it’s GPT-4 until patch if you need “long System” type interactions.

Oh I get it I am not sure if the order mattered my api call loads up each of the three roles prior to making the call. This five minutes I am only using the system and user not the assistant. So my concatenated variables are the last GPT response and what I have found in the text file along with users questions

There is an order to loading up the curl call didn’t think that mattered

It might help. Give it a try. Otherwise, GPT-4 in the meantime.

1 Like

So you would reverse system and user?

Perhaps it will make a difference

Well I flipped them and broke it

I had one question that was rock solid in its response that no longer worked

I flipped it back and it started working

So the row that was telling the bot what to say was last apparently and was ignored

When it is closer or first it works

Sounds like exactly what you described I just was fortunate to have the correct configuration originally.

So therefore the bot wants to be told what to say before processing the user request

Makes sense to me now thank you

Not automatically but manually got them off talking I might try and automatically do this

Got my API workout

Now to get them talking to each other and opening up a debate room where others can join in