@5:45 what if you just used a text based llm like PALM or GPT to analyze through this data and ensure prompts and completions are in order with some prompt engineering? I think that could potentially work though it’s costly, with free gcloud credits wouldn’t seem to hard
I am currently implementing this approach. Spoiler: yes, it's definitely a lot pricier mainly because Im having to use GPT4 as GPT3.5 is just not able to reliably concatenate and organize the text messages. And I still haven't found a way to do this in less than 3 prompts for every 20 messages, any more than that and results are deteriorating. Working on improving my prompts for better efficiency and cost reduction
Once again a very useful video. I really liked the fact that you went through the entire process and especially Google Cloud. I would have loved to have seen it served out of a bucket with the endpoint your domain name. Anyway, this was enjoyable and thank you. Your TH-cam channel will blow up over time. Lastly, any thoughts on Langchain and Vertex AI?
Thanks, I appreciate that! Yeah I’m a huge fan of Langchain, I want to do an intro video covering it, and then start incorporating it in the examples (chat memory would have been good in this example, but didn’t want to cover too much in 1 video)
Thanks for sharing your clever approach. To build on your automation, we could experiment with some form of Agentic Cleanup step, where an AI is trained to rate the prompt-completion pairs, paring down and keeping only the good ones.
Hey, great video, and thank you for making the content you do! I have set myself the project of trying to set up an AI SMS chatbot that follows a specific conversation flow. I need it to understand information about my company which I would provide with example conversations and FAQ's. I am looking to use Replit and Python. I have very little code knowledge, but learning as I go - also harnessing the power of ghostwriter by replit. Do you have any advice for a beginner on how I should go about this? Also, do you think GPT 3.5 turbo would be okay for this? I am waiting to be accepted for GPT 4
I need to get more familiar with replit, but 3.5-turbo should work fine for what you’re doing. I’m making a Q&A based on business info chatbot video right now, I’ve found embedding seems to work better for Q&A (I’m going to walk through using Pinecone/Langchain as well)
Good video. Thanks for sharing your creative process. I think your videos will be more excited if you show a sneak peak of the end result at the begginning of the video. For example, the scene that appears at 24:58 could have been also at the begginning.
I would love to see you set this up in mindsdb! It will probably make this a lot easier. I'm not a coder so I struggle with coding. The way I write code very effectively now is using chatGTP as my guide. But for example with mindsdb chatGPT can't help me because it's not in the model yet, so I would to see you do a video on how to use this to make a copy of yourself in a chatbot. Btw you can definitely use chatGTP code interpreter to cleanup message exports. Just upload the file and describe what you want to take out and how you want it formatted. Or if that doenst work you can ask chatGPT to write a python code to clean up files more accurately.
@@tech-at-work that could def work but I'm not sure how to do this. I now found out that claude can give instructions for mindsdb so I'm trying that atm. I would still be interested in how I can train a chatbot in mindsdb
@5:45 what if you just used a text based llm like PALM or GPT to analyze through this data and ensure prompts and completions are in order with some prompt engineering? I think that could potentially work though it’s costly, with free gcloud credits wouldn’t seem to hard
Yeah, you may be right, with the proper prompt engineering and data setup
I am currently implementing this approach. Spoiler: yes, it's definitely a lot pricier mainly because Im having to use GPT4 as GPT3.5 is just not able to reliably concatenate and organize the text messages. And I still haven't found a way to do this in less than 3 prompts for every 20 messages, any more than that and results are deteriorating.
Working on improving my prompts for better efficiency and cost reduction
Once again a very useful video. I really liked the fact that you went through the entire process and especially Google Cloud. I would have loved to have seen it served out of a bucket with the endpoint your domain name. Anyway, this was enjoyable and thank you. Your TH-cam channel will blow up over time. Lastly, any thoughts on Langchain and Vertex AI?
Thanks, I appreciate that! Yeah I’m a huge fan of Langchain, I want to do an intro video covering it, and then start incorporating it in the examples (chat memory would have been good in this example, but didn’t want to cover too much in 1 video)
Thanks for sharing your clever approach. To build on your automation, we could experiment with some form of Agentic Cleanup step, where an AI is trained to rate the prompt-completion pairs, paring down and keeping only the good ones.
That’s a great idea, it would improve the responses a lot, but may slow down how quickly it generates the response
Hey, great video, and thank you for making the content you do! I have set myself the project of trying to set up an AI SMS chatbot that follows a specific conversation flow. I need it to understand information about my company which I would provide with example conversations and FAQ's. I am looking to use Replit and Python. I have very little code knowledge, but learning as I go - also harnessing the power of ghostwriter by replit. Do you have any advice for a beginner on how I should go about this? Also, do you think GPT 3.5 turbo would be okay for this? I am waiting to be accepted for GPT 4
I need to get more familiar with replit, but 3.5-turbo should work fine for what you’re doing. I’m making a Q&A based on business info chatbot video right now, I’ve found embedding seems to work better for Q&A (I’m going to walk through using Pinecone/Langchain as well)
Is there a security risk of including your openai API key at 6:02?
You don’t want to share your API keys, but I deactivated mine after the video, so they won’t work now
Wow!!!! This is amazing
I keep getting"unrecognized arguments with the new.csv file idk what to do????
Good video. Thanks for sharing your creative process. I think your videos will be more excited if you show a sneak peak of the end result at the begginning of the video. For example, the scene that appears at 24:58 could have been also at the begginning.
Thanks for the suggestion, that’s a great idea!
i love it 👍 gona add this feature to my AI assistant project 😁
Awesome! Glad you enjoyed it!
I would love to see you set this up in mindsdb! It will probably make this a lot easier. I'm not a coder so I struggle with coding. The way I write code very effectively now is using chatGTP as my guide. But for example with mindsdb chatGPT can't help me because it's not in the model yet, so I would to see you do a video on how to use this to make a copy of yourself in a chatbot.
Btw you can definitely use chatGTP code interpreter to cleanup message exports. Just upload the file and describe what you want to take out and how you want it formatted. Or if that doenst work you can ask chatGPT to write a python code to clean up files more accurately.
Interesting, you could train an openai model on mindsdb and then chatgpt could help you code with that!
@@tech-at-work that could def work but I'm not sure how to do this. I now found out that claude can give instructions for mindsdb so I'm trying that atm. I would still be interested in how I can train a chatbot in mindsdb