How to Build Custom GPTs with 100% Accurate Answers

In this video, we're tackling a common frustration with AI - getting accurate answers from your own documents. Forget about ChatGPT's hit-or-miss responses; we're diving into the world of Pinecone Assistant, a game-changer for document-specific AI interactions.
Discover how to:
Create a near-perfect knowledge base from your own files
Get pinpoint accurate answers to complex document questions
Integrate this system with custom GPTs for enhanced capabilities
Set up a simple automation for streamlined document processing
Whether you're new to AI or a seasoned pro, you'll learn practical, actionable techniques to revolutionize how you interact with your documents using AI. I guarantee you'll walk away with valuable insights that will transform your AI document queries.
🚀 Pinecone Assistant Mastery Kit: bit.ly/4gJr9CR
📞 Build Your AI Receptionist With Us: bit.ly/4e0sS4A
🌐 Visit My Agency Website: bit.ly/4cD9jhG
🚀 Work Together on Fiverr: bit.ly/3XorT7R
📅 Book a Consultation: bit.ly/3Ml5AKW
📰 Join My Newsletter: bit.ly/3WVEHlK
👋 About Me: I'm Mark, owner of Prompt Advisers. For the past two years, we've been helping companies across industries integrate AI effectively into their workflows. With my experience in AI automation, I'm here to help you maximize the potential of your AI projects.
Don't miss the special reveal at the end - a super simple Make automation that lets you upload files to Google Drive and get incredibly accurate responses in seconds!
#PineconeAssistant #AIDocumentAnalysis #CustomGPT #AccurateAI #DocumentIntelligence #AIKnowledgeBase #MachineLearning #TechTutorial #AIInnovation #DataScience #AIAutomation #DocumentProcessing

[@jannismoore]

That's a tutorial of my liking! Love the depth!

[@tonynolan9428]

Hi Mark. Just started watching and your intro used to have you commenting on your previous years in ai b4 your business. There plenty out there with 2 years experience now but your point of differnce is your education in this field. Keep this in your intro for newbies. It gave you my subscribe.

[@Mark_Kashef]

Tony, this is a fantastic callout and I really appreciate it -- I tend to think that constantly mentioning the 10 year background is a bit of a 'subtle flex' so I typically stray away from it. Wording it the way you did though, I think there's some merit for me to mention it in the context as to 'why you'll find this more valuable than someone else's video who doesn't have the background' which I didn't consider as much -- noted and TY for reminding me!

[@themank945]

You make it look so easy! Great work, and I love the content. Thanks for your generosity!

[@Mark_Kashef]

I'm glad it comes off that way haha - and really appreciate your feedback 🦾

[@Ethiventure]

This sounds amazing. BTW I watch lots of KZbin and you’re the best at properly explaining AI. I’m immediately thinking about how to organise my files. Is there a number of files at which it doesn’t choose the correct file reliably? I’m wondering about whether to have an AI staff team folder that’s all their files - like a human staff shared drive rather than each customGPT having separate files? At the moment they have knowledge files, some unique to the customGPT , some which most of them have 🤔

[@Mark_Kashef]

first of all, thank you so much for the feedback! Really appreciate it :) It has a limit on the front-end of 10 files (can be very long files) -- I would organize the files to be bundled by theme ideally if you want to load many files. Via the API, I think you can get upto 20 files; it's really good at retrieval even with 10 large files from my testing. I would personally have a different google drive per person so there isn't too much accidental cross-pollination,

[@peterbell2446]

Thanks Mark, great content. Keep it coming!

[@Mark_Kashef]

Peter! Thanks so much for this feedback as well as your generous support for the channel -- very much appreciate both. Cheers!

[@ammoacid]

This is too amazing! I love it. Been digging into RAG and thought NotebookLM is the next best thing but this is on another level!

[@Mark_Kashef]

Thanks for the feedback my friend! NotebookLM is awesome, I’d say this is a cool way to supercharge any existing application or GPT, where is NotebookLM is good for some quick and dirty RAG tasks 🦾

[@Reflowflow]

Mark!! Again, top tier content! Can you do a video on Vercel?

[@Mark_Kashef]

Thanks so much! Did you mean Vercel as a product as a whole or just v0?

[@andreisas1621]

Great video brother, thank you!

[@Mark_Kashef]

thanks so much!! appreciate your feedback

[@vinoe__7473]

That's exactly what I search 🙌🏻❤

[@Mark_Kashef]

Glad I could drop this for you just in time!

[@pandipatipavan3804]

​@@Mark_KashefPLEASE do provide the code files and the links in the description as soon as possible for free 🙏 thank you so much.

[@Mark_Kashef]

you can just enter $0 and not pay :)

[@pandipatipavan3804]

@@Mark_Kashef I didn't get you brother. Where I need to enter 0 dollar and where is the link for that ?

[@Mark_Kashef]

@@pandipatipavan3804 this is the link to the resources that's already in my description: bit.ly/4gJr9CR

[@LacrimaLucis]

I was looking for a working example of custom GPTs that connect to a user RAG server for a long time. Thank you for this recipe! I want to host that vector db (+ key store) locally (no replit/pinecone servers needed), and your workflow here is still a good practical model to contemplate. But for those without local servers, wondering if you can show something similar using the Supabase free tier as the OpenAPI endpoint in custom actions?

[@Mark_Kashef]

I'm glad the recipe tastes good aha -- interesting question, are you trying to run custom GPTs locally using the ChatGPT Desktop app? I use OpenAI and Qdrant locally, so curious to hear your stack to see if I can help!

[@tonyblack2141]

Thanks mate. Really enjoyed this one as it gave me ideas. Is it possible to somehow bring in LlamaParse into this mix where it handles technically complex documents (with tables, charts, illustrations, formulars/maths etc.) as well as non-true pdfs requiring OCR first, before the extracted Markdown then gets passed to pinecone for embedding via Google drive?

[@Mark_Kashef]

happy to hear you liked it! In terms of your idea, what I would do is extend that Make scenario I showed at the very end to enable an OCR-friendly workflow. Perhaps you wait for any document in the Gdrive, then you use the PDF4Me module with the PDF OCR trigger OR run it through a custom code python module that runs your llamaparse script, and then follows the existing workflow I outlined. Totally possible, just will need some automation love to get it to work.

[@lakergreat1]

Does pinecone have Canadian instances? Also any recommendation on Canadian hosted LLM 's for business use cases?

[@Mark_Kashef]

It’s hosted on AWS and GCP environments so I think it’s primarily US based regions

[@lakergreat1]

Is notebook LM poor with their RAG methods as well? I found it to be pretty good

[@Mark_Kashef]

NotebookLM is actually pretty good! This use case is super helpful to improve existing apps that need an extra bit of juice for extra performance!

[@florianrolke]

Hey Mark! Great video. Would you be willing to make a tutorial how to set up the relevance ai in-built vector database?

[@Mark_Kashef]

hey Florian! thanks so much -- frankly I stopped using Relevance AI a long time ago after running into many scalability issues with it. Not a big fan

[@florianrolke]

@@Mark_Kashef Got it! Thanks for that perspective. Could you share briefly what is likely to go wrong with relevance as you scale? I think a lot of people in the comments might want to know.

[@Mark_Kashef]

@@florianrolke for sure -- again my experience is 4-5 months old. When executing a relevance AI workflow, I many times received timeout errors when the execution of that workflow exceeded 1-1.5 minutes at a given time. Also, the way they do RAG (retrieval augmented generation) 'was' fairly basic, so when it comes to more complex true PDFs, it struggled. Not to mention that it was hard to pinpoint exactly where the information was coming from, and whether a wrong answer was because the LLM was wrong, versus the knowledge base, or both. Again, this is a multiple month old personal review :) could have gotten way better since then.

[@florianrolke]

@@Mark_Kashef Thank you for that detailled response. Super valuable ! Last question if you don't mind: I am right now wrestling o1 to give me an open api schema to deploy with agentive instead of replit. That's technically possible, right, or am I missing something?

[@Mark_Kashef]

@@florianrolke agentive and replit are not interchangeable; agentive has specific endpoints you can deploy to, replit allows you to build those endpoints from scratch

[@mzafarr]

Isnt there a way to give prompt (instructions) to pinecone assistant? So we can get a response following some instructions.

[@Mark_Kashef]

In a few weeks to a month, I believe this is coming! 🦾

[@philosafi]

The Pincone Kit link in description is broken.. Thanks for timely video.

[@Mark_Kashef]

I just tested it, seems to work fine for me! Here’s the link again: bit.ly/4gJr9CR

[@philosafi]

@@Mark_Kashef It works, thanks. Instead of PDF, can pinecone accept relational CSVs? I hope to hear from you more about the economical alternative to GraphRAG someday soon. After a decade, you're the first resourceful KZbinr I have started to comment and interact with. Keep up the great work!!

[@Mark_Kashef]

@@philosafi thanks so much again for the complement, super appreciated :) The Pinecone Assistant API only accepts PDF or txt files currently -- I usually format CSVs into PDFs so that I can process them.

[@philosafi]

@@Mark_Kashef Formatting the CSVs into PDFs could lose the dataset's relational nature I guess.I have spent months on my custom dataset and also have a pretty good panel-of-experts CustomGPT (thanks to you again) but being apprehensive, I've never uploaded my dataset onto OpenAI platform. The advantage of the latter is I can access the chatbot on my other devices, therefore I'm now looking for a local CustomGPT powered by GraphRAG level accuracy using my relational CSV dataset.

[@tuaitituaiti1565]

Value 💣ing again. So greatful. 👊 Thank you, Uncle Mark & Co 🙏🎯🫡🤙

[@handfacts]

How about how to use make to access the pinecone vector database?

[@Mark_Kashef]

Did you mean the ‘old fashion way’? One of my older videos is just about that with some custom code that still works today (my infinite memory video from a few months ago)

[@janalgos]

Any way to do this within Poe or Claude?

[@Mark_Kashef]

from what I've seen, I don't think Claude Projects has access to custom actions yet -- POE bots don't seem to accept API requests either from an initial scan

[@we.bastos]

My ChatGPT is giving this error "405 Method Not Allowed" when call the action, do you know what could be?

[@we.bastos]

I found the problem; it was because there was a '/' at the end of the address that I copied.

[@Mark_Kashef]

Was about to literally mention that, glad you spotted it!

[@Desire398]

This was a very informative and helpful video. Thanks for your help 🥹

[@Mark_Kashef]

pleasure to hear that Desire! My pleasure as usual 🦾