First, I don't speak English, so I apologize if this is poorly translated. Your video is by far the best one I've found about n8n. It's very clear and includes many details that others don't show. Please accept my humble congratulations, and thank you so much for your excellent work. I'm new to this and want to start a business to help me earn some extra money. Sending you a big hug and many thanks. Please keep making your videos!
@@leonvanzyl I wanted to share that thanks to you, I was able to implement my first functional RAG with n8n. I am a Spanish speaker, and I give instructions to the agent in English. At first, the agent was responding to the tool and not delivering the answer to the user, as if it was replying to itself, haha. Do you know why this could be happening? Thank you so much for your videos.
Thank you so much for this Leo! You got a like and subscriber 👍 I am stuck at 18:16 trying to share my bot publicly, because it says "This workflow has no trigger nodes that require activation". What did you do for that step? Thanks in advance!
Thank you for the like and sub 🙏. You need to run n8n in the cloud (either self hosting of their paid service) in order to share the charbot. If you are running it on the cloud then you might just have to check the settings on the chat node. There should be a toggle available to make it publicly available.
@@leonvanzyl Thank you for answer, but I have the same issue on n8n cloud trial - the same massage for activation and and this massage in Chat parameters "This node is where a manual chat workflow execution starts. To make one, go back to the canvas and click ‘Chat’", and no toggle.
@@oleksandrbabiy5737 Here is how you can get it to work: - Click the + icon to open the nodes panel - Search for chat... - Select Chat Trigger Replace the other chat with this and you should be good to go. HTH
@@oleksandrbabiy5737 You may have figured this out already, but I was having the same issue and it was extremely aggravating. For your trigger, do not select "On chat message". Instead, type "chat" in the search nodes field and "Chat Trigger" will be the first option. Use this node and you will have the ability to make it publicly available. If you check the settings of the original "On chat message" node, you will see that it is deprecated. Why it is still available as an option, I have no idea.
Great video, many thanks. A question: For me the response come back in a Jason format in the chat like: { "response": { "text": "The emergency contact is Jack Daniels." } } instead of plain text? If it is embedded or hosted, is the same. The Chat model response format is set to TEXT. Thank you!
Very good content! - I would add brief info that the problem for this type knowledge base (RAG) is/can be the limitation of database for this type ... when you want to add a lot of data there then you need to also take care the RAG retrieval - optimize and supposedly use prefix data - for example name of article, topic etc before every chunk and rag need to check if text is relevant at all. plus I would rather also focus on dekerized n8n usage as its more robust (can be) and of course It would be perfect to have your thought for n8n vs flowise vs others to know what to use and with which ind for what.
Did you use any guardrails on chats - setting public acces (especially on websie) can atract many who try to use "free public endpoint" just by asking own questions not from RAG
I typically try to instruct the model (via system prompting) to only answer questions about the business and nothing else. But you make a good point. Things like rate limiting and approved domains are important. I'll try and cover these in a video as well. Great comment 👍
Thank you it is very interesting. I’d like to clarify the costs of using embeddings with a chatbot. If I embed a set of documents (e.g., store them in Pinecone), I understand there is an initial cost for the embedding process. If these documents don’t change over time, do I need to re-embed them every time the chatbot is used? Or can I pay for embedding once, 'freeze' the embeddings, and reuse them indefinitely for queries without additional embedding costs?
Great tutorial. Taking your example further, could the N8N chatbot display images? Say the bot was asked "Show me a photo of today's special" and the dish's images were stored on Google Drive?
That's a great question! I actually managed to "accidentally" return images before! Since the responses are in markdown, it's possible to instruct the model to return image links as embedded images. It looks absolutely crap though and styling is all over the place 😀. There might be some bright spark in the community who knows how to make this work.
Great video 👍 Assuming the menu changes every month, what is the best way to handle the updating of the vector store? Is there a way to automate the deletion of the outdated data? Also, instead of documents, what if the context should come from the website which is refreshed regularly? What is the optimal way to handle this? Thank you!
Thank you for the feedback 🙏. I actually have a video on both Record Manager (which removes duplicates and stale data) as well as auto refreshing document stores. BUT, Flowise released a new API for handling the refresh side of things. I'll release a video on it in the next few days.
Hey Leon, thanks for your video. Just one thing... How can you manage my token usage? To be more specific... how can I extract the use of tokens during a conversation (I used Tools Agent .... practically the same as you)?
Hi Leon, Thank you for the awesome video, as usual. I have a question: for beginners without coding experience, which is easier to start building a RAG chatbot for my own documents between this n8n or flowise?
Thank you! I think both platforms make it really easy. The Document Store in Flowise is great if you've got many different data sources that need to be managed as a single knowledge base. Either way, both are great to learn if you're new to RAG.
Hi @Leon, thanks again for the valuable work you are doing, you got me hooked on AI. I would like to use Flowise to do customer care ticket analysis. My idea was to upload the tickets to Airtable, connect it to pinecone and then connect an Agent. but it seems that the agent is not able to analyze all the tickets and give me some overview information such as “what are the most frequent requests” or it is not able to calculate how many complaints there are as a percentage of the total. It seems to analyze more the individual index record...am I doing something wrong? even with n8n...the total number of tickets it can give me but it can't do analysis...how come? Thanks a lot in advance!!!
First off, I have to say that you do the best tutorials of any I find on TH-cam. Second, I have a question. I have a workflow similar to the one in this tutorial. Do you have any issues with the Google Drive trigger having trouble when you add a lot of documents to the watched folder at once? I find that it sometime submits the same file that was triggered earlier or it misses a lot of them. I see this trigger used in tutorials all over youtube so I'm either doing something wrong or most people only feed it a small number of documents at once. I'm curious if you've noticed anything similar.
This is great, thanks for posting. I only got snagged when trying to add a filename property from the metadata. I have a single Google Doc in my drive folder but when i run the workflow and access the schema, there is nothing to choose. The vector DB contains the doc contents. No errors. Just no metadata in the data loader. It says "No fields - item(s) exist, but they're empty"
Hi Leon thanks again for such a comprehensive tutorial! However when I try to activate this flow, I get "This workflow has no trigger nodes that require activation"
Your tutorial has been like a ray of light for me! Thank you so much. I have one question. The prompt I’m planning for my chatbot is quite long, but when I try to input it as it is, I get a message saying it’s too long and that I need to shorten it. Is there any way to extend the character limit for the prompt?
Hi Leon, amazing tutorials like always. Actually I always see it, but this is the first one I am doing it. Question How I reset the AI Assistance to start a new conversation? And not still having the last conversation on memory
Hi Leon, wonderful video. Thanks. I want to understand what will be the approximate costs involved for this setup and where all subscriptions will be required?
Check out the first two videos of my n8n playlist where I answer both questions 😊. In summary, you can self host for free, or pay like $7 to host it on Render. You could also pay for n8n's hosting if you don't want to deal with any of this.
Great video! I have installed n8n in a Docker container on my NAS using docker compose. Now I wanted to rebuild your chatbot, except that I wanted to use a local file (from my Macbook download folder). Unfortunately, I always fail when accessing the local file. I have already tried the entry “- /Users/hpr/Downloads/N8N:/files” under Volumes in docker-compose.yaml, unfortunately without success. Is it possible to show how this works in a video?
Great vid!! Could you please make a video about using openwebui as front end, and this n8n workflow as backend orchestrator? I'm having some issue...thanks! :)
Hi @leonvanzyl, great video. I have a question: how do I add an additional final step that saves me a record with user question and chatbot response to database? I have already connected the postgres database and when I write something, it actually saves the output. So so far so good, however then on the chatbot side it returns the postgres output to me, so the record I just entered ({id: 12, input: “test”, output: “something”}) And not just the llm output message. I put this item as an item linked to the “AI Agent” item. Maybe this is where the mistake is? I tried to move it but I don't understand where to put it. In other words, I want the chatbot to work exactly like yours, but in addition it saves my conversations to database. Thanks
Awesome stuff like always, many thanks! Using this structure, what would be the most efficient way to embed both documents from Google drive as you've shown here and a public website (for example an API documentation page that gets updated regularly) at the same time?
Hey there! Thanks for the feedback. n8n provides a couple of nodes for scraping website data, so you could have a workflow that runs on a schedule, extracts the content from those pages and then calls the Pinecone vector store, just like we did in this video.
@@leonvanzyl Thank you for your prompt reply, really appreciate it! Could I store both the scraped data and the google drive with the documents in the same pinecone vector store, or should I use separate vector store tools?
@@leonvanzyl Hi. I have done the tutorial and get one csv file into the vector store. I am not getting a Google Drive Trigger. It is set to check every minute but nothing happens. The workflow is set to active. I Drag another csv file into the folder no response. Anyone else having this issue ?
Great video Leon, really undecided between this and floiwise, both are cool. I think n8n might be easier. However, despite everything, I get the best results using Assistant API and with GPT-4o-mini it is affordable. Would love a tutorial using Assistant API on n8n
Thank you Leon, I loved the tutorial, I followed it step by step, writing down details, and I was finally able to understand part of the process. As for the final result of the chatbot, it was frustrating, maybe I did something wrong, but I had a simple document that informed the age range of children who could enroll in a school, from 4 to 11 years old. I repeated 10 variations of the same question with varied answers, always informing the same age range and when asked, the chatbot always answered that it was from 2 to 7 years old. The basic text was in Portuguese, but the clarity of the answers indicated that it understood the question perfectly, but it always got the answer wrong, repeatedly. Do you have any suggestions on what could be the cause?
It could very well be the language. Maybe you could try the same thing with an English data source, just to prove this? These models are small for a reason 😊
Can I resell n8n chatbots under the self hosted license? Can you do a video on how to remove the powered by n8n logo from Hosted Chat widget or using the embed Chat method?
Fantastic! I've been looking for a simple way to add a basic Q&A chatbot to our website to distill lots of Policy and Guidance notes for our students to refer to. The Open AI Assistant we were using worked well but with all the docs uploaded was costing several dollars per conversation whereas the same conversation is costing single digit cents through your approach! Our next step is attempting to use functions to fetch specific data from a student db based on match of their Google auth credentials. Not yet sure whether it's possible yet through n8n but would be great to see a tutorial on this as well. Many thanks for this Leon! 🙏
hi Leon wheni try to share the URL of the chat it say "his node is where a manual chat workflow execution starts. To make one, go back to the canvas and click ‘Chat’" how do you enable this as a trigger?
Nice, can you build full stack SaaS AI project using n8n, Nextjs, Supabase and Lemon Squeezy as subscription payment gateway? It's gonna be huge. Some projects like PhotoRoom Clone, ChatPRD Clone, with credit system for multi pricing plans will be awesome
Hey Leon, I followed along the tutorial, but i do seem to be having problems with the retrieval or the pinecone vector indexing. the model keeps saying it does not know, even though I give it the precise text that I have in my QandA document. Have you had this problem before where the retrieval is not correct. I have made sure my models are the same for both saving and retrieving in from the index, and the embeddings is 3-small. what else can i check here
Hi Marius, Retrieval can be quite tricky to get right, so don't get discouraged. The quality of the results can vary greatly on the vector store used, the embeddings model, the type of data, the text splitter used, the chunk sizes, etc. So it's a little bit tricky to give you guidance based on the comment alone. Try using one of the more advanced models for a start, like GPT-4o or Claude 3.5 Sonnet. Ensure that there's nothing in the system prompt that might force the model to say it doesn't know. Use a high quality vector database, like Pinecone. Create larger chunks (like 2000 chars with 200 char overlap).
How do i make a setup, where the Google trigger watches for fileUpdated - updates the vector store, but also deletes the previous records pertaining to that document? To avoid duplications?
How about creating an advanced chatbot that acts like an RPG game, with variables to keep progress, long term memory and rag to store information such as items list, world info?
Hey, what about returning the documents that it used for the answer? I have, for example, in the metadata the URL of the document, and I put it in the instructions, but it is not working, I see that the Pinecone store receives the metadata, but it looks like it is lost in Vector Store Stool, do you know how to do it?
Thank you for your great tutorials! I don't understand why, if I upload 3 files in my Google Drive folder, this agent only downloads the last one I uploaded, not all the 3 files.
Whoops. I mostly used this solution to add files ad-hoc, but I think you're right. If you load multiple files at once, the solution might only grab the last file uploaded. You could adjust the solution slightly to load all files in a given folder when there's a change, and clear the index before doing so. Darn, it's always hard to create a "one size fits all" solution in these tutorials without making them hours long 🤣.
Hello, I seem to always have a cors error when adding these chat bubbles onto my site. no matter if its n8n or Make or Flowise, i can never get past this cors error. Am i missing a step?
hey man im trying right now to do it step by step, and i have run into a problem witch says that i exceeded the plan, its because of the new plans they made with openai?
Hi Leon, I seen your AI video's, it's Awesome. I could like to connect with you regarding the project support. Please let me know you availability. So we can connect and discuss further.
@leonvanzyl Hi It can't make the vector data based for JSON file with around 100 records and with the small ones it makes but I can't find any info Thanks again 5:06
I like to combine the two actually. Use n8n for the automation functionally, and Flowise as the AI Builder. You can call Flowise from n8n using the HTTP node. I will be creating tutorials on this 👍
@@leonvanzyl I was using this approach, but calling Flowise via http needs around 5 to 6 seconds to come back with a response. I'm your opinion, are there really any benefits of using Flowise if you can set up your RAG flow using n8n?
n8n is a workflow automation tool, whereas Flowise is an AI App builder. You can use n8n to create workflows that are triggered by events, and it also offers simple integration with hundreds of 3rd party apps. n8n also offers some AI capabilities, but it's really not their focus point. Flowise on the other hands offers plenty of features for building super advanced AI workflows. It's extremely common to combine n8n and Flowise to build an end-to-end application.
sorry... another thing. Why sometimes the AI Agent didn´t use Tools (Vector DB)? I create the same that you but the AI Agent isn´t using the Vector DB. All my messages are being answered by GPT only
It's really a terrible platform for this. It abstracts too much. Gives you virtually no control over the serialization and tool selection. The prompts generally only work with OpenAI because of the fact that different models have different (and strict) formats for how they want to see the schemas.
Ok nicely presented. but oh how TRAGIC that there are SO MANY MANY MANY MANY STEPS (!!!!) in accomplishing something that fundamentally can and should be achieved in just a few clicks. ie "Here are the folders containing things to be vectorized, and to be refreshed regularly. NOW GO!". ok post implementation tuning would be expected.. but the complications in just getting started is ridiculous! Come on n8n.. you can do better than this!
You might be interested in my series on VectorShift then. It offers automation capabilities like n8n, but the Knowledge Base solution is exactly as you just described.
How are you planning to use AI Agents in your workflows?
Please remember to hit the like button 👍.
Thank you for the patience. You’ve earned another like and subscription.
First, I don't speak English, so I apologize if this is poorly translated. Your video is by far the best one I've found about n8n. It's very clear and includes many details that others don't show. Please accept my humble congratulations, and thank you so much for your excellent work. I'm new to this and want to start a business to help me earn some extra money. Sending you a big hug and many thanks. Please keep making your videos!
Thank you very much for the kind words! 🙏
@@leonvanzyl I wanted to share that thanks to you, I was able to implement my first functional RAG with n8n. I am a Spanish speaker, and I give instructions to the agent in English. At first, the agent was responding to the tool and not delivering the answer to the user, as if it was replying to itself, haha. Do you know why this could be happening? Thank you so much for your videos.
Love your tutorials Leon! So easy to follow! Packed with really usefull knowledge and no unnecessary chatting. Amazing work!
Much appreciated!
Thank you my man. This is fantastic. A concise and excellent step-by-step tutorial.
You're welcome 🤗
I didn't add the chat bot to a website but I was able to follow from start to finish. Great tutorial.
Great to hear 💪
we love youuuu leeon
Love is in the air it seems 😅
Thank you so much for this Leo! You got a like and subscriber 👍 I am stuck at 18:16 trying to share my bot publicly, because it says "This workflow has no trigger nodes that require activation". What did you do for that step? Thanks in advance!
Thank you for the like and sub 🙏.
You need to run n8n in the cloud (either self hosting of their paid service) in order to share the charbot.
If you are running it on the cloud then you might just have to check the settings on the chat node. There should be a toggle available to make it publicly available.
@@leonvanzyl Thank you for answer, but I have the same issue on n8n cloud trial - the same massage for activation and and this massage in Chat parameters "This node is where a manual chat workflow execution starts. To make one, go back to the canvas and click ‘Chat’", and no toggle.
@@oleksandrbabiy5737 Here is how you can get it to work:
- Click the + icon to open the nodes panel
- Search for chat...
- Select Chat Trigger
Replace the other chat with this and you should be good to go.
HTH
@@oleksandrbabiy5737 same here, did you end up figuring this out?
@@oleksandrbabiy5737 You may have figured this out already, but I was having the same issue and it was extremely aggravating. For your trigger, do not select "On chat message". Instead, type "chat" in the search nodes field and "Chat Trigger" will be the first option. Use this node and you will have the ability to make it publicly available. If you check the settings of the original "On chat message" node, you will see that it is deprecated. Why it is still available as an option, I have no idea.
Awesome content! Not enough people make information about n8n! Thanks so much!
You're welcome 🤗
Great videos 10/10. Straight to the point and providing value. Thanks Leon
Thank you 🙏
Thanks for your tutorial! Helped me a lot.
You're welcome 🤗
Great video, many thanks. A question: For me the response come back in a Jason format in the chat like: { "response": { "text": "The emergency contact is Jack Daniels." } } instead of plain text? If it is embedded or hosted, is the same. The Chat model response format is set to TEXT. Thank you!
Very nice video Leon. Keep going with AI and n8n please.
Thanks, will do!
As always, the best and most useful videos on youtube
Thank you 🙏.
Very good content! - I would add brief info that the problem for this type knowledge base (RAG) is/can be the limitation of database for this type ... when you want to add a lot of data there then you need to also take care the RAG retrieval - optimize and supposedly use prefix data - for example name of article, topic etc before every chunk and rag need to check if text is relevant at all. plus I would rather also focus on dekerized n8n usage as its more robust (can be) and of course It would be perfect to have your thought for n8n vs flowise vs others to know what to use and with which ind for what.
Congratz Leon! Nice Tutorial Video! 👏👏👏
You are a genius! Many thanks Leon
you are the best youtuber ever, thank you
Thanks 😁
This helped me get set-up. Thank you!!!
You're welcome
Did you use any guardrails on chats - setting public acces (especially on websie) can atract many who try to use "free public endpoint" just by asking own questions not from RAG
I typically try to instruct the model (via system prompting) to only answer questions about the business and nothing else.
But you make a good point. Things like rate limiting and approved domains are important. I'll try and cover these in a video as well.
Great comment 👍
Another awesone video Leon Thank you
Very welcome
Thank you it is very interesting.
I’d like to clarify the costs of using embeddings with a chatbot. If I embed a set of documents (e.g., store them in Pinecone), I understand there is an initial cost for the embedding process. If these documents don’t change over time, do I need to re-embed them every time the chatbot is used? Or can I pay for embedding once, 'freeze' the embeddings, and reuse them indefinitely for queries without additional embedding costs?
Amazing video man. I'd love more long videos on n8n and Flowise like this one!!!
Will do 👍
Great tutorial. Taking your example further, could the N8N chatbot display images? Say the bot was asked "Show me a photo of today's special" and the dish's images were stored on Google Drive?
That's a great question!
I actually managed to "accidentally" return images before!
Since the responses are in markdown, it's possible to instruct the model to return image links as embedded images.
It looks absolutely crap though and styling is all over the place 😀.
There might be some bright spark in the community who knows how to make this work.
Great video 👍
Assuming the menu changes every month, what is the best way to handle the updating of the vector store? Is there a way to automate the deletion of the outdated data?
Also, instead of documents, what if the context should come from the website which is refreshed regularly? What is the optimal way to handle this? Thank you!
Thank you for the feedback 🙏.
I actually have a video on both Record Manager (which removes duplicates and stale data) as well as auto refreshing document stores.
BUT, Flowise released a new API for handling the refresh side of things. I'll release a video on it in the next few days.
Really nice! Very thank for share 🙏
You're welcome 🤗
Hey Leon, thanks for your video. Just one thing... How can you manage my token usage? To be more specific... how can I extract the use of tokens during a conversation (I used Tools Agent .... practically the same as you)?
Hi Leon, Thank you for the awesome video, as usual. I have a question: for beginners without coding experience, which is easier to start building a RAG chatbot for my own documents between this n8n or flowise?
Thank you!
I think both platforms make it really easy.
The Document Store in Flowise is great if you've got many different data sources that need to be managed as a single knowledge base.
Either way, both are great to learn if you're new to RAG.
Nice work this is a super tedious task so it is really nice to have this step by step ❤
You're welcome 🤗
Hi @Leon, thanks again for the valuable work you are doing, you got me hooked on AI. I would like to use Flowise to do customer care ticket analysis. My idea was to upload the tickets to Airtable, connect it to pinecone and then connect an Agent. but it seems that the agent is not able to analyze all the tickets and give me some overview information such as “what are the most frequent requests” or it is not able to calculate how many complaints there are as a percentage of the total. It seems to analyze more the individual index record...am I doing something wrong? even with n8n...the total number of tickets it can give me but it can't do analysis...how come? Thanks a lot in advance!!!
I noticed that "OpenAI Assistants" are using chunk 800 with overlap 400 - which surprisingly works really well for larger datasheets with text.
Definitely try increasing the chunk size for larger data sets.
My sample was small for the tutorial, hence that small chunk size.
@@leonvanzyl yes, thank you - if I remember "k" value parameter - returns similar results (chunks) and then extracts the info?
First off, I have to say that you do the best tutorials of any I find on TH-cam. Second, I have a question. I have a workflow similar to the one in this tutorial. Do you have any issues with the Google Drive trigger having trouble when you add a lot of documents to the watched folder at once? I find that it sometime submits the same file that was triggered earlier or it misses a lot of them. I see this trigger used in tutorials all over youtube so I'm either doing something wrong or most people only feed it a small number of documents at once. I'm curious if you've noticed anything similar.
thank you bro for this very nice cocktail 🍹🍸 ❤❤❤
This is great, thanks for posting. I only got snagged when trying to add a filename property from the metadata. I have a single Google Doc in my drive folder but when i run the workflow and access the schema, there is nothing to choose. The vector DB contains the doc contents. No errors. Just no metadata in the data loader. It says "No fields - item(s) exist, but they're empty"
had this error aswell its a bug in n8n. try to drag and drop when you are not in Schema mode
Thank you for the good video.
I have a question to ask you.
I don't have an option in the message received node like you.
Do you know why?
Amazing videos, thanks for this work !
My pleasure!
Interesting. Is there a way so that the chatbot returns also the documents, or context, used for the answer?
@leonvanzyl - Your videos are great. Getting an error in Window buffer memory "No session ID found". Please help.
Very usefull!
Thank you for sharing 👍
My pleasure 😊
Hi Leon thanks again for such a comprehensive tutorial!
However when I try to activate this flow, I get "This workflow has no trigger nodes that require activation"
Thanks, Leon for a very useful video, as usual. I'm joining @mortu90 on this.
Your tutorial has been like a ray of light for me! Thank you so much.
I have one question. The prompt I’m planning for my chatbot is quite long, but when I try to input it as it is, I get a message saying it’s too long and that I need to shorten it. Is there any way to extend the character limit for the prompt?
Hi Leon, amazing tutorials like always. Actually I always see it, but this is the first one I am doing it. Question How I reset the AI Assistance to start a new conversation? And not still having the last conversation on memory
Great demo..thanks
You're welcome
Hi Leon, wonderful video. Thanks. I want to understand what will be the approximate costs involved for this setup and where all subscriptions will be required?
Check out the first two videos of my n8n playlist where I answer both questions 😊.
In summary, you can self host for free, or pay like $7 to host it on Render.
You could also pay for n8n's hosting if you don't want to deal with any of this.
Great video! I have installed n8n in a Docker container on my NAS using docker compose. Now I wanted to rebuild your chatbot, except that I wanted to use a local file (from my Macbook download folder). Unfortunately, I always fail when accessing the local file. I have already tried the entry “- /Users/hpr/Downloads/N8N:/files” under Volumes in docker-compose.yaml, unfortunately without success. Is it possible to show how this works in a video?
Unless you're technical, it can be very hard to work with docker and volumes. Check out my n8n getting started video and follow the setup instructions
Great vid!! Could you please make a video about using openwebui as front end, and this n8n workflow as backend orchestrator? I'm having some issue...thanks! :)
Hi @leonvanzyl, great video.
I have a question: how do I add an additional final step that saves me a record with user question and chatbot response to database?
I have already connected the postgres database and when I write something, it actually saves the output.
So so far so good, however then on the chatbot side it returns the postgres output to me, so the record I just entered ({id: 12, input: “test”, output: “something”})
And not just the llm output message.
I put this item as an item linked to the “AI Agent” item. Maybe this is where the mistake is? I tried to move it but I don't understand where to put it.
In other words, I want the chatbot to work exactly like yours, but in addition it saves my conversations to database.
Thanks
Hello, thanks for the tutorial - maybe you know Leon - why we didn't use ada for embedding - but "3-small" one .. is that one better?
Ada is an older model.
@@leonvanzyl Ok, I see. Thank you.
Awesome stuff like always, many thanks! Using this structure, what would be the most efficient way to embed both documents from Google drive as you've shown here and a public website (for example an API documentation page that gets updated regularly) at the same time?
Hey there! Thanks for the feedback.
n8n provides a couple of nodes for scraping website data, so you could have a workflow that runs on a schedule, extracts the content from those pages and then calls the Pinecone vector store, just like we did in this video.
@@leonvanzyl Thank you for your prompt reply, really appreciate it! Could I store both the scraped data and the google drive with the documents in the same pinecone vector store, or should I use separate vector store tools?
@@ramsesvandensteen6386 you can use the exact same Pinecone Index and Namespace 👍.
Hi Leon, would love a post on posting to Facebook and LinkedIn.. really struggling
It is amazing that n8n is free to use
Agreed
says starter pack is $20 a month. Am I missing a hack to get it for free ?
@profitpurpose
I think you are right. I was hoping there would be a free at least.
Great video. Referring to 18.27 to make the chat public, is there a way to do this on the self-hosted n8n on render?
Absolutely!
I'm self hosting on Render
High Leon. Thanks for another well planned video 😊. Does this automatically handling the update if documents?
In the video we are watching for new files, but you can change that you update if you wanted
@@leonvanzyl Hi. I have done the tutorial and get one csv file into the vector store. I am not getting a Google Drive Trigger. It is set to check every minute but nothing happens. The workflow is set to active. I Drag another csv file into the folder no response. Anyone else having this issue ?
Great video Leon, really undecided between this and floiwise, both are cool. I think n8n might be easier. However, despite everything, I get the best results using Assistant API and with GPT-4o-mini it is affordable. Would love a tutorial using Assistant API on n8n
Both are worthwhile to learn.
In fact, I like combining n8n and Flowise.
@@leonvanzyl agreed that is the way to go
Thank you Leon, I loved the tutorial, I followed it step by step, writing down details, and I was finally able to understand part of the process. As for the final result of the chatbot, it was frustrating, maybe I did something wrong, but I had a simple document that informed the age range of children who could enroll in a school, from 4 to 11 years old. I repeated 10 variations of the same question with varied answers, always informing the same age range and when asked, the chatbot always answered that it was from 2 to 7 years old. The basic text was in Portuguese, but the clarity of the answers indicated that it understood the question perfectly, but it always got the answer wrong, repeatedly. Do you have any suggestions on what could be the cause?
It could very well be the language. Maybe you could try the same thing with an English data source, just to prove this?
These models are small for a reason 😊
Hi, thanks for sharing your knowledge, Getting an error in Window buffer memory "No session ID found". Please help
Can I resell n8n chatbots under the self hosted license? Can you do a video on how to remove the powered by n8n logo from Hosted Chat widget or using the embed Chat method?
Fantastic! I've been looking for a simple way to add a basic Q&A chatbot to our website to distill lots of Policy and Guidance notes for our students to refer to. The Open AI Assistant we were using worked well but with all the docs uploaded was costing several dollars per conversation whereas the same conversation is costing single digit cents through your approach! Our next step is attempting to use functions to fetch specific data from a student db based on match of their Google auth credentials. Not yet sure whether it's possible yet through n8n but would be great to see a tutorial on this as well. Many thanks for this Leon! 🙏
Thanks Paul!
Thank you for your n8n videos. Great!!!
I know you are going step-by-step on difficulty but is it possible to have a video using custom JS code nodes?
Great idea. The JS code nodes are awesome. Will definitely create a video on them.
@@leonvanzyl Thank you🤩🥳
demasiado bueno, gracias
Thank you ❤️
Hi Leon, how would I do this with an S3 bucket, with multiple files please? The node options are a bit different to Gdrive.
hi Leon wheni try to share the URL of the chat it say "his node is where a manual chat workflow execution starts. To make one, go back to the canvas and click ‘Chat’" how do you enable this as a trigger?
Nice, can you build full stack SaaS AI project using n8n, Nextjs, Supabase and Lemon Squeezy as subscription payment gateway? It's gonna be huge. Some projects like PhotoRoom Clone, ChatPRD Clone, with credit system for multi pricing plans will be awesome
I have configured vector store tool as you said in the assistant work flow, but it is not retrieving any documents
The issue could be with the data / document that you're trying to upload?
Hey Leon, I followed along the tutorial, but i do seem to be having problems with the retrieval or the pinecone vector indexing. the model keeps saying it does not know, even though I give it the precise text that I have in my QandA document. Have you had this problem before where the retrieval is not correct. I have made sure my models are the same for both saving and retrieving in from the index, and the embeddings is 3-small. what else can i check here
Hi Marius,
Retrieval can be quite tricky to get right, so don't get discouraged.
The quality of the results can vary greatly on the vector store used, the embeddings model, the type of data, the text splitter used, the chunk sizes, etc.
So it's a little bit tricky to give you guidance based on the comment alone.
Try using one of the more advanced models for a start, like GPT-4o or Claude 3.5 Sonnet.
Ensure that there's nothing in the system prompt that might force the model to say it doesn't know.
Use a high quality vector database, like Pinecone.
Create larger chunks (like 2000 chars with 200 char overlap).
How do i make a setup, where the Google trigger watches for fileUpdated - updates the vector store, but also deletes the previous records pertaining to that document? To avoid duplications?
How about creating an advanced chatbot that acts like an RPG game, with variables to keep progress, long term memory and rag to store information such as items list, world info?
i just wanna say, i love you man. you are a fucking master!
I tried other free AI models, but none of them worked. Will the OpenAI Plus plan be sufficient?
Hey, what about returning the documents that it used for the answer? I have, for example, in the metadata the URL of the document, and I put it in the instructions, but it is not working, I see that the Pinecone store receives the metadata, but it looks like it is lost in Vector Store Stool, do you know how to do it?
Hello Leon, can you share the Word Doc and menu, then I can try to buy the same RAG Chatbot
Hey man, I actually simply used ChatGPT to generate a QA document for a fake restaurant 😊
Hey man, I actually simply used ChatGPT to generate a QA document for a fake restaurant 😊
@@leonvanzyl good idea
Thank you for your great tutorials! I don't understand why, if I upload 3 files in my Google Drive folder, this agent only downloads the last one I uploaded, not all the 3 files.
Whoops. I mostly used this solution to add files ad-hoc, but I think you're right. If you load multiple files at once, the solution might only grab the last file uploaded.
You could adjust the solution slightly to load all files in a given folder when there's a change, and clear the index before doing so.
Darn, it's always hard to create a "one size fits all" solution in these tutorials without making them hours long 🤣.
Hello, I seem to always have a cors error when adding these chat bubbles onto my site. no matter if its n8n or Make or Flowise, i can never get past this cors error. Am i missing a step?
Can we make RAG Voicebot for cold calling in n8n?
hey man im trying right now to do it step by step, and i have run into a problem witch says that i exceeded the plan, its because of the new plans they made with openai?
I'm guessing you're using OpenAI. You might be out of credits
Hi Leon,
I seen your AI video's, it's Awesome.
I could like to connect with you regarding the project support. Please let me know you availability. So we can connect and discuss further.
Hi, could you please make a rag for JSON file , I tried to find any help on how to chat with JSON file , but there's nothing, thanks again
Ok, will see what I can do.
Does the recursive character splitter not work for you? What's the issue that you're experiencing?
@leonvanzyl
Hi
It can't make the vector data based for JSON file with around 100 records and with the small ones it makes but I can't find any info
Thanks again 5:06
Which is better, using N8N with Flowise or just using N8N?
I have the same question, should we combine flowise with n8n or n8n can replace and expand flowise ???
I like to combine the two actually.
Use n8n for the automation functionally, and Flowise as the AI Builder.
You can call Flowise from n8n using the HTTP node.
I will be creating tutorials on this 👍
@@leonvanzyl I was using this approach, but calling Flowise via http needs around 5 to 6 seconds to come back with a response.
I'm your opinion, are there really any benefits of using Flowise if you can set up your RAG flow using n8n?
Whats difference between n8n and flowise?
n8n is a workflow automation tool, whereas Flowise is an AI App builder.
You can use n8n to create workflows that are triggered by events, and it also offers simple integration with hundreds of 3rd party apps.
n8n also offers some AI capabilities, but it's really not their focus point.
Flowise on the other hands offers plenty of features for building super advanced AI workflows.
It's extremely common to combine n8n and Flowise to build an end-to-end application.
sorry... another thing. Why sometimes the AI Agent didn´t use Tools (Vector DB)? I create the same that you but the AI Agent isn´t using the Vector DB. All my messages are being answered by GPT only
wow!!!
Hm, the prblem could be that upsert happends every tiem - and vector store just keep growing ..
The document will only be added once to the Vector Store, when the file is added to the Google Drive.
@@leonvanzyl yes, I was testing "when the file is updated" - and then it is upserted again. I wonder how we would solve , if file is updated.
Thank you so much.....
First!
Darn, you even beat me 😝
Hello Sir
I am from Pakistan
Please mention that is it free to use all these or wo have to buy anything, like keys?
I do mention this in the video. I'm using OPENAI, which is not free (but super cheap).
@@leonvanzyl sir you answered me, it is really awesome for me.
Please use some which is free for practice, as students cannot afford.
aitutorialmaker AI fixes this (AI driven Tutorials). Create RAG chatbot using n8n
Fixes what?
noticed that the ai are not consistent in repeated runs probably openai is the only perfect one for it
I noticed if answers are not consistent - you can use larger chunks - try 800 chunks with overlap 400.
It's really a terrible platform for this. It abstracts too much. Gives you virtually no control over the serialization and tool selection. The prompts generally only work with OpenAI because of the fact that different models have different (and strict) formats for how they want to see the schemas.
You should try Flowise if you want more control of the AI side of things.
I like using Flowise for the AI logic, and n8n for the automation bits.
Ok nicely presented. but oh how TRAGIC that there are SO MANY MANY MANY MANY STEPS (!!!!) in accomplishing something that fundamentally can and should be achieved in just a few clicks. ie "Here are the folders containing things to be vectorized, and to be refreshed regularly. NOW GO!". ok post implementation tuning would be expected.. but the complications in just getting started is ridiculous! Come on n8n.. you can do better than this!
You might be interested in my series on VectorShift then. It offers automation capabilities like n8n, but the Knowledge Base solution is exactly as you just described.