I was learning RAG with Lama index and thinking about what kind of project I'll be building. I have used NotebookLM now I think there is possibility i will be trying to replicate NotebookLM or Think of idea on how we can go beyond with multimodal. - Chat with PDF and Senatigation (old stuff) Experimenting with Features : - Save/pin generated content. - Repurposing Saved/pinned content along with new instructions. - Additionally, Audio or Multimodel. Thanofor sharing your views on this I really like it. Again I'm liking more of your video and approach. Let's me know if I can assist you in buulding and managing paid community along with course.
Thanks for creating this video. I totally agree that plain vanilla RAG solutions feel outdated in comparison to NotebookLM. This product is a nice experiment on how RAG-based solutions can be better. I am really looking forward to the newest release with audio capabilities because this added modality definitely improves the interaction with a RAG-based solution to a whole new level.
i think that function calling now has become qute normal , as we can use jupyter client to execute the funcitons locally and return the output as well as the results.... etc .. allowing for the model to execute code infact without calling a funciton ... it will execute it ? ... but the funciton calling is still used for external apis instead of funciton calling .... also generating code to run in python ikernal can be trained ...
Sam, have you found any online forums where people are discussing business applications for AI? It's dizzying how fast things are changing and how any kind of product development seems like it could be undermined.
for scientific work it is sadly useless to have tons of citations (1, 2, … n) after an entire article as it doesn’t show, which paragraph or sentence references to which citation. or am i missing something here?
This video is really valuable as you see it as a RAG service. How do you feel about this project based on the degree of enhancements it’s delivered so far?
I think it is great!! When I made the video I knew they were adding voice, but I didn't expect it would hit as much as it has. It is great to see that it has got traction etc.
@@samwitteveenai it’s really interesting to me as the RAG aspect creates the ‘trusted’ centricity I need, but preserves quite a bit of creative synthesis! :). It was cool to find your video on this that was so keyed on that aspect only after that’s the conclusion I drew! As that RAG aspect is really not being driven in the current product messaging.
@@jncolligan1 Yeah this really wasn't get much attention at all when I made the vid. I wanted to support the team that created it because I thought it was a really good experiment of giving people their RAG system for their own docs.
At the moment, I agree. In my case, I get accurate output only when I include data and text for analysis as part of the prompt. RAG is not reliable enough for me.
I disagree. Infinite context windows assume everyone can fine-tune models on specific data and retrain them for every business update. RAG allows quick info retrieval and easy documentation updates without constant retraining, which is more practical for many real-world applications, especially with frequently updated info or limited resources.
@@CreamyBootyJoose i don’t think you know what infinite context means. Infinite context means you can fit everything in model context. You can even have multiple shot examples. Imagine instead of having RAG running on 10 PDFs where the retrieval might ignore references or other miscellaneous information. You can have all 10 PDF files in the context length and perform inference on it. This is not only superior to RAG, it also means any miscellaneous information will also be processed by the model and if the model sees fit to notify you about said information, it will. So Infinite context doesn’t mean fine-tuning. It simply means the model sees everything all at once.
Google and Microsoft are all about the cloud . They want people to abandon physical software and pay monthly fees for their cloud software. Microsoft want to know all about you, this is why Microsoft own LinkedIn. A potential employer can ask Microsoft more private details about you, your psychological profile Base on behaviours or political inclinations. That's enough to get a long silence from potential employers.
I was learning RAG with Lama index and thinking about what kind of project I'll be building.
I have used NotebookLM now I think there is possibility i will be trying to replicate NotebookLM or Think of idea on how we can go beyond with multimodal.
- Chat with PDF and Senatigation (old stuff)
Experimenting with Features :
- Save/pin generated content.
- Repurposing Saved/pinned content along with new instructions.
- Additionally, Audio or Multimodel.
Thanofor sharing your views on this I really like it. Again I'm liking more of your video and approach. Let's me know if I can assist you in buulding and managing paid community along with course.
Thanks for creating this video. I totally agree that plain vanilla RAG solutions feel outdated in comparison to NotebookLM. This product is a nice experiment on how RAG-based solutions can be better. I am really looking forward to the newest release with audio capabilities because this added modality definitely improves the interaction with a RAG-based solution to a whole new level.
Only available in the US 🙄
Nice overview, amazing how these tools are coming together
Now it's available in Thailand (and 200 countries), I come back to learn again.
i think that function calling now has become qute normal , as we can use jupyter client to execute the funcitons locally and return the output as well as the results.... etc .. allowing for the model to execute code infact without calling a funciton ... it will execute it ? ... but the funciton calling is still used for external apis instead of funciton calling .... also generating code to run in python ikernal can be trained ...
Sam, have you found any online forums where people are discussing business applications for AI? It's dizzying how fast things are changing and how any kind of product development seems like it could be undermined.
for scientific work it is sadly useless to have tons of citations (1, 2, … n) after an entire article as it doesn’t show, which paragraph or sentence references to which citation. or am i missing something here?
The new update has the citation inline now. It points you to the exact location of the PDF or document it's referencing from within the response.
This video is really valuable as you see it as a RAG service. How do you feel about this project based on the degree of enhancements it’s delivered so far?
I think it is great!! When I made the video I knew they were adding voice, but I didn't expect it would hit as much as it has. It is great to see that it has got traction etc.
@@samwitteveenai it’s really interesting to me as the RAG aspect creates the ‘trusted’ centricity I need, but preserves quite a bit of creative synthesis! :). It was cool to find your video on this that was so keyed on that aspect only after that’s the conclusion I drew! As that RAG aspect is really not being driven in the current product messaging.
@@jncolligan1 Yeah this really wasn't get much attention at all when I made the vid. I wanted to support the team that created it because I thought it was a really good experiment of giving people their RAG system for their own docs.
hmm, i'm not sure if that citations feature is still around. and also, certainly the voice usage is non-existant.
Google is so slow to ship this stuff. I just noticed that ChatGPT built a module to use Google Docs which I have not tested yet.
And google is going to launch LearnLM
Without allowing us to crawl and scrape public online sources like online docs, blogs etc. this RAG product will fail I think.
Chatcasts
RAG is a feature destined to die. Infinite context is the future.
At the moment, I agree. In my case, I get accurate output only when I include data and text for analysis as part of the prompt. RAG is not reliable enough for me.
I disagree. Infinite context windows assume everyone can fine-tune models on specific data and retrain them for every business update. RAG allows quick info retrieval and easy documentation updates without constant retraining, which is more practical for many real-world applications, especially with frequently updated info or limited resources.
@@CreamyBootyJoose i don’t think you know what infinite context means.
Infinite context means you can fit everything in model context. You can even have multiple shot examples. Imagine instead of having RAG running on 10 PDFs where the retrieval might ignore references or other miscellaneous information. You can have all 10 PDF files in the context length and perform inference on it. This is not only superior to RAG, it also means any miscellaneous information will also be processed by the model and if the model sees fit to notify you about said information, it will.
So Infinite context doesn’t mean fine-tuning. It simply means the model sees everything all at once.
It may be far superior to RAG, but it way also require much more resources than RAG. Combine it with RAG, and you have something more efficient.
So wrong...
i don't like it's in the cloud. They need to make a super-chip, not server farm gimmick of movie Her, which is not scalable anyway.
Google and Microsoft are all about the cloud . They want people to abandon physical software and pay monthly fees for their cloud software.
Microsoft want to know all about you, this is why Microsoft own LinkedIn.
A potential employer can ask Microsoft more private details about you, your psychological profile Base on behaviours or political inclinations.
That's enough to get a long silence from potential employers.
2min in and still not gotten to what the intent is for RAG. Not quite effective vid establishing the concept before adding support info