Graphify Interview Setup
0:00
If you are using Claude Code, you've definitely heard about the new plug-in called Graphify, which is essentially a
0:06
memory solution for all AI users. I already did a full breakdown on Graphify itself, so if you haven't seen that yet,
0:12
please watch that here first. But this plugin already has over half a million downloads in just a few weeks that it
0:18
was set up. And boy, do I have a treat for you guys today. You're going to have to excuse some of the audio mixups, but
0:24
I have the creator of this plugin here live with me. And together on this video, I'm going to be asking him a lot
0:30
of the same questions that you guys have in mind about this plugin. It's going to be interview style. We'll go through his
0:36
best use cases, his inspiration on why he created this plug-in, and pinpoint the truth of whether this plugin can
0:43
actually save you 70 times on token. So, without further ado, grab your cup of Joe, grab a snack, and let's get into
0:49
this. All right, guys. I got Safi here with me. He is the creator of Graphify.
0:55
And today we're going to go through some common questions, kind of reveal the intentions and inspiration behind
1:01
creating this awesome plug-in. So Sophie, what I'd love for you to give us is an elevator pitch. I got some
1:08
questions on my side, so you guys might see me looking over. Uh, but give us an elevator pitch, a TLDDR of what you
1:14
currently do and how that work really led to the creation of Graphify. Hi guys, Sophie here, the creator of
Safi Elevator Pitch
1:20
Graphify. So I'm I'm an AI engineer been working on knowledge graphs since over two years now. I've got two research
1:27
papers on graph networks and then hybrid rock systems have combined vector
1:33
databases with graph architectures and I've been working on the increasing the
1:38
process memory systems of large language models and that's where the graphify came in as a link to my research and
1:44
then my AI engineering experience. Yeah. So graphify actually started as a link
Origin Story Karpathy Tweet
1:50
by from a post of Andre Karpati. Uh so he tweeted about like a structure when
1:56
you can use raw data files and then you can extract them later on so that you can minimize the tokens usage of your uh
2:02
coding assistants and that's where the idea the whole idea clicked in from my experience of have been been working on
2:08
graphs. I shipped that feature within two days of Karpati's uh uh idea pitch
2:13
on to X and that's how it all started and here we are with 500K plus downloads
2:19
on Pipi and nearly 43 43K stars on GitHub. I
2:25
absolutely wild how that actually all went down. Um what I'm trying to
Unexpected Viral Growth
2:30
understand now right is like did you did you expect the kind of traction that you got when you guys when you initially
2:37
launched this product? Was it kind of expected or were you really surprised? Oh, I'm totally humbled by this totally.
2:43
So I I did think like it will go past 10k stars maybe because it was a need as
2:50
a tool it was in need of an hour and the the timing was to totally correct and then right after fe of kapati I did
2:56
expect like it should it will blow up a bit but not this far and and I'm totally humbled by this uh the the people the
3:04
sgestions of people and then people have been using this and putting it publicly as well how they are using my my my
3:10
software graphify they've been making posts on LinkedIn they've been making posted eggs and and pretty much uh what
3:17
do we say I'm I'm in dilemma of like it's it's something I had never experienced before and I'm pretty much
3:23
excited about this part. Yeah, that's key. You know, I I think before honor comes humility and keeping
3:29
the humility is always key. It's what makes us likable and from what I've known you I've known you that this past
3:35
month, you are a super humble guy and I think you deserve everything that you have coming for you. and looking at what
3:42
we're looking at what for sure looking at where things have come like walk us through well not this part
3:49
yet what I want to ask you is when you were looking at this was this you per like were you scratching a personal itch
3:56
when you kind of put this together um was this was this aimed at a specific
4:01
category do you did you have something in mind when creating this or was it more like you saw a problem you solved
4:07
it and you were kind of like testing your skills Tell me a little bit about how that looked. So it all started as a personal as you
Personal Itch Memory
4:13
told as I told you I'm an AI engineer. I've been working in this in this whole system since over two years now. So I've
4:18
been facing the same problem of like losing the context once I switched from one session to another session and and
4:24
and AI labs like anthropic and open are continuously working on fixing this problem of context when you switch from
4:30
one session to another session and they've been giving a compact conversation feature as well. you must have seen and and the whole reason I
4:37
came up with this idea was like let's do something into it. let's come up with a good good feature good good software
4:43
that can fix this problem for just not me but for others as well that's necessary and I came up with the open source software idea to do this and then
4:50
eventually when I saw people picking up my tool I I went full on it and then started shipping as fast as possible
4:56
started integrating more features did some R&D and then went into production after that it's
5:02
it's doing pretty well at at this point uh it's been like a month now and I've got like decent decent uh response from
5:09
the audience Yeah, you definitely have gotten a really good um amount of traction on this. So, now that I have a
5:14
little bit more of an understanding, which I get because that's usually where all things start and the the market demand was there. You know, we all knew
5:21
it. Memory is a huge problem. It's what a lot of the AI users are looking at solving right now. Nobody has quite
Top Use Cases
5:27
solved it yet. I wanted to segue into something that I think a lot of us are
5:32
really curious about because there's beginners, there's intermediate, you
5:37
know, uh, AI users, and then there's more advanced. And I want to know, you
5:43
know, what are the what are the if you had to pinpoint the three best use cases of this product, maybe for a business
5:50
owner, you an everyday college student, you know, like different categories of people, like what would you say are the
5:57
top three use cases for Graphify in your head? Yes. So, let's say Graphifi's main
6:03
criteria is it's it works pretty well with your coding assistance. Let's say you've got a new uh contractor employee
6:09
who's going to come as a senior engineer in your in your company and then he wants to go through the whole codebase of your of your of the whole
6:15
architecture of your company's codebase. So if he were to get in deep into the whole architecture it will take him like
6:21
two days three days and plenty of slack messages to the previous employer or something like that. So just to prevent
6:27
this whole scenario to come like to spark this whole scenario graphify it
6:32
what it does is it creates a map of the whole codebase using abstract syntax tree a libraries it's it's free of cost
6:39
as well there is no lms involved into it so it creates a map it creates god nodes
6:45
and then it connects with the coh different parts of a codebase together and then you can pretty much pretty much
6:51
easily understand them in one shot it saves you time saves you money. Let's say you you you are hiring someone for
6:58
like $150 per hour just to go through your code base and you can ship that you can you can do that for free in like few
7:05
minutes and it's a big ROI just by a tool. Secondly, uh let's say
7:11
you are working for an enterprise and then there's a ticket on your back end some some users facing a problem right
7:17
and then you want to fix that ticket as soon as possible. What are you going to do? uh maybe one of your engineers will sit or fix this for the whole day or two
7:24
days or come up with a solution after 3 days right so to to like bring the time down here you all you have to do is you
7:31
have got a brain here you got the map of the whole code piece you'll just look for the dependency and you will see how
7:37
the dependency is connected to all other files in the same codebase and and it's one shot it's boom you've got the whole
7:43
solution from there you can easily find the solution from the god nodes of the whole architecture thirdly uh so
7:48
graphify isn't just about the code bases It's also there to create persistent memory. So we create a map of your
7:55
documents, of your images, on of your YouTube urls, of your audios using whisper. So I'm converting all those
8:02
videos and audios into transcripts and then creating those text from those transcripts in the form of a graph and
8:07
you can pretty much retrieve everything. Let's say you had some lecture or you want to see you you want to prompt a lecture of of of Stanford let's say and
8:15
then what you're going to do is just put that link get the get the lectures uh transcription and then you can have the
8:21
map of the of the lecture with each section of the lecture divided into into
8:26
pre various graphs and you can retrieve anything much from there. It saves you
8:31
it saves you a lot of time like go through the whole video rather than like prompt the whole video in one shot. So
8:37
it saves you time and money both of these things and gives you persistent memory as well. You can keep on adding context to the same graph time after
8:44
time with less tokens rather than like making the coding agent go through the whole corpus uh or call multiple times
8:51
for a simple task then going at once and then shipping you the final response.
8:57
Right on. And what what I'm looking at is for me what's really really exciting
9:02
is the ability to create the semantic relationships meaning the hidden meaning between the different assets that is
9:09
super huge because when I have something searched in my clo my claude files it's like how does this file really connect
9:16
to this file you know claude can search and find the different file names but I
9:21
think a really good important use case especially on number three which is for everyday day-to-day users users.
9:28
That relationshipbased mapping is absolutely huge. Now, what we're looking
9:35
at is essentially a local rag system. You know, it it's a database. It's a way
9:41
to look through all that information. And now with the development of of all
9:46
these different plugins like GSD you know about you know you might have seen some of the stuff we've built internally
9:51
on our side Paul and seed you have non-engineers non-developers building
9:57
develop you know actually developing products so this is also a way for
10:03
people to use these these plugins and these orchestration layers to build
10:08
products match it with something that allows them to sift through their code bases more effectively with less tokens
10:16
and and to have those relationships tied together creating a a way more effective
10:22
and I even fun like I mean using these tools are so much fun just in exponentially increasing our
10:30
output and our our capacities capabilities [snorts] so on that note
10:35
let me pull up my questions again um so we understand I have something to say here uh that's
10:41
the best best use case here. So there's too much of AI slop coming in. People are shipping fast using AI and they are
10:48
actually aware what what's what they are shipping the code. There's a lot of junk code there. So what if you like map the
10:55
junk code and then you can you can see what's what's wrong in your code as well by just using a small coding tool coding
11:01
assistant graphify. So it can help the non-coders with no technical background as well to like retrieve and then fix
11:08
their fix the issues with their code bases. And that's one of the use cases I I forgot earlier on. Yeah, there's definitely a lot of AI
11:13
slop out there, bro. Um, so it it that question kind of tied into my next
11:19
question, so I might not have to go deep into it, but maybe there's a uh specific
Business ROI Digital Twin
11:25
ROI that you can outline. you know, for most business owners, most of my clients, you know, these are guys that
11:30
are doing at least a quarter million plus a year in their business and they want to know, you know, because we've
11:36
put out Charlie OS and that's, you know, oneclick install for claw code. You know, the frameworks, the file
11:43
structure, you know, how important file structure is using clawed code internally. If that's not set right,
11:48
like everything else is just going to be a mess. You you you need the right frameworks. So for those of for those of
11:56
for those business owners that might be watching this or anybody looking to start a business um is there anything
12:01
that you might have missed in terms of graphify like how how is this save we
12:06
know how it's saving time but is there anything that you might have missed in what you said about the particular ROI
12:12
financially for businesses in this case is there anything you want to expound on there yes so as an enterprise let's say uh you
12:19
have hired a new person and then uh so basically The whole concept is about a digital brain. So we create a digital
12:27
twin of your whole enterprise. Let's say uh see the context uh doesn't last for
12:33
long. Either it says in your head or it says on Slack messages. It is people
12:38
forget forget what they talked about yesterday. So graphify has a capability
12:44
of giving you a persistent digital twin of your whole corpus which you can retrieve at any certain point of time
12:50
rather than wasting your time and money hiring someone to search things for you or do things for you uh which you can do
12:57
just by using your own coding assistance. So you're going to save a lot of like uh unnecessary money having
13:03
been spent on employees doing very trivial task for you where you have your
13:08
own digital twin your whole enterprise brain working also graphify has an incremental update feature which updates
13:16
the whole graph automatically with every new injection of documents you have so
13:21
it's an intelligent enterprise brain which gets smarter with time so I'm also
13:27
working on a new layer of graphify which will make the whole brain smarter with
13:33
self-arning layer I'm going to implement there so graphifi's brain will adapt to
13:38
the users's use case like a certain domain let's say you are working in legal tech or rather you're working in
13:44
real estate so the whole graph will start adapting as per your own requirements and that's my uh coming
13:50
vision future vision is to help the enterprises as on bigger levels yeah I I love how excited you are about
13:56
this product because a lot of the guys that I've been meeting on this journey, uh, posting on social media, it's only
14:02
been about 5 months for me, but meeting you and seeing this product really, really push far and especially the one
14:09
video I did, it has a little under half a million views. That was a huge accomplishment for me. Um, so thanks to
14:15
you and your product, you know, I was able to accomplish that. So, you know, it's really exciting hearing about the
14:21
new developments. I know you're very excited about the enterprise solution and I I think with the right push and uh
14:28
hopefully guys if you guys are watching this and this gets enough traffic definitely like the video. If we get
14:34
over a thousand likes on this video, he hasn't promised to do another video with me yet, but if this gets over a,000
14:40
likes, he might be tempted to do another one with me so we can go deeper on the workflows and see what he's got going
14:46
on. Um, so anyways, the first version that dropped, right, and then we're
Whats New Since V1
14:52
going to go into something a lot of you guys are probably curious about, like
14:57
how would a new user use this if you're starting from scratch? We'll get into that. Uh, but I want to ask you with the
15:04
version one that dropped versus what you have now, what are some of the key differences that you've actually put
15:10
into the news version that make this even better than when it first launched? Yeah, so there there are plenty of differences now. So I have given access
15:17
to local backend system as well. So you can use Olama back end as well. Now earlier on it was just like you you were
15:23
using CL cloud-based LMS which were given by your coding assistance like GPD or so models or like Opus models. So
15:30
they were like causing a lot of problems to users on the burns of to burn of
15:36
tokens. However, now you can call on your back end using llama uh local model
15:42
as well the SLMs and it's pretty much cost effective and I've also got like 27 plus languages now earlier on we started
15:49
with like 50 plus languages so we have gone past 26 plus languages or for the as you can extract a lot of languages uh
15:56
coding languages now uh so backend system with uh a local local SLMs or LLMs and then uh with with more coding
16:04
languages thirdly I have released another feature You can connect your Google workspace or even your AWS
16:10
bedrock models there with graphify as well. Now it's it's a big solution for the enterprises they were looking for.
16:16
So you can even collect bring your documents from Google workspace directly into graphify create a map of those
16:23
documents and then you can retrieve whatever you want. See the relationships on your offse as well if you want. It's
16:28
more interactive. That's crazy. And visually appeal. Yep. I I I'm using GWS CLI. It used a
Local Models And Privacy
16:35
term called SLM. I'll have to research that. I know what LLM is. What is that? Small language model. Yes. All right. That makes sense.
16:44
So basically large language models do not fit in smaller RAM size. So you need larger RAM or so GPUs like CUDA kernels
16:52
to fit them in. So rather than like going for LMS, you can just go for SLMs. And SLMs are the future by the way
16:57
because everything is going to be local and and that's what people are keen on. so that their data isn't shared to cloud-based LLMs in the future. And
17:03
that's my vision is at the moment. Yeah. The the the good news is that the uh the small language model, the local
17:09
stuff is not any worse. Yeah, it's a local first AI memory system. That's what Graphy will be in
17:15
coming days or months. Pretty much none of your file will be shared with cloud-based LMS. It'll all run locally
17:22
on your system. And that's the best uh win for the enterprises in the coming days. Yeah, that that's exactly what I'm
17:28
what I'm seeing is when we get to a point where the large language models become they're going to keep getting
17:34
better, right? Always better and better. But what's going to happen as a byproduct is that these small local models, this is the worst they'll ever
17:41
be and they're only going to get better. So, a year from now, a small language model locally hosted might be as good as
17:48
Opus 4.6 or maybe two years. I don't I don't know. What would you say like when are we going to have a local model as
17:54
powerful as let's say Opus 4.6? six or 4.7 like what is your take on that perhaps maybe a year or so with the pace
17:59
we are moving forward I think so we are going to have pretty smart SLMs pretty soon so basically what I've done is I
18:06
I've fine- tuned SLM for my use case and they work decently well as as large
18:11
language models as well so it's like how could you fine tune your SLMs for your use case so what I want to ask you now is I want
New User Setup Tips
18:19
to look at what do we uh what would be the best way for a new user
18:25
maybe who's never used this before to effectively set this up, you know, because I know the biggest for me was
18:31
like I set it up and it ran through my whole daily limit and then I I spent like an extra like 25 or like 30 bucks
18:38
off of and I'm on the $100 monthly plan with with Claude. Um what would be the best way to effectively run this without
18:45
killing your whole limit and just exceeding it? Like what would you recommend? would you say um one of the
18:50
gentlemen I did an onboarding call with before this he's like do I uh do I
18:56
pretty much like say x amount of files do create a handoff do the rest of them do I just power bang you know my whole
19:04
database um and am I even ready to run graphify for a new user it's like is it is it
19:10
effective for them to do it with nothing in their codebase especially if they're not like developing things that's uh
19:16
that's a lot I just threw at you um but does That question makes sense. The main problem if you are running if
19:22
you are calling an LLM to parse your code base that doesn't make sense at all. I've given I' we have got like a
19:28
there. So so you don't have to like call a cloud-based LM rather than like it's free of cost. The a call is free of
19:35
cost. There's no API call included for an LLM. So a pretty work pretty much well for the code bases. You have to
19:42
call the LLM back end only for your documents. Abstract syntax trees. It's a free Python library which can connect
19:49
various parts of a codebase like functions and all of those sections of a
19:54
code all together and create uh relationships between them in different files even in multiple repository. So
20:01
it's it's free of cost. So basically what you have to do is with the new feature I've added in graphy with backend system is Olama. So you've got
20:08
Lama as well as a back end which is free of cost as well. you've got locally running LLMs or SLMs for you or using
20:15
Olama's API. Uh so what you have to do is call the code base using as which
20:21
which which works pretty well for my even doers as at the moment and then if you have some documents you can use an
20:26
LLM. So which creates a semantic extraction of the corpus of the of the chunks within a document and that's how
20:33
you can save a lot of money. Secondly, when you are querying that the the graph the map, make sure that you first ask
20:41
you you prompt to give you the god nodes from the whole corpus first of all. So so if you have large amount of god nodes
20:48
that means that something has gone wrong with the cohesion or or maybe with your codebase. Uh if you are looking for an
20:54
extraction of your codebase as well. uh and if you have a a decent number of god notes maybe for documents as well things
21:00
have worked fine and then you have the whole architecture ready right in front of your eyes of even of your documents as well of your codebase just by
21:07
prompting for god nodes and secondly from god nodes you can easily see where you have to prompt and write to minimize
21:13
as much tokens as possible uh third also if when you are like ingesting more data
21:20
make sure you use the word uh update as well graphy update so that it doesn't like start fresh from the beginning
21:26
rather than I've used hashing strategy h 256 in here so it will start from where
21:32
it left and then I've also got some dduplication techniques so there will be very less collusions of the entities
21:38
within the same document or multiple documents so I've taken everything uh taken care of everything all you have to
21:44
do is follow the right tactics for the right hooks I have put in the readme file or maybe you can just ask cla code
21:50
to explain how to make the best use of graphy uh as the initial shot and then you work from there as well as a as a
21:56
work tree workflow for for your whole use cases. Yeah, that makes sense. Yeah. So, it's a always remember as for
22:02
code bases, LLM's for documents, extract backend using llama or if if you want to
22:08
save cost or if you want to go with all your cloud-based LLMs, you can choose cloud code, gemini, openis, any of the
22:15
models GPD. Uh and then after that always go for the god nodes and from
22:20
there you can work work through into anything. Also make sure that you don't have to like when you are prompting the
22:25
graph don't ask the LLM to go through the graph rather ask the LM to extract from the graph. So graph is the memory
22:33
graph is the context. So always make sure and you can work around a bit and I've seen people so there is no ceiling
Token Savings Reality Check
22:39
or floor in token uh savings here. What I've got 71.5 in my read me is for the
22:47
repository I tried on and I've totally mentioned that it I've seen people who are getting like 90x token reduction in
22:54
their retrieval as well and there are people who are getting 20x token reduction in their retrieval. So it's totally corpus dependent and and there
23:01
is no ceiling or floor here. You know what's interesting? I have a lot of hate in my comments and I'm sure
23:06
you've seen it. Oh 70 times what are you out of your mind? I'm like, "Guy, relax." Like, all these all these
23:14
different situations are circumstantial. It's like from what you're saying,
23:19
right? You and let me let me just let me can I I'm going to try to put this in layman terms because, you know, you're you're a
23:25
super smart guy. For me, a lot of what you're saying, I'm reading in between the lines. Like, I hear big words and I
23:31
chalk it up to like, okay, that he's just saying like as for code. So if if
23:38
uh GSD builds an app, okay, that's perfect for that. So in a n like if you
23:43
had to explain it in like an an to a ninth grader, you know, as having it
23:50
look through codebase, great. Using um you said use an for your documents
23:57
and you mentioned uh llama back for documents which are also free of cost now. So it it gives you a
24:04
local model running on your own RAM on your own whole machine rather than calling a cloud-based LM. And I and I know that a lot of this is
24:11
already hooked into Graphify. So if you guys have any questions on this, you guys can drop a comment below. And what
24:18
I would suggest to get the best use or to get the best out of this, take the
24:23
transcript of the video. What are you waiting for? plug it into your clawed code and when you guys are setting up or
24:29
using Graphify, this is going to be perfect for you guys and um I think this is going to be great for anybody looking
24:35
to use Graphify. So, on that note, I I I want I don't want to take up too much
24:40
more of your time. I know you're excited about the growth. You know, I can go on to the the marketing side. I don't think
24:46
you need help with the marketing right now. That's it's exploding on its own with the uh the creator outreach. So, we
24:53
talked a little bit, maybe we didn't, about the uh the future of AI memory.
24:59
So, there's two more questions I want to ask you. And then there were some some of my actual followers and school users
25:07
that had particular questions and that isn't actually super long. Um, but
25:13
before I get into these two questions, I just want to expound on what he said because he said something super
25:19
important here. Sophie Safi Sophie was saying that each of these situations
25:24
when you're using Graphify, it's circumstantial. There's no floor. There's no ceiling. So now when you guys
25:32
are using this tool, it's dependent on how you're leveraging it. That's how you're going to start seeing
25:37
the savings. So if you're not following this this playbook here, you know, he's
25:43
talking about god nodes. It's like highlevel uh nodes, the god nodes, right? If you guys aren't leveraging
25:50
that and knowing how to input that when using claude, it might be running and looking into all the specific nodes. So
25:56
if you want highle information, you can then scalpel like go into the graph and get the information you need with those
26:03
semantic relationships tied in. So the proper understanding and context and
26:08
what he's saying of 70 times usage is key when understanding whether that's
26:14
actually valid or not. So, I just want to leave that there and I want to tie
26:20
into the next two questions here before we go into the the user questions. Where do we see the future of AI memory going?
Future Of AI Memory
26:27
And I just want to hear your thoughts there and I want to understand after you share that, what is the biggest takeaway
26:35
you're trying to give a user who's watching this video? what is the biggest takeaway you can give them that'll 20
26:41
times output their usage with with their their language models maybe let's just say cloud code so two questions there
26:48
yep so the current future uh with AI memory uh always be a problem to be
26:55
honest if you have worked with the whole structure of how neural networks are set and how retrieval systems work so
27:02
basically there are some people some startups who are working on the same problem like super memory they are based in SF hydrab And then there's another
27:09
thing called Gbrain by Gary Tan who is also working on the same problem of large context for LLMs and and and it's
27:17
the market is very bright in in in person memory systems and I believe
27:23
there's too much to come in next three years and uh the problem will never be solved until unless we have a new
27:29
architecture other than transformers which are the base of NLM at the moment
27:36
because there always be hallucination there there will always be a loss of context and uh they but they they never
27:42
will be 100% efficient as as a human and that's something uh will always exist
27:48
and persist but you can always optimize with the right tools with the right techniques as I said graphy is one of
27:54
those tools if you optimize it the right way you can always have a persistent memory of your code base of your
28:00
documents with incremental updates and features okay now I hear you on that I think how
28:05
we use things will row like how much MPG in a car miles per gallon that has we're
28:11
never going to get unlimited miles per gallon but we can always raise that ceiling. We can change the source of the
28:18
gas you know the the the power source from gas to battery. We can extend those
28:24
miles with different uh infrastructure and and setups. So I see where you're
28:30
going with that and I wouldn't disagree at all. So what what would be your biggest takeaway like to anybody
28:37
watching this? What do you want them to know that's in your brain that would allow them to use graphify and totally
28:43
just change like total game changer using graphify? What would that takeaway be?
28:49
Takeaway would be first of all you should know what you're putting into the whole system and if you're putting
28:54
codebase you should always aware that graphify has uh in there so you have to like uh spend your tokens just to fetch
29:01
the the code base. So, play it smart. Uh, and then you can have a lot of savings in there. For documents, maybe
29:07
use a as I said earlier in in the interview as well. You can have a SLM
29:13
now working for you using Llama. Uh, and it's free of cost as well rather than like going for a cloud-based LM that
29:19
takes a lot of money. Secondly, you can split your repository or your documents as well and then you can reingest them
29:26
uh in a way they they will just get added in the best way possible rather
29:32
than like putting the whole chunk of documents in one shot. So basically we say that claude has 1 million of context
29:39
and everyone is going towards like maximum millions of context in in one session right but it doesn't make sense
29:45
at all. the more the context would be there will be a massive chance of
29:52
hallucination or dilution in the reasoning as well and that's where you lose and that's where the whole graphite system comes in. So graphify isn't just
Neuro Symbolic Graphs
29:59
uh supporting neural networks it's giving a s it's giving a a rise to neurosymbolic AI systems where you have
30:06
a map or symbols to support the neural networks to come up with a response. So, and that's the whole graph vision is
30:13
neurosyolic AI. Neuro symbolic AI. I love that. I it it's a way to simplify the language of
30:22
the data being processed in a way that makes it efficient and and more relatable. No,
30:29
not really. Yeah. So, neural networks, right? And then neural networks can
30:35
hallucinate as well. So all the retrieval all of those uh path pathways
30:40
can can hallucinate any response. So to make this whole problem uh with solvable
30:47
we have symbols. So symbols are in the form of graph structures and then if you go deep dive into this graph theory
30:53
you'll know about the new symbolic AI systems and even graphs also can give you an opportunity of clustering various
30:59
communities in there which other tools like obsidian would never would never able would never give you like I've seen
31:06
like some people are asking can't we just like make obsidian do do this and that it's not possible at all obsidian
31:13
can't do clustering for you obsidian can't do cross community interaction for
31:18
it will just give you an visualization which is which looks pretty much well but there is nothing credible you can
31:25
take from there so you always need you always need a new symbolic structure or a tool that can solve it for you in one
31:31
shot and that's where graphy is yeah I love that because that was literally what what a lot of my guys are
31:36
saying to me bro they're like well can I just use obsidian by itself but this is
31:42
this is basically putting in a way that makes sense like when it's in graphify it's just ass is this is just what
31:49
already is existing versus having it organized and sifted through with your
31:54
tool. So, I love that. I think you hit the nail on the head with that. And what
31:59
I want to look at, you know, we we talked briefly now about the
32:04
uh I'm working on another research for graphify, but I'll be doing hyperbolic embeddings uh using point ball theory,
32:12
which is a pretty much a famous problem in mathematics from 2002. So this will
32:18
solve the problem of retrieval hierarchal retrieval as well. So basically when they are god knots on top
32:24
so they are baby nodes the child nodes at the bottom and then when you when you go down into into a tree there's always
32:31
a context drift. You lose the context the more deep you go into a tree. And to solve that problem, I'm working on
32:36
hyperbolic mings which will give you an exponential
32:42
downward flow into a graph into a tree without losing context. And I think I will be implementing this whole research
32:49
uh what I'm doing with graphifi's code base as well which will help and
32:54
leverage the use cases of my users currently on on the open source side as
33:00
well. Yeah. What what we're going to have to do man is when you come to the states we'll have to meet up. Um, I'd love to
33:06
do like a um even if it's maybe I'm I'm gonna be in Florida at the time. I don't know when you're coming, but if we have
33:11
something upstate in New York or something in New Jersey, uh what we could do is we can plan ahead. We can
33:17
invite people out. We can maybe rent an area kind of bring some awareness next month because I' I've won a
33:23
hackathon last last week and they are taking me to SF uh maybe in June or July
33:29
something. Okay. Yeah. Well, whenever you come, man, we we'll have to figure that out because I I definitely want to do some local meetups. Uh but to kind of wrap
Obsidian And Shell Scripts
33:36
this thing up, I have some um some basic questions here. We talked briefly about the Obsidian overlap. Um it's not a hard
33:44
dependency. We know that it it is I I would say it's a decent recommended
33:51
pairing because it's visual. Yes. Contact people obviously I think is is
33:57
is a decent way of visualizing your stuff. and then it totally makes it more interactive rather than like going
34:03
through the whole just on files and then extracting stuff. Yeah. Yeah. I I would agree with that.
34:08
Um we know that Obsidian by itself is not going to get the best out of their data. Um so I what I want to ask you is
34:17
we talked about the token economics that was a big question. Um something anal ad
34:23
in my school he asked why does why does graphify not support
34:28
scripts shell scripts are very flat I think I believe and they're very indiscriminatory I would be adding those
34:34
as well in the future but for now if you want you can just pass them to an LM so what LM would do is rather than like going with an A which isn't capability
34:40
at the moment for shell scripts what you can do is just put those as documents and then make the LM do a semantic
34:47
extraction of a shell script so it'll be able to fetch you all the nodes from there. That's a that's a fix for now, a
34:53
temporary fix. I'll I'll get it sorted maybe in coming releases so you can have shell scripts as well in there. But to
34:59
me, shell scripts are a bit flat and that's the reason I didn't put them in the beginning. If if it's a requirement, it's a hard requirement, I I'll for sure
35:06
put that in and I'm I'm more open for PRs and issues being submitted on the GitHub repository. I I I fix them on
35:13
time and I'm pretty much active at the moment. Bro, I know you're always moving. every time I talk to you, you're always you're
35:19
always moving. Um, so we kind of we kind of hit briefly, right? Like what would you do in your first three minutes of
35:25
Graphify with a fresh uh client codebase which you've never seen? Which views or
35:30
curies do you hit first to decide whether the architecture is sane versus a landmine?
35:35
Yes. So you're talking about codebase, right? Mhm. Let's say you got a new hire as I
35:41
told you in the beginning as well who has never seen the architecture and then you have been running graphify on your
35:46
own code base even of on your multiple repositories of your whole enterprise where it has been incrementally
35:51
incrementally updating the nodes and even the god nodes even in the child nodes and then clustering them even
35:58
merging or connecting various communities together how they are dependent on on various dependencies in
36:03
the codebase. So it's pretty easy for a new hire or maybe a new contracted senior engineer who wants to go to the
36:10
code base and then have a look at the at the whole arch whole band architecture in in one shot,
36:16
right? And I I think it's pretty self-explanatory when it when it's ran. It's like you're best off just running it through and kind of just seeing the
36:23
output. I mean, if you have a subscription, it's not like you're paying for API cost. So yeah, that was a great answer. Um token economics we've
36:30
went through. Um, for the people who have asked why is it that they're not seeing any token savings, it's because
36:36
of the way that you're using it. We we if you know we've discussed that earlier in this video. Um, so I think token
36:43
economics practical first run obsidian overlap we've went through. Um, and the
Workflow And Connectors
36:50
last question is what is your current workflow like? What is it that you do like using this tool? Like what's your
36:56
day-to-day look like with graphify for yourself? So current workflow is uh as I as I said I've been using local lms
37:03
since earlier days but the thing is I didn't put them in since I was doing some hardcore testing if local are going
37:09
to work good in a good way or not and and I've seen after some testing it's
37:15
working working pretty well. So for the documents as I mentioned earlier on as well so you you use a local lm and then
37:21
you can extract the information from there. Also by the way I'm adding a connector to various uh even even your
37:27
one node even your slack I'm working on connectors as well. So you can connect your slack messages and then you can
37:33
have them uh as as a map as well from from now onwards in in the coming
37:39
releases even your slack messages even your meetings the meeting transcripts
37:44
everything will be in a brain and as and you can work as an enterprise it will be on one short retrieval. So my current
37:50
workflow is as I said local LM for the extraction and then always go to the god
37:56
nodes first and then start retrieving from there and make sure you pron the graph and not like the LLM to go through
38:02
the whole corpus once again and that's one thing also I'm adding some more features so that the the like the the
38:10
agent would be hardcoded or hard configured to go through the graph first
38:15
rather than doing anything else. what if what respect to what you prompt when graphy is active in a session and that's
38:21
something I'm I'm looking forward to as well. Yeah, I I think finding a way to properly get it to work maybe via a hook
38:28
or maybe in the cloud MD those instructions [clears throat] with it effectively like if if there was a way
38:35
to have that integrated with an optimized [clears throat] flow to save tokens and it worked great like that
38:41
would be phenomenal. I think some of the for me at least right now it's like anything the static information the
38:48
information in your codebase or for me using it as like an operation center like my clients my data like everything
38:54
and anything I I do I have MCPs CLI tools but my assets become stale over
39:00
time so the process of having to remember to you know groom your data and go through a lot of that does
39:06
[clears throat] um for most people um become a little bit of a constraint. So I'd love to see you know a lot of these
39:12
things that you're doing over time to solve. I know always one more thing just hit my brain always make sure you update you you you
39:20
put graphy update as well when you are like ingesting new files or code bases or documents as well. So make sure that
39:27
it gets the right injection rather than when you you ask a question it doesn't have to go through the whole code once
39:33
more. So you have to like update it as well. some of the techniques. Uh maybe I can share you a file and then you can
39:39
share it with your audience as well. Like the best way of using graphify would be pretty much helpful like the
39:45
way I do from the creator itself. What I want to do and I didn't mention mention this earlier but he's dropping a
Book And Final Thoughts
39:52
book or he already has it out your new book. Yeah, it's out on gum road. It's called the memory layer. So it it talks about
39:58
the whole journey as well how I've been working on graphs and also tells you about how graph has been in in like in
40:05
computer science you can you can learn about knowledge graphs and graph theory spectral graph theory the matrices the
40:10
llassian matrices and everything the community detection strategies using clustering algorithms lead in low vein
40:17
so everything every theory about graphs is there and I think it's it's a it's a
40:22
much must something must you can go for the link is going to be in the description below
40:28
And if you did have an additional PDF or playbook that you put together, I can host it on my website. Um, so if you
40:35
guys check the description, the book will be there. I might have a resource depending on, you know, whether he wants to drop that sauce or not. Um, but I
40:43
know Sophie, Sophie, I know you're busy and you've answered so many questions
40:48
for me personally. I'm going to take this transcript and throw it into my quad so that I can self-improve. What I
40:55
love what you're doing is connecting the different like grid AI transcripts or you're you're pulling in Slack and
41:01
you're creating you know nodes for that and you're going to create that on like an enterprise level. I think that's
41:08
absolutely huge and [snorts] I think there is going to be a lot of ways for us to collaborate in the future. Uh but
41:14
I don't want to hold you up. So if there's anything that you want to add that you want to leave people with
41:19
before we conclude this thing I I would love for you to you know share your last words again I know you're busy so floor
41:26
is yours so to other fellow developers I just I've just got one thing to say you can
41:32
just do things and if you just have like the best determination in you that's it
41:37
you can just do things yeah we could just do things with Claude now the best way to put it so guys uh if
Wrap Up And Goodbye
41:44
you can please like the video. Subscribe if you like the kind of content that you're seeing here. And we're going to
41:51
Sophie and I are going to try to do some more videos together if he's got time. But I'm also going to link his social
41:57
medias below. We'll link Graphy. Um but again, everybody just leave a comment below. Thank him for joining us here
42:04
today because he's really busy. He's doing a lot of great things in the AI memory space and you got to remember his
42:10
name. You got to remember this guy. He's going to go far. So Safi, thank you for joining us, brother. Greatly appreciate
42:16
you. Thank you so much, Charles, for having me here as well. All right, brother.
