I completed this complete playlist, Feeling much confident. Obviously needs to practice it a lot. Thanks for your time to build such a remarkable playlist
Thank you for your feedback, for a small channel like this it really makes a difference! :) I enjoyed making this playlist and learning myself, glad it was useful to you!
Thank you so much. You have done an awesome presentation, using sketch notes to convey the concepts and diving into Colab Notebooks to explain the nitty-griddy details. I love it :-)
Thank you for taking the time to leave a comment, I really enjoy making those sketches when I am learning something, I like to have an entire topic summarized in a single infographic! Now I'm sharing these, I'm glad people are finding it useful!
It is indeed one of the most interesting applications in practice, as metadata search is quite effective and we spent years extracting metadata from data, so we might as well use those!
Could you please also make an atleast overview (Detailed if possible) video on Mistral? And also on how midjourney work? some other interesting topics can be GAN, same character in different scenario etc.
Thanks for the suggestions. I'm planning a part 2 for LLM Chronicles, which will include a section on architectural improvements for transformers (MoE, Flash attention), Agents (ReAct), multi-modal LLMs and one on other generative AI methods (variational auto encoders, diffusion models, GANs). It'll take some time, but I'm looking forward to it :)
Hi. Firstly, I would like to thank you for this video, it's very informative. Secondly, I would like to ask you something about the ParentDocument technique, which I already tried to implement. Sometimes I get 0 documents retrieved by the ParentDocument when I give it a query, why is that? is there a filter that limits the retrieved documents by their similarity score? How is the similarity score calculated?
Hi! Thank you so much for your kind words. 😊 Some of the answers to your questions can be found in Part 1 of the RAG video, which is in the LLM Chronicles series. In a nutshell, the similarity score is calculated as the distance between the embedding vector generated from your query and the stored embedding vectors in the vector database. This means the embedding model you're using plays a big role. Also, make sure you're using the right measure to calculate distance: some embedding models work better with cosine similarity, while others with Euclidean distance. If you’re getting 0 documents, it might be due to a threshold filter on the similarity score.
@@donatocapitella Thank you so much for your quick response. Your answer confirmed my suspicions, but I still don't know the reason why is that happening. I didn't establish any similarity score filter nor a similarity score type of calculation in the RAG pipeline, I've just replicate your code with another embedding model (bge-large-en) and the Chroma vectorstore, so I think there must be a internal parameter in the Parent Document retriever that sets that similarity score. But I don't find it anywhere in the documentation. Do you have any information about it?
I completed this complete playlist, Feeling much confident. Obviously needs to practice it a lot.
Thanks for your time to build such a remarkable playlist
Thank you for your feedback, for a small channel like this it really makes a difference! :) I enjoyed making this playlist and learning myself, glad it was useful to you!
you are doing incredible job! you deserve more recognition. keep up the amazing work - Pakistan
Thank you for your kind words! 🙏
Thank you so much. You have done an awesome presentation, using sketch notes to convey the concepts and diving into Colab Notebooks to explain the nitty-griddy details. I love it :-)
Thank you for taking the time to leave a comment, I really enjoy making those sketches when I am learning something, I like to have an entire topic summarized in a single infographic! Now I'm sharing these, I'm glad people are finding it useful!
Thank you for the video:) Metadata + semantic search looks intersting.
It is indeed one of the most interesting applications in practice, as metadata search is quite effective and we spent years extracting metadata from data, so we might as well use those!
Thank U From Pakistan. Love Your Work.
Thank you for watching and supporting, it gives me motivation to continue!
Could you please also make an atleast overview (Detailed if possible) video on Mistral? And also on how midjourney work?
some other interesting topics can be GAN, same character in different scenario etc.
Thanks for the suggestions. I'm planning a part 2 for LLM Chronicles, which will include a section on architectural improvements for transformers (MoE, Flash attention), Agents (ReAct), multi-modal LLMs and one on other generative AI methods (variational auto encoders, diffusion models, GANs). It'll take some time, but I'm looking forward to it :)
Hi. Firstly, I would like to thank you for this video, it's very informative. Secondly, I would like to ask you something about the ParentDocument technique, which I already tried to implement. Sometimes I get 0 documents retrieved by the ParentDocument when I give it a query, why is that? is there a filter that limits the retrieved documents by their similarity score? How is the similarity score calculated?
Hi! Thank you so much for your kind words. 😊 Some of the answers to your questions can be found in Part 1 of the RAG video, which is in the LLM Chronicles series.
In a nutshell, the similarity score is calculated as the distance between the embedding vector generated from your query and the stored embedding vectors in the vector database. This means the embedding model you're using plays a big role.
Also, make sure you're using the right measure to calculate distance: some embedding models work better with cosine similarity, while others with Euclidean distance. If you’re getting 0 documents, it might be due to a threshold filter on the similarity score.
@@donatocapitella Thank you so much for your quick response. Your answer confirmed my suspicions, but I still don't know the reason why is that happening. I didn't establish any similarity score filter nor a similarity score type of calculation in the RAG pipeline, I've just replicate your code with another embedding model (bge-large-en) and the Chroma vectorstore, so I think there must be a internal parameter in the Parent Document retriever that sets that similarity score. But I don't find it anywhere in the documentation. Do you have any information about it?
What to do to improve retrieval on tables