Відео

Thanks for the kind words! I'll be doing a video on the WhatsApp trigger and response setup in the next week or two and we have more RAG / Knowledgeable content coming out for this type of WhatsApp chatbot use case so stay tuned!

@TheAIAutomators so thanks dear

Thanks ... And great question! I have three concerns over the Flowise Multi-agents. For refererence, here is the Flowise supervisor prompt "You are a supervisor tasked with managing a conversation between the following workers: {team_members}. Given the following user request, respond with the worker to act next. Each worker will perform a task and respond with their results and status. When finished, respond with FINISH. Select strategically to minimize the number of steps taken." The first is that it can't handle standard greetings. This is a weird one, but basically the Superviser agent is a master delegator. Even though it is an LLM, it can't actually respond itself without delegating the task. So in response to a "Hello, how are you?" - it delegates the task of forumlating a response to its various workers. And it ends up in this strange dialogue with its various workers, where theres a debate about how to respond to this!!! Eventually it surfaces a response and sometimes that response is not an answer to the original question, but to a question asked by a worker agent! Weird! - I tried to have a "Greetings & Small Talk" worker, to handle this but that didnt' really work! So it's not ideal for a conversational style interface like what you'd have in Whatsapp. The second is that the supervisor is just that ... a supervisor. It's not really a director. As you can see from the system prompt, its the workers that are having the dialog between themselves and the supervisor is faciliating it and deciding which worker agent to call next. I found that the amount of back and forth between the worker agents, through the supervisor can be extreme for simple use cases. There's basically too much dialog and too much delegation! There's also an element of trying to do everything possible to satisfy the users query - which can result in workers, calling every tool. The third is I hit a bug where the system glitched and triggers a tool 150 times in response to a question and crashed the Flowise instance. I got from that, that the Multi-agent features may be a little on the Beta side! I'd say this system is brilliant for certain use cases that require detailed chain of thought style back and forths between Workers to reach a conclusion. However for a production grade, converstaional style multi-agent system, I've found using Flowise Assistants (calling other Assistants via Tools) or N8N Agents (calling other Agents via Tools) to be more reliable. Both of these are as good as each other. The only reason I didn't use Flowise Assistants to build this is that the Cloud version has a 60 second timeout on tool calling. So you'd need to run your own server and change the Nginx settings to extend that timeout to allow the sub-agents to do what they need to do.

@@TheAIAutomators Wow thank you so much i really appreciate the detailed breakdown.Your insights are incredibly helpful. Keep up the great work

You're very welcome!

@@TheAIAutomators Do you have a LinkedIn account by chance?

That's a great use case. I did a news automation video before ua-cam.com/video/d2MK-xYGSQw/v-deo.html so you could set this up on a daily basis to run and gather data from lots of different local news sources and trigger the director agent via a webhook to kickoff the interaction!

Glad you found the video useful! The easiest way would be to use a router and just duplicate the flow and add filters for your selected board in airtable. That could get clunky if there are a lot of boards. The most efficient way would be to map the values directly to the one Pinterest module within Airtable. It should be possible to do this. Check the "map" button next to Boards on the Pinterest. Pinterest is expecting an ID. If you uncheck the map button, then select your board, then check the map button again, then you should see the ID for that board (that's a lazy way of getting the IDs - you could also get them via API probably). You could have those IDs as in the single select but that won't be very user friendly, so you could alternatively create a separate table for Boards and just have two columns (Name, ID) and then link those to your Pin records (kind of like how I linked the records and got their values in this automation on our second channel: ua-cam.com/video/4sv-GlnaBKk/v-deo.html). Hope this helps - Alan

@@TheAIAutomators Alan, thanks for the reply. Yes, this helps, although I have 42 boards, so it would get clunky... Not sure how it would work with the 'map', I understand how to get all the ID's and I could figure out how to match them to 'board names' in Airtable, but the problem remains, that from the pinterest module, you can't 'read' the airtable info for 'boards'. So for now, I think I will stick to the cloning, and maybe make 3 identical scenarios, with just different boards and filters and see how I can make it work that way... It should also work with the Pinterest API, but unfortunately, the original email address I used to sign up to Pinterest no longer exists, and they send the verification email to that address, eventhough I changed the address in my business account. So can't use the API (for now). Anyway, I learned alot in a short time and appreciate your reply. Thanks again!

You're very welcome! If you've any questions on our community feel free to reach out!

Absolutely! I was lazy cause I already had a lot of the tools built in Make.com!

@ nice ill try to replicate all in n8n. Wish n8n gives more features for devs in selfhosting

I don't think theres much difference between community and cloud. But yeah, more dev features would be nice across the board

Hi, I just responded to your email. I don't see an account under your name or email address within our community. If you want to sign up for membership, you can do it via the link in the video description. Let us know if you run into any issues. Thanks! Alan

Yeah that's a great idea! One of the main considerations I'd have is timeout time and how to gracefully handle hitting these. For N8N cloud, on the starter plan the max execution of a workflow is 5 minutes. On the pro plan it's 40 minutes and the enterprise is unlimited. If you're self hosting you can set it to whatever you like. So it you have lots of nested agents as tools, all themselves executing different tools, then you could easily go over the 5 minute limit. But at least there's a way around that (upgrade or self host). The standard timeout for the HTTP Request node is 5 minutes. I haven't yet hit a timeout on the AI agent tools execution so I'll definitely do more testing on that on my self hosted version to see where the limits are (if any)

For the images, if you're constructing correct <img> tags within the HTML there's no reason why they wouldn't show up - unless your Worpdress theme or plugins are doing something weird. It wouldn't make any difference what image model you're using - as long as it's a valid image it should work. Check the HTML output within the automation that you're passing to Wordpress to see if you're seeing correct img tags. Also, btw, either use the Auto Upload Images plugin within Wordpress or adapt the scenario to upload the images directly within this scenario by using this tutorial: ua-cam.com/video/P58OvNdpPYk/v-deo.html. For the youtube URLs, generally on Wordpress if you pass a plaintext UA-cam URL it will auto-embed. This should happen if you insert a URL manually within a post on your site. If not, then maybe there's something messing with the auto-embed feature. Alternatively, you could construct an iFrame or the youtube URL similar to how I did it in the shopify version of this automation on our second channel: ua-cam.com/video/YWgxlYTl0Og/v-deo.html Hope this helps - Alan

@@TheAIAutomators Thanks for the reply! but still I am not able to get the generated images in the blogpost. This is the data passed into the wordpress create media module : " {"detail": "Cannot access application 'fal-ai/flux'. Authentication is required to access this application."}". What do you think might be the problem?

Great question... A few reasons. Timeouts - I believe there's a 3 minute timeout with open AI tools calling so calling multiple subagents sequentially would be a problem if they had any decent work to do themselves with their tools. Subagent wrappers - you'd need to wrap the subagents in make.com scenarios to have them callable by the supervisor agent. Not a deal breaker, just clunky. Limitations on the OpenAI Assistants in Make - the maximum number of sequential calls an OpenAI assistant can trigger in make.com is 2. This limitation is with the make.com implemention.... Not with OpenAI assistant itself ... But it's a problem

Oh yeah - apologies, I didn't show the actual settings part but you can see at 13:54 that we're passing the article array. For the iterator, I've enabled the "map" setting, and I pass in the "article" item from the parsed JSON outline (module 59 in the scenario), the code for it is {{59.article}} in my scenario. You can then see the settings the text aggregator at 24:09 in the video which closes off the loop.

@@TheAIAutomators i did that and now it’s iterating the first section heading multiple times….. any solution to this problem ?

if you follow the prompting within module 56 then it should output an array, if you select that array within the iterator then it should be iterating through each of those items rather than just looping through the first one. Within write article section (module 60) make sure you're selecting the values from the iterator, not from module 56. Mistakenly selecting data from module 56 instead of the iterator is the only reason why I can think that it's iterating through the first section every time. When you're running this, if the number of iterations matches the number of sections you have in your outline then you've likely set up your iterator correctly, it's just you're probably mapping the modules after that incorrectly. If you're still having issues after that, maybe read up a bit more about how iterators work and mapping within them as you should be able to fix the issue fairly quickly once you understand the concepts of those a bit more. Hope this helps!

Hi, yes we have a walkthrough video of how to create a slack app and integrate via webhooks within the Community. We previously used to use the "New event" module within Make.com slack App, which was an instant trigger that would wait for messages, but this was buggy, stopped working, then Make removed the option entirely, so creating a Slack app is the way to do it now. Swapping out Slack instead of Telegram should be pretty straightforward in these scenarios. We don't currently have other slack specific blueprints that use slack apps as mentioned above, but it's definitely on our to do list, particularly we're going to be creating a lot more AI agent content - Alan

Apify is definitely a better fit where they have pre-built scrapers already available. We use Apify a lot on this channel and love it! The problem is if there isn't a pre built scraper on Apify for the website you want to scrape, then you'd need to code your own in Python or JS (and deploy it to Apify ... Or run locally). So the option in this video, along with platforms like Browserflow and Axiom is to provide a No-Code means of doing this. Of course if you're a developer then it definitely makes sense to just code it!

You can just send plain text in the webhook response by itself. So the image_url for the image tool. Or for the research, then the text response from Perplexity. You can also send escaped JSON either via the "create JSON" module or by using the nested formulas as I show in this video on our other channel: ua-cam.com/video/18wgBZxOp0U/v-deo.htmlsi=HBXQsV3HGFTrOduw

Exactly, you’ll need to split the content into smaller chunks and test different chunk sizes to see what works best for retrieval accuracy. I'd start with a few paragraphs as it gives enough context in each chunk without losing too much meaning. 1 sentence is probably too granular and 1 page might be too broad and miss detail. The embedding model you use-like the one I use in this video, has 1,536 dimensions ... So it'll always generate 1,536 numbers to represent the text, no matter the chunk size. So, it’s all about finding that sweet spot where the vector can represent the text with the greatest meaning. For vector IDs, you could go with docID_pageNumber_chunkIndex. It keeps things organized and makes it easy to trace back to the original section of the document. You could even load the full page into an LLM later if you need additional context that wasn’t captured by the specific vector that was retrieved. You can also overlap your chunks slightly-like including the last sentence of one chunk at the start of the next. This comes out of the box in Langchain with the likes of Flowise. Also it's good to store metadata, like docID, pageNumber, or sectionName, alongside the vector ID in Pinecone. This makes it easier to filter and query your data later. Plenty to work with there!

Thanks Philipp, glad it was helpful! I really like your channel btw - some great stuff on there. Yeah Make is great for building out the flows for agent tools in this way - it's like a micro-services architecture, and a nicely cross-compatible approach, as you can call the scenario webhooks from Flowise, Relevance, etc! I think it's a good approach particularly as this is such a fast changing field, so we don't have to rebuild all the tools if we want to jump to a different agent building platform later! - Alan

You could, but the Buildship + Google sheets batch update approach I'm using here is just waaaaay more efficient on both operations, and far less manual effort for creating other similar reports. You could use the make.com DataForSEO native integration and then pass the data to Buildship instead of querying the data within Buildship, you'd just need to add a transform to JSON module to create a JSON string of all the data so you can send it to Buildship.

Hi! Yes, I go through most of the process in this video here. (I actually used a hybrid of exports from the platform and this) ua-cam.com/video/A1wmrpaCiWA/v-deo.html Organizing the data is as time consuming as setting up the RAG system unfortunately! Garbage in garbage out

Are you definitely passing the UA-cam URLs into the "choose video" prompt? If so, have a look through the execution history to make sure this is happening. We're passing in the previously chosen UA-cam URLs here so that the LLM will exclude those when picking the next URL. Also, variable modules are quite counter-intuitive in Make.com, so make sure you're getting/setting variables correctly and choosing the correct one.

@@TheAIAutomators yeh have everything copied exactly, but still wont remove duplicates, no doubt something super minor. I think I definitely need some more focused coaching. Are you pretty quick to respond in your community? Happy to join, i am part of 2 other paid skool communities but find they jsut dont provide enough support to troubleshoot and resolve issues.

Yeah we provide great support in our community. It's super important for us that we can be responsive to members and resolve issues. We know that our members have businesses to run and clients to serve! If we ever feel like our support standards are starting to dip, we'll be closing the doors to new members and operate a waiting list. We have a money back guarantee in case you're not satisfied, but I'm confident you'll be happy with our support! Alan

absolutely, PDF.co responds with a list of invoice lines as an array, so if you use an iterator to iterate through those, you'll be able to output those to the Google Sheet. You can get GPT4o to do the same but respond as a JSONList and then iterate through those. It's a bit more advanced and might be a bit more prone to errors in the data, but it's very doable.

You're very welcome!

Thanks! Good question - Yes it should be straightforward enough. If using WooCommerce then products are stored as custom post types so it might work directly from the wordpress module as I'm doing it in the video (just select your product post type if it appears there) as that would likely update the product description. It would probably be better straightforward to update the product name too!

@@TheAIAutomators Thanks bro!

Thanks so much!!!

It's up to you! You could load the full HTML but that'll be a lot of noise and would probably result in poor quality retrieval. You could load the markdown of the page and strip out all irrelevant HTML code. You could get an LLM to summarize the page and index that. Or you could create a custom scraper for structured data extraction and index that. It's really up to you and will need to be specific to your exact use case. Chunk size is also a consideration, one vector to one page or multiple vectors to one page. For long pages, a single vector might not have the depth of meaning to fully represent it so maybe smaller chunks are needed. Trial and error is key really.