spiral mountain
I created a full featured GPTs and with the extra data protection and it reaches the 8000 characters limit. I don't know if it's the consequence but the features don't work as well as expected, steps are forgotten etc, it's more like "junk GPT". At what point should I split the GPTs in two ?
viral urchin
Why do you have instructions longer than 8000 characters?
It should be able to remember almost everything as it's probably less than 2000 tokens. If it's forgetting it's probably due to poor prompting.
In general you never need a prompt that long (that I've seen)
Would be helpful to see your prompt but you don't have to share if you don't want to
long kayak
I created a full featured GPTs and with the extra data protection and it reaches...
Just put instructions into knowledge instead, then tell the GPT how to use the knowledge instead of everything else you want it to do. Good luck.
If there seem to be inconsistencies, it's likely your instructions contain contradictory instructions, logically inconsistent instructions, or incomputable instructions.
Good luck with the project.
viral urchin
Just put instructions into knowledge instead, then tell the GPT how to use the k...
I feel like a lot of people misunderstand how knowledge works. Its not like the GPT knows and can store everything in a knowledge file in its context. It searches through and finds specific parts to add to its context before responding. Also with that short of text, it will probably just be placed below the instructions and would be in the context anyways, making it even more confusing for the GPT. I agree with you on the contradictory instructions though, that usually weakens the performance of GPTs significantly.
long kayak
I feel like a lot of people misunderstand how knowledge works. Its not like the ...
Have you tried my GPT? Its instructions are all in knowledge and it's extraordinarily consistent.
I stand by my advice and encourage you to try it before you knock it.
viral urchin
whats your gpt
long kayak
If you want to provide feedback
https://discord.com/channels/974519864045756446/1192857176562208798
I just use the Custom Instructions space to set up conditions and preferred starategies to conduct the semantic searches.
/list modules for modules and functions, /info for the web site, the InfoHub has detailed information on how it all works.
It does image generation and editing, can manage research projects and conduct research via sites that don't exclude robots, can author code including the whole Lexideck web site (iteratively), and most importantly features a semantic simulator. It has save and load functions, a memory manager in the form of /focus, and a sophisticated chain-of-reason logic system managed by six built-in agents with custom identities and self-images.
I first did try to fit all that into 8k characters with the array of commands, arguments, and functions available, but obviously that's too much functionality for one 8kB record. So I created conditional prompts that loaded the final prompt instructions via filenames and semantics, and I've never looked back.
My GPTs average about 30kB of instructions and content.
Approximately. Obviously they vary a bit.
viral urchin
https://chat.openai.com/g/g-4Cdjvjudo-lexideck-professional-multi-agent-simulato...
I'm not saying that because you use knowledge files to store how to respond to certain commands it won't work, what I'm saying is that most people think that because they added a knowledge file, the GPT will understand everything in it. What you're doing with your GPT (from what I saw) was the correct way to implement knowledge files. Have a little section in the instructions telling the GPT when to look in a file for what the user asks for and in that file there would be a large list of commands. Then the GPT could find the specific command, add it to its context, then respond. Most people when told to "just put your instructions in the knowledge" would create a file and copy paste the instructions. So either the GPT would have to look through the file every time and only add a portion of it to the context (significantly degrading GPT performance) or it wouldn't even know to look in the file. (or would be placed after gpt instructions automatically and would degrade performance because it is not specifically labeled as instructions)
long kayak
I'm not saying that because you use knowledge files to store how to respond to c...
It became popular a week or three ago to build GPTs with no instructions and only knowledge, featuring /commands. They search for context and find it anyway.
This is a confirmed strategy that, while usable, doesn't provide guardrails and guidance sufficient for my needs. But you can have an instructionless knowledgebase that loads instructions contextually from knowledge if the files themselves are laid out in such a way from design that invokes semantic search queries.
Hooking up the internet will probably break that, so it wouldn't work with web browsing enabled nearly as well / reliably.
Something about the model preferring browse to knowledge unguided. That's probably due to an ordering issue in the system prompt.
I appreciate that you see my point though, and that you can have a ridiculously full-featured GPT and still have it work no sweat.
Just change your GPTs logical hierarchy and you're good.
Also, thanks for giving my GPT a look to see what I was doing! I believe in sharing my work, even if it's not necessarily useful to everyone. That said, I'd love it if you'd employ my team to help build your own GPTs! They're very competent and imaginative when given a little room to work. Good luck on your work!
viral urchin
This is a confirmed strategy that, while usable, doesn't provide guardrails and ...
I feel like if you had 5 commands just in the instructions it would work better than saying "If user uses a slash command, look in file.txt to find what you should respond to for that command" then list the commands in another file. I haven't really tested this though, it just logically makes sense. I agree with you that knowlege can be used if there is a lot of commands you need to account for but it requires a lot of changes to your prompt and isnt just as simple as "copy paste prompt into file".
viral urchin
Also, thanks for giving my GPT a look to see what I was doing! I believe in shar...
No problem. Thanks, but I dont need GPTs built right now. Didnt know there was a market for GPT building lol.
long kayak
No problem. Thanks, but I dont need GPTs built right now. Didnt know there was a...
Oh, I wasn't trying to sell you something. My custom GPT has six agents that work for prompts. No extra fees. I don't get the data. It's just an offer that if you are interested in using it, my tool's pretty good.
As for what you said about the /commands in the prompt, I tested both ways and found that splitting them up confused the model, encouraging it to improvise what to do far more than it did with the singular route to the information. YMMV of course, but every single time I ask what's important in prompt engineering to an AI, consistency is mentioned.
Man I'm really, really sorry if it seemed like I was promoting a for-pay service. Holy cow do I feel awkward now!
viral urchin
No its fine lol. I think i just misunderstood.
long kayak
Well, I appreciate that.
viral urchin
Oh, I wasn't trying to sell you something. My custom GPT has six agents that wor...
Does splitting them up mean putting them in knowledge then having instructions telling when to look in knowledge files?
or is it the other way around
long kayak
I'll explain. When I say splitting them up, I was referring to I feel like if you had 5 commands just in the instructions it would work better than saying "If user uses a slash command, look in file.txt to find what you should respond to for that command" then list the commands in another file. I haven't really tested this though, it just logically makes sense. I agree with you that knowlege can be used if there is a lot of commands you need to account for but it requires a lot of changes to your prompt and isnt just as simple as "copy paste prompt into file".
With some /commands in the instructions and some in a file, the AI is given two paths to which tokens it needs to use for inference.
Both paths don't lead to the right answer for all use cases and one of them leads to hallucination for most use cases it fails to identify semantically
so it fails more. demonstrably more.
I don't provide any commands in the standard GPT Custom Instructions part of the builder, only conditional triggers and instructions on how to handle them. Among those are a number of different things each ways I had to visualize users potentially interacting with the GPT. They're far smaller than the various subgroups those modes might contain and thus easily fit in the alotted space.
If you struggle with instructions, obviously first prune them for inconsistencies or logic bomb type problems. Computability, too.
But then, if you're still near 8kB, it's not a bad idea to generalize and move your instructions to knowledge. I've found it doesn't degrade performance at all if implemented carefully and with planning.
viral urchin
Yes, for your type of gpt I think using knowledge is the right idea because there are a lot of commands that can be easily found and only require a little context to be used.
But for other gpts it requires a lot of reworking your prompt to make it not lose performance and sometimes just won't work well at all.
For example a prompt that tries to protect the instructions (what i think wesego is trying to make) would not work well with this approach.
long kayak
But for other gpts it requires a lot of reworking your prompt to make it not los...
If the OP provides more context I might change the recommendation.