OpenAI Assistants API - Part 1: information retrieval functionality with working code

No video

OpenAI Assistants API - Part 1: information retrieval functionality with working code

Рет қаралды 14,538

9 ай бұрын

This a 3 part series to explain and demo how OpenAI's Assistant API works:
- Part 1: Information Retrieval ( • OpenAI Assistants API ... )
- Part 2: Adding Agents with Function Calling ( • OpenAI Assistants API ... )
- Part 3: Code Interpreter with Run Steps walk through ( • OpenAI Assistants API ... )
To get a copy of the code on google colab notebook, come find and follow me (Lily Cheng) on Linked-in and I'll send it to you!
OpenAI just announced the brand new Assistants API on DevDay and it's so hot, it's taken down the whole chatGPT and API service! Assistant API basically enables you to easily build a custom chatbot on top of the GPT4 LLM that knows how to retrieve additional information from files that you upload and also can take actions (like agent functionality) by being able to call up to 128 functions that you give it. You would've heard about "GPTs" from DevDay - this is the same functionality but exposed via API
In this video, I will explain the new lingo and how it all works conceptually, then work through a retrieval demo and also show you how it works on the no-code interface.

Пікірлер: 61

@chimoji608 Ай бұрын

I think for non programmers like me... this might still be too next level XD

@narindermahil6670 9 ай бұрын

The best video on Assistance AI

@samsquamsh78 9 ай бұрын

very instructive! really well explained, and nice walk through of your mental model of the flows of the assistant and the thread. 5/5 stars🙂

@briancoder7621 9 ай бұрын

Very succinct and clear. Thank you very much, Lily.

@ranymagdy4217 9 ай бұрын

That was very informtive , thank you so much and diffenetly following you and looking forward to connect with you and to use the colab file

@ManiSaintVictor 9 ай бұрын

Thank you! This is very clear and useful. I finally understand how to implement the code.

@pindopurba 9 ай бұрын

Thank you very much, i manage to work it out following this.

@XCmdr007 9 ай бұрын

What an amazing guide Lily thank you! I just followed you and sent you a connection invite. Looking forward to connecting and hopefully having productive discussions 😃

@hubel-labs 9 ай бұрын

Thanks!

@jaysonp9426 9 ай бұрын

For their RAG it's just simple retrieval. They said they're looking at adding more strategies but I'd still use my own rag system with function calls for now

@hubel-labs 9 ай бұрын

Yes. There are some file number and size limitations as a RAG system that is quite restrictive. I made a second video to demo the more advanced agent and function calling functionality: check it out OpenAI Assistants API - Part 2: using Agents with function calling together with info retrieval kzfaq.info/get/bejne/eLxdkpiQ15qam3k.html

@oscaralbani6303 8 ай бұрын

Nice explanation!. Thank you

@eternalia4059 8 ай бұрын

Thanks, great job

@rluijk 9 ай бұрын

Thanks for both. Very clear explanations.

@Capjuancode 8 ай бұрын

My hero thanks loved your video.

@shoesholic 9 ай бұрын

Thank you! Amazing explanations!

@jozonas 9 ай бұрын

Thank you for the Video.

@joexia1957 9 ай бұрын

Great tutorial, thanks!

@balrajvishnu 9 ай бұрын

Why would you want to use Google Collab when we can use the playground

@yifeitong 9 ай бұрын

Many thanks for the video ! I am wondering about the data privacy when uploading files and what if the file(s) would have too big size like a relational database. I am working on a project using GPT API (0613 version) to do some data analysis over local files. Having concerns about data privacy, I only use prompt describing the file and get python code through API. Without having the uploaded file, GPT itself cannot make a self-check whether the code reaches what I asked for. Do you have any suggestions on this ?

@hubel-labs 9 ай бұрын

I would be cautious about uploading anything too sensitive unless your company data security team has vetted it. I work on the assumption that OpenAI’s core focus is foundational model research and development and not enterprise cloud - so…. While they say they are SOC compliant and do data encryption etc…. Many companies still prefer to use Azure’s enterprise offering of the openAI services. If you are concerned, you can use assistant API to do a function call to your backend that retrieves only the relevant pieces of into to pass to openai. You are still passing it data but at least not everything

@kevinkyosunlee 9 ай бұрын

Thanks for the video. It is really helpful. I just followed you on LI as well. Is it possible to get codes for this video? Thanks!!!

@hubel-labs 9 ай бұрын

I haven’t seen you on LI yet - drop me an LI message if u still don’t have the colabs

@elrecreoadan878 9 ай бұрын

Hi! What differences are there between GPTs and Assistants? When is it adequate to use one or another?

@hubel-labs 9 ай бұрын

Conceptually quite similar but GPTs currently have some advantages over the Assistants API. Who knows how quickly that will change. Firstly, GPTs can be built using a no code interface (though feels very buggy right now) - I was quite impressed that I can even put in my own APIs which require authentication in a no code way. Secondly, GPTs can fetch additional data from the web in a real-time search to augment the response, Assistants API, I believe cannot so far. Thirdly, GPTs can do image generation from Dalle, I don't think Assistants can do that yet. The downside... other than having less flexibility.... is that you need to have OpenAPI (note... this is totally different from OpenAI, confusingly) aka Swagger specification in either YAML or JSON format for your API - whereas the Assistants API seems a little less fussy about that. Arguably, you should already have that if your API was developed / documented properly.... but not sure how many people do that in reality. I'm planning to do a video on GPTs in the next day or two .... as soon as the system stops being so buggy. Anyone else have any insights, please chime in! All ears!

@elrecreoadan878 9 ай бұрын

@@hubel-labs wow awsome response. Thanks a lot!

@Happyfortunestudio 9 ай бұрын

😍😍😍😍😍😍

@KB-047 9 ай бұрын

I've been trying to find a sample template to feed to the assistant. What should the format look like? For example can i feed product specs sheet in CSV file?

@hubel-labs 9 ай бұрын

You mean as a file? It can be almost anything - here is a list of supported files. platform.openai.com/docs/assistants/tools/function-calling

@KB-047 9 ай бұрын

@@hubel-labs Thanks. I decided to go with word doc, it seems to be working

@oscaralbani6303 8 ай бұрын

Hi Lily, s it possible to get codes for this video? Thanks!!

@hubel-labs 8 ай бұрын

Find me on LinkedIn, search “lily cheng Hubel” and I’ll drop u the colab links there!

@MannyBernabe 9 ай бұрын

excellent visual

@LearnWithShajeel 9 ай бұрын

Thanks, which tool did you use this illustration?

@hubel-labs 9 ай бұрын

Excalidraw

@itsjustmeemman 9 ай бұрын

Do threads automatically expire or get deleted ? I cannot find documentation about handling threads. Though its still in Beta I'm worried about the current context length of a thread. The simplest I could do now is to limit the number of messages inside a thread and if it exceeds that delete it and create a new one kinda like langhcains ConversationalBufferWindowMemory.

@hubel-labs 9 ай бұрын

My understanding is that the threads do not expire. If the thread gets so long that it exceeds context length, there is some “intelligent handling” by openAI - e.g sliding window that only reads the most recent that fits into window. This might be helpful community.openai.com/t/questions-about-assistant-threads/485239/2

@itsjustmeemman 9 ай бұрын

@@hubel-labs Thank you for this resource and your video, really appreciate it !

@itsjustmeemman 9 ай бұрын

@@hubel-labs I've managed to steer these llms through prompt engineering and created a prompt below to prevent them from answering unrelated questions. " Insert the scope of the assistant HERE (usually I insert a numbered list of what can and it cannot do) " "Please refrain from engaging in personal conversations or responding to open-ended questions beyond the specified scope. If faced with a query unrelated to__________________ , respond with “I do not have that information” and gently remind the user of your dedicated purpose in ______________________." Placement of these "guardrail" prompts matter as well so I repeat it a few times mainly in the middle and in the last part of the instruction. It prevents the assistant from answering basic questions like what is 1+1 ? , who is elon musk , why is the sky blue , etc. and stick to its "assigned" instruction. It works well in production ( currently building AI assistants for e-commerce brands ) Hope this helps.

@robr9313 9 ай бұрын

thanks for your instructions. I followed you on linked in but it won't let me message you about the code

@hubel-labs 9 ай бұрын

Let me see if I can find you

@cescfabregas13 9 ай бұрын

Hello Lily, I followed you on LinkedIn and am unable to message you as well.

@hubel-labs 9 ай бұрын

Instead of follow, click “connect” in the drop down menu - then I will see u and can message u

@RedCloudServices 9 ай бұрын

How do you think this compares to AutoGen from Microsoft?

@jaysonp9426 9 ай бұрын

It's easier and more customizable. The downside is that with autogen you can use local models. Hopefully some open source stuff comes out soon using the new OpenAI format. I wouldn't trust their RAG though.

@hubel-labs 9 ай бұрын

I haven’t played around autogen enough yet. Will take a look when I have time!

@AI-Wire 9 ай бұрын

Could you please share and drop a link to your Colab notebook?

@hubel-labs 9 ай бұрын

Sure - connect with me on LinkedIn - search for “lily cheng Hubel” and I’ll drop u some links there!

@Hygionature 9 ай бұрын

I found and followed you on LinkedIn, but I couldn't find a way to request the code. Could you please send it to me?

@hubel-labs 9 ай бұрын

Done!

@Hygionature 9 ай бұрын

I am impressed! You did it. Thank you very, very much. I am eager to be able to run that; it is full of tricks to set up this assistant, and frankly, I admit that I need a tutor.