Supercharge Your RAG with Contextualized Late Interactions

Рет қаралды 11,498

Күн бұрын

ColBERT is a fast and accurate retrieval model, enabling scalable BERT-based search over large text collections in tens of milliseconds. This can be used as a potential alternative to Dense Embeddings in Retrieval Augmented Generation. In this video we explore using ColBERTv2 with RAGatouille and compare it with OpenAI Embedding models.
🦾 Discord: / discord
☕ Buy me a Coffee: ko-fi.com/promptengineering
|🔴 Patreon: / promptengineering
💼Consulting: calendly.com/engineerprompt/c...
📧 Business Contact: engineerprompt@gmail.com
Become Member: tinyurl.com/y5h28s6h
💻 Pre-configured localGPT VM: bit.ly/localGPT (use Code: PromptEngineering for 50% off).
Signup for Advanced RAG:
tally.so/r/3y9bb0
LINKS:
Google Notebook: github.com/PromtEngineer/Yout...
ColBERTv2 Paper: arxiv.org/pdf/2112.01488.pdf
ColBERT Github: github.com/stanford-futuredat...
RAGatouille: github.com/bclavie/RAGatouill...
TIMESTAMPS:
[00:00] Problem with Dense Embeddings in RAG
[01:52] Colbert v2 for Efficient Retrieval
[04:55] RAGatouille to the rescue
[05:32] Semantic Search in Action: A Practical Example with ColBERTv2
[09:33] Comparing Retrieval Performance: Colbert vs. Dense Embedding Models
[12:54] Enhancing Retrieval with Increased Chunk Size
All Interesting Videos:
Everything LangChain: • LangChain
Everything LLM: • Large Language Models
Everything Midjourney: • MidJourney Tutorials
AI Image Generation: • AI Image Generation Tu...

Пікірлер: 48

@engineerprompt 4 ай бұрын

If you are interested in leanring more about Advanced RAG Course, signup here: tally.so/r/3y9bb0

@qwertyntarantino1937 4 ай бұрын

that's definitely a hot topic

@JosephCardwell 4 ай бұрын

by 51 seconds we have the most direct explanation of embedding on youtube.

@hl236 4 ай бұрын

Thanks for this. There is a lot of obsession over LLMs but I RAG has huge room for innovation that will multiply the performance of ai applications.

@engineerprompt 4 ай бұрын

I agree, I am personally really interested in RAG and see that as the main application that will assist people in their workflows before we see anything else

@TeamDman 4 ай бұрын

Thank you for the great walkthroughs and insights! RAGatouille interface looks great, can't wait to mess around with it

@engineerprompt 4 ай бұрын

thanks, have fun :)

@maxlgemeinderat9202 4 ай бұрын

nice! Yes another video which uses this in langchain would be cool!

@hl236 4 ай бұрын

Yes please!

@mowlanicabilla5002 3 ай бұрын

Thanks for the clear and concise explanation.! What metrics can be used to evaluate the output of these models.?

@LoveWorldamineK 4 ай бұрын

yes please make the next video with RAG and integrate it and also please can you create for us a video tutorial demonstrating how to build a chatbot that inputs in XLS or CSV format, prompts the user for input, and provides charts as output. using OPENAI API

@utkarshtripathi9118 4 ай бұрын

Hii have you figured out solutions for this ??

@LoveWorldamineK 4 ай бұрын

@@utkarshtripathi9118 Still m working on it

@yusufersayyem7242 4 ай бұрын

Go Ahead Sir..... ❤

@engineerprompt 4 ай бұрын

thank you :)

@sanoussabarry4218 3 ай бұрын

Gread job !!

@PoGGiE06 2 ай бұрын

Super interesting. I want to use dspy with ragatouille/colbert2 for embedding and retrieval. I’d like to use llama index with a different vectordb, e.g. chromadb, pinecone, or qdrant. I want to use ollama with llama 3 to then summarise my retrieved rag data, and combine with some basic analysis of my own dataset. How feasible is that now? I assume that i can use dspy to finetrain on my specific analysis cases if necessary.

@henkhbit5748 4 ай бұрын

Thanks, would like to see a combination of colbert and langchain optimal chunking method.

@nirmalthacker8566 4 ай бұрын

me too please

@youngchrisyang Ай бұрын

Great content, thanks! Also curious what tool did you use to come up with such beautiful graphs on the "blackboard"

@engineerprompt Ай бұрын

I use excalidraw.com

@nicholasdudfield8610 4 ай бұрын

Nice!

@VenkatesanVenkat-fd4hg 4 ай бұрын

Can you discuss newly pdf handling with tables & docx files parser....

@VenkatesanVenkat-fd4hg 4 ай бұрын

Can you discuss on tables in Pdf files for RAG & other .docx files loader as pdf parser but some os there......

@shubhamvijayvargiya4119 3 ай бұрын

Please make a video on how to handle dynamic tabular data in pdf to feed in llm and query on tables data, as tables structure gets messed up when creating vectors.

@sohelshaikhh 4 ай бұрын

Nicely explained! also, wanted to know about time comparision between embedding retrievers and colBERT

@engineerprompt 4 ай бұрын

From my experience, colBERT is usually faster.

@ShreyasVaishnav 2 ай бұрын

How can we use this with Chroma ?

@mohsenghafari7652 4 ай бұрын

hi. please help me. how to create custom model from many pdfs in Persian language? tank you.

@JMai-ci9nl 3 ай бұрын

Thanks for the video and sharing, I can't seem to pass the loader.load_data("Orca_paper.pdf") line in the colab notebook. The load_data call complains about 'str' has no 'name' attribute.

@JMai-ci9nl 3 ай бұрын

fixed, you need documents = loader.load_data(pathlib.Path("Orca_paper.pdf")), the load_data expects a Path object, not str.

@JMai-ci9nl 3 ай бұрын

BTW, the load_data() method by default parses the pdf page by page into multiple documents, in case you are wondering like I do.

@Abdoana 4 ай бұрын

So We can try this with local gpt?

@THE-AI_INSIDER 4 ай бұрын

Please make a video on Rag with a UI where input is a file pdf or csv + Colbert behind the scenes

@engineerprompt 4 ай бұрын

will do!

@borisrusev9474 4 ай бұрын

So what's the disadvantage of using CoBERTv2? Or are you saying it's strictly better?

@engineerprompt 4 ай бұрын

At the moment, the number of vectors store supports are limited, I think only FAISS supports that. You will need a GPU to run this. In THEORY, it should perform better than dense retrieval but probably need better evals.

@dheerajsai236 Ай бұрын

Whenever I am doing Rag.search ,I am getting the name of the document in contents rather than answers for the query . how do I solve it ? Please kindly help

@jaysonp9426 4 ай бұрын

Wait for the second example you used GPT4 for embeddings instead of ada? Did I miss something?

@engineerprompt 4 ай бұрын

Its the tokenizer not the LLM. Probably can replace that with tiktoken package to get tokens.

@shameekm2146 4 ай бұрын

Thank you so much for this... :). I deal with large number of documents. I find dense retrieval is very bad at it. Let me check this approach and comment back.

@engineerprompt 4 ай бұрын

Please do share your experience. Would love to see what you find.

@utkarshtripathi9118 4 ай бұрын

Please bring next video fast

@AdamTwardoch 3 ай бұрын

@engineerprompt Is there a reason why you design your videos so that they must be viewed on a large screen? The font used on the diagram slides is obviously completely unreadable on a phone.

@aghast666 4 ай бұрын

As I dive into the world of storytelling and creative expression, VideoGPT emerged as my trusted ally, subtly enhancing the quality of my videos without stealing the spotlight.