Why your ChatBot is stupid - Focus on the important parts! - The ultimative RAG-Tierlist

Поделиться
HTML-код
  • Опубликовано: 28 дек 2024

Комментарии • 41

  • @codingcrashcourses8533
    @codingcrashcourses8533  3 месяца назад +1

    You see anything different? Change my mind in the comments ;-)

  • @itsalexpaul
    @itsalexpaul 3 месяца назад +3

    This helped me a lt sorting some concepts in my brain. You have a huge talent for explaining things simply. It's clear that you have real hands-on experience. keep it up! 🥂

  • @KristianSchou1
    @KristianSchou1 3 месяца назад +3

    As a person who is developing an internal RAG application, multiquerying/rephrasing seems useless for one who understands the system, but I've found that users, especially new GenAI users, love it as it lowers the entrybarrier for usage.
    "Remember the stand user could be anyone"

    • @codingcrashcourses8533
      @codingcrashcourses8533  3 месяца назад

      For multiquery i totally agree (that why I put it into D tier). But rephrasing is not useless. How do you handle follow up questions without rephrasing in a RAG System?

    • @KristianSchou1
      @KristianSchou1 3 месяца назад

      @@codingcrashcourses8533 it is very early days, im primarily an SCADA engineer and this is a side project at work, so it dosen't handle multiple queries at the time.
      The project scope is iterating through BID/NOBID questions for every tender we plan to bid on, and then use some LDA magic to sort chunks into project roles to minimize wasted time reading through materials

  • @davidtindell950
    @davidtindell950 3 месяца назад +1

    Thank You ! A good and very timely review !!!!

  • @bastienmoenaert6940
    @bastienmoenaert6940 3 месяца назад

    Hey man, love yor vids!! I am trying to build my own RAG application and with my background in backend/data engineering I think I'll be able to come up with a pretty good RAG system. My big problem is building I nice front end, any plans on doing a tutorial (or udemy course) on building professional front ends to go with a RAG application?

    • @codingcrashcourses8533
      @codingcrashcourses8533  3 месяца назад

      No sorry I don´t plan on that. There are tons of tutorials on building frontends from people who are much better at that than I am. I would learn that somewhere else and than you will be good enough to come up with solutions for that on your own ;-).

  • @juanes1793
    @juanes1793 3 месяца назад

    You’ve helped me a lot by explaining all these topics related to RAG so well :) . Keep it up, you're a pro, man! I have a question: How is it possible to keep a Langchain course updated knowing that almost every week new things come out in the documentation or others become outdated? You are the best :D

    • @codingcrashcourses8533
      @codingcrashcourses8533  3 месяца назад

      Thank you! It´s actually not possible, especially since I a have a full time job and a life beside my job and this channel. I currently work on an LCEL Deep dive update for my first course. My courses do not cover EVERYTHING too, I think good courses are not bloated and don´t cover every single edge case or in the LangChain World the integration of 1000 different VEctorstores or AI models.

  • @arslanabid2245
    @arslanabid2245 3 месяца назад

    Thankyou soo much, I really needed this video. It just cleared so much confussions.

  • @Investwithwisdome
    @Investwithwisdome 3 месяца назад

    I hope you're doing well! I’ve been really excited about your LangGraph course and was wondering if you could provide an estimated release date. I’m looking forward to diving into it!
    Thanks so much!

    • @codingcrashcourses8533
      @codingcrashcourses8533  3 месяца назад

      Hello! Sorry for the late response, there is currently a lot of spam on my channel. My estimation is the beginning of november, since I will have one week holiday then

  • @micbab-vg2mu
    @micbab-vg2mu 3 месяца назад

    Thank you - I will try your courses -)

  • @rorycawley
    @rorycawley 3 месяца назад

    Incredible content! Danke!

  • @pabloguatibonza5992
    @pabloguatibonza5992 3 месяца назад +1

    Great Video ! Which vector store would you recommend to use? I was using supabase, but I will use azure OpenAI, so I’m not sure if change to an azure vector store

    • @codingcrashcourses8533
      @codingcrashcourses8533  3 месяца назад +2

      You can use Azure AI Search which is fine. I use PgVector as a container there, since it makes me independent from the cloudprovider.

    • @pabloguatibonza5992
      @pabloguatibonza5992 3 месяца назад

      @@codingcrashcourses8533 thank you very much :D

  • @tanyongsheng4561
    @tanyongsheng4561 3 месяца назад

    What about technique like query decomposition?

    • @codingcrashcourses8533
      @codingcrashcourses8533  3 месяца назад +1

      Better than something like Multiquery! But I have not experimented too much with it.

  • @dumbol8126
    @dumbol8126 3 месяца назад

    what would rate 4o-minis chunking capabilities, what about distilling some of 4o's chunking to a smaller opensource model wonder if that would work and make a lot cheaper

    • @dumbol8126
      @dumbol8126 3 месяца назад +1

      since small models would require like exponential (or quadratically?) less vram and compute to make the chunks i think it would be a lot cheaper

  • @kukubanana
    @kukubanana 3 месяца назад

    So it is better to use llm to for intent classification and entities extraction instead of using tool like RASA?

  • @Sar2393
    @Sar2393 3 месяца назад +1

    You mentioned gpt4o is a great semantic chunker, please can you share the notebook for the same

    • @codingcrashcourses8533
      @codingcrashcourses8533  3 месяца назад +2

      That´s something I covered in-depth in my udemy course, where I build a custom chunker on top of the LangChain interface.

    • @Mostafa_Sharaf_4_9
      @Mostafa_Sharaf_4_9 3 месяца назад

      @@codingcrashcourses8533 please make a free video here on youtube about that topic (gpt4o as a great semantic chunker) 😁😁❤❤

  • @ayushjhabak2851
    @ayushjhabak2851 3 месяца назад

    with the new hnsw indexing, supabase pg vector seems better than pinecone and for scalability too , or am i missing something ?

    • @codingcrashcourses8533
      @codingcrashcourses8533  3 месяца назад

      Interesting question, but I honestly can´t answer it. I never ran into issues with my normal PgVector and I doubt the current similarity search approach is suited to support that many docs that hnsw indexing or not really matters. But I am not 100% sure about that.

  • @ramishelh9998
    @ramishelh9998 3 месяца назад

    Hi, I am a big fan of your youtube videos. Wanted to ask if you offer any consultation / project dev services?

    • @codingcrashcourses8533
      @codingcrashcourses8533  3 месяца назад

      Sorry, I only make youtube videos and courses, no consulting or active development.

  • @張博文-n6l
    @張博文-n6l 3 месяца назад

    S:
    rawdata, chat model
    A:
    document chunking,
    prompt enginnering
    GuardraiIs
    B:
    embedding model
    agent RAG
    rephrasing
    tool calling
    document compression
    C :
    vector store
    reranking
    D:
    base routing
    Multiquery Retrieval
    RAPTOR
    graphRAG
    finetuning

  • @dimensionalcookie
    @dimensionalcookie 3 месяца назад

    graph rag like trying to apply NP hard problem to production, should be F tier..

  • @harshithdr3952
    @harshithdr3952 3 месяца назад

    Well informative video. But I will recommend some good images, instead of shitty AI generated images which are unreadable.....