LangChain 101: Working With The New ChatGPT API

[@YugKhatri-ht8kd]

How's LangChain's ChatGPT API or model whatever you called is different from using simple Chat API from openAI? like if i use LC's model/API than it will use less tokens or what?

[@ilovenaturesound5123]

What’s the benefit of using LangChain’s openAi wrapper over the OpenAI itself? It looks like the OpenAI can achieve everything the wrapper does anyway. Please correct me if I’m wrong.

[@henkhbit5748]

Great intro to the new Chatgpt api 👍 Langchanin is really versatile and quick in responding new developments in LLM. Love to see new videos about LangChain and the new chatgpt api.

[@DataIndependent]

Nice, glad to hear it.

[@junaidmughal3806]

nice subtle lo-fi music in the background.

[@DataIndependent]

I had a bunch of users say it didn't sound good when they sped up the video so I removed it from current videos.

[@nattapongthanngam7216]

Great tutorial!

[@adriancaor]

Does it work well with vectorstore? Like pinecone?

[@DataIndependent]

Pinecone is just used as a tool to get relevant documents and text. So you can grab that text independent from this process and use it how you'd like. I'm sure there will be more first class support in the future.

[@adriancaor]

@@DataIndependent Makes sense, I have been using "VectorDBQAWithSourcesChain" so far but without OpenAIChat; I guess, as you say, it will be in the langchain roadmap to include first-class support to this kind of chains for OpenAIChat. Thank you for your reply.

[@chriszuidema]

Quote of the day @ 6:07 : "The moon is not out, yep, that's interesting!". Not sure why, but I started laughing like crazy haha! Great vids about LangChain btw, I learned a ton going trough them. Much appreciated!!

[@DataIndependent]

Ha glad they’re helpful! One day I’ll do more editing.

[@shaunpx1]

love your video series', id like to know of you think it is possible to set template where a user can answer a series of limited questions. where we can step the persesion of user messages based on a countdown loop. say if the user says" I have 5 people joining me sessions' got can the use a template to get the name and email of each user separately, user 1 user 1 email, then loop back to get user 2 and user2 email... and so on and know to stop at 5 since that was the initial prompt?

[@DataIndependent]

Hm, that is a good example and very sound. I haven't seen a best practice to do that yet. You could of course do it hardcoded, but then it would be flexible to other situations

[@catyung1094]

Love the video, but also curious to know why the : messages = [ SystemMessage(content="Say the opposite of what the user says"), HumanMessage(content="I love programming."), AIMessage(content='I hate programming.'), HumanMessage(content="What is the first thing that I said?") ] chat(messages) The output is not following the SystemMessage , it should output the opposite, instead of AIMessage(content='You said, "I love programming."', additional_kwargs={})? Does it means we should specify the SystemMessage everytime before the HumanMessage ?

[@DataIndependent]

It's still a bit of an unknown territory from what I can tell. Everyone is trying to figure out the best way to use this new functionality. I believe that if I would have worded the System Message even *stronger* then it might have obeyed. But as it's written right now it's a loose command. I bet there is some prompt engineering to make it more strict on responses.

[@henkhbit5748]

I think the answer is correct if u look at the chat history list. The chat history: 1 I love programming, 2. I hate programming, 3.What is the first thing that I said...ergo first sentence is I love programming.

[@juhu3709]

I was wondering the same thing

[@sysadmin9396]

What would be the difference of using “batch messages” vs “templates” ?

[@ddd12343]

In the documentation there is section about "Memory" and there are examples there of how to build knowledge graphs (ConversationKGMemory), but all examples seem to be using this old "text in text out" approach. Can I somehow use this class also with ChatGPT API in this conversation based manner?

[@DataIndependent]

Thanks for the comment and yes you can do it with memory. I'll add that to the video list

[@albus-hc]

Can we combain this API with a custom dataset yet? (i.e. upload my documents, make API Call questions about that data)

[@DataIndependent]

Oh ya - big time. You could do that through prompt templates or a variety of other ways. Check out langchain documentation for more

[@THCV4]

Are you intentionally spelling “incorporate” as “incorportate” for some reason (like an AI optimization or something)? I’ve noticed a pattern of oddly misspelled words with a lot of AI KZbin videos lately and I’m curious… P.S. Thanks for putting together this video and sharing the information!

[@DataIndependent]

Nice! Whoops, no. That's an interesting side effect of LLMs. Just updated the code on the github

[@caiyu538]

great, great, great

[@berkkilicoglu6587]

I have one related question on Embeddings. So every time a user asks the bot a question, the bot searches the knowledgebase and answers. the following use case impossible: USER: How can I register for starter training ? BOT: If you meet all the conditions, you can register from the link … USER: What does it cost ? BOT: I dont know. how can i deal with this problem.Since the user specified the course name in the previous question, but not in this question, there will be no match from the knowledgebase. How can I search while keeping the context from the previous question?

[@DataIndependent]

That is a great question. There was a bunch of talk about it on the Q&A webinar today. Check out the langchain twitter and see if you can snag the recording. The answer was on there.

[@user-wr4yl7tx3w]

What is the purpose of streaming?

[@DataIndependent]

Other than UX I'm not sure yet. It certainly makes the app feel more special and alive.

[@bartzhang4798]

Reduces the appearance of latency by returning a response token by token

[@AITwo_co]

Can you create a video on explanation of visual ChatGPT?

[@DataIndependent]

You mean using the image recognition functionality? I don't have access to that yet, I don't think anyone does outside of researched and a closed beta

[@DhananjayMudgule-f4g]

can I use : agent with this NEW chat GPT API with serpapi as tool ??? How?? llm = OpenAI(temperature=0,openai_api_key=open_ai_api_key) tools = load_tools(["serpapi"],llm=llm,serpapi_api_key=serp_api_key) memory = ConversationBufferMemory(memory_key="chat_history") agent_chain = initialize_agent(tools, chat, agent="conversational-react-description", verbose=True, memory=memory) with this NEW chat GPT API ...????? If YES --- How..???

[@DataIndependent]

You totally could, check out the langchain documentation for agents in the new chat api

[@DhananjayMudgule-f4g]

@@DataIndependent "role": "system" not able to define system role as per the documentation

[@khairulhaaziq2332]

Hi greg! thank you for these videos, learnt aton! can you possible remake the langchain+pinecoin tutorial adapted with the new chatgpt api? how would you do it differently? maybe we dont need pinecone?

[@DataIndependent]

Nice! And you would definitely still need pinecone (or a different vectorstore) because of prompt limits. Ex: You can't feed a whole book into the new chatgpt api. I'll add this to the list of videos!

[@khairulhaaziq2332]

@@DataIndependent I see.. so I guess there isnt any way to leverage the chatgpt api and better to just use the previous method?

[@khairulhaaziq2332]

@@DataIndependent may I have a quick question? In the langchain function the params are the docs and the query, does that mean we run through the docs with openai api each time? because that would be very costly. or no? how does it work?

[@juhu3709]

@@DataIndependent I would be interested in that as well

[@aaroldaaroldson708]

what’s wrong with just substituting the previous opeai class with the new chatopenai class?

[@RealEstate3D]

I cant imagine “Bacon and shrimp Sauerkraut“ as German food though.

[@larrysankey6976]

My comments keep disappearing.

[@DataIndependent]

Really? Sorry to hear that, I don't know why. What's up?

[@larrysankey6976]

​@@DataIndependent I appreciate your videos. They really helped me understand langchain. I got through the ChatGPT API examples and am trying to figure out how to do agents and memory in the same configuration. There was something like that for the GPT-3 models. Can't figure it out for ChatGPT. Are there examples somewhere, or do you plan to cover it? Thanks!

[@DataIndependent]

@@larrysankey6976 It's so new that I haven't done the videos for it yet. I'll add it to the list

[@larrysankey6976]

@@DataIndependent Thanks!