can someone explain that this is basically instead of using similarity search, here we are instead using llm to decide which database is more appropriate to query from? am i correct here
Thanks so much, youre talented. I really interested in the way you share your useful knowledge, information. I love it, the best wishes to you. Im looking forward to your video everyday, once again, thanks a lot.
This will help you in case when you have multiple documents or subsections and you are storing them into different vector databases. In that case, based on user's query, agent will understand which vector database model needs to access.
4:23 it doesn't have to decline to answer. It probably is able to answer such questions by virtue of it already being a sufficiently big enough LLM. It could say "Though this isn't part of our internal database (creating a clear signal that it goes beyond its domain) the world series was won by..."
Yes but it's best to decline for most customer facing use cases, right? For example I wouldn't want my Amazon bot to answer questions about Superbowl, it's best if it just declines to answer. Also reduces legal exposure?
thank u!! One question: Do you know how to evaluate an agentic rag? Do I have to take anything into account compared to a normal rag evaluation? (e.g. with RAGAs..) Best regards
I asked Perplexity how best to decide if seperate vector DBs is preferred over a large vector DB with metadata to contextualize the pool of vector data. I guess it's tantemount to asking if we should normalize vector tables similar to OLTP or use a data warhouse approach. It recommended the metadata approach to help the LLM decide which portion of the vector data to contextualize, which makes sense to me. Maybe it's the same thing described in this video in a different way.
Semantic routing uses more straightforward methods like cosine similarity and other predefined rules to make the decision on which route to take. The example I drew used an LLM agent, which can understand and interpret more complex/nuanced queries, understand context, but is much heavier (compute and latency) because it is using an LLM. -David
IBM’s main task now is to create the catchup teaching videos. No innovation or breakthroughs are coming out from them. I see so many IBM old timers are watching with amazement how smaller open source innovators are moving lightning fast in GenAI and then they start recording teaching videos… I do not see any new things / new demos coming out from IBM.
Got a particular question? A standard rag pipeline just takes your question convert it into a vector and searches the data source to find content that is similar to it. By adding an agent in front of it, the question can be interpreted to figure out which data sources should be used to fetch that content. It's like the difference between asking a single book a question or asking a question to a librarian who can find the right book for you before looking into the book for the answer to that question. Its hard with all the unique fancy naming they give these architectures, it abstracts the simplicity of what is really happening underneath.
Why do you say "We'll" instead of "We".. are you just starting to understand it theoretically just yet? This stuff is more than 2 years old already... And on top of that - this stuff is not only old but outdated as well...
Overkill for most consumer facing applications and, as the answer to generalist queries, not specific enough a system for internal tooling. Lots of noise introduced here.
You make the best LLM + RAG explainers in the world. Thanks so much.
Fantastic description, it seems using agents to help you route to potentially different databases is a game changer, thank you.
Thanks for this well explained video. This is the most simplified explanation of agentic rag that sank into my grey matter
The most amazing thing about all this is that he can write so easily in reverse.
Excellent explanations! Simple without additional fluff. Thank you.
Thanks David, well done presentation.
can someone explain that this is basically instead of using similarity search, here we are instead using llm to decide which database is more appropriate to query from? am i correct here
Thanks so much, youre talented. I really interested in the way you share your useful knowledge, information. I love it, the best wishes to you. Im looking forward to your video everyday, once again, thanks a lot.
Thanks. This is pretty useful and much better than the naïve approach of overloading the model with lot's of irrelevant data from the vector DB.
So the Agent is also a pretrained LLM with those 2 vector db ???
I want to know about the display you are using . I like the way this white board works and gives the illusion of being there . What tool is that ?
Excelllent concept -- many thanks.
Where are these agents installed/located? How do they get into the pipeline connected to the LLM?
This will help you in case when you have multiple documents or subsections and you are storing them into different vector databases.
In that case, based on user's query, agent will understand which vector database model needs to access.
Makes sense I’ll be implanting this into our agent workflow
Simple, clear explanation! Thanks!!
What is the logic or rule to help llm decide where to go ?? Is it only key value pairs or looking for keywords , context based understanding....
This this would be good for training an agentic swarm for a trading system based on on depth guides curated from pdf?
4:23 it doesn't have to decline to answer. It probably is able to answer such questions by virtue of it already being a sufficiently big enough LLM. It could say "Though this isn't part of our internal database (creating a clear signal that it goes beyond its domain) the world series was won by..."
Yes but it's best to decline for most customer facing use cases, right? For example I wouldn't want my Amazon bot to answer questions about Superbowl, it's best if it just declines to answer. Also reduces legal exposure?
Please do RAG + Big Data next
An interesting suggestion, any specific element or question you have about the topic?
Easy to Understand explanation. 👍🏼
thank u!! One question: Do you know how to evaluate an agentic rag? Do I have to take anything into account compared to a normal rag evaluation? (e.g. with RAGAs..) Best regards
We need to know how to implement agent and what is this? is this another llm to determine the context and route to right db
Really good explanation.
I asked Perplexity how best to decide if seperate vector DBs is preferred over a large vector DB with metadata to contextualize the pool of vector data. I guess it's tantemount to asking if we should normalize vector tables similar to OLTP or use a data warhouse approach. It recommended the metadata approach to help the LLM decide which portion of the vector data to contextualize, which makes sense to me. Maybe it's the same thing described in this video in a different way.
Que bien está explicado este video, gracias.
There should be huge response delay due to additional step?
Where can I learn this approach?
Im confused ... Andrew Ng says just RAG is Agentic AI. But RAG does not need LLM during the query phase. So what's correct here?
What is the difference between this and semantic routing?
Semantic routing uses more straightforward methods like cosine similarity and other predefined rules to make the decision on which route to take. The example I drew used an LLM agent, which can understand and interpret more complex/nuanced queries, understand context, but is much heavier (compute and latency) because it is using an LLM.
-David
Since the agent make use of LLM in decision making, it would be more clear to show the connection in the drawing.
What framework support this? Ag2?
How about a citation that shows it works?
In the reflection I think that’s the teleprompter… or maybe I am hallucinating 🤣
retro encabulator vibes 😊
every day new concept keep popping up, too hard to get the concept and follow before the concept gets into another form :)
“More responsible, more accurate, more adaptable, “
plus more secure as well?
I'd love a recommendation for an LLM that actually admits it does't have an answer instead of just making something up.
IBM’s main task now is to create the catchup teaching videos. No innovation or breakthroughs are coming out from them. I see so many IBM old timers are watching with amazement how smaller open source innovators are moving lightning fast in GenAI and then they start recording teaching videos… I do not see any new things / new demos coming out from IBM.
Yeah. On the other hand, your company is booming… 😆
So the agent will acts like a controller here .
yo
This pipeline doesnt make a lot of sense.
skill issue
Got a particular question?
A standard rag pipeline just takes your question convert it into a vector and searches the data source to find content that is similar to it.
By adding an agent in front of it, the question can be interpreted to figure out which data sources should be used to fetch that content.
It's like the difference between asking a single book a question or asking a question to a librarian who can find the right book for you before looking into the book for the answer to that question.
Its hard with all the unique fancy naming they give these architectures, it abstracts the simplicity of what is really happening underneath.
@@scycervery cool example. Thanks
@@scycerreally great explanation. Thanks!
@@scycertotally going to swipe this librarian analogy and may even give you credit on occasion 😊
👏👏🥰🥰🕊🕊🙏🙏👍👍🪻
Why do you say "We'll" instead of "We".. are you just starting to understand it theoretically just yet? This stuff is more than 2 years old already... And on top of that - this stuff is not only old but outdated as well...
Overkill for most consumer facing applications and, as the answer to generalist queries, not specific enough a system for internal tooling. Lots of noise introduced here.