Transcript
00:00:00Graphify just solved Claude Code's memory problem.
00:00:03It's able to turn any repository and turn it into a wild knowledge graph,
00:00:06just like the one you see here.
00:00:08And in the process, it allows Claude Code to give you more accurate answers
00:00:12at a fraction of the token costs.
00:00:14It's able to do this by traversing your entire code base,
00:00:17mapping all the connections, and discerning the why behind the connections.
00:00:21And the best part is, it's also open source and completely free.
00:00:24And so today, I'm going to show you how you can get this working yourself
00:00:27and what's actually going on under the hood,
00:00:30so you can start leveraging it right away.
00:00:32So Graphify came out a couple months ago.
00:00:34It's at nearly 60,000 stars.
00:00:36And what it does is it allows your AI coding assistant,
00:00:39doesn't have to be Claude Code, but that's what we'll be using today,
00:00:41to map your entire project, code, docs, PDF images, and videos
00:00:45into a knowledge graph that you can query instead of gripping through the files.
00:00:49So we are able to take Graphify and point it at any sort of repo we want,
00:00:54and it creates this sort of knowledge graph.
00:00:55The reason we care about this is when we create a knowledge graph,
00:01:00it allows Claude Code to more easily answer questions about that repository
00:01:04because everything's already mapped out.
00:01:06It's very clear how A connects to B, how B connects to C,
00:01:09and why those connections matter.
00:01:11This is in contrast through gripping through files,
00:01:13which is how AI coding assistants like Claude Code normally work.
00:01:16Kind of a simplistic analogy, but it's as if it's just doing Control-F
00:01:19and trying to search for it versus having a clearly mapped out path of how everything's going, right?
00:01:25This gives Claude Code a map while gripping through files doesn't at all.
00:01:29So because of that, it costs less tokens to get more accurate answers with something like Graphify.
00:01:35Now, how significant are those token savings?
00:01:37Well, some people are claiming up to 70x, which I found to be a little on the high side.
00:01:41And as you'll see when we demo it today,
00:01:42it's a bit lower than 70x, but still significant.
00:01:45So that's the why you should care.
00:01:47Now let's talk about how it actually works.
00:01:48How do we go from a code base to some sort of knowledge graph like this,
00:01:51which looks very, very similar to something like a graph RAG knowledge base.
00:01:56Are they the same?
00:01:56How does this relate to RAG?
00:01:57We'll talk about that.
00:01:58Well, the way it works is through three different passes.
00:02:00On the first pass, we are looking at the code structure,
00:02:03and this is completely free.
00:02:05Everything you see right here, this is just through pass one.
00:02:09This is deterministic.
00:02:10This is an AI doing a guessing game.
00:02:12It is literally going through the code itself and saying,
00:02:15this piece of code relates to this second piece of code.
00:02:18And that's literally how the code base is written out.
00:02:20These are established connections.
00:02:22As it says here, a tree sitter parses your code files and extracts classes,
00:02:26functions, imports, call graphs, and inline comments.
00:02:29This runs locally with no LLM involved.
00:02:31On pass number two, it's looking at video and audio,
00:02:34if those files exist at all.
00:02:36And if they do exist, they're going to be transcribed with faster whisper.
00:02:39And so once they're actually broken down into text,
00:02:41they will also be injected into the knowledge graph.
00:02:44Lastly, it does a third pass on docs, papers, and images.
00:02:47So if your code base includes things that isn't true code,
00:02:50whether that's just like PDF files, documentations, images, whatever,
00:02:54this gets hit on pass number three.
00:02:56And this is where the large language model actually comes in
00:02:58and does some sort of like semantic analysis,
00:03:00aka what does this document actually mean
00:03:03and where should it fit in this larger knowledge graph?
00:03:06This third pass is kind of similar without true embedding
00:03:10to what a RAG system does.
00:03:12Once it does all that,
00:03:13it then begins to create the actual knowledge graph itself.
00:03:17It goes into a little bit more technical detail in here,
00:03:19but all you need to understand is it's going to create nodes,
00:03:23nodes, which are these little circles, right?
00:03:26Each one of these circles is a node.
00:03:28We then have edges, which are the line between two nodes,
00:03:33two things that are connected, and then communities.
00:03:35Communities are simply large groupings of nodes
00:03:38that are similar in nature.
00:03:39What you see here are 486 communities.
00:03:43So that's kind of the overview of how the data is actually extracted
00:03:46and turned into a graph.
00:03:47And remember, we care about turning into a graph
00:03:49because for all intents and purposes,
00:03:51it's a map to cloud code,
00:03:52so it can more quickly answer questions.
00:03:54Now, you probably have a few questions at this point.
00:03:56One, what if there is no code structure?
00:03:58What if I'm pointing at a repository full of markdown files?
00:04:01It's just like a bunch of documents
00:04:02that I want to create a knowledge graph of
00:04:03and I don't want to go full RAG.
00:04:05Can I do that?
00:04:05Yes.
00:04:06In fact, you can actually turn it into an obsidian vault
00:04:08through Graphify.
00:04:09We'll talk about that a little bit at the end.
00:04:11The second question you probably have is,
00:04:13yeah, this actually does look super similar
00:04:15to something like GraphRAG.
00:04:17What's actually the difference
00:04:18and when should I use one or the other?
00:04:21Well, the biggest difference between Graphify
00:04:23and a GraphRAG system like LightRAG
00:04:25or RAGanything or Microsoft GraphRAG
00:04:28is really going to be the embeddings, right?
00:04:29Graphify isn't using any embedding system whatsoever.
00:04:33The second biggest difference is the use cases.
00:04:35So Graphify is best and we get the most out of it
00:04:37when we're talking about code bases.
00:04:39But if we see some sort of huge repo,
00:04:40whether it's a new one or one we've been working on
00:04:42and we want to figure out how it's wired,
00:04:44Graphify is perfect for that.
00:04:46GraphRAG, on the other hand,
00:04:48is great for something that's more unstructured.
00:04:50Let's say you have tens of thousands of documents
00:04:52that are all PDF files or Markdown files
00:04:55and you just want to ask about them.
00:04:57You know, imagine they're all policy documents
00:04:58and you're asking like,
00:04:59what does the policy say about X, right?
00:05:01It could be anywhere amongst any of these documents.
00:05:04They aren't necessarily connected.
00:05:05It's very unstructured.
00:05:06That's where GraphRAG or really any RAG system shines.
00:05:09That being said, the division between those two here
00:05:13is kind of murky
00:05:14because like I mentioned on that third pass,
00:05:16we can kind of do that with Graphify.
00:05:18It's almost like a RAG light system in that sense.
00:05:21So that's what Graphify is,
00:05:22how it works and why you should care.
00:05:24Now let's talk about actually installing this thing
00:05:27and using it for real.
00:05:27But before we jump into that demo,
00:05:29a quick word from today's sponsor, me.
00:05:32So not too long ago,
00:05:33I released the Cloud Code Masterclass
00:05:35and it is the number one way to go from zero to AI dev,
00:05:37no matter your technical background.
00:05:39This course gets updated weekly
00:05:40and it also includes additional masterclasses
00:05:43like the Codex Masterclass
00:05:45and the Cloud OS Masterclass.
00:05:48So if you're someone who wants to take this
00:05:49a little more seriously,
00:05:51definitely check it out.
00:05:52You can find it inside of Chase AI+.
00:05:53There is a link in the pinned comment.
00:05:55So installing Graphify is relatively simple.
00:05:58We have a few prerequisites
00:05:59as well as instructions for how to install it.
00:06:02If you're using Cloud Code,
00:06:03I suggest you make it very easy on yourself.
00:06:06Just go to the Graphify GitHub link.
00:06:08I'll put that down below.
00:06:09Copy it, paste it into Cloud Code
00:06:11and just tell it,
00:06:12hey, install Graphify for me.
00:06:14But if you want to do it manually,
00:06:15you can just follow the steps
00:06:16as they are laid out.
00:06:18And again, Graphify is platform agnostic
00:06:20and it works with any coding agent out there.
00:06:22And once you have Graphify installed,
00:06:23the next question becomes,
00:06:24okay, how do I use this?
00:06:25What are the commands?
00:06:27Well, there are quite a few commands
00:06:30and there's so many commands.
00:06:31In fact, you are not going to
00:06:32remember any of these.
00:06:33Luckily, when you install Graphify,
00:06:35it's going to come with a Graphify skill.
00:06:38The skill is going to teach Cloud Code
00:06:39how to use Graphify
00:06:41and when it should use which commands
00:06:42depending on the natural language you use.
00:06:45So that being said,
00:06:47I suggest you take a look at the GitHub repo,
00:06:49somewhat familiarize yourself
00:06:50with what is possible
00:06:51because there is a lot.
00:06:52But understand,
00:06:53you don't have to have this memorized.
00:06:54Cloud Code understands what to do.
00:06:56But there are a few
00:06:58we should be aware of.
00:06:59If I do forward slash Graphify,
00:07:00that's going to run the whole thing
00:07:02on whatever directory I'm currently on.
00:07:04There are also Graphify commands
00:07:05for querying the knowledge graph.
00:07:07So if I do Graphify query
00:07:09or Graphify explain,
00:07:10it's going to explicitly tell Cloud Code
00:07:12or whatever coding agent you're using
00:07:13to, hey,
00:07:14take a look at the knowledge graph
00:07:16when you answer this question.
00:07:17Don't be lazy
00:07:17and just try to answer it on your own.
00:07:19Furthermore,
00:07:19we have commands
00:07:20to make sure it's always on.
00:07:21So if I do Graphify Cloud install,
00:07:23that means it's always going
00:07:25to use Graphify
00:07:26to answer the questions.
00:07:27I don't have to be explicit.
00:07:28It literally becomes a hook.
00:07:29And there are some other
00:07:30interesting flags
00:07:31like the obsidian flag,
00:07:32which will,
00:07:33with one command,
00:07:34create an entire obsidian vault
00:07:35for you
00:07:36and fill it with
00:07:37whatever Graphify comes up with.
00:07:39But again,
00:07:40remember the skill is installed.
00:07:41So if you ever get confused
00:07:42about what makes sense,
00:07:43just ask Cloud Code.
00:07:44It will understand.
00:07:45So now let's actually run this.
00:07:47For the demo,
00:07:47we are going to be pointing
00:07:49Cloud Code at OpenDesign,
00:07:51which is a relatively large code base.
00:07:53If you've never used OpenDesign,
00:07:55it's essentially Cloud Design,
00:07:57but open sourced.
00:07:59So I've cloned it on my machine
00:08:00and I'm going to open Cloud Code
00:08:02inside that directory.
00:08:03So we're inside the directory
00:08:04and all I'm going to do
00:08:05is forward slash Graphify
00:08:07and then dot.
00:08:08It's now going to run Graphify
00:08:10on this entire folder.
00:08:12So after running for six minutes,
00:08:13this is what we got.
00:08:15It took a look at 203 files.
00:08:17We got 1,907 nodes,
00:08:203,447 edges in 109 communities
00:08:24and output tokens
00:08:25was just under 120K.
00:08:27So it lists the God nodes.
00:08:29The God nodes are pretty much
00:08:30like the most prominent nodes,
00:08:32the most prominent connections
00:08:33inside whatever it traversed.
00:08:36We have surprising connections
00:08:37that I didn't expect
00:08:39and suggested questions.
00:08:42So if we want to take a look
00:08:42at the graph,
00:08:43I can say,
00:08:44go ahead and bring up
00:08:47the graph for me.
00:08:49So here's a look
00:08:50at the knowledge graph
00:08:51it built
00:08:52and you can kind of see
00:08:53the communities there.
00:08:54It created 109 communities
00:08:56and that's really just
00:08:56all of these clusters.
00:08:58As we scroll in on them,
00:09:00we can see the nodes
00:09:01which are the actual dots
00:09:03and then the edges
00:09:05are the connections between them.
00:09:06When I click on the node,
00:09:07you can see over here
00:09:08on the top right,
00:09:10it's type,
00:09:11so it's a code node,
00:09:12it's community,
00:09:13it's source,
00:09:14as well as its neighbors.
00:09:15But remember,
00:09:16as cool as this visualization is
00:09:17and it does look neat,
00:09:19the real value here
00:09:20isn't the knowledge graph.
00:09:21This is cool looking,
00:09:23but the actual value
00:09:24is the fact that
00:09:25now we have handed
00:09:26Claude Code a map
00:09:27to the open design repository
00:09:29and I can now ask questions
00:09:31about it
00:09:31and get accurate responses.
00:09:33So what we'll test now
00:09:34is we'll ask it a question
00:09:35about something to do
00:09:36with the repo
00:09:37and we're going to have it
00:09:38use Graphify,
00:09:39so have it actually
00:09:40use the knowledge graph
00:09:41and then we'll ask
00:09:42pretty much the same question
00:09:43not using Graphify,
00:09:44so just have it like
00:09:45grab the answer
00:09:46and we'll take a look
00:09:47at what the token difference
00:09:48looks like.
00:09:49So to take a look
00:09:49at the token difference
00:09:50with and without Graphify,
00:09:51we're going to ask
00:09:52the same question
00:09:53to Claude Code
00:09:54about the repo.
00:09:55The first one is
00:09:56trace how a design request
00:09:58flows from the web app
00:09:59to a coding agent
00:10:00and back.
00:10:00So we're trying to understand
00:10:01how this application
00:10:03actually works
00:10:03and in the first tab
00:10:04we're going to say
00:10:05use Graphify
00:10:06and in the second tab
00:10:07with the same question
00:10:08we're saying
00:10:09do not use Graphify.
00:10:10So we can see
00:10:11the Graphify skill
00:10:11being loaded right away
00:10:13and then we can see
00:10:14commands like
00:10:15graphify query
00:10:16asking the question
00:10:17we just gave Claude Code.
00:10:18Over here
00:10:19on the non-graphify side
00:10:20we see that Claude Code
00:10:21has spawned
00:10:22to explore agents
00:10:23to take a look
00:10:25at the code base
00:10:25and right off the rip
00:10:27we've already used
00:10:27100,000 tokens
00:10:28between them.
00:10:29Now in terms of
00:10:30the actual answers
00:10:30we got
00:10:31they were the same
00:10:32they both identified
00:10:32how this app
00:10:34actually works
00:10:35but with the
00:10:36non-graphify version
00:10:37we needed to run
00:10:38those explore agents
00:10:39so we were looking
00:10:40at about
00:10:40150,000 tokens
00:10:42give or take
00:10:43with the explore agents
00:10:44plus an additional
00:10:4550,000 tokens
00:10:46on the main session
00:10:47so you know
00:10:48about 200,000 tokens
00:10:50total
00:10:50versus over here
00:10:52on the non-graphify version
00:10:54we only used
00:10:55about 80,000
00:10:58so about
00:10:5840%
00:11:00of the total cost
00:11:01of the non-graphify
00:11:02which is significant savings.
00:11:03Now since
00:11:04this non-graphify version
00:11:06has now sort of
00:11:07crawled through
00:11:08the repo itself
00:11:09if I ask additional questions
00:11:11the token cost
00:11:12won't be as
00:11:13off
00:11:14however
00:11:14since we have
00:11:16the knowledge graph
00:11:16built
00:11:17whenever we want
00:11:18to ask questions
00:11:18about it
00:11:19via graphify
00:11:20well we're not
00:11:21going to have to
00:11:21deal with that
00:11:22token cost
00:11:22of going through
00:11:23it again and again
00:11:24and that kind of
00:11:25leans into the
00:11:26whole memory piece
00:11:26like we've built
00:11:27it out already
00:11:28we can always
00:11:28query it for cheap
00:11:29now the question
00:11:30then becomes
00:11:31if this is a
00:11:31living breathing repo
00:11:32what happens
00:11:33when we make
00:11:34updates to the repo
00:11:35will this knowledge graph
00:11:35also be updated
00:11:36well the answer
00:11:37is yes
00:11:38we see this spelled
00:11:39out in the workflow
00:11:40in the readme
00:11:40if we run
00:11:41graphify hook install
00:11:42it's going to
00:11:43auto rebuild
00:11:44after each commit
00:11:45and that is the
00:11:45AST only
00:11:46there's no API
00:11:47cost associated
00:11:48with that
00:11:48it's literally
00:11:49just looking at
00:11:50what actually
00:11:51changed
00:11:51what is it now
00:11:52connected to
00:11:53and it rebuilds
00:11:53that tree
00:11:54but it's at no
00:11:54cost to you
00:11:55like this is
00:11:56all done
00:11:56in a deterministic
00:11:57way
00:11:58furthermore
00:11:59this also works
00:12:00in a team
00:12:00setup
00:12:01so if you had
00:12:01two devs
00:12:02working on
00:12:02the same repo
00:12:03in parallel
00:12:04it also deals
00:12:04with that situation
00:12:05so in the end
00:12:06you get this
00:12:07persistent yet
00:12:08living map
00:12:09of whatever repo
00:12:09you want
00:12:10that you can give
00:12:10the cloud code
00:12:11so you can get
00:12:12more efficient
00:12:13answers
00:12:14and lastly
00:12:14we hinted at it
00:12:15a little bit here
00:12:16with the obsidian flag
00:12:17we can do all this
00:12:18with the repo
00:12:19that is not code based
00:12:19it's a little bit
00:12:20different and we are
00:12:21actually going to do
00:12:22that in another video
00:12:23where we drill down
00:12:23on graphify and obsidian
00:12:25and sort of what
00:12:26that connection looks like
00:12:27but just understand
00:12:28we aren't pigeonholed
00:12:29into code only
00:12:30this is a pretty
00:12:31flexible tool
00:12:32but that is where
00:12:33I'm going to leave
00:12:33you guys for today
00:12:34I think this is a
00:12:35really cool tool
00:12:36and when you look
00:12:37at the spectrum
00:12:37of sort of these
00:12:39like memory adjacent
00:12:40applications and plugins
00:12:42that we can use
00:12:43alongside things
00:12:43like cloud code
00:12:44and codex
00:12:44I think graphify
00:12:45sort of falls
00:12:46somewhere in between
00:12:47obsidian
00:12:48and a true rag system
00:12:49and I think that's great
00:12:50the more options we have
00:12:52the more tools we have
00:12:53at our disposal
00:12:53the better we can choose
00:12:54the right one for the job
00:12:55we don't have to only
00:12:56use obsidian
00:12:57you know we might not
00:12:58just be doing something
00:12:59in markdown
00:12:59and we don't have to go
00:13:00crazy and generate
00:13:02some huge rag
00:13:03infrastructure
00:13:04this is again
00:13:04it's a cool little
00:13:05middle ground
00:13:05that I think
00:13:06is worth exploring
00:13:06so as always
00:13:08let me know
00:13:08what you thought
00:13:09make sure to check out
00:13:10Chase AI Plus
00:13:11if you want to get your
00:13:11hands on the
00:13:12cloud code masterclass
00:13:13speaking of obsidian
00:13:14I'm actually going to be
00:13:15running a free
00:13:16live webinar next week
00:13:17about obsidian
00:13:18and cloud code
00:13:19I'll put a link to that
00:13:19down there as well
00:13:21and besides that
00:13:22I'll see you around
Community Posts
No posts yet. Be the first to write about this video!
Write about this video