Wow! I finally understood everything. I am a student in ML. I have watched already half of your videos. Thank you so much for sharing. Greetings from Jerusalem
and learning becomes more interesting and fun when you have an Teacher like Umar who explains each and everything related to the topic so good that everyone feels like they know complete algorithms. A big fan of your teaching methods Umar.. Thanks for making all the informative videos..
Thanks Umar. I look forward for your videos as you explain the topic in an easy to understand way. I would request you to make "BERT implementation from scratch" video.
Thank you very much for a detailed explanation on RAG with Vector Database. I have one question: Can you please explain how do we design the skip list with embeddings? Basically how to design which embedding is going to which level?
This was a wonderful explanation! I understood everything and I didn't have to watch the Transformers or BERT video (I actually know nothing about them but I have dabbled with Vector DBs). I have subbed and I will definitely watch the transformer and BERT video. Thank you!❤❤ Made a little donation too. This is my first ever saying $Thanks$ on UA-cam haha
Hey, big thanks for this awesome and super informative video! I'm really intrigued by the Siamese architecture and its connection to RAG. Could someone explain that a bit more? Am I right in saying it's used for top-K retrievals ? Meaning, we create the database with the output embeddings, and then use a trained Siamese architecture to find the top-K most relevant chunks computing similarities ? Is it necessary to use this approach in every framework, or can sometimes just computing similarity through the embeddings work effectively?
The siamese network he talked about just provides details of the sentence-bert that is used for encoding. The connection to RAG is that the sentence-bert model is used to do the encoding for both the query and the rest of the document chunks fed into the DB. In the case, Umar is providing some additional information regarding how the sentence-bert model was developed and why it is better than the natural BERT. I think it's important to understand the distinction. The top-K retrievals is done by the vector search. Using the HNSW, example, the query is compared with a random entry and then you proceed to the neighbors of each of the vectors until you get to a local minimum. You save this point. You do this a few times (> k) and retrieve the top-K ones sorted by their similarities. So the embeddings from S-BERT are used but not directly. The retrieval of the top-K embeddings is done at the vector DB search level. And doing this multiple times (via a different entry into the HNSW graph) you will get different results. And then you retrieve the top-K from there. I hope this is clear.
Amazing presentation! I have a couple of questions though... What size of chunks should be used when using Ada-002? Is that dependent on the Embedding model? Or is it to optimize the granularity of 'queriable' embedded vectors? And another thing: am I correct to assume that, in order to capture the most contexts possible, I should embed a 'tree structure' object (like a complex object in C#, with multiple nested object properties of other types) sectioned from more granular all the way up to the full object (as in, first the children, then the parents, then the grand-parents)?
Awesome I completely understand the RAG just because of you, Now I am here with some questions let's I am using the Llama2 model to where my main concern is I am giving him the pdf for context then user can ask question question on this, but this approach took time, during inferencing. so after watching your video what i undersatnd using the RAG pipeline is it possible to store the uploaded pdf into vector db then we will used it like that. I am thinking right or not or is it possible or not? Thanks,
Are we storing the sentence embeddings together with the original sentence they were created? If not how do we map them back (from the top-k most similar stored vectors) into the text they were originated for, given that the sentence embedding lost some information when pooling was done.
Yes, the vector database stores the embedding and the original text. Sometimes, they do not store the original text but a reference to it (for example instead of storing the text of a tweet, you may store the ID of the tweet) and then retrieve the original content using the reference.
Lets say I want to create a Online semantic search tool , that uses vector DB, and RAG performance. just like bing tool . will it follow the same procedure and what new things I will be adding it to integrate to Internet? Plus nicely put video Umar . can you do a coding session for this one like you do for all others , like make something with real time output with rag ? or anything . will be a pleasure to watch.
Umar, great content! Around 25:00, when you say that we have a target cosine similarity. How is that target's cosine similarity calculated? Because there is no mathematical way to calculate the cosine similarity between two sentences. All we can do is only take a subjective guess. Can you please exlain in detail to me how this works?
When you train the model, you have a dataset that maps two sentences to a score (chosen by a human being based on a scale from 1 to 10 for example). This score can be used as a score for the cosine similarity. If you look papers in this field, you'll see there are many sofisticated methods, but the training data is always labeled by a human being.
@@umarjamilai Understood! Thanks very much for the prompt response. It would be great if we can identify a bias free way to do this as the numbering between 1 - 10, especially when done by multiple people and at scale, could get biased.
@@rkbshiva The numbering is not done by random people. Usually, some specialists, e.g., language specialists are employed to get this dataset, and this reduces the noise in the label (but you'd still get some bias but should be small). Google does this for the search quality. They have a standard search quality evaluation document that is provided to the evaluators and they use the document as a guide and how to score the different documents returned for a give query.
Great video, keep up the good work! :) Around 19:25 you're saying that the embedding for "capital" is updated during backprop. Isn't that wrong for the shown example / training run where "capital" is masked? I always thought only the embedding associated with non-masked tokens can be updated.
You're right! First of all, ALL embedding vectors of the 14 tokens are updated (including the embedding associated with the MASK token). What happens actually is that the model updates the embedding of all the surrounding words in such a way that it can rebuild the missing word next time. Plus, the model is forced to use (mostly) the embedding of the context words to predict the masked token, since any word may be masked, so there's not so much useful information in the embedding of the MASK token itself. It's easy to get confused when you make long videos like mine 😬😬 Thanks for pointing out!
why is the context window size limited? Is it because these models are based on transformers and for a given transformer architecture, long distance semantic relationship detection will be bounded by the number of words/context length ?
I think he is mostly explaining how the skip-list data structure works. In general, with HNSW, you are not looking for a particular value (those values are cosine similarity scores) but rather you are traversing the graph to find neighbors with smaller similar scores until you get to a local minima, then that is the node that is returned. You then repeat it again from another entry point.
Salam Mr Jamil, i was wondering if it was possible to use the BERT model provided by apple in coreml for sentimental analysis when talking to siri then having a small gpt2 model fine tuned in conversational intelligence give a response that siri then reads out
![BERT explained: Training, Inference, BERT vs GPT/LLamA, Fine tuning, [CLS] token](http://i.ytimg.com/vi/90mGPxR2GgY/mqdefault.jpg)
![BERT explained: Training, Inference, BERT vs GPT/LLamA, Fine tuning, [CLS] token](/img/tr.png)
This is what a teacher with a deep knowledge on what is teaching can do. Thank you very much.
Wow! I finally understood everything. I am a student in ML. I have watched already half of your videos. Thank you so much for sharing. Greetings from Jerusalem
and learning becomes more interesting and fun when you have an Teacher like Umar who explains each and everything related to the topic so good that everyone feels like they know complete algorithms.
A big fan of your teaching methods Umar.. Thanks for making all the informative videos..
Man, your content is awesome. Please do not stop making these videos as well as code walkthroughs.
Among all videos about hnsw, the best. Others dont understand. They just pretend. This one, on the other hand, is honest and thorough.
The 🐐. I wish I had you as my teacher in real life, not just through a screen
You are the best teacher of ML that I have experienced. Thanks for sharing the knowledge.
This is one of the best explanation i ever seen in youtube.... Thank you.
Amazing teacher! 50 minutes flew by :)
Wow, thanks a lot. This Is the best explanation on RAG I found on UA-cam
The way you've explained all these concepts has blown my mind. I won't be surprised to see your number of subscribers skyrocket. Channel Subscribed !!
Best explanation I found on UA-cam, thank you!
This is the type of content we really want
The best explanation of RAG
Waited for such content for a while. You made my day. I think I got almost everything. So educational. Thank you Umar
What an exceptional explanation of HNSW algo ❤
Amazing !! I finally understood everything. Good Job, all your videos have in-depth understanding
Impressively intuitive, something most explanations are not. Great video!
Thanks Umar. I look forward for your videos as you explain the topic in an easy to understand way. I would request you to make "BERT implementation from scratch" video.
One of the best channels to learn and grow
Awesome context sir, it was the best explanation I found till now!
This was fantastic (as usual). Thanks for putting it together. It has helped my understanding no end.
Hello sir i just want to say thanks for creating very good content for us. love from India :)
total 48 minutes worthy vedio
This was fantastic and I have learned a lot from this! Thanks a lot for putting this lesson together!
The explanation of HNSW is excellent!
Amazing content and what clear explanation. Please make more videos. Keep making this channel will grow like anything.
Thank you so much for sharing. Looking for more content about NLP and LLMs
Thank you very much for a detailed explanation on RAG with Vector Database. I have one question: Can you please explain how do we design the skip list with embeddings? Basically how to design which embedding is going to which level?
Thank you so much - this is a great video. Great balance of details and explanation. I have learned a ton and have saved it down for future reference
Wow! You explained everything great! Please make more videos like this
This was a wonderful explanation! I understood everything and I didn't have to watch the Transformers or BERT video (I actually know nothing about them but I have dabbled with Vector DBs).
I have subbed and I will definitely watch the transformer and BERT video. Thank you!❤❤
Made a little donation too. This is my first ever saying $Thanks$ on UA-cam haha
Very nice explanation. Many thanks!
Just love ur videos. Soo much Details but extremly well put together
Awesome paper. Please keep posting more videos like this.
This video is really good, subscribed! You explained the topic super well. Thanks!
Cannot thank enough! Awesome content!!
Really amazing content!!, looking forward for more such content Umar :)
Glad I've subscribed to your channel. Please do these more.
This was such a great explanation. Thank you!
Please bring some more content !
Wow, I saw the Chinese knotting on your wall ~
This was super insightful, thank you very much!
Super explanation. Thank you
Great video!! Shouldn't 5 come after 3 in skip list?
Amazing explanation!
Wooo you are the best I have ever seen
wow wonderful explanation thanks
Great explanation! Thank you so much
Hola, coming back with a great content as usual
Thanks 🤓😺
amazing work very clear explanation ty!
Hey, big thanks for this awesome and super informative video!
I'm really intrigued by the Siamese architecture and its connection to RAG. Could someone explain that a bit more?
Am I right in saying it's used for top-K retrievals ? Meaning, we create the database with the output embeddings, and then use a trained Siamese architecture to find the top-K most relevant chunks computing similarities ?
Is it necessary to use this approach in every framework, or can sometimes just computing similarity through the embeddings work effectively?
The siamese network he talked about just provides details of the sentence-bert that is used for encoding. The connection to RAG is that the sentence-bert model is used to do the encoding for both the query and the rest of the document chunks fed into the DB. In the case, Umar is providing some additional information regarding how the sentence-bert model was developed and why it is better than the natural BERT. I think it's important to understand the distinction.
The top-K retrievals is done by the vector search. Using the HNSW, example, the query is compared with a random entry and then you proceed to the neighbors of each of the vectors until you get to a local minimum. You save this point. You do this a few times (> k) and retrieve the top-K ones sorted by their similarities. So the embeddings from S-BERT are used but not directly. The retrieval of the top-K embeddings is done at the vector DB search level. And doing this multiple times (via a different entry into the HNSW graph) you will get different results. And then you retrieve the top-K from there. I hope this is clear.
Cool video about RAG! You could also upload into Bilibili, as you live in China, you should know that. :D
Good explanation, thanks
simply impressive
Amazing presentation! I have a couple of questions though... What size of chunks should be used when using Ada-002? Is that dependent on the Embedding model? Or is it to optimize the granularity of 'queriable' embedded vectors? And another thing: am I correct to assume that, in order to capture the most contexts possible, I should embed a 'tree structure' object (like a complex object in C#, with multiple nested object properties of other types) sectioned from more granular all the way up to the full object (as in, first the children, then the parents, then the grand-parents)?
Thanks for making these videos🎉
Thank you for the excellent content!
Nice lecture, Thank you!
Thanks for sharing, really a great content 👏
Thank you so much. Such a nice explanation. 😀
Awesome I completely understand the RAG just because of you, Now I am here with some questions let's I am using the Llama2 model to where my main concern is I am giving him the pdf for context then user can ask question question on this,
but this approach took time, during inferencing. so after watching your video what i undersatnd using the RAG pipeline is it possible to store the uploaded pdf into vector db then we will used it like that.
I am thinking right or not or is it possible or not?
Thanks,
Man, keep it up! Love your content
This is really awesome session. Off course it is a lengthy but nice. Seems problem with git. Unable to access python and pdf files.
Thanks for making this video!
Very informative, thanks
Hi Umar,does RAG also has context window limitation as prompt engineering technique
When the algorithm runs to store the k best scores, it uses a pop operation on the list to remove the nodes that have already been visited?
Can you tell how that hierarchial structure supporting HNSW is created?
One of the best videos
Excellent content!
Are we storing the sentence embeddings together with the original sentence they were created? If not how do we map them back (from the top-k most similar stored vectors) into the text they were originated for, given that the sentence embedding lost some information when pooling was done.
Yes, the vector database stores the embedding and the original text. Sometimes, they do not store the original text but a reference to it (for example instead of storing the text of a tweet, you may store the ID of the tweet) and then retrieve the original content using the reference.
Lets say I want to create a Online semantic search tool , that uses vector DB, and RAG performance. just like bing tool .
will it follow the same procedure and what new things I will be adding it to integrate to Internet?
Plus nicely put video Umar .
can you do a coding session for this one like you do for all others , like make something with real time output with rag ? or anything . will be a pleasure to watch.
I am so glad I am subscribed to you!
awesome as usual! ty
so helpful! thx for sharing
Thank you, awesome video!
awesome content
Best video ever!
You are legend
at 44:00 , the order of linked list is incorrect... isn't it? because it should be 1 3 5 9
Even I have the same doubt. It should have been sorted as per the definition
Great content , keep doing it .
Umar, great content! Around 25:00, when you say that we have a target cosine similarity. How is that target's cosine similarity calculated? Because there is no mathematical way to calculate the cosine similarity between two sentences. All we can do is only take a subjective guess. Can you please exlain in detail to me how this works?
When you train the model, you have a dataset that maps two sentences to a score (chosen by a human being based on a scale from 1 to 10 for example). This score can be used as a score for the cosine similarity. If you look papers in this field, you'll see there are many sofisticated methods, but the training data is always labeled by a human being.
@@umarjamilai Understood! Thanks very much for the prompt response.
It would be great if we can identify a bias free way to do this as the numbering between 1 - 10, especially when done by multiple people and at scale, could get biased.
@@rkbshiva The numbering is not done by random people. Usually, some specialists, e.g., language specialists are employed to get this dataset, and this reduces the noise in the label (but you'd still get some bias but should be small). Google does this for the search quality. They have a standard search quality evaluation document that is provided to the evaluators and they use the document as a guide and how to score the different documents returned for a give query.
Great video, keep up the good work! :) Around 19:25 you're saying that the embedding for "capital" is updated during backprop. Isn't that wrong for the shown example / training run where "capital" is masked? I always thought only the embedding associated with non-masked tokens can be updated.
You're right!
First of all, ALL embedding vectors of the 14 tokens are updated (including the embedding associated with the MASK token).
What happens actually is that the model updates the embedding of all the surrounding words in such a way that it can rebuild the missing word next time. Plus, the model is forced to use (mostly) the embedding of the context words to predict the masked token, since any word may be masked, so there's not so much useful information in the embedding of the MASK token itself.
It's easy to get confused when you make long videos like mine 😬😬
Thanks for pointing out!
I see, didn't know that the mask token is also updated! Thank you for the quick response. You really are a remarkable person. Keep going!
Thanks
You are the BEST!
why is the context window size limited? Is it because these models are based on transformers and for a given transformer architecture, long distance semantic relationship detection will be bounded by the number of words/context length ?
Excellent video! 👏👏👏
Thank you so much man..
How would you find number 3 at 44:01 ? The algorithm you said will go to 5 and then since 5 is greater than 3, it won't go further. Am I right?
I think he is mostly explaining how the skip-list data structure works. In general, with HNSW, you are not looking for a particular value (those values are cosine similarity scores) but rather you are traversing the graph to find neighbors with smaller similar scores until you get to a local minima, then that is the node that is returned. You then repeat it again from another entry point.
Thank YOU :)
💪👍 good introduktion
Salam Mr Jamil, i was wondering if it was possible to use the BERT model provided by apple in coreml for sentimental analysis when talking to siri then having a small gpt2 model fine tuned in conversational intelligence give a response that siri then reads out
Legend
Master peice :)
how to do get target cosine similarity at first place?
There is an annotated sentence-sentence scored by experts. This what is used to compute the loss.
Thanks!
So how llm converts vector to text ?
So how LLM converts vector to text ?
Thanks bro