Thanks for sharing, you're totally right. I just read through their code, it was interesting to see their prompt methods and how they implemented their memory store! Cool feature
We were just talking about this issue. All of these chat UI's treat content as if it's disposable. What works in text messaging between two people doesn't translate as well when working with computers, not if the information has utility value. This is great!
This is so great: structured data extraction from conversations is something I’m also working on. And by the way: congrats: you have a better like-to-view ratio than MrBeast, at 4%. 🚀
Hey thanks for this. Very nice to see an actual application build up like this. I will most certainly come back for more. I would be interested in seeing you set up a more corporate oriented use case. What if a free text field in a form contains information relevant to a sign up for a a services or something like that
I think I missed this in the memgpt paper but summarizing and storing attributes in long term memory and then refetching it for the context window is also likely going to increase the latency of the main response
Agreed, that's something you would need to monitor for sure. For this version, I chose to just pass the memories along in the request so they wouldn't delay the response, but that will break down once the memory store gets to a point where you need to trim it. I think you could always pre-fetch the entire condensed form of the conversational memory ahead of the user's request, but it might be hard to pre-fetch just the memories relevant to the new query in the small window of time between when the person finishes their message and submits their request. Someone pointed out AutoGen's Teachable Agent approach to memory management, and they warn: "Note that this adds extra LLM calls involving a relatively small number of tokens, which can add a few seconds to the time a user waits for each response." But it looks like their approach adds multiple calls before and after the main request, which are all blockers to a response.
Hello Deploying AI, such an amazing tutorial and it was really useful in my project. Can you please share the code for the Supabase's Postgres database backend and the UI please?
Great, thanks for watching! I'll need to clean it up a bit, but I'll share it when I have a chance. I used Radix UI and Tailwind for this demo, so they made this very easy to throw together.
I have one query here. Conversation memory and context length are different ? . If the input context length of the LLM is 32k. then Prompt Input + conversation memory context length should not exceed 32k right ? please correct me if I'm wrong.
Wow, this is super cool. Could something like this be applied to using an LLM to code a web app? One of my problems right now is that I'm using gpt4 to help code a project but before I can get through it, the context window creeps in and it starts to return incorrect code.
Cool channel & video!! May I ask how long it took in order to get some views on your videos? Did you got views within 24 hours? Or did you start to get some views after X uploads?
How do you store the memories ? I prabably implemented it wrong, I did a while loop with input, now trying to come up with a solution of how to actually store the messages in the memory
Can you share the details about how you implemented vectorDB in the loop? Would love to see the code. Much love, appreciate the effort of making the video. thanks!
In my demo, I set up a backend using Supabase's Postgres database. The agent writes the memories directly to the database, and I stream the memories from the database to the frontend to instantly see updates. And since I have the memories available already in the frontend, I just use those in my requests to the LLM so that I can avoid having to make a request to my db before every new request to the LLM.
Ha! Is it weird that I feel guilty about it now? Thanks for pointing this out, you sent me down a rabbit hole of great papers on prompting techniques and a new idea I want to try out. If it works I'll share a video! I really liked this paper on eccentric prompts: arxiv.org/pdf/2402.10949.pdf
@@deployingai Can’t wait to see what else you come up with! After reviewing that document on eccentric prompts, it actually has me thinking I need to employ more ToT prompting into my own applications… 🤔
Great code walk through. I'm working on something similar so it was cool to see how you approached it. Thanks for sharing.
That's awesome! I'd love to hear what approach ends up working for you
Nice. Look similar to Autogen's teachable agent. Appreciate your work.
Thanks for sharing, you're totally right. I just read through their code, it was interesting to see their prompt methods and how they implemented their memory store! Cool feature
Dude, your video help me a lot, THANKS!!!
That's awesome, glad to hear it!
Superb content, Thanks for sharing ✨
We were just talking about this issue. All of these chat UI's treat content as if it's disposable. What works in text messaging between two people doesn't translate as well when working with computers, not if the information has utility value. This is great!
Great explanation, thank you!
your miro drawing skills are next level thing
i have a suspicion that @deployingai is creating AI videos to showcase his miro skills.
Haha thanks!
Thanks for the excellent idea and explanation.
Thanks for watching!
This is so great: structured data extraction from conversations is something I’m also working on. And by the way: congrats: you have a better like-to-view ratio than MrBeast, at 4%. 🚀
That's cool, thanks for sharing! I'd love to hear the approach that ends up working for you.
Excelent content! Thanks for sharing, man :D
Hey thanks for this. Very nice to see an actual application build up like this.
I will most certainly come back for more.
I would be interested in seeing you set up a more corporate oriented use case. What if a free text field in a form contains information relevant to a sign up for a a services or something like that
Hello this is so cool, is it possible to share the full code including backend/front-end code as well, would love to try this, thanks so much!
lol
Hahahahahahah
Can you share the vite front end? or how you setted up the front and the backend?
Super loaded tutorial 🎉
Thank you, Aiming to build a similar one for a different use case. If you can share the code with appropriate licensing would be helpful !
I think I missed this in the memgpt paper but summarizing and storing attributes in long term memory and then refetching it for the context window is also likely going to increase the latency of the main response
Agreed, that's something you would need to monitor for sure. For this version, I chose to just pass the memories along in the request so they wouldn't delay the response, but that will break down once the memory store gets to a point where you need to trim it.
I think you could always pre-fetch the entire condensed form of the conversational memory ahead of the user's request, but it might be hard to pre-fetch just the memories relevant to the new query in the small window of time between when the person finishes their message and submits their request.
Someone pointed out AutoGen's Teachable Agent approach to memory management, and they warn: "Note that this adds extra LLM calls involving a relatively small number of tokens, which can add a few seconds to the time a user waits for each response."
But it looks like their approach adds multiple calls before and after the main request, which are all blockers to a response.
Hello Deploying AI, such an amazing tutorial and it was really useful in my project. Can you please share the code for the Supabase's Postgres database backend and the UI please?
Hey, awesome work! Can you please share the UI code?
Interesting. Subscribed. By the way, can I get the source code for the UI. It looks pretty neat and clean
Great, thanks for watching! I'll need to clean it up a bit, but I'll share it when I have a chance. I used Radix UI and Tailwind for this demo, so they made this very easy to throw together.
Hello@@deployingai possible to share the code now?
Hello @deployingai - amazing tutorial. Thanks a lot! Can you please share the source code for the UI as well please it will be really useful? Thanks.
This is amazing, I creating something similar, Now I will use sentinal approach. btw, I would like ask, where can i find the frontend?
Would you be able to share the doc of Mem GPT?
Great work, don't you mind to share the code for front end? ;)
I have one query here. Conversation memory and context length are different ? . If the input context length of the LLM is 32k. then Prompt Input + conversation memory context length should not exceed 32k right ? please correct me if I'm wrong.
Thanks for sharing 👍
Did you forget to add the graphQl frontend code to Github ?
Wow, this is super cool. Could something like this be applied to using an LLM to code a web app? One of my problems right now is that I'm using gpt4 to help code a project but before I can get through it, the context window creeps in and it starts to return incorrect code.
Great content, not only this visit. A shame you ghosted. :(
Cool channel & video!! May I ask how long it took in order to get some views on your videos? Did you got views within 24 hours? Or did you start to get some views after X uploads?
Hey thanks for watching! I've only published a handful of videos so I don't have much info for you, but the first few were definitely pretty quiet.
How do you store the memories ? I prabably implemented it wrong, I did a while loop with input, now trying to come up with a solution of how to actually store the messages in the memory
Can you share the details about how you implemented vectorDB in the loop? Would love to see the code. Much love, appreciate the effort of making the video. thanks!
In my demo, I set up a backend using Supabase's Postgres database. The agent writes the memories directly to the database, and I stream the memories from the database to the frontend to instantly see updates. And since I have the memories available already in the frontend, I just use those in my requests to the LLM so that I can avoid having to make a request to my db before every new request to the LLM.
“I won’t go through how I actually built the application” what’s the point ? I switched off here.
great, now do this with an open source LLM
13:38 - Did... Did you just bribe and / or threaten your LLM to produce better results? lmao
He was also compassionate and asked the LLM to take a deep breath
I loved that part, as well! lol@@akashdeb9823
Ha! Is it weird that I feel guilty about it now?
Thanks for pointing this out, you sent me down a rabbit hole of great papers on prompting techniques and a new idea I want to try out. If it works I'll share a video!
I really liked this paper on eccentric prompts: arxiv.org/pdf/2402.10949.pdf
@@deployingai Can’t wait to see what else you come up with! After reviewing that document on eccentric prompts, it actually has me thinking I need to employ more ToT prompting into my own applications… 🤔