RAGAS - Evaluate your LangChain RAG Pipelines

[@melika1725]

I am working on RAG systems for my master thesis. Thank you for this video. really thank you!

[@codingcrashcourses8533]

You´re welcome. On my channel I have more RAG videos and I also offer an Advanced RAG course on Udemy :)

[@seallyolme]

This is awesome! Great and clear video :)

[@mosheragomaa2]

So simple, helpful and clear! Very interesting. Thanks for the video

[@M10n8]

Excellent timing ;-) Thanks for video

[@MMO-g2w]

Bro is on fire this month!

[@codingcrashcourses8533]

You guys give me so many requests on topics 😀

[@MMO-g2w]

@@codingcrashcourses8533 i was, i am and i will support you till the end. Ur videos helped my sooooooooooo much.

[@maxlgemeinderat9202]

Nice one! Also a big fan of RAGAS, however there are still many bugs that come with RAGAS, especially when trying to evaluate with local llms

[@codingcrashcourses8533]

yes, it´s still far away from perfect, but good that frameworks like these are developed

[@andreypetrunin5702]

Огромное спасибо за видео!!

[@Challseus]

Another banger! :)

[@mohammed333suliman]

Great video , thank you

[@codingcrashcourses8533]

thank you for your comment :)

[@robertputneydrake]

Nice, Meister! Machste irgendwann das Thema Code RAG ggf. mit Knowledge-Graphen?

[@codingcrashcourses8533]

Currently no plans on working with knowledge graphs, since I don´t have experience with these. But maybe in the future :)

[@alexandershevchenko4167]

Thank you for the video! Yeah, It will be really intereseting to know how to perform RAGAS in CI/CD pipline. Can you record video for this one please? Will be really helpful

[@codingcrashcourses8533]

Maybe in a few weeks

[@nguyenquynghia9755]

I switched to using RecursiveCharacterTextSplitter, but my context relevance is still low. Do you know why?

[@kumarrajaakula9064]

I want to know one thing that even the ground truth is generated by an LLM, how can we determine whether it is correct for a particular query?

[@codingcrashcourses8533]

You probably want to create your own dataset for that. I also dont want the llm to define a ground truth

[@GenerativeAI-Guru]

I was waiting for this thank you so much, is it possible to add how to evaluate accuracy using F1 scoring or other methods

[@codingcrashcourses8533]

Not out of the box, F1 scores can be easily caculated with pandas (to_pandas) like this: F1 = 2*precision*recall/(precision+recall)

[@GenerativeAI-Guru]

@@codingcrashcourses8533 thanks

[@maxlgemeinderat9202]

you could also calculate the RAGAS score which is the mean across all metrics

[@fire17102]

It's there an ai pipeline to auto optimize the rag quality? Seems like the obvious next step... Great video 🙏👍

[@codingcrashcourses8533]

You probably would have to build something like that on your own, since there are so many ways how a pipeline could look like. You could also work on your prompt and so on.

[@fire17102]

@@codingcrashcourses8533 I'd always want to manually make changes I think are best, but I'd still like to see a full matrix of hyperperameters to remove alot of the guess work. Chunk size for example. More over I'd like to benchmark everything and add scoring functions. For example a score for fact checking - see Lucidate's last video. And also IndyDevDan last video battle royal of models, I suggested to combine it with something like you do with rag params and what I suggest for full pipeline benchmark with ai suggested optimization

[@doggydoggy578]

Does this need an open ai key ? i have zero in my account

[@codingcrashcourses8533]

@@doggydoggy578 yes

[@doggydoggy578]

@@codingcrashcourses8533 did you mention this in this video ? i think it's super important

[@doggydoggy578]

@@codingcrashcourses8533 very nice, not at all an important detail to conveniently leave out