Hi Trellis, this is my favorite channel now. I love the topics and approach. I am also interested in the same things. However, I am more of a hobbyist. So I am grateful to learn from you
Awesome video! In your previous video, Mastering Retrieval for LLMs, you used llama 3 8B for the RAG. How well does this technique works with llama 3, or does it need a larger foundational model?
works well with Llama 3 8B for short questions. For more complex longer answers, probably 70B or more is needed. You can test it out quickly for free at endpoints.trelis.com (as of me writing this, that's got a llama 3.1 8B model in the background).
Sorry I haven't been following your channel in a while so not sure if you've covered this already, but have you been doing much with hybrid RAG (KG + Vector)? I'd be interested to see this methodology tandem with KG.
My favorite Ai channel
Hi Trellis, this is my favorite channel now. I love the topics and approach. I am also interested in the same things. However, I am more of a hobbyist. So I am grateful to learn from you
Cheers. Appreciate the comment
With a bit more time we could generate data to fine-tune for this use case. Thanks for the hard work!
Awesome video! In your previous video, Mastering Retrieval for LLMs, you used llama 3 8B for the RAG. How well does this technique works with llama 3, or does it need a larger foundational model?
works well with Llama 3 8B for short questions. For more complex longer answers, probably 70B or more is needed.
You can test it out quickly for free at endpoints.trelis.com (as of me writing this, that's got a llama 3.1 8B model in the background).
Sorry I haven't been following your channel in a while so not sure if you've covered this already, but have you been doing much with hybrid RAG (KG + Vector)? I'd be interested to see this methodology tandem with KG.
No bother. Yeah I havne't done Knowledge Graphs yet but there's a vid on cosine + BM25 called "Clean up that Dirty RAG" that may be of help.