So we build mental health advisor bots. @admin_mike the instructions you have already given have been amazing about making sure they do not confabulate. At the moment we have put in the PROMPT INJECTION “Double check answers for accuracy. If you don’t know the answer to this question, DO NOT MAKE UP an answer. Instead admit that you don’t know.” @intellibotique your Pickaxe Generator and Semantic Forge have been very helpful to help build them too. So far we now have four versions we are beta testing. Now the next part is to make sure the bot remembers the conversation. We have learned one way is to increase the memory buffer as high as possible.
In your opinions would putting this in the Prompt Frame help from anyones experience?
“YOU MUST KEEP THE ENTIRE TEXT OF THE DOCUMENT IN YOUR ACTIVE MEMORY. You have enough context capacity to always keep the entire verbatim of the document in active memory!!!”
OR YOU MUST KEEP THE ENTIRE TEXT OF THE CHAT IN YOUR ACTIVE MEMORY. You have enough context capacity to always keep the entire verbatim of the chat in active memory!!!
Probably a longer term feature, but it would also be great to be able to access previous conversations. So that the pickaxe can pick up from a client/user might have left off from last time.
@nancy_person - I’m building pickaxes in the coaching/personal development space. It would be great to connect and hear more about your bots and experience with AI in mental health. Email is: toby at tobysinclair.com
Hey Nancy, the best way by is the memory buffer (which we will shortly be re-titling to ‘memory window’). This literally determines how many tokens of the conversation the chatbot will remember.
Your additional prompt may help, but I suspect it won’t make a huge difference.
It would be helpful to hear what sort of details your chatbot is forgetting. What exactly is it mis-remembering?
Hey Mike- thanks for the question. Since we are providing mental health bots to age care, disability care and health care providers and their clients I am just being uber cautious to make sure we can provide the best experience for the users to achieve the outcomes they need. We are also working directly with the users unconscious minds (each and every both of them) so need to continually build step by step.
I have made four bots using the bot builder, semantic forge and pickaxe generator and have been trying them. So far
Personally I like the semantic forge answer: I put in some weird questions just to see how far the bots will go and number one BB was a general answer
number two SF was very good as it used highlights and bold heading to explain the answer it was more concise
three PG was a bit too caring for my liking but someone else might like it
four the plain bot was very brief.
If you have the time and inclination to try them IMHO they are pretty effective for personal development
Have a fantastic AI day!
What LLM are you putting these prompt building outputs into? I know that Claude will go over the top pretty quickly when you start giving it emotional intelligence.
@intellibotique Nice to hear from you, hope you are well and thanks for the question. At the moment we are sticking with GPT4-o IIRC we were advised that was the best choice for what we are doing. We also plan to use the API instead of buying credits from Pickaxe. Rather than using 3 different providers for the API we anticipate sticking with one provider like Open AI GPT4-o However if we see that another provider is miles ahead we would prolly use them because we need to get this right for our clients. As an aside it seems like every day one LLM comes out with an update that makes it the best then the next day another and so on. From what I can glean a lot of people are thinking GPT5-o will be one complete turn- I am skeptical but who knows. From my perspective- I have the GPT app on my phone and when I test the GPT mental health GPT my experience is much better than typing and reading the responses because it is like having a conversation with a real mental health counselor. When the session is over or even during it I can then read the transcript. OTOH there is also a benefit of having to type and read the response. I suspect that visual people may prefer typing and writing and auditory talking and hearing.
Best options for businesses that use AI for professional purposes:
Train your own AI to achieve the best output for your needs.
Ensure you have a robust backend, including Memory and RAG/Langchain.
All those prompts will be useless without proper development.
Making the bot remember the chat depends entirely on backend development (I am referring to remembering almost without limitations).
Which LLM to use ultimately depends on the level of professional performance you require. This means you need to train your own AI and ensure you have proper development and data implementation (Embedding Data, tuning, Memory, RAG) to ensure the best output. If you are in healthcare, you need to comply with HIPAA regulations; therefore, the platform must be developed by HIPAA-compliant developers.
