GPT 4: Superpower results with search

[@shamaldesilva9533]

it's literally been one day since the release of GPT 4 and we already have a video about how to practically use it by james ... 🤩🤩🤩🤩🤩 mad respect for you man keep it up 🥳🥳🥳

[@jamesbriggs]

thanks man appreciating your comments 🙏

[@antoninleroy3863]

and we don't even have access to the api 🤣

[@blackenedblue5401]

Also fireship watches your channel lol

[@jamesbriggs]

@@blackenedblue5401 really how do you know this? 👀

[@ferguing]

EXCELLENT video! Subscribing + sharing with all my colleagues

[@jamesbriggs]

awesome, thank you!

[@jackmartin1146]

Great video, my friend. Are there any good advanced resources you have come across to help GPT-3.5/4 answer "followups". For example, the first generated message is good using Prompt Engineering but for generating next message to user's msg requires another Prompt template or context to be added. If we change the prompt template again and again, it does seem like we are only doing Rule based programming (If/Else) rather than using the Capabilities of an AI System. Would love to know your thoughts/resources on how to handle such cases. Please do share! Thanks

[@JasonMelanconEsq]

You did exactly as you said! Thank you! This will help out my devs tremendously! ⭐⭐⭐⭐⭐

[@jamesbriggs]

glad to hear it!

[@StephenPasco]

Great video. I wish you would have had a token counter so we could see the cost incurred throughout each step.

[@paraconscious790]

Great Demo, the last part is crucial, open to imagination about the infinite powerful tools. Thank you very much!

[@jamesbriggs]

there's a ton of possibility, it's really awesome

[@6lack5ushi]

I respect the AI researcher keeping up with the pace of innovation in the space its scary to learn anything its terrifying to learn something that will be out of date with weeks...

[@jamesbriggs]

yeah I have no idea how those people are doing it, it's wild

[@Phasma6969]

@@jamesbriggs They're probably hyper intelligent and neurodivergent. That does it for me.

[@henkhbit5748]

As always great video James. Yes, for specific information or information which chatgpt is not trained a vector database like pinecone is needed. Btw: Alpaca, a new llm based on llama 7 billion and finetuned with chatgpt is also very interesting development where with little training good results are possible...

[@eRiicBelleT]

Do you apply GPL to the data before embedding? What chunking techniques do you use apart overlap chunks? You and David Shapiro are the best at explaining ChatGPT-related stuff. Your videos are amazing, thanks

[@rabanal_josh64]

Exactly What I've been looking for. Amazing

[@jonatan01i]

Based on how Language Models work, isn't it better to put the question before all the information you ask it about? So that it can "read" those having in mind what it needs to pay attention to.

[@adamal93]

Thanks James! How do you properly query the Pinecone database when a user asks a follow-up question? Can you prime a chat model to rephrase the question and pass that output to the database query?

[@jamesbriggs]

we can use the react agent in langchain, will cover this in the langchain series very soon, but in the meantime you can find the code we'll be going through here: github.com/pinecone-io/examples/blob/master/generation/langchain/handbook/06-langchain-agents.ipynb

[@TheCocktales]

the ideas are great but the code is very tangled into one big chunk and its functions are very coupled so it is hard to repurpose it for anything other than this exact use case. it would be nice if in the future you could have a tutorial with a more modular approach

[@EleanorPrashamshini]

Hi This is a great intro. Thank you! Where can I find the code for this?

[@GayathriG-h5h]

Can we use gpt4 in llmchain call instead of chat completion?

[@wdgxhitman3423]

Can we save it as a updated pre-trained model? So that next time we can use directly without feeding the new data?

[@jamesbriggs]

unfortunately there's no way to do this right now using gpt-4 - you could try 'distilling' the knowledge from gpt-4 into a smaller local model, to do that you'd just create a ton of training data with gpt-4 and then use that to train your smaller local model

[@ylazerson]

awesome video - thanks again! Keep em coming!

[@sergedadesky5638]

James. Brilliant demo. Great explanations. Was able to replicate most of this. Ran into snag trying to implement mode InvalidRequestError: The model: `gpt-4` does not exist. I am signed up for gpt-4 , I believe. Has this model been deprecated ?

[@FelipeNovaesRocha]

Make sure your openai package have the last version.

[@tajwarakmal]

Awesome video! Question: is there a way to set it up such that you can ask clarifying questions? As in continue the chat to dig in deeper?

[@BigFattyNat]

bro this is such a goofy huge amount of work...just copy pasta the api docs into gpt4 and then ask your question on the next prompt

[@jamesbriggs]

you're limited to 8K tokens with GPT-4, so a few pages only - they will release a 32K model (eventually...), so that will help but still far from being the full docs

[@slayermm1122]

your code works. thanks very much for sharing.

[@hvbris_]

Thanks for the great video James, is it possible for this type of bot to have memory?

[@johnwallis1626]

keep up the great work james

[@absta1995]

So does this auto search by itself and create embeddings or do you have to search yourself and create the embeddings beforehand?

[@midnightmoves7976]

I think the market is well and truly cornered now.

[@0730pleomax]

Will James update LLM-related sections to the Udemy course?

[@imran.shabbir]

Amazing, great explanation

[@mikemansour1166]

Hello I was wondering about Chunking and chunks overlap , Does the language model connect the chunks , and what benefits are there for smaller and bigger chunk sizes ?

[@lutune]

Great content!

[@jamesbriggs]

thanks Luke!

[@chatwithdata]

Good stuff 🙌

[@DarioMader]

Such a good video🎉

[@mariodelatorre6047]

Excelente video! q soft usas para hacer los gráficos con el lápiz?

[@jimmynguyen3386]

James is the man!

[@venkateshr6127]

can we use T5 as a generator in RAG QA model if yes then how can we use it ?

[@arnoldlai822]

this way of text cutting is mechanical, may loss contextual meaning. Is there any better way?

[@Phasma6969]

Vector database and embeddings

[@arnoldlai822]

@@Phasma6969 it is using pinecone vector db and openai embeddings right now

[@Phasma6969]

@@arnoldlai822 owh oops yes that's the way for now, until we get longer context length through other mechanisms. I saw some on the OpenAi subreddit

[@chrisalmighty]

A better way is to actually do proper text processing into reasonable breaks by heading or topic.

[@sersoage]

One question regarding the splitter what if i do not want to split by chunks just by a separator?

[@jamesbriggs]

you could just do "your text goes here".split(" ") -- but replace " " with whatever separator you'd like to use

[@WereOutOfBuziness]

Kudos to you!

[@ricardocosta9336]

That wacon pen. I was here before it was cool.

[@jamesbriggs]

you've been here from the start haha

[@DarrenTarmey]

Were can I find all instructions to do this.

[@jamesbriggs]

for the openai side of things, I think I used their ChatCompletion docs ( platform.openai.com/docs/guides/chat ), for Pinecone the pinecone docs ( pinecone.io/docs ), and if you just want the code, I put it all here: github.com/pinecone-io/examples/blob/master/generation/gpt4-retrieval-augmentation/gpt-4-langchain-docs.ipynb

[@KyleRosebk]

Good video

[@JohnMcclaned]

Copy pasting the OG GPT-3 videos with a GPT-4 variation 😆

[@TheChocolatBlanc]

Can’t wait for the same video but for GPT-5

[@fullcrum2089]

[@jamesbriggs]

🤫

[@Dr.Trustmeonthisone]

nice