Great video buddy. Since you can do pretty much the same setup on Flowise, what would be deciding factors that would make you choose between Flowise or n8n?
Hi Rolando, that's a great question. There's some overlap in functionality. Flowise is AI native -- built ground up on AI use cases. n8n is built with automation as main use case with AI being added recently. Chat with your data is a use case that is common to both. But what you'll see in next week's tutorial is how easy it is to integrate into tools like Slack and later Discord. It's built into n8n, so those types of use cases are much easier. Flowise has strengths in the amount of integration in all the different types of vector stores, LLMs, service providers etc... There are use cases that one will use the other: for example, Flowise calling into n8n flows as tools to get access to Google tools like drive and docs. And n8n flows that call into Flowise to access all the different LLMs that Flowise has.
Hey Derek, first, thank you for that tutorial - gave me some pretty new ideas, how all that AI and storing data works, thank you! I have a question - it works as expected with the Question and Answer chain. Now I was actually trying, to use an AI Agent, to also use memory. But there I have no chance, to use the Vector Store Retreiver. Can you lead me into the direction, how to actually also use a similar workflow, but by using also memory? Because I instantly realised, that it will always answer pretty good on specific querstions, but when I ask something without being specific, I get an answer, that it doesnt know, what I am talking about ;) As an example I ask "How many Yachts are listet" - and it gives me the correct answer. But a follow question like "what is the cheapest?" will result in "what do you mean?" (I guess, because I didnt specifically talk about yachts) - and also, I would like to be able, to use the options how the AI has to respond etc. Thank you!
Derek, this tutorial is truly fantastic (helped me get a basic proof of concept in place for something I've been trying to build). Question for you: are you aware of a way for n8n to pass in other variables on upserting the data? My case is I'm doing a HTTP post to a webhook in n8n and it's indexing member content using this workflow. I added a field in the documents table of Supabase called "userid" and I'm trying to log that at the time of upsert but don't see a way to append additional info. Wondering if this is best accomplished by some type of trigger in supabase that tries to extract it from metadata... or more ideally there would be a way to append additional params from the insert step in n8n. Any suggestions?
Thanks Sean for your feedback! One easy approach for the use case you mentioned is to use a slightly different vector store QDrant. QDrant might be easier in this case because if an index is not created yet, it will automatically create it. So in your upserting flow, you can pass into the webhook the userid and have Qdrant create an index based on that id. When you add, you specific the user id again and it will add to that index.
@@derekcheung2598 thx so much. Will investigate Qdrant. I came up with a solution though I'm not sure it's ideal- I was able to inject the userid as metadata into Supabase and then I also stuffed it into the indexed content itself so that the LLM has the userid and I can query the LLM and return ID's via structured output parser. I can think of reasons why this approach is less ideal (seems more brittle and less deterministic, no surefire way to return all indexed content by userid, potential privacy issues with LLM as a downstream provider now having PID). But this is at least a semi-workable solution for now. Will check out Qdrant and see how hard to swap out the vector store. It sounds like the abstraction of LangChain should make that fairly straightforward which is great.
hello could you explain to me when and why to use this or that node and also explain to me and the workflows which are made available on the site, there are very few explanations and I do not understand not
It's a good question on scaling. First is also cost of vectorizing 100K pdf pages. It might get quite expensive. Second is performance, vector databases will vary quite a bit in terms of performance at scale, so you'll want to do your due diligence there. There's a lot of database options to pick from. You'll also want to add in meta data so that you can filter out your similarity search. In terms of customize chat interface, yes. In fact, there are different channels you can plug into: slack, discord, telegram, teams, whatsapp, and webpage embed.
Hi Derek, thanks for your video, and your explanation. Do you know if there is any way to embed the caht in wp site, or any html file? Thanks very much for your help
I see all the workflows that use Gdrive to download a single file, what about accessing the local fs and index/vectorize a directory full of files (for example code)?
hello again - anyway to stop the upsert reuploading the same documents every time you run this? Or is it standard older langchain where you just create a chat flow with retrieval instead? Too used to flowise! thanks And how can you scrape a website and add to the vectorDB?!
Thanks! I did notice a Whatsapp node in n8n. I haven't looked in detail, but it has a couple credential settings that I think you need to go through the Whatsapp for business application process to get an access token
Could I somehow use the assistant? My idea is to use pgvector to store the threads and then the assistant retrieves them in the context to be larger than the model's context capacity.
![How to Build Slack AI Chatbot with custom PDF in 15 mins with n8n [Part 2]](http://i.ytimg.com/vi/GRXWasQVxes/mqdefault.jpg)
![How to Build Slack AI Chatbot with custom PDF in 15 mins with n8n [Part 2]](/img/tr.png)
Brilliant this is exactly what I've been looking for. The potential for this is limitless how exciting!
Me too
Very useful! It'd be great to see a tutorial with Buildship. As always, thanks for sharing Derek!
Great video buddy. Since you can do pretty much the same setup on Flowise, what would be deciding factors that would make you choose between Flowise or n8n?
Hi Rolando, that's a great question. There's some overlap in functionality. Flowise is AI native -- built ground up on AI use cases. n8n is built with automation as main use case with AI being added recently.
Chat with your data is a use case that is common to both. But what you'll see in next week's tutorial is how easy it is to integrate into tools like Slack and later Discord. It's built into n8n, so those types of use cases are much easier.
Flowise has strengths in the amount of integration in all the different types of vector stores, LLMs, service providers etc...
There are use cases that one will use the other: for example, Flowise calling into n8n flows as tools to get access to Google tools like drive and docs. And n8n flows that call into Flowise to access all the different LLMs that Flowise has.
@@derekcheung2598 thanks for the reply.
thanks four your simple but valuables and helpful videos. Can you load multiples pdf but using N8N?
Hey Derek, first, thank you for that tutorial - gave me some pretty new ideas, how all that AI and storing data works, thank you!
I have a question - it works as expected with the Question and Answer chain.
Now I was actually trying, to use an AI Agent, to also use memory.
But there I have no chance, to use the Vector Store Retreiver.
Can you lead me into the direction, how to actually also use a similar workflow, but by using also memory?
Because I instantly realised, that it will always answer pretty good on specific querstions, but when I ask something without being specific, I get an answer, that it doesnt know, what I am talking about ;) As an example I ask "How many Yachts are listet" - and it gives me the correct answer. But a follow question like "what is the cheapest?" will result in "what do you mean?" (I guess, because I didnt specifically talk about yachts) - and also, I would like to be able, to use the options how the AI has to respond etc.
Thank you!
Hey, I am having the same problem, did you find a way to solve it?
Great video! Would I be able to use this workflow with my google sheets files?
Yup
Derek, this tutorial is truly fantastic (helped me get a basic proof of concept in place for something I've been trying to build). Question for you: are you aware of a way for n8n to pass in other variables on upserting the data? My case is I'm doing a HTTP post to a webhook in n8n and it's indexing member content using this workflow. I added a field in the documents table of Supabase called "userid" and I'm trying to log that at the time of upsert but don't see a way to append additional info. Wondering if this is best accomplished by some type of trigger in supabase that tries to extract it from metadata... or more ideally there would be a way to append additional params from the insert step in n8n. Any suggestions?
Thanks Sean for your feedback! One easy approach for the use case you mentioned is to use a slightly different vector store QDrant. QDrant might be easier in this case because if an index is not created yet, it will automatically create it. So in your upserting flow, you can pass into the webhook the userid and have Qdrant create an index based on that id. When you add, you specific the user id again and it will add to that index.
@@derekcheung2598 thx so much. Will investigate Qdrant. I came up with a solution though I'm not sure it's ideal- I was able to inject the userid as metadata into Supabase and then I also stuffed it into the indexed content itself so that the LLM has the userid and I can query the LLM and return ID's via structured output parser. I can think of reasons why this approach is less ideal (seems more brittle and less deterministic, no surefire way to return all indexed content by userid, potential privacy issues with LLM as a downstream provider now having PID). But this is at least a semi-workable solution for now. Will check out Qdrant and see how hard to swap out the vector store. It sounds like the abstraction of LangChain should make that fairly straightforward which is great.
hello could you explain to me when and why to use this or that node and also explain to me and the workflows which are made available on the site, there are very few explanations and I do not understand not
Excellent video!
Can we do this with 100k pdf pages?
And how can we customize the chat interface?
It's a good question on scaling. First is also cost of vectorizing 100K pdf pages. It might get quite expensive. Second is performance, vector databases will vary quite a bit in terms of performance at scale, so you'll want to do your due diligence there. There's a lot of database options to pick from. You'll also want to add in meta data so that you can filter out your similarity search.
In terms of customize chat interface, yes. In fact, there are different channels you can plug into: slack, discord, telegram, teams, whatsapp, and webpage embed.
Hi Derek, thanks for your video, and your explanation. Do you know if there is any way to embed the caht in wp site, or any html file? Thanks very much for your help
Yup, here is how: www.npmjs.com/package/@n8n/chat?activeTab=readme
I see all the workflows that use Gdrive to download a single file, what about accessing the local fs and index/vectorize a directory full of files (for example code)?
hello again - anyway to stop the upsert reuploading the same documents every time you run this? Or is it standard older langchain where you just create a chat flow with retrieval instead? Too used to flowise! thanks
And how can you scrape a website and add to the vectorDB?!
Great vid! Would love to see a WhatsApp n8n bot..it looks like there’s no direct WA API integration?? And it still needs twillio?? Thx
Thanks! I did notice a Whatsapp node in n8n. I haven't looked in detail, but it has a couple credential settings that I think you need to go through the Whatsapp for business application process to get an access token
Could I somehow use the assistant? My idea is to use pgvector to store the threads and then the assistant retrieves them in the context to be larger than the model's context capacity.
Hi Arthur, can you expand a bit more on your use case? What data are you thinking to store in your database
What about the reliability of the data?
Great video !Can you show me txt file ?Thanks