Building Agentic RAG From Scratch in Pure Python

追加: 2026-05-12

[00:00:00]So in this video, we're going to build an agentic rack system from scratch in pure Python. And not to build another coding agent, but rather to replace or extend semantic rack in scenarios where we want company data, private data, any kind of information that we want to make available to a large language model when we're automating with AI. I'm going to walk you through a couple of code examples and we're going to start with some simple tool definitions all the way to some production best practices that you can immediately apply in your projects today. Now, if you're new to the channel, welcome. My name is Dave Abalar. I'm an AI engineer with over 10 years of experience in the field. I'm also the founder of Dat Luminina, which is an AI development company. And on this channel, I share all of the engineering best practices on what's working right now in the field for the client solutions that we're working on.

[00:00:43]Now, classic semantic rag is definitely not that as some people on the internet may claim. If you're looking at very low latency scenarios where you need raw speed or you really need to optimize for costs, semantic rag is probably still a better starting point for you to look into. But if you have the time and if you have the budgets, agentic rag will generally outperform semantic rag just because of the feedback loop that the agent can use. This is very simply explained in the example over here where generally semantic rack is a more linear process where we leverage the intelligence of the large language model only once after we have all of the information and we make a single LLM API call. Whereas with agentic rack, we have this loop where we have search tools and read tools that can get results, get back to the language model and we can utilize its intelligence in a loop multiple times and therefore it can self-correct in order to get to the right information where it may at first did not find what it was looking for.

[00:01:45]And now the simplest way to get started with aentic rag is by using simple markdown files that live on the file system in combination with three tools.

[00:01:52]a tool that can list files, a tool that can search files, and a tool that can read files. And these are the exact same tools that all of the popular agent harnesses are built around nowadays. So, let's explore what that looks like in practice. And if you want to follow along, the link to the GitHub repository is in the description. This is part of my AI cookbook. If you go in there to the knowledge folder, you'll find the Agentic Rack folder. And now I am in the build tool files. That's where I'm going to start. And the only thing that you need in order to follow along is an OpenAI API key. or you can also simply replace it with another provider that you like to use. Okay, so let's walk through this together. The goal of this first file, build tools, is to show you and to really help you understand how these tools work under the hood. Because once we understand that, once we understand how to list, how to grab, and how to read files, we can bring that together and we can start to engineer the agentic loop around this. Here we're going really to pure Python, nothing else. and we're going to explore the core primitives in order to show you what this looks like. We're going to import regular expressions and path from the podlift library and we're going to first set our notes directory. So, as I said in the introduction, we're going to work with simple markdown files. So, as you can see over here, I created a bunch of hypothetical markdown files about an engineering team. This is the use case that we will be going through for your own situation and project. You of course want to replace this with whatever knowledge you want to make available to your system. So you have to note that this is a knowledge agentic rag and then notes. So let me close this out and come back to the first line over here where we specify the directory. So what I'm going to do is I'm going to use the path library to say we're going to take the path of the current file and then we're going to take the parent and then we add nodes to that. So let's see what that looks like. If I take the current file, you can see over here that we're in the build tools.py. Now, if I expand that to parent, the path library will know that we are now in the agentic rag folder.

[00:03:50]So, it will go one level up to the parent. If we now append notes to that, you'll find that it now correctly identify the notes folder. And then the resolve method at the end is just a safety precaution to make sure we actually have a clean path in there without any sim links or other weird characters. All right. So that gives us the clean notes directory folder as you can see over here. Now now we're going to start with step one listing the files. So we're going to walk through this step by step and then we'll bring it together in a function that we're later going to provide to our agent. So we'll start with the nodes directory and now we're going to use a method that is called glop. and globe. What that is, I created a list over here with some of the uh methods and functions that we're using in the Python programming language for you to reference. So here you can see if we're using glo it find files matching a pattern in a directory. So for example, we can search for markdown files or we can also let it search for a particular phrase within a file name. If I now come back over here and you see the path is sorted nodes and we do a glob and we simply filter for MD files and I take this path over here you can see that we found all of the nodes that are in that folder because all of the files are simply MD files. So that's how we can find the files. And now if we take that same path and we apply the relative to which is another method that we can use what we'll find is that if we run this and we print that we can see that now we only get the actual file name. So let me show you the entire list over here. So you can see that by using glo we found all of the full paths in here. But for the AI agent we don't want to provide all of the context because it doesn't really need to know the full path. So that's why we can use relative to. So it has a shorter hand of going through all of the files which will save us tokens in the long run. If we now bring that together, you can now see that through this simple function at once, we can get all of these files in the relative notation that is short enough to provide to our agent. All right, so without going into an actual database, this is how an AI agent can find files on your system. All right, now let's get into function number two and that is grab. So with grap we can search for patterns that we can define through regular expressions. So it can be a clever search algorithm that we can create. So let's say again through this example we want to search for the pattern connection pool. So within our development engineering team documents we want to search for all dimensions of connection pool. So we start with that string. What we then do is we use the regular expression library and we compile a pattern and we also say ignore the case. So we want to match connection pool whether it is with a capital C or any type of other capital letter. We ignore that. We ignore the case. What we then do is we turn that into a regular expression pattern that we can now reuse. So how this works if I run these two examples over here you can see that for the first string with connection pool in there it will say it found a match and then in the second result over here it will not find a match because there's no connection pool in here. So you can see I negated with not. So there's no match in here. Okay. So that is how we can perform these searches.

[00:07:09]Now if we continue with the next step if we take a file. So let's say we take a particular file that we want to look into. Let's say we want to start with a billing runbook. MD we can grab the text from that. So let's see what's in here.

[00:07:22]Now we get the entire string of everything that is within the markdown.

[00:07:27]Now we can use another Python method, a string method which is split lines. And what this will do, it will take all of the lines that are within that file and it will create a list out of it. So it will break on a new line as you can see in here. So here we have the entire document but now in a way where we can loop over things. Now the next step that we're going to do is we're going to create a an empty list where we can store the results if we're going to loop over everything and try to search for the results. So let's create that empty list. And now here's one more piece of code that may be a little bit hard to understand in the beginning, but we'll take the lines over here and we're going to use the enumerate method over here where we can also not only loop over a list in this case, but also start a counter. So we start at one otherwise Python will be zero index and it will for example reference to this first line. It will say line zero. But this is more human readable. So you can see that right now we're applying that same regular expression search pattern that we were applying over here, but now we're simply doing it in a loop over all of the results that we got from the split lines. It will then proceed and if it finds a match, it will first store the file name with the method we covered earlier in this tutorial, which is the relative to to get the file name. And then you see we append the list with the file name, the line number, and then also what's on the actual line. So let me actually run this for you to get a better visual understanding of what this now looks like. So we can see the final result where we find connection pool on line 46 in the billing runbook.md.

[00:09:10]So this is how we're step by step defining the search pattern, then split up the file that we're going to search over, bring that together in a loop, store the results, and if we bring all of that together and take all of the files that were in the folder, right, that we looked at, we now get our grab function, our grab tool. And I wanted to really split this up because if you look at this over here, like it's a lot. Like even for me trying to read through this, you really need to break it down step by step. So I also highly recommend you to kind of like go for these examples.

[00:09:44]That's why I made it interactive with all of the print statements and to really be modular. And now once we have that stored in the function over here and I run that, you can see that if we loop over all of the files in our notes, it finds two occasions where connection pool is mentioned. First in the billing runbook and then also in the architectural decisions. You can also see the line numbers in here as well as the particular line. And that is exactly the same way that your coding agents like cloud code or cursor can navigate your codebase and find what code it actually needs to update and edit when you ask a question. All right, then we can continue to step number three that is reading the files. This is now an easy one. Grab was definitely the most complicated in here. So if we now say let's say we want to read this particular file, the architecture decision. So we start with a string then we make it a target. So we turn it into a path. You should now be familiar with what that looks like. We have our target and here is another check that we are that I'm going to apply. We have another method which is called is relative to and this checks if the target is actually within the nodes directory. So this is a safety precaution that we can add to contain the agent within a particular folder. So if I run this check you can see this is now true. So target is a path and notes there is also if I put that correctly is also a path and then if we plug them together so target path is relative to insert another path we get a boolean that we can check for then we can read the text we just take the first 200 characters in there and we can bring all of the together in our read file. So we again we check the target and we then have a value in error in there that we raise if we get a false. So we simply flag look this file is not inside our notes so we cannot use it. Bringing that together let me clean this up for you. We can then just simply print that and now we have another simple function a simple tool that we can later use with our agent to also read the file. All right so now zooming out a little bit again if we come to our project you can see that in the utils folder there is a tools.py by and here are the exact tools that we could just created. So we have the list files, we have the grab and we have the read files. If we then continue over here with file number two here you can see how we can use that simple import statement to get the tools from the utils folder. So we import grab list files and read files. And now you can see with all of the logic that we already built and also more importantly now understand we now have a very lean file over here where we can use all of these functions. So if I now run all of this, you can immediately see all of the outputs with just a few lines of Python.

[00:12:28]Okay. So if we then continue and go to file number three over here called basic agent, we can now actually start to bring this together and start working with this. So in this example, I'm using paidic AI and that is simply because it simplifies the way that we can make the tools available for a language model and perform that execution loop. But you can do this with any agent framework and also build your own using the model APIs directly working with OpenAI cloud whatever you want to use. So I set up an AI agent which pretty much means we're using OpenAI GPT 5.5 in this case and here are the tools that I make available. So this is now how we're actually going to set up a simple agentic rag system. So with the agent in memory, I can now start to execute this.

[00:13:12]And we're going to ask a question over all our engineering wiki knowledge. And we ask why does our nightly deploy job run specifically at this time? And then if we run this, this will go on and on in a loop. And it will come back to us with the question and answer. So why does the deploy job run at the specific time? Well, here is the answer to that.

[00:13:31]Apparently there is an oper over overlap with the European batch ETL and it also describes where that is coming from and it gives us a result with the usage and the tokens and also the tool calls which are five in total in this case which means that in this example this agentic rag loop over here with list files grab and read files was executed five times.

[00:13:51]But there's one problem right now and that is we cannot really see what's going on behind the scenes. So let's see how we can fix that. And for that we're jumping over to file number four which is the streaming steps.py. And what we have over here is we have the same simple agent setup with pidenticai same model but now what we're doing instead of running the agent directly and making the lm API calls. We have a little bit of a function around it where we use the agent. Method from pyantic AI. Now, this is framework specific, and I'm not going to bother with you with how this specifically works, but it's a way to intercept all of the tool calls to get a look behind the scenes as to what is going on. And this is going to be very helpful for debugging so we actually know what our agent is doing. So, I'm going to put this into memory over here because again, this is not a tutorial on how Pentic AI specifically works. This is a tutorial to show you what our agent will do, what our agentic rack system will do. So now you can see what is going on behind the scenes. Before we saw it made five tool calls, right? Now we can see look it starts with a question and then here are the agent steps. So it first decides to use the grab tool that we created and here this is where it gets interesting. Now you can see these parameters over here and you can see these parameters are what the language model actually decides to put into the tool call. This is what makes a gentic rag and aentic search so powerful because it uses all of these terms in here and it will then go off.

[00:15:25]It will list the files. It will read the files depending on the results that it got back. So it will simply go on in this loop until it has all of the right answers. Now one extra uh parameter that I added for you here is the debug parameter. So if you put this in file number four to true and we run this one more time. Now you will also literally see all of the results that it got back.

[00:15:47]So here you can see also when we do the grab command. So let me show you before it just showed you this is a little bit messy just showed you. Okay, I'm doing a grab command. So it only showed you what it did not what it got back. So now with this we can actually also see look if we do a grab. So here it runs grap you can see that it got 14 results 14 times the document the line number and then that particular line for this particular search over here. So these again were all the input parameters that the model decided to put in there. So this is super important to understand and that's why I want to spend some time here because when you are building and optimizing these information systems and you're trying to decide okay is it working is it not working and you're debugging you need to understand from first principles really how these tools and these patterns work behind the scenes. You need to be able to look in here and then see look what is the agent searching for and then check are the right documents coming up and how you can also improve this. If we come back to our tools what you should also understand that everything that you put in here your tool definition and even the dock string in here this is all information that the large language model will use. So based on this dock string it will understand okay this is what I can use the tool for these are the parameters that I can put in here and then here you can give in additional instructions or even domain knowledge to steer it in a particular direction so that when we are using it it knows what to put into the function it knows what to search for. So, this is really how you demystify what's going on behind the scenes. And luckily, most of these state-of-the-art models right now are super good at this particular task. And that is because all of the big labs are optimizing their state-of-the-art models really to also be really good coding agents. And what makes a good coding agent is not only understanding program itself, but also really mastering this agentic loop with these tools to search over information which then is your codebase. But we apply that same principle to general knowledge, markdown files, anything we need for our application. And then real quick, a lot of the lessons that I share on this channel, also this one, are coming from the real world where I built and implement solutions for my clients. So, I've been working as a freelancer for over 6 years already, and right now I focus primarily on selling AI solutions to small to medium-size companies. Now, if you've been thinking about starting as a freelancer as well, but don't really know where to start or how to find that first client, then you might want to check out the first link in the description. There you'll find a website where I talk more about data freelancer, which is the program I've been running for three years already, where I bring together developers from all over the world that want to do exactly that, make more money, work on fun projects, and create freedom. We have all the steps and the blueprints to help you get started, as well as a community with support to actually get you started with freelancing and land that first client instead of just thinking about it and watching year after year go by without taking any action. So, if that sounds like you and you've been thinking about it, but just putting it off, make sure to click on that link and watch the video on the page for more information.

[00:18:57]Okay, let's continue to step number five, that is structured output. So here we're just going to make our search agent a little bit more robust by providing it with a default output model that it needs to adhere to. So here we have a search answer which is the actual answer in plain English to the question.

[00:19:14]But we also add these citations and a citation is built up of it's a list and buildup of citations which is another data model that we specify over here which has the file the quote and the line number. Everything else in here is the same. We have that same agentic uh setup over here. We provide the tools but now we specify the output type. So if we bring all of this together and we run this over here, we now already have a minimal setup for an agentic search system. When you are running this, you can also see right now that it's taken a couple of seconds. It will probably take 10 to 15 seconds for it to complete. It needs to do the five tool calls. We're using a relatively fast but heavy model on this. And then you can see it comes back with the answer and now also these citations. And this is perfect for putting it into a another system in a software environment where downstream processes or functions can rely on this particular data format. So we can show the answer to the user and then we can have some type of front-end component maybe that we show these citations in that the user can also click on. So that is the foundation of how you set up an agentic rack loop in pure Python which can replace or extend a semantic rack system. But of course keep in mind the differences especially when it comes to latency and cost. Okay. But now you might be wondering but how does this work in a production environment? How does this work on a VPS container app or when you're using a serverless function?

[00:20:42]So what I did is I have a last file for you the file file number six which is called production. And in here I took the simple starting files that we use to understand the basics and added some production best practices to them that you can literally actually use in a real system which is all modeled after for example what codecs open code cursor cloud code all the agent harnesses are really using. So let's go over a couple of the things that I want you to be aware of. So let's come over here to file number six production. close this out and let's begin with the end in mind. Let's go all the way to the bottom and run this because what I want you what I want to show you is that it's exactly the same as what we previously did but now just with added safety precautions and depth really to it. So you can see we have some logging over here to figure out what's going on behind the scenes. But we can still ask that exact same question that we have been using. It will go through all of our markdown files. it will answer correctly and it will also site where it came from. And here you can see also again the total time it took and also the tool calls. So that's the end.

[00:21:56]That's how it works. Now let's reverse engineer to see how we got here because this is a little bit more of an extensive file. And if I scroll all the way to the top, first of all, you'll find we have some extra parameters in here. We'll add some safety precautions.

[00:22:12]for example, an agent request limit, readmax lines. This is all to avoid either the agent getting stuck in loops or opening files that are really large and still blowing up the context window by reading everything. Then we get a logger in here. And one of the biggest changes that we are doing over here in this system is we're using a library or tool called rip grab. Rip grab is built in rust. So, this is a grab tool that all of the pretty much all of the modern agent harnesses are using because it's really fast and it has some out ofthe-box built-in things where it doesn't go through hidden files. It will ignore files that are within git ignore.

[00:22:50]So, there are just a lot of benefits to this. So, we're using rip wrap and this is a simple he helper function to check if it's installed on your system. On Mac OS, you can install it via brew. On Windows, you can run the following command if you don't have that because you need to run that to uh you need that to run this code locally, but you would also need that in your production environment in order to run this. So, it would be a dependency. Okay. So, then we have the function to safely check the path. This is already something we covered. But then the biggest change is here the grab function. So, I'm not going to go into every line over here and how this works because this is just how the rip grab library works. But we're you what's interesting is we're using Python's subprocess. So what we can do over here, we can say we are within a Python file. We're running things and then Python can spawn a subprocess where it uses that rip grab tool. So that's what's going on over here. And then here you can see the parameters that we're plugging in there.

[00:23:49]So this is the command to actually start things. And we're giving all of these flags. We say we want the line numbers.

[00:23:54]We want to ignore the case. No modification. And all of this is documented over here if you really want to go deep into this to figure out why we're doing certain things. And this is all modeled again by literally doing research into the best coding tools right now and taking the best practices.

[00:24:12]So here you can see it actually fires the Python subprocess and runs everything. If I then scroll down the list files is pretty much the same. The only thing that we added to it is a little bit of extra like try accept. And one thing that you'll also notice is that we have implemented a lot of human readable errors. And we do not raise them, we return them. Because if we were to raise them, it would stop our process. So if our AI agent, for example, would make an error, it would put something into a function that results into some type of error because a file does not exist or any other edge case that is going on. It would simply stop. with the return error. It will simply return a human readable error message that the model can then intercept and course correct on. So that is also documented in the production best practices over here along with some other tips. Again, if you really want to understand this, I recommend you just go read through this and find where it's at in the codebase. But really, this is all you need for now in order to start playing and start experimenting with this. You now understand the three simple tools that make up such a a gentic rack system from simple all the way to these production best practices.

[00:25:29]How to make it a little bit more extensive. And now all you need to do to apply this into your own project is simply use this production.py file over here. I would recommend probably to do a little bit of refactoring, split it up, put it in your project, and then you need a folder or file system, ideally with markdown files because that's just the easiest to to work with for this whole agent loop to run over. You simply combine it with your favorite agent framework or you build your own loop and you use your favorite model and this will simply work. Now, finally, here is another list of some questions that you may have right now, right? So, does this only work with local markdown files? How does this work on a VPS? How does this work on container apps? And the answer pretty much to all of this is this can all work and tailor to whatever situation you are using. You may just need to adjust a few things. So, for example, you could just also put all of your markdown files in a postcore SQL database. The concepts will still be the same. You just slightly need to adjust the functions in order to create that loop to actually get the information.

[00:26:35]Same is true for the VPS or the container app. So I left some tips over here depending on how you want to deploy. This is also very common for us at data luminina working as an AI development agency where we need to tailor to different clients where some clients want to use container apps, some want to deploy on a VPS, some use serverless functions and then we also take the same principles that already work but then just make simple adjustments to make them work in the new environment. All right and then that's it for this video. Now, as always, if you found this valuable, please leave a like and also consider subscribing. And then next, if you are also interested in working on these AI projects, finding freelance work, then I highly recommend to check out this video where I essentially give my entire road map of the past six years, what I've learned freelancing in AI, and what I would do if I had to start over from scratch today and quickly get up and running again. So, if that sounds interesting, make sure to check out this video