Latest / Agentic Conversations (formally mlops.community) / RAG Has Been Oversimplified // Yujian Tang // #206

Transcript

RAG Has Been Oversimplified // Yujian Tang // #206 · Agentic Conversations (formally mlops.community)

• 0:00
  My name is Yujin. I'm a developer outlet Zillis
• 0:05
  and yes, so I take my coffee, two shots of espresso over ice
• 0:09
  with sparkling water and a pump of caramel, and then some milk
• 0:14
  to mix. Welcome back to the Community
• 0:19
  Podcast. I'm your host Dimitrios, and
• 0:22
  today we're talking with Eugen. What a conversation.
• 0:25
  We went real deep when it comes to rags.
• 0:28
  I know it's everyone's favourite topic.
• 0:30
  Or at least it was in 2023. Let's see how 2024 shapes out.
• 0:34
  Maybe you are bored of them or tired of them and now, but I
• 0:38
  promise you we got into some caveats when it comes to rags.
• 0:43
  There's a million and one courses or pieces of content
• 0:49
  that have come out around rags and so how can we differentiate
• 0:53
  ourselves? Well, we talked about some
• 0:55
  optimization methods that you might want to consider when you
• 0:58
  are using Rags. We went from what the hell is a
• 1:02
  rag to how do I optimise it? And what do I need to be
• 1:06
  thinking about at the deepest level?
• 1:08
  How am I building my app so that I can take full advantage of the
• 1:14
  pros and cons of each design pattern and these dense vectors
• 1:18
  and sparse vectors and all that fun stuff that all database
• 1:25
  vector database people like talking about?
• 1:26
  I guess so. If you are doing anything with
• 1:32
  rags, I would love to hear from you.
• 1:33
  To be honest, I would love to hear what you're working on.
• 1:36
  And if you decided against Rags, I would also love to hear from
• 1:41
  you because the beauty of this is that rags, as we talked about
• 1:45
  with Eugene, they're not for everything.
• 1:49
  Although you may go on Twitter and hear that you can just throw
• 1:54
  some AI at it and have it be a rag or use a rag with it, you
• 1:58
  don't need to. And the last part, I think
• 2:02
  Eugene has some cool insight around multi modal rags and what
• 2:06
  we can do when that starts taking off and when support
• 2:11
  happens. For that more O If you are on
• 2:17
  the market for anything vector database, check out Zillas Cloud
• 2:22
  and check out Zillas Pipelines, their new feature that they just
• 2:28
  have 4:00 early access for the community and listeners of this
• 2:32
  podcast. You can find it by clicking in
• 2:35
  the show notes. And huge thanks to those for
• 2:37
  sponsoring this episode. Of course, if you liked it,
• 2:43
  share it with just one friend. That's all I ask.
• 2:46
  All right. I'll see you dude.
• 2:55
  Well, I'm stoked to have you on here.
• 2:57
  I want to do like a huge deep dive into Rags and basically
• 3:02
  where we're at right now. It's 2024.
• 3:04
  Rags have become the hero of the LM space.
• 3:08
  I think it's more or less. The QA bots are like the Hello
• 3:12
  Worlds that you create with LM's and I know you've been spending
• 3:16
  a lot of time working on Rags and and their implementations
• 3:21
  and all that and stuff. So we should probably start with
• 3:25
  we had you on last time and we'll link to the last podcast
• 3:29
  that you came on in case anyone wants to basically hear.
• 3:32
  I'm not going to say it's part one, because we're going in a
• 3:34
  totally different direction where it's not like you need to
• 3:36
  listen to that one first before you get to this one.
• 3:38
  But we did have a pretty insightful conversation around
• 3:43
  vector databases and the idea of how much AI has been accessible
• 3:51
  by the masses now because of things like vector databases and
• 3:55
  Ellms. And so whereas before it took
• 3:59
  quite a lot of effort and a lot of engineering time to get some
• 4:04
  kind of ML model into production and that was really the thesis
• 4:09
  for the ML OPS community from like 2020 to 2022.
• 4:15
  Now it feels like were in a place where we can get something
• 4:21
  at least something quick and dirty U and running real quick
• 4:25
  and validate assumptions, validate different ideas that we
• 4:29
  might have O Before we go any further, the vector database LLM
• 4:37
  embedding design pattern is probably the one that everyone
• 4:41
  thinks about when it comes to Rags.
• 4:44
  Am I correct? And correct me if I if I've
• 4:47
  gotten anything mistaken there. Yeah that sounds about right for
• 4:52
  most of like what you what people like ask about when it
• 4:56
  comes to like the rack sack is like you know what our
• 4:59
  embeddings. How do we get embeddings into
• 5:01
  vector databases? Like what do I are embedding
• 5:03
  models in Elms, the same thing and they are not the same thing.
• 5:08
  And this is like something that quite often gets asked and
• 5:13
  people like, you know, once they look at these different
• 5:16
  applications like using like, if you're gonna use, like chain, if
• 5:20
  you're gonna use on the index and you're gonna do rag, you're
• 5:23
  going to have an embeddings model and you're gonna have a
• 5:26
  LLM. And people will ask like, oh,
• 5:28
  like do they have to be the same?
• 5:30
  And the answer is no, they don't have to be the same.
• 5:31
  It's like, Oh well, how do I pick one?
• 5:33
  And then, you know, it gets a lot more complicated than just
• 5:37
  like the very basic like here is use the default.
• 5:39
  Here's Open AI. Boom.
• 5:42
  So yeah, and embedding models it.
• 5:45
  It's probably worth just a quick summary too, for the two people
• 5:49
  that are listening that have been under a rock and do not
• 5:51
  know how embeddings work and vector stores work with
• 5:55
  embeddings and what you need to do there.
• 5:59
  Uh, yes. OK.
• 6:01
  So yeah, I should probably should have more clarified what
• 6:06
  these pieces of the puzzle were. So embeddings models are the
• 6:10
  models that you use to turn your data that you're going to be
• 6:14
  working with. This is your proprietary
• 6:15
  internal data or whatever. And you turn that into vector
• 6:18
  embeddings so that you can well, so that you can work with it.
• 6:23
  Basically because embeddings are numerical representations of
• 6:26
  this input data that is like a text or image or something.
• 6:32
  It's like a machine learned using a machine learning model
• 6:36
  representation of this kind of data and embeddings embedding
• 6:41
  models are trained well. Now we've moved into we've moved
• 6:46
  far long enough into the ragged staff that people who are
• 6:49
  working on the more like cutting edge stuff.
• 6:52
  I think this was probably this probably came to light a few
• 6:55
  months ago. Like the whole like you need to
• 6:57
  be training your embeddings model on specific types of data
• 7:00
  that you want to work with. Like you can start off by going
• 7:03
  to hugging face and getting like a very basic embeddings model
• 7:08
  and it will work decently fine. But when it comes to like
• 7:10
  enterprise production and it when it comes to working with
• 7:12
  very specific data, you're gonna need to make sure that you're
• 7:15
  training your beddings model on the right data.
• 7:18
  And then what these embeddings models do, or what we do with
• 7:21
  the embeddings models. What you need to do with
• 7:22
  embedding models in order to get the embeddings from them is you
• 7:25
  actually after you train the model, you need to ensure that
• 7:29
  you're only taking the output from like the second to last
• 7:31
  layer. So traditional models,
• 7:32
  traditional natural image processing models will, they
• 7:37
  will or even even image rattles. They'll do some sort of like
• 7:42
  image models do some sort of segmentation.
• 7:43
  Natural image processing models might do some sort of
• 7:47
  classification or some sort of prediction, something like that.
• 7:51
  But with the vector embeddings, that's what we're looking for
• 7:53
  actually. We just want to know what the
• 7:57
  semantics the model has pulled out of the input is.
• 8:01
  And so that's why instead of taking the output of the model,
• 8:04
  where we actually do is we take the output from the 2nd to last
• 8:06
  layer and then just to touch on the LMS's and the factor
• 8:12
  database in case there are other people who have not heard about
• 8:15
  these yet. LM's are the large language
• 8:18
  models, right GPT, everybody kind of has heard of that one.
• 8:21
  There are many others, right? Like mixture came out.
• 8:24
  Perplexity is I think perplexity is coming out with another one.
• 8:27
  I don't know. There's like a huge number of
• 8:29
  Owens now, Falcon Llama, basically everything you can
• 8:33
  think of is an LM. I also like actually I have like
• 8:37
  a personal grief on this is at this point like I don't even
• 8:39
  know like what makes something a large language model anymore.
• 8:42
  How large does it have to be, you know, 7 to 1320, a trillion
• 8:47
  parameters. I'm talking about billions with
• 8:49
  these like first few numbers. And then like I heard someone
• 8:51
  saying like, oh, you know, actually I think anything under
• 8:54
  7 billion is a small language model.
• 8:56
  And I was like, I don't know, it seems pretty big to me.
• 9:00
  Yeah, like, just like, like. Literally three, yeah.
• 9:04
  Yeah. Like literally like last year
• 9:05
  we're like ohh large language model with 5 billion parameters
• 9:09
  and now it's kind of like, yeah, if you don't have more than 10
• 9:11
  billion parameters is a really large number if I don't know and
• 9:14
  then. You guessing, you don't kick
• 9:16
  them, you don't get the stamp of approval of being large.
• 9:20
  You're not large enough. And yeah, it's funny because a
• 9:23
  lot of people are coming through and saying, yeah, this year it's
• 9:26
  all about the small language models.
• 9:28
  And so now it's like who's deeming what is large and what
• 9:34
  is small? Where have we come to a
• 9:37
  consensus on that yet? I don't.
• 9:38
  I don't think so. So I know I don't.
• 9:41
  Know so you were and these people are like, yeah we're
• 9:43
  talking about small language models and they're gonna be
• 9:45
  talking about like 3 billion parameter models like, I don't
• 9:48
  know man. And then the last part is the
• 9:50
  vector database piece, which is where you store your vector
• 9:53
  embeddings. Once you get your vector
• 9:54
  embeddings and you basically use this as a semantic store, as a
• 9:59
  way to store the meaning behind the image and the text and the
• 10:02
  audios and the videos that you have, and a way to work with
• 10:06
  that semantic meaning in a mathematical manner.
• 10:09
  Yes and no. And the plane?
• 10:11
  What does it look like when you're actually building the
• 10:16
  application? The end user will whatever type
• 10:20
  into their Q and A. Bot what does the user actually
• 10:22
  have to do? Well, I I think I'm more just
• 10:26
  thinking the end user will type into the QA bot right?
• 10:30
  Like I'm looking for more examples of this or whenever it
• 10:35
  says that and or tell me tell me about the if I'm if I'm reading
• 10:42
  a book where I have a bunch of PDF documents, tell me about
• 10:46
  this paragraph more or where it talks about this more right then
• 10:53
  behind the scenes what happens. Ohh, yeah yeah so there's.
• 10:59
  OK. So actually the way that a user
• 11:01
  is expected to interact with the app also plays into like the way
• 11:04
  that you think about designing your app.
• 11:05
  But we can we'll touch more on that in a bit.
• 11:09
  So what actually happens behind the scene is.
• 11:11
  So there's a few ways that you can structure your architecture
• 11:13
  and it depends on how you would like to build it.
• 11:16
  But what I would suggest, this is my suggested architecture is
• 11:20
  what happens behind the scenes is the query comes in, it goes
• 11:22
  to the LLM for the LLM to determine what the actual
• 11:27
  question is. So perhaps you have, like, you
• 11:29
  know, here's the you want to give us some context.
• 11:31
  You want to say like I'm pretend that I'm in, I don't know, like
• 11:35
  an astrophysicist. And I'm reading this book about
• 11:37
  the cosmos and I want to know about black holes or whatever.
• 11:41
  So then, you know, it gives us some context that pulls out like
• 11:43
  what it actually is. And so when you say that the LM
• 11:46
  has to know that, OK, well, actually what I want to go look
• 11:49
  for in the database, let's say you have a database of books is
• 11:52
  I want to go look for and I want to philtre on the book Cosmos
• 11:55
  and I want to look for black holes.
• 11:57
  And So what it will do is then it will say, OK, it will form a
• 12:02
  way to query the vector database.
• 12:04
  And so vector database queries are typically formed in like,
• 12:07
  you know, like a call, right, like either some sort of like
• 12:11
  API call, some sort of HTTP://GERC, some sort of call
• 12:15
  like that. So it's passing data along, it
• 12:17
  passes it, It makes some sort of, you know, call to the
• 12:20
  venture database and it says, OK, let's philtre on this book
• 12:25
  and Cosmos, we're only going to look for text about black holes.
• 12:29
  OK, so then maybe you get, I don't know, 10 responses back.
• 12:33
  Right. So you get 10 responses back and
• 12:35
  then you send those responses back to the LM and and you know,
• 12:39
  you give the, remember the first query, it's like, OK, we want to
• 12:42
  know about black holes from this book.
• 12:45
  And then it says, OK, here's my response, Here's what it is in
• 12:48
  the back of database. Here's the book name.
• 12:50
  Here's the, I don't know the section name, chapter name, the
• 12:53
  author name and like all this information.
• 12:56
  And then what I'll do is I'll take that information, we'll
• 12:59
  punch it down and say OK, here you go, black holes are, I don't
• 13:02
  know supermassive clusters of gravity or something.
• 13:06
  I'm not really sure what they are.
• 13:07
  But yeah, then it will return you the The answer to you
• 13:09
  basically was. So why?
• 13:10
  What goes on in the background is basically the Allen first
• 13:13
  takes your input and that interprets it, turns that into a
• 13:18
  query, queries the vector database, takes that context,
• 13:22
  and then turns that context into something as human readable and
• 13:25
  gives you back something that makes sense.
• 13:28
  I was really hoping you were going to give me some crazy
• 13:33
  definition of a black hole right there.
• 13:35
  I miss, but we can hopefully edit it.
• 13:39
  Attack in in post production. I don't have any any crazy
• 13:42
  things, all my goals. Let me think if there's anything
• 13:44
  I know a lot about that would be like fun to know about.
• 13:48
  The fun fact I do appreciate this is just like a little bit
• 13:51
  of a side note and a tangent. Since we are going on a little
• 13:56
  tangent here, I I ask speakers of the podcast to fill out a
• 14:02
  speaker form when they come on to this podcast.
• 14:06
  You filled that out and one of the questions on there is a fun
• 14:10
  fact about you. Do you remember what you put for
• 14:14
  your fun fact? You said I can build a fire.
• 14:17
  Is it like Boy Scouts fire where you're just kind of in the woods
• 14:21
  and you have Flint and that stuff, or is it with a whole
• 14:24
  bottle of kerosene? Yes.
• 14:26
  So I don't know why, but that was like one of the first things
• 14:29
  that came to mind for that question.
• 14:31
  And I am really proud of being able to build a fire because of
• 14:34
  like one specific instance, which was like I was down in
• 14:38
  Coronado Beach with some friends and we brought like Korean
• 14:43
  barbecue and we've brought like something to cook it on.
• 14:45
  And we brought like like a bunch of things.
• 14:49
  But ohh and we and we have like firewood and all this stuff.
• 14:52
  And then we're having a lot of trouble starting the actual
• 14:55
  fire. Right.
• 14:55
  You gotta imagine, like you have three kids who spend most of
• 15:01
  their time in front of computers, like, you know,
• 15:03
  things like, you know, very like nerdy, like, I mean I'm, I come
• 15:06
  from a software background, right.
• 15:07
  So you've got these three kids who spend basically like 20
• 15:10
  hours a day in front of a computer, and they're just
• 15:12
  sitting at a beast trying to figure out how to, like, start
• 15:14
  fire. And so we actually get the fire
• 15:16
  started. So first we figure out like, OK,
• 15:20
  we're just going going to go around and find like dry leaves
• 15:22
  and like stuff like that and it finally works.
• 15:25
  We Oh my It was probably like my first
• 15:50
  time ever actually building a fire and I was like just like
• 15:52
  one random thing that I came to my when I was filling out the
• 15:55
  form. I was like, OK, here we.
• 15:56
  Go so the mental picture that I have in my head of the three of
• 16:00
  you trying to build this fire, and you're used to writing unit
• 16:05
  tests and all that fun stuff, and there's no unit tests for
• 16:08
  these fires. It's.
• 16:09
  Like there's no. We need, we need more kindling.
• 16:12
  Do we need more gasoline? Let's just go get gasoline and
• 16:16
  dump it on this and see if that'll work.
• 16:18
  So this as as you were telling this story, basically I was
• 16:25
  thinking about the parallels in what I just did to get more
• 16:29
  context out of you on this fire. It was it was like I was I was
• 16:35
  raging right there. Yeah, I was.
• 16:38
  I was your I was a human rag. Cause you get you said I can
• 16:43
  build fire. But then I said give me more
• 16:45
  context, I need more context. And I went to the vector
• 16:50
  database, which is you. And I've got more context.
• 16:54
  And now I can plug it into the LLM and I can say, all right,
• 16:58
  here's all the context you need about Eugene's fire
• 17:01
  capabilities. So we went on a huge tangent
• 17:04
  there, man, I want to get back on track.
• 17:07
  And for those that are listening, I promise we're going
• 17:09
  to get deep into the rags and rag stack.
• 17:14
  I like that word that used. It's good.
• 17:16
  One huge question that is on my mind when it comes to rags.
• 17:22
  I'm sure you've heard a lot of people talk about this is the
• 17:26
  rags Eliminate the ability for LM's Toulouse Senate because you
• 17:33
  give it that context. No, it doesn't.
• 17:36
  So this is something that we talked about a lot in like many
• 17:40
  of our earlier presentations and I think a lot of people talk
• 17:42
  about this when it comes to tarak stuff is like rats
• 17:46
  hallucinate. How do you get rid of that?
• 17:48
  While you should use some sort of contacts to inject your data,
• 17:52
  and injecting your data does it reduces like a certain like type
• 17:57
  of hallucination? Like, you know it's not gonna
• 18:00
  just like make things up about your data anymore.
• 18:03
  But that doesn't mean that it's not possible for the rag to like
• 18:07
  still like put out the wrong answer.
• 18:09
  Because in the end, like the way the LM's work is still using
• 18:15
  statistics, and it's still just like, here's the context.
• 18:19
  I'm gonna predict based off the context.
• 18:22
  But if the training data was weird and the contact
• 18:24
  contradicts the training data, you might get some weird, some
• 18:26
  weird results. And so an important thing to do
• 18:29
  to kind of deal with that is basically to have your
• 18:33
  observability tools and have like you know, your your
• 18:36
  evaluation tools afterwards. And I think that's part of like
• 18:39
  the whole reason that there's so many tools that have recently
• 18:45
  kind of like moved from what was originally like ML OPS into now
• 18:48
  what they're calling LM OPS. Yeah, and do you feel like that
• 18:52
  is a critical part of the rag stack?
• 18:54
  Oh for sure. Like if you're going to be doing
• 18:56
  this in production like you don't want like, you gotta know
• 19:00
  like that your your answers are like decently, like, right,
• 19:05
  right, like, not like even humans.
• 19:07
  Like, you know, no, no human is right 100% of the time.
• 19:11
  But it's you know like you gotta make sure that your rags are at
• 19:16
  least matching up to what people are capable of.
• 19:17
  And so like examples of like you know some of the things that
• 19:21
  people talk about in this space like make sure that the, the the
• 19:24
  context that you get is relevant, right.
• 19:26
  Just because some of these semantically similar doesn't
• 19:29
  mean that it may necessarily be semantically relevant and make
• 19:33
  sure that you cite your sources right.
• 19:35
  So this is something that I was building earlier last year using
• 19:39
  like Llama index was like how do you build a rag app with
• 19:44
  citations and how do you ensure that, like you know the actual
• 19:50
  answer given doesn't deviate too far from the the data that's
• 19:55
  pulled back from the vector database?
• 19:57
  And how did you do that? Building citations is actually
• 20:02
  very, very easy. What it kind of requires you to
• 20:06
  do is to realise that you're going to build an application
• 20:08
  that's gonna need citations, and so you have to store the text
• 20:12
  that you're going to use along with your embeddings.
• 20:16
  So you're going to store more data, but you're gonna in
• 20:19
  return, you're going to get that groundedness in that context and
• 20:22
  that relevance and that ground truth data back.
• 20:27
  And then actually, so this was something that wasn't originally
• 20:32
  available on Lama Index that I actually ended up writing.
• 20:35
  I ended up writing a PR to make this available.
• 20:38
  Nice. Yeah, it's almost that trade off
• 20:40
  of, hey, we're going to store more data, but we're going to
• 20:43
  have more trust. So yeah, it's worth it 100% of
• 20:47
  the time. Yeah.
• 20:48
  And you said something else that's interesting in there that
• 20:50
  I wanted to just dig into for a second, the context and
• 20:55
  relevance, like what's the difference between those two
• 20:58
  there? Right.
• 21:03
  I think it's probably more it's probably, it probably makes more
• 21:06
  sense to differentiate between like relevance and similarity.
• 21:09
  I guess these are both like attributes of contacts.
• 21:11
  I was I would say, but context is like you know like the the
• 21:17
  the additional data or the data that the original form of the
• 21:21
  data or like the data that you know like the surrounding data
• 21:24
  for something. So for example, like an apple a
• 21:27
  day keeps the doctor away, like you need the contacts that the
• 21:30
  apple you're talking about is a fruit and not like, you know,
• 21:33
  the Mac, right? Laptop.
• 21:36
  So context is kind of like that. And then similarity and
• 21:39
  relevance is like let's say I wanna find something for an
• 21:45
  apple. A day keeps the doctor away, and
• 21:48
  maybe a similar sentence to that might be like a banana a day
• 21:53
  keeps the chiropractor away. But you know, is that.
• 21:57
  You just made that shit up, didn't you?
• 21:59
  You just made that shit up, but is a banana a day?
• 22:03
  Keeps the chiropractor away. Relevant to my question about,
• 22:07
  you know what keeps the doctor away and not really.
• 22:10
  So that's kind of like the difference there is like, you
• 22:12
  know, like what would be relevant might be like, you
• 22:15
  know, an apple a day keeps the doctor away because it makes you
• 22:19
  E fibre. I don't know, something like
• 22:21
  that. Yeah, fibre, it's a big one
• 22:23
  these days. Get that gut health right and
• 22:26
  you won't have to see the doctors, Yeah.
• 22:28
  Yeah, I just heard a podcast about it, so that's why I was
• 22:31
  thinking. About was it the Peter ratio 1?
• 22:34
  No, no. It was like one of the ones with
• 22:36
  Angie Huberman. No way.
• 22:39
  All right, I gotta go watch that.
• 22:40
  Now that's on my list. I just watched the ERA Tio 1.
• 22:43
  So, OK, so now I want to just make sure that I'm getting this
• 22:48
  right. Just a little pause because the
• 22:51
  rag stack, we introduced something new into it that we
• 22:54
  didn't have in the beginning When you said, oh, these are the
• 22:57
  things that you need for the rag stack, which is the embedding
• 22:59
  model, the LM and the vector database.
• 23:04
  But then we just said, Oh yeah, and by the way, you might want
• 23:07
  to have some kind of and evaluation guardrails tool on
• 23:11
  there. And funny enough, I'm sure
• 23:14
  you've been seeing some of this stuff OU on social media where I
• 23:17
  think it's Chevrolet. If you go to the Chevrolet
• 23:20
  website, they use an LLM for their customer service bot and
• 23:26
  it's just straight plugged into GPT 4.
• 23:29
  And so I've seen people talking about how, like why pay for open
• 23:35
  AI when you can just get it through a Chevrolet website?
• 23:39
  Or I've also seen people asking like, are Tesla's better than
• 23:44
  Chevys And it will come back like Teslas are known to be
• 23:48
  better than Chevys. Yes, on the Chevrolet website is
• 23:51
  just like the amount of brand damage that you're doing right
• 23:54
  now is incredible. So it's worth its weight in gold
• 24:00
  to get some kind of tool that will make sure you're rag is not
• 24:05
  spewing out things it should not be spewing out.
• 24:09
  Yes, it is definitely worth looking at doing something like
• 24:13
  that. And you definitely need
• 24:15
  something like that when you're gonna be putting this into
• 24:16
  production. I guess like since we're on the
• 24:19
  topic of tools that you can add to your rack sack, other things
• 24:25
  to think about could be like data pipelines, like for
• 24:29
  example, how do you pipe new data into your vector database.
• 24:34
  And this is actually something that we have been building a
• 24:38
  Zillis. So Zillas recently released our
• 24:40
  own pipelines thing. And so one of the things that
• 24:43
  people always ask me like I I'm pretty sure I did, I didn't
• 24:48
  really like 14 talks last year and I got this question at least
• 24:51
  like 12 or 13 of them, which was like how do I get my data into a
• 24:54
  vector database? And it's like, ohh, well you
• 24:56
  need to use the embeddings model, but you can make it
• 25:00
  easier by building a pipeline around it or something like
• 25:03
  that, right? Where so far we have for Zillis.
• 25:06
  Right now we have an open source.
• 25:09
  We have an open source model as our embeddings model on
• 25:11
  pipeline, but uh, you know, like that's kind of a just a one way
• 25:17
  to get into it, right? We want to offer people a way to
• 25:19
  get into it. And the thing about using
• 25:21
  pipelines, because when you're in production, you probably
• 25:23
  don't want to. I mean, you could, I guess
• 25:26
  people do actually do this in production.
• 25:28
  You could just every, you know, day or so, execute a script,
• 25:32
  runs for 10 minutes, put your data into your vector database.
• 25:36
  Yeah. Hey everyone, My name is Aparna,
• 25:38
  founder of Verizon. And the best way to stay up to
• 25:40
  date with Emma Lops is by subscribing to this podcast.
• 25:46
  But if you want like a, you know, more automated version of
• 25:50
  that, like if it starts taking you too long to go and, you
• 25:53
  know, do the script and then get a pipeline.
• 25:55
  Right. And Zoe, pipelines aren't the
• 25:57
  only pipelines that are available.
• 25:58
  But you know, this is like one way that you can use a pipeline
• 26:00
  to pipe your data from some sort of data source into your
• 26:04
  database in a more, let's say, streamlined, automatic, easy,
• 26:08
  whatever kind of manner. Yeah, so, so then basically, if
• 26:13
  I if I understand this correctly, it's like as we build
• 26:17
  out this RAG stack in my head, we've got the LLM, we've got the
• 26:20
  vector database, we've got the embedding model.
• 26:23
  We need a way to connect that and connect all of these along
• 26:28
  with on the output. We need a way to make sure that
• 26:32
  there are guardrails and that we validate the output before the
• 26:35
  end user sees it. Is there anything else I'm
• 26:39
  missing there? I feel like probably potentially
• 26:41
  some kind of prompt tooling? Like the, I mean like like you
• 26:47
  know lane chain, Lama index, these are orchestration
• 26:50
  frameworks like as Pont tooling prompt engineering kind of
• 26:52
  frameworks also kind of like fits the if the shoe fits, you
• 26:56
  know I guess um so that's yeah that would probably be like one
• 27:00
  more thing in there. I I have seen that a lot of
• 27:03
  people like to also do their own prompt engineering and I think
• 27:08
  that that is actually something that.
• 27:11
  So I I would I would really like to see this is a feature request
• 27:15
  for the link chain law people are still listening.
• 27:17
  I'd really like to see a way to to mess with the prompts inside
• 27:21
  of the apps which is kind of I guess it's kind of weird because
• 27:24
  it kind of defeats the purpose of the apps themselves like you
• 27:26
  know the frameworks themselves are supposed to be there to give
• 27:28
  you this this way to interact with the island.
• 27:30
  I'm not sure, but that's something that I think is like
• 27:34
  if you're working in production like, you're probably gonna want
• 27:37
  to tweak your own prompts. Yeah, yeah.
• 27:40
  You want easy access to that. You don't have to dig around,
• 27:43
  that's for sure, yeah. So.
• 27:45
  The other piece that I I'm fascinated by is I think some
• 27:50
  people have come and asked about why would you need to use RAG
• 27:59
  for some use cases. Is it not good enough just to
• 28:04
  use LIKE search? Like keyword search.
• 28:07
  Or similarity search for that matter.
• 28:10
  Like do you need the whole throwing it into an LLM and
• 28:14
  getting a bunch of output, or can you just do search and have
• 28:18
  it give you back something you know?
• 28:20
  Like is LM over engineering yet just because of the height?
• 28:25
  Yeah, I mean like, so for this, I would like invite you and
• 28:30
  other people to kind of consider like a thought experiment, which
• 28:33
  is let's say I have semantic search and I asked the question,
• 28:42
  you know, what ingredients are in an apple pie.
• 28:48
  When I do a similarity search for that, I'm not gonna find,
• 28:53
  like an answer that says you need three apples, a pound of
• 28:57
  butter, you know, a pound of flour or whatever.
• 29:02
  What I'm gonna find is things like what are the ingredients
• 29:05
  for a blueberry pie, what are the ingredients for a pecan pie?
• 29:09
  And that's what I'm going to get back.
• 29:11
  And so if you think about, if you think about it in that
• 29:15
  manner, you'll quickly recognise that you need something that's
• 29:19
  going to take your natural language, query your question
• 29:22
  and turn it into how do I find the things that's going to be
• 29:25
  similar to the answers of this question?
• 29:27
  Yes, I see. That makes 100% sense.
• 29:31
  Yeah. And it's like we want the
• 29:33
  answer, we don't want the similarity of this question.
• 29:38
  Correct. I should have known that.
• 29:40
  I can't believe I didn't know that.
• 29:41
  And I'm asking you that. Now that I think about it, I
• 29:44
  kind of feel like an idiot itself.
• 29:48
  Is a very common question that I also get.
• 29:52
  Good. So one other piece that I'm
• 29:55
  wondering is like what kind of RAG optimization do you do?
• 30:02
  And should we be thinking about like let's say we've got the RAG
• 30:06
  up and running, but now we want to make it better.
• 30:09
  What are some ways that you can turn some knobs and levers you
• 30:12
  can pull? Yeah, so I think this is part of
• 30:15
  where some of the preprocessing stuff comes in.
• 30:19
  And this is also something that kind of this comes before the
• 30:22
  embeddings model, which is thinking about how you want to
• 30:25
  chunk up your data, how do you want to split up your data.
• 30:27
  And earlier when we were talking about like the end user and how
• 30:30
  the end user is going to use the RAG app, this is I referred to
• 30:33
  this. I alluded to this like, oh,
• 30:35
  we'll cover this later. Yes, foreshadowing.
• 30:40
  When you think about the way the user is going to use an app, it
• 30:45
  also makes you think about the way you want to check up your
• 30:47
  data. So for example, if you want your
• 30:50
  user to use your app in a conversational manner, like it's
• 30:52
  gonna be a chatbot, then you probably can chunk up your data
• 30:56
  into smaller sections like. It's like OK, like we probably
• 31:02
  need to find very specific, very like small kind of things.
• 31:04
  But if you perhaps want your user to use your application as
• 31:09
  something like something that helps them complete, like blogs
• 31:16
  or writing things, kind of like, you know, kind of like what
• 31:18
  Jasper does or copy AI you know, then maybe what you want is you
• 31:22
  want to give it more data. That is an example of bigger
• 31:26
  blocks of paragraphs or maybe also like depending on not even
• 31:32
  not even getting to the part where we talk about what the
• 31:34
  user is going to do. But maybe let's think about what
• 31:36
  data you are working with to begin with and was it look like.
• 31:39
  So if you have data in the format of a conversation of like
• 31:43
  a, you know, like a text back and forth, then you're going to
• 31:46
  want smaller chunks and you're probably going to want to chunk
• 31:48
  on specific characters. Maybe you have specific
• 31:50
  characters to separate new text blobs and you're gonna want to
• 31:56
  check on that. And then you know, if you have a
• 31:59
  big document then you're gonna want bigger chunks.
• 32:02
  If you have maybe a QA or a lecture where there's going to
• 32:05
  be like a short and then long answers, you're going to want a
• 32:08
  way to chunk on specific characters.
• 32:10
  And possibly also you're going to want to be able to save
• 32:16
  coupled texts. And then one other thing to
• 32:20
  think about with that is how much context overlap do you need
• 32:23
  between between your data chunks, right.
• 32:26
  So maybe you have a document and you know these paragraphs are
• 32:30
  like 500 characters, but you need the sentence before and the
• 32:34
  sentence the sentence, the last one of the, the last sentence of
• 32:36
  the paragraph before, and the first sentence of the paragraph
• 32:38
  after. Then you're gonna need some sort
• 32:40
  of chunk overlapping, right? So there's a bunch of different
• 32:43
  things to think about when it comes to how am I gonna set up
• 32:46
  my data To even embed this is like prior to building your rag
• 32:49
  app. Like you can build an example
• 32:51
  one, and then once you build an example one you probably want to
• 32:53
  think like OK, there's definitely ways I can make this
• 32:56
  better. How do I make this better?
• 32:57
  The first step you're going to want to look at is what does my
• 32:59
  data actually look like inside of the application?
• 33:01
  And then of course the eval stop.
• 33:03
  Yeah, that is really interesting when it comes to this idea of
• 33:09
  like start from the end and then go backwards and recognise how
• 33:14
  you can optimise that way. And I also look at like other
• 33:18
  ways that we are interacting with LMS as far as like the
• 33:24
  point and click with that, we're doing it too, right?
• 33:27
  Like I think about Notion AI and that's a great interface of it.
• 33:31
  We don't have to use it like a chat bot, but we can if we want
• 33:34
  to. So if you have all of those
• 33:39
  options open, how are you then thinking about optimising in the
• 33:43
  background? You know what I mean?
• 33:44
  Like do you do both and then just sacrifice on the storage
• 33:48
  like we talked about before where yes you are going to store
• 33:52
  more but you're also going to have more trust or is it
• 33:56
  something like you're you're finding that happy medium?
• 34:00
  Yeah. So there's actually, I have
• 34:02
  something to touch on that as well.
• 34:03
  So there's a couple of ways you can even think about storing the
• 34:06
  way you store your chunks and embeddings can also affect the
• 34:09
  way that your application works. So a couple of techniques that
• 34:12
  I've talked about in some of my talks have been like, you can
• 34:17
  have, you can store larger chunks of data with smaller like
• 34:22
  with only like maybe one of those sentences embedded and you
• 34:25
  can use that as the embedding. You can store the entire chunk
• 34:27
  with it. Then you get more context every
• 34:29
  time you have, you know, specific things you're looking
• 34:31
  up. The other way is that you do the
• 34:33
  reverse. So you sorry, you embed an
• 34:36
  entire paragraph and maybe you only store a sentence and that
• 34:39
  way you can get more specific responses to a more generalised
• 34:44
  query perhaps. So there's a couple different
• 34:47
  ways to kind of address that as well, just like the way that you
• 34:49
  store your data. As for the trade off there, you
• 34:54
  know, I I actually haven't seen this in production.
• 34:56
  So I'm not, I'm not entirely sure exactly how the trade-offs
• 35:00
  work, but it sounds like, yeah, like you know, kind of once
• 35:04
  again, one of the many, many trade-offs that we make here is
• 35:08
  like compute versus storage or storage versus latency or
• 35:13
  storage versus quality or you know, things like that.
• 35:15
  And this is just another one of those things that we'll have to
• 35:18
  optimise for when we build in production.
• 35:22
  And along those lines, what are things that you feel like rags
• 35:27
  are not suited for? Because it feels like you,
• 35:32
  you've seen a lot of square pegs trying to be fit into round
• 35:37
  holes, right? Just because it rags our rags
• 35:40
  and they're the height these days.
• 35:43
  Yeah, what a rise. Not suited for.
• 35:47
  I mean really, you can use a RAD for anything you want.
• 35:50
  I don't know how well it works. Just throw all your data in the
• 35:54
  vector database and use it for whatever you need.
• 35:56
  We got you covered. Yeah, exactly.
• 35:59
  Throw all your data in the middle of this.
• 36:00
  Use it for whatever you need. We got you covered.
• 36:03
  Someone asked me the other day, they were like, oh, I wanna be
• 36:06
  able to map all my users and I want to be able to track all my
• 36:08
  users and I want to map them based on specific
• 36:11
  characteristics that I've predefined for them.
• 36:13
  And so I'm like, well you know, then you're not even looking at
• 36:16
  unstructured data, you're looking at structured data like
• 36:18
  go use a SQL database and write an automation.
• 36:22
  That's what you should do. Yeah, don't come.
• 36:24
  That's. That's not that's that's a great
• 36:26
  example of like when to not to use rag.
• 36:28
  OK, there you go. I don't get a lot of people
• 36:32
  asking me about so like a lot of people just ask about RAG
• 36:35
  because it's like familiar, but it's also like different, right?
• 36:39
  People are like, oh this is like, we've seen this before in
• 36:42
  like 2013 and like 2000 whatever.
• 36:45
  Like the chat bots that pop up and people like Ohh, it's a chat
• 36:47
  bot and it has that familiarity to it.
• 36:51
  But then when it comes to actually using them and building
• 36:53
  them and it's like how do you make these better than it's kind
• 36:55
  of interesting. It's kind of different people
• 36:56
  like that. So I would say anything that you
• 36:58
  can think of that you know relates back to chat bots from
• 37:03
  2013, then use RAG to replace that.
• 37:06
  Things like, oh, actually I have a great example of something
• 37:09
  that might not need an LM for like similarity search.
• 37:12
  So I recently talked to somebody who's building some sort of like
• 37:15
  fashion AI tool and I build like a couple of these myself for fun
• 37:20
  and one of the. Things from Wait, wait wait
• 37:22
  wait. You gotta go a little deeper
• 37:25
  into that fashion. I I didn't take you much as a
• 37:29
  fashionable guy There. Not.
• 37:30
  No offence, but I feel like you wear a uniform of a T-shirt and
• 37:38
  jeans or or sweatpants maybe every day and you're often on
• 37:44
  your free time building fashion apps.
• 37:46
  What's going on there? I need more context.
• 37:48
  I need to do the LLM thing. The rag thing again?
• 37:52
  I'm not necessarily a fashionable person, but I, I, I,
• 37:58
  I just like, it's just like a fun thing to mess around with.
• 38:00
  You know, people, everybody wears clothes and a lot of
• 38:03
  people wear clothes and they look kind of foolish.
• 38:06
  And, you know, yes, me too. Like, I get, you know, I I get
• 38:11
  it. So.
• 38:12
  But, but, you know, some people wear clothes and they look
• 38:14
  really nice. Like, I, when I was at a WSI saw
• 38:17
  this one lady very dressed up. She had like this nice pic code.
• 38:20
  Like, I don't know, I actually don't know a lot of the words
• 38:22
  for some of the other things that she was like, wow, this
• 38:24
  person looks really put together and you like, look, everyone
• 38:26
  else and everyone else is wearing like a, like a text
• 38:28
  shirt. Like I'm wearing a tight shirt
• 38:30
  and like jeans, sweatpants walking around and like, I'm
• 38:33
  like, oh, OK. I don't know.
• 38:35
  It's just like, it was just like a thing that kind of like popped
• 38:37
  into my mind, you know, like it would be cool, it'd be fine.
• 38:40
  So I'd build wine and. Now we have more context.
• 38:44
  Yeah. Give me give me what you were
• 38:46
  gonna say. What was the takeaway?
• 38:47
  Sorry, I totally derailed that. I had to know.
• 38:50
  What is the contact there? Yes, yes, yes.
• 38:53
  You you don't need an LM to convert your picture of a dress
• 38:57
  into the description that this is a picture of a dress.
• 39:00
  If you have pictures of dresses in your vector database.
• 39:03
  So that's an example of something that's not a rag app,
• 39:05
  doesn't use an LM, still uses the Metro database similarity
• 39:09
  search. OK, but you're not using Clip or
• 39:12
  anything like that to get the metadata.
• 39:16
  Around that you can the the person who built this fine-tuned
• 39:22
  her own like segmentation model. OK.
• 39:24
  Nice. I was like, wow, I I I
• 39:27
  personally wouldn't put that much effort into building a
• 39:29
  fashion app. But I'm glad that you did,
• 39:31
  because this is cool. All right, so wait, what did
• 39:34
  your fashion app do? I basically found like an open
• 39:38
  source like site, former like you know segmentation
• 39:40
  transformer thing and it's just like segments out like like like
• 39:44
  clips like your shirt and your pants and whatever and then it
• 39:47
  matches it to other pictures. I had a picture, a bunch of
• 39:50
  pictures of Taylor Swift, a bunch of pictures of like random
• 39:52
  celebrities and just like put them up there was like match
• 39:54
  your outfit to Taylor Swift or match your outfit to some random
• 39:57
  celebrity and just like ohh room in there and I was like,
• 39:59
  alright, here you go you know. Ohh, I'm gonna go see if I can
• 40:03
  match my Taylor Swift. Now give me the link for that.
• 40:07
  What is it? Is it live?
• 40:11
  All right. We're going to put this in.
• 40:12
  Ohh. No, it's for anybody.
• 40:13
  OK. It's on.
• 40:15
  It's on the back. We will put.
• 40:16
  That out here in this Fashion V it's actually kind of it's it's
• 40:23
  yeah, here we go and drop in here.
• 40:24
  Yes, I love it. This is all.
• 40:27
  It's not super simple to run. You have to run like the
• 40:30
  ingestion and then yeah there's like 2 things you have to run.
• 40:32
  I think like one is like ingest and then the other is like ohh
• 40:35
  no one is set UK radio fashion the other one's radio fashion
• 40:41
  and then just you get a nice UI drag and drop.
• 40:45
  See pictures of Taylor Swift compared to you.
• 40:48
  Ohh, I love it dude. That's cool.
• 40:50
  That is really cool. And you didn't need any LM to
• 40:53
  deal with that, and you don't need to do rags at all for.
• 40:56
  That no LM not a. Rag need to overcomplicate
• 40:59
  things you but it's not, is it? Giving you a recommendation of
• 41:02
  like what you should wear if you want to look like.
• 41:06
  It's just like, it's just like it's just like your outfit looks
• 41:09
  kind of like this one. Ohh.
• 41:11
  It's not that far along. It's not like Ohh.
• 41:14
  I see. I see.
• 41:15
  Alright. Cool.
• 41:16
  I I don't really think that like I'm not really sure that like I
• 41:19
  would be qualified to be telling anybody like what they should be
• 41:22
  wearing, you know, need someone with like some some background
• 41:25
  on that. You know, like, ohh, this
• 41:27
  person's a fashion designer. OK, that makes sense, right?
• 41:29
  There you go. And so if you were to make this
• 41:34
  into a rag app, I feel like you could do a few things where
• 41:38
  maybe you're recommending it. Is that recommendation on how to
• 41:42
  you make this outfit look better or look more like Taylor Swift?
• 41:46
  And then you send it to the LM and it will give some kind of a
• 41:51
  recommendation. Yes.
• 41:53
  So first up, this could be, this could be easily a first step
• 41:57
  into launching A multimodal rag app, right?
• 42:00
  Like you've got images and you want to do rag, so now you've
• 42:03
  got text, you've got images. This is perfect to get into
• 42:05
  multimodal. What you could do is you could
• 42:07
  have some sort of LM or clip fit or something like that and be
• 42:11
  like, describe the pictures to me and now you can search the
• 42:14
  pictures. So instead of putting a picture
• 42:15
  of yourself and saying finding like, one that's close to
• 42:18
  whatever Taylor Swift, you can be like, ohh, show me a picture
• 42:21
  of Taylor Swift performing at a concert in a green dress.
• 42:24
  And it'll pull like, it'll look at the pictures and it'll pull
• 42:27
  like, oh, OK, here you go. Or you can give it a picture of
• 42:30
  you and be like, describe what I'm wearing in this picture and
• 42:33
  then I'll describe what you're wearing.
• 42:34
  And then you can be like, find me a picture of Taylor Swift
• 42:36
  wearing something similar to it, and then it will find a picture
• 42:40
  of Taylor. So for me, something similar to
• 42:41
  whatever it is you're. And are there any considerations
• 42:45
  that we need to be thinking about when it comes to
• 42:47
  multimodal rags versus just the LLM rags?
• 42:52
  Ohh man, I think that we are just on the tip of the iceberg
• 42:56
  with multimodal rags. I think the 2024 right now we're
• 43:00
  talking about multimodal rags. We're you know no one's, there's
• 43:03
  not a lot of these have really been built.
• 43:05
  I haven't seen any in production yet.
• 43:07
  I think this is the year that a lot of people going to build
• 43:09
  them. We're going to see a lot of
• 43:10
  things that people are going to run into.
• 43:12
  I think one of the problems that we're gonna run into with
• 43:14
  multimodal rags that I have no solution for that I don't, I
• 43:17
  haven't heard any solutions for yet is just the fact that there
• 43:21
  is a likelihood or a possibility or probability.
• 43:24
  Let's say there's a probability that your LLM is also going to
• 43:28
  hallucinate on what's in the image.
• 43:30
  And so now you have two steps of hallucination.
• 43:32
  So you have a hallucination of what may be returned and a
• 43:36
  hallucination of what may be in the image.
• 43:38
  And so I think this is going to be something that is going to be
• 43:41
  kind of the next like as we as we use the LLM more and more in
• 43:45
  an application in LLM applications.
• 43:48
  I think one of the main issues is going to be this issue of
• 43:50
  reducing compounded hallucinations.
• 43:54
  Yeah, because if it hallucinates once, then it's going to create
• 43:58
  downstream effects and it's going to be.
• 44:02
  Exactly. Less reliable?
• 44:04
  Yeah, that makes a lot of sense. And then to debug that also is
• 44:08
  not as straightforward as where did it hallucinate?
• 44:10
  You have to. You have now two points of
• 44:12
  failure, not that single area where it could be eliminated.
• 44:17
  All right, all right. And any other ideas on AS that
• 44:23
  are of interest when it comes to using multimodal rags?
• 44:27
  Like, what are some other use cases that you've seen or that
• 44:30
  you are thinking about? I've heard requests for video
• 44:34
  search or catalogue search from people.
• 44:37
  I think video search, if you're gonna do true video search, oof,
• 44:40
  that's gonna be computationally expensive.
• 44:42
  That's gonna be tough. Why do you say that?
• 44:44
  Just cause frame by frame. Yeah, I mean, you know, how many
• 44:49
  frames do you have and then not. It's not just like the frames,
• 44:52
  right? Like when you have a video file,
• 44:55
  it's not just the images, you also have the audio and they
• 44:58
  have to match up. And so you can't, like you can't
• 45:01
  individually do image, audio stuff on videos.
• 45:04
  You have to do on video files, right?
• 45:06
  I mean, you can try to do it image, audio stuff separately,
• 45:10
  but you have to work it on the video files themselves and they
• 45:15
  just take up more space. They're they're they're memory
• 45:17
  heavy and they have a lot of numbers, so they're naturally
• 45:21
  compute heavy. So I think that's going to be an
• 45:24
  interesting space. It's going to be kind of tough.
• 45:27
  And just so just so we're clear on that one, I feel like the
• 45:31
  multimodal can be one. Imagine we are having a chat bot
• 45:38
  that can chat with a movie and you have different parts of this
• 45:45
  multimodal which is 1 all the frames in the scene.
• 45:49
  And then you have another which is all of the audio that comes
• 45:53
  out of that. And then you have another which
• 45:55
  is like the script that I imagine you could try and get
• 45:59
  right. So there's all those different
• 46:01
  pieces that come along with it. Yes, there's the frames, the
• 46:06
  audio, the script. You can turn the audio into
• 46:10
  images. Using MFCC you can one thing
• 46:15
  that I I really would would like to see and I think this is going
• 46:19
  to be quite tough for people to build is like video to video
• 46:23
  search. Like I want to know like, hey,
• 46:26
  like, I see this, alright, this is going to sound, this is going
• 46:30
  to sound really nerdy. OK, so I'm watching this like
• 46:32
  Naruto like music video, right? I wanna see where these clips
• 46:36
  are coming from. What episode is this from?
• 46:39
  What did I miss? Because a lot of times I'm like,
• 46:40
  I haven't watched all of this series, but like sometimes like,
• 46:43
  you know, like, Oh yes, this is cool.
• 46:45
  Like, we'll watch one of these and then I'll be like, ohh,
• 46:47
  where's that from, you know? Nice.
• 46:50
  Yeah. So it's almost like the what we
• 46:52
  were talking about before where you have citation, you have
• 46:56
  citation in the. Videos.
• 46:58
  I'll see citations for videos. Yeah, there you go.
• 47:02
  Dude, Awesome. A little bit away from that, but
• 47:04
  it'll be cool, yeah. But I mean a little bit away as
• 47:08
  in this year or this decade. Ohh, I think it'll I think I I
• 47:14
  would be surprised if this wasn't solved in this decade.
• 47:19
  I would not be surprised if they wasn't solved this year.
• 47:22
  OK. Cool or salt?
• 47:25
  Mostly Salt, let's say. Yeah, usably.
• 47:27
  Salt. Usably.
• 47:28
  Salt. Yeah, that first iteration,
• 47:31
  dude. Well, this has been fascinating.
• 47:34
  For everybody out there who wants to play around with the
• 47:37
  pipelines that you were talking about, we can just go to
• 47:41
  silas.com/pipelines or Google. Pipeline for You can just
• 47:47
  cloud.com sign up for a free account if you spin up one of
• 47:51
  the serverless options, The serverless options, the free
• 47:54
  option, you get immediate access to pipelines and it's in the UI.
• 47:58
  You can click around and build your own.
• 48:00
  Excellent. And if you do that, make sure
• 48:03
  they send out. I think you mentioned you send
• 48:06
  out a feedback form to anyone that does that.
• 48:11
  And so in that feedback form let them know that you listen to
• 48:15
  this crazy chat that you did and I just had and let them know
• 48:19
  that we sent you over that way. It would be great to to hear
• 48:25
  that. Man.
• 48:27
  I am still impressed that you can build fire and I appreciate
• 48:31
  you coming on here and talking with me.
• 48:33
  This was brilliant. Hey, Lazlo here.
• 48:39
  If you're serious about Emma Lops, you hit subscribe, right?
• 48:42
  Now.