n8n AI Agents Masterclass | Complete guide to all AI nodes in N8N

there are so many videos out there teaching how to create this AI workflow that AI workflow but that doesn't give you the knowledge to create AI workflows on your own or leverage the full power of any done instead what you need is a complete guide that explains all the AI related nodes present in aniden what's their purpose why when and how to use them in this master class we will Deep dive into anon's AI capabilities whether you're a beginner or have some experience by the end of this video you will be equipped to create powerful

eii workflows on your own let's start by breaking down what AI agents are in simple terms AI agents are autonomous programs that can perform tasks using artificial intelligence think of them like little digital workers they perceive their environment they process information and then they take action to achieve specific goals it's like having your own personal assistant but powered by AI they can do everything from sending automated emails analyzing data or even holding conversations with users and here's the best part you don't need deep technical knowledge to use them with tools like n10 you can put

these AI agents to work without writing a single line of code now let's talk a little about nen if you are not familiar nen is an open source no code workflow automation tool that empowers you to connect different applications and services it's a really good alternative for applications like make.com or zapier you can watch this video to see how to install it locally or host it on your own cloud VM now let's dive into the n10 dashboard and explore all the a related notes available now the first note has to be trigger node for now

let's select trigger manually now if I click on plus button here we see all the nodes present in an and in this video we are mainly interested in advanced AI nodes if we go inside this we'll see all of these different node categories at the top you will see AI templates if you click on this you will be taken to a website page where you can find the pre-created AI workflows then there is AI agent then open AI then there are three pre-created chains basic LM chain question and answer chain and summarization chain and at

the end there are a bunch of AI nodes if we go inside these you'll see a long list of categories of nodes that we can use in our workflows now this may seem overwhelming but don't worry we'll go through all of these notes in just a few minutes and understand why when and how to use these nodes and by the end of this video everything will be Crystal Clear if you haven't subscribed the channel yet do subscribe the channel and hit the Bell icon so that you don't miss any future updates in the next couple

of videos I'll show you how to set up anyon for production applications how to use external databases how to attach the external memories how to use the nodes in a correct way and we'll also develop some really interesting AI workflows so stay tuned for that now here we will understand all of these notes in three categories first we'll talk about the AI agent in the second category we have these pre-created chains and in the third category we will Deep dive into all the other AI nodes let's start with AI agent if we select this node

we will see a window like this now in the drop down we can see there are multiple agent types we have tools agent conversational agent open functions agent plan and execute agent react agent and SQL agent we'll talk about each of these but before that let's talk about some common features that all of these agents provide so here in the prompt we have two options take from previous node automatically or Define below most of the time we want to take the input from the previous node however you can Define your prompt here as well next

we can Define the specific output format that we want so we can enable this flag here and it will give us an option to Define our output format and then it will convert the response that It prepares to match our format before it outputs it to the next node now there are a bunch of options that we can configure for these agents we can set system message to guide the agent's decision making Define the maximum number of iterations for generating response and choose whether to include intermediate steps in the final output these settings give us

control over how the agent operates within our workflow and at the bottom we see these four options now not all of these options may be present in different asent types we'll talk about that in a bit the first option enables us to select the model that will be used by this AI agent so if I click on this we will see a list of language models that we can use with this particular asent type next there is an option to add the memory to this agent now certain agent types may need memory to process the

request in those cases we can click on memory and we can add these memory nodes now the first option is the easiest one window buffer memory it will use the memory of the VM where this workflow is running it is not recommended if you expect a large number of workflows to execute in parallel on your host because it will consume the memory of your host and make the workflows slower if you're trying to deploy a production application with a large number of workflows executing in parallel you want to use an external memory like redish chat

memory next we have an option to add tools to this agent now these are the different tool types that we can add we have calculator custom code tool HTTP request tool sub API Vector store tool and so on the most interesting node out here is the custom code tool now in this node we can give a name to this tool then we need to give a detailed description of what this tool does and when to call it then we have two language options JavaScript and Python and we can put our code here now e agent

will have access to all the tools attached to it and it can execute them if required and finally we have the output parser now if we have this required specific output format enabled we will have this option here and we can attach an output parser we have three options for this autof fixing output parser it automatically fixes the output if it is not in the correct format then we have item list output parser and structured output parser the third one is pretty amazing it allows us to give adjacent formatt output that we expect this agent

to return and the agent will return the response exactly in that format so if I select structured output parser we can Define the Json here now let's go through each of these agent types conversational agent is designed for human-like conversations this agent maintains context over multiple queries and understands user intent it's ideal for building chat Bots virtual assistants and customer support systems next we have open a functions agent it utilizes open ai's function calling feature to select the appropriate tool and Arguments for execution it is best suited when integrating with open function models so if

we click on the models we will see just two options as your openi or openi chat third plan and execute agent accomplish an objective by first planning what to do and then executing the subtasks it is perfect for complex tasks that need to be broken down into manageable steps and execute them one by one it doesn't use the memory and instead focuses on executing the current task so we don't have a memory option here next next we have the react agent it implements reasoning and acting logic reasoning about tasks determining necessary actions and executing them

iteratively it's suitable for complex problem solving tasks that don't require conversational context or memory of previous interactions then we have SQL agent this agent interfaces with SQL databases converting natural language queries into SQL statements and executing them it allows users to query data without any SQL knowledge it works with mySQL SQL light and postgress databases and at the bottom we see two options model and memory and finally we have tools agent it uses external tools and apis to perform actions and retrieve information deciding which tool to use based on the task so at the bottom

we see all of these four options for this asent type model memory tool and output parser now let's see when to use which agent if maintaining conversation history is important the conversational AI agent or open functions agent are preferable for tasks not requiring memory you can consider the react AI agent for complex tasks requiring planning and step-by-step execution the plan and execute node is suitable for problem solving without the need for conversational memory you can use react AI agent you can use SQL AI agent when interacting with SQL databases you can choose tools AI agent

for workflows involving external tools and apis as an example let's select tools agent for now and then let's select the model to be open chat model next let's attach the memory I will just select the easiest one window buffer memory for now and let's attach one tool for now calculator and this is how our AI agent looks like now we can attach the next set of noes to take action on the message written by this AI agent so let's say I want to save the output in Google sheet so I'll add a Noe to upend

or update the row in a sheet let's say I also want a notification message on a slack channel so I'll add this send a message slack node over here and connect it with this node now we can keep adding more and more nodes depending on our requirement but you get the idea right so this was our AI agent and let's talk about the chain provided by an0 now before that let's take a quick look at open a node you can message an assistant or GPT or analyze images you can generate audio Etc with this node

and we can see all of these actions we can create assistants list assistants message or update an assistant we can directly message a model we can even analyze and generate images we can also generate audio transcribe or translate recording and we have the file actions to delete a file list a file or upload a file let's say I want to generate an image so I can select this then I can select the credentials select the resource type as image because I want to make a call for this resource and then the operation name it can

be generate an image analyze image or custom API call I can select the model from here and give a prompt now let's talk about the basic llm chain it's a really simple chain to prompt a large language model so what's the difference between this openi node and the basic llm chain so if you look at this node we have option to require specific output format it means that we can control the format of the output another difference is we can add multiple prompts here in the prompt type we can select AI system or user now

this node will receive an input it will make a call to the llm and return return the output next let's talk about question and answer chain now this is basically a rag internally if you have certain files or documents and you want to create a workflow that takes the users questions searches for the answer in those documents and then return the customized response to the user then you can use this chain we will look at the detailed usage of this chain in just a few minutes while going through the other AI nodes and finally we

have the summarization chain it transforms your text into concise summary we have the option to select the node input the input can come as a Json or binary data from the previous nodes or we can use the document loader document loader will load a document and provide it as input to this node then we have the chunking strategy we can keep it to simp simple or Advanced with simple chunking strategy we can select the number of characters per chunk and the chunk overlap if we select Advanced chunking strategy we get an option to add the

text splitter if we click on text splitter we get these three options we can split the document as per the number of characters or the tokens and here we can select the model for this summarization chain and this is how our simple summarization chain looks like now let's move to the most important part of this tutorial other AI nodes let's bring everything we have learned together by walking through a practical example a document question answering system we will use this diagram to First visualiz the workflow and explain each component involved this will help us to

understand how all these pieces fit together in an actual an workflow this Q&A system will have two parts first is upserting and second is saring upserting flow will process our files and make them searchable by our system in this flow first we will load our documents using a component called document loader after loading the documents we will split those documents into small chunks after splitting them we will use the embeddings model and convert those chunks into vectors and finally we will store those vectors in a database called Vector store now we can query this database

for a question and we can get the related text now in the second flow we have the chat flow where a user will ask a question using a chat component and then his query will be converted into a vector using the embeddings model and then the retriever component will do a similarity search of that query in the vector store once the system gets the similar results it will build the context and finally it will make a call to the L LM along with that context and llm will return a personalized response on the basis of

the entire context that it received and finally the response is returned back to the user now if we want to return the response in a particular format we can use output parsers now let's build this entire system in any 10 I'll start with the trigger component now let's say our documents are in Google Drive so I'll add a Google Drive node here to download the file from there and here we can select the credentials and the resource type is file and operation is download and then we can select the file from here now we need

to convert this document in vectors and store it in a vector store so I'll go here and in the vector stores I will select Pine con Vector store you can select any other Vector store as per your requirement here I can select the credentials and I want to insert the document so I'll will select insert documents and here we can select the pine con index now I'll connect Google drive to this node and here you can see we need to connect embedding and document so we need to connect the embedding model I will select embedding

open a you can select any one of these next we need to connect the document loader so we can click here and select one of the document loader from here we have two options default data loader and GitHub document loader default one will load the data from the previous step whereas GitHub document loader will use GitHub data as input to this chain I'll select the first one and here the type of data will be the binary I'll leave the data format as it is and now we see an option to add text splitter to this

node I'll click here and to keep this thing simple I'll select character text splitter you can split the file as per the characters or the tokens here I'll keep the chunk size as th000 and chunk overlap as 50 and this is how our workflow looks like now now if we look at our diagram you can see we had a document loader then we were splitting the text then we were converting those chunks into vectors using embeddings model and then finally storing it in the vector store and now if we look at this workflow here we

are downloading the file from Google Drive and passing it to this Pine con Vector node now this node has access to the embeddings and it has access to the document loader and document loader has access to the splitter so this document loader will receive the file downloaded from Google Drive and it will split as per this splitter node and this Pine con Vector store node will use this embedding mod model to convert it into vectors and store it in the pine cone Vector store now let's work on our second part of the system where a

user will start a chat he'll ask the question then we'll convert that question into vectors using the embedding model then we will use the retriever node to do a similarity search on the vector store that we just created then create a context out of it and make a call to the llm now to do all of that we will use question and answer chain and it will receive the prompt from the previous node so we can add a chat trigger to this and now we need to add the model and retriever so I'll select the

model as open a model and select the retriever from this list here the contextual compression retriever enhances efficiency by compressing the retrieved documents before passing them to the language model it identifies and retains only the most relevant information necessary to answer the query and this way it improves the response accuracy and speed the vector store retriever uses Vector embeddings to find documents that are semantically similar to The user's query the multiquery retriever enhances the retrieval process by generating multiple variations of the original query by doing this it is able to retrieve a broader range of

relevant documents from the vector store and in turn it improves the quality of the response and then we have workflow retriever now instead of fetching documents from some external database or vector store this node can pull information from the previous nodes or steps in your workflow in our case we want to use Vector store retriever now in this node we need to select the vector store from where we want to retrieve the information so I'll click here and select pine cone Vector store because that's where we stored our information and here this time we want

to retrieve the documents instead of inserting them and we need to attach the embedding model to this Pine con Vector store so we can click here and select the embedding model and this is how the second part of the system looks like now what we want to do is move this first part of the system where we are storing the documents in the vector store to a separate workflow and execute it once because we don't want to convert the same documents into vectors again again so we will execute this workflow only once and then this

workflow can be executed by using chat as the trigger so user will ask the question it will be passed on to the question answer chain then chain will use the vector store retriever to extract the similar content from the pine con Vector store and pine con Vector store will use the embeddings to convert the user query into vector and then search it in the vector store and then it will return the response back to the question answer chain then Q&A chain will put the response in the context and make a call to this openi model

to get the response and finally it will return the response back to the user and this is how all of these different nodes work together to build really powerful workflows we have seen all of these nodes in action in our Q&A system document loaders are used to load the documents memory stores the conversation history or state between interactions text splitter is a tool that splits large documents into smaller manageable chunks output parsers parses and structures the output from the language model if necessary they are used to extract specific information or format the response before sending

it back to the user retriever nodes enables us to find the most relevant document chunks based on the similarity between embeddings they are used to retrieve only the most relevant pieces of information to answer the user's question accurately embeddings generation is the process of converting text chunks into numerical vectors that capture the semantic meaning of the text embeddings allow us to perform similarity searches between the users questions and the document chunks Vector store is a database optimized for storing and retrieving embeddings it is used to store the embeddings so we can quickly find and retrieve

relevant document junks when a user asks a question and this pretty much sums up everything we have related to to AI in anyon by now you should have a solid understanding of anon's AI capabilities and feel confident in building your own AI workflows even without any technical background in upcoming videos we'll explore how to use any tenin production how to use external database and memory and how to update any1 version without losing the data we'll also build some really complex and interesting workflows in any1 if you found this video helpful please give it a thumbs

up and consider subscribing so that you don't miss out on future content feel free to leave comments and questions below I'd love to hear about the AI workflows you are building with anyon thank you for watching and see you in the next video