Hi all,
I’d like to do the following:
Is this possible, and what’s the best mechanism to do this completely within n8n?
Thanks,
/mike
Hello @trigrman welcome to the 
community!
Sure, that’s possible.
The first two steps would be a workflow that starts with the form (as you mentioned, the user fills it in), then it gets analyzed by the agent, and finally the user is directed to the chat link, at that point a second workflow begins..
But if I think about it, I’d actually keep it all within a single workflow and just start directly with the chat (for example, on Telegram).
It could begin with a welcome message, then the user fills in the form, and continues within the same chat.
That way everything stays in one place and within one workflow.
Thanks @mohamed3nan, I appreciate knowing it’s possible, but it’s the actual execution of this that’s difficult, and that’s what I need help with.
My specific questions:
I’ll have more, but these seem to be the most pressing questions. Thanks.
/mike
Hi @trigrman,
For the first question, since you’re not planning to use Telegram or anything similar, you’ve already answered yourself in the second question.
You’re saving the form data in the database with session ID then redirecting the user to the chat link with that session ID as a parameter.
The next step in the workflow would be to get the form data based on session ID and then provide it to the agent as part of the context or the prompt..
Side note: keep in mind that the link is exposed and potentially vulnerable, since any user could modify the session ID and start a chat with the agent. You may want to secure it.
I guess my current challenge is getting caught up in the particulars of how agents and chats operate, and it seems difficult to make them behave seamlessly with the form. I would like the user experience to be:
I’m struggling to have the chat start this way after the end of the form–it’s not at all clear what the necessary combination of settings are in the chat trigger, agent, and/or response nodes to enable this. Right now it just sits there until the user types something.
Where should the query node go to retrieve the saved form content–between the trigger node and the agent? Or should this be a tool node instead?
What triggers the chat to end once the agent thinks it has enough info, so that the workflow can continue? How do I connect the agent to the next step in the workflow?
/mike
Also, it appears that I cannot actually pass data to a hosted chat in the URL and have that be accessible downstream from the chat trigger node so that I can retrieve the form with a query. Any ideas how to transition seamlessly from a form to a chat, while also passing session-specific data between them?
/mike
Let’s clarify a few things regarding the chat node.
There are two modes:
Hosted mode:
You won’t be able to pass parameters here, so you’ll need to ask the user for the info and proccess it after that.. (Feature request)
Embedded mode:
You can pass parameters in this way:
Is there a way to pass query parameters in the Chat trigger? - #4 by LucBerge
Now, if you’re using the n8n Hosted Chat, which I guess, you’ll need to ask the user for their session ID that you created from the form, then you can use it to load the form data from the database and use that in the following nodes.
Since you want this with “no open discussion”, you should keep the flow responding through the Respond to Chat node until the AI agent steps..
Once the AI agent sends a reply, you pass that reply back to the user and continue in the same way.
I think the missing piece here is the Respond to Chat node, along with setting the Chat Trigger node option Response Mode to “Using Response Nodes”
here is an example of what i mean:
Actually, I’m now embedding the chat in a web page with a webhook per the embedded mode instructions you gave above.
So, I should be able to pass the session ID to the chat now. However, I’m having trouble configuring the chat page, as there’s no entry field showing up for the user to type anything. The prompts coming back from my agent are not correct, either.
What’s really difficult here is testing–there are multiple disconnected workflows and I can’t look at the data in test mode because half my workflow doesn’t run unless the workflow is active. I can’t see what it’s doing when it’s not behaving as expected (which right now is all the time).
Right now it looks like this:
/mike
Hi @trigrman
Yes, you need to set the workflow to production because there are multiple triggers in the same workflow, and this will also ensure everything is running correctly.
Now, from the screenshot you sent , since you’re following the mentioned workaround, something is still missing which is (the HTML generation)
in that HTML template you’ll need to set:
also I’ve updated the lib links and also enabled automatic reading of the meta data.
You can use this workflow directly:
Note (to make the logic clear):
At the end of the form, you’ll send the user the webhook URL (which will have the query parameters) that is generating the HTML page that contain the chat that will triggered and you will find the parameters under the metadata..
Hope this helps, let me know if it’s not clear.
Actually, I put the HTML directly in the webhook response node, and that works fine. But I tried yours, also, and they behave the same.
However, neither version gives me a working chat, so something is wrong in my chat configuration. I don’t get a working chat window because there is no user entry field for some reason–I’m not sure why that’s not there.
Also, I cannot get the agent to do its thing and analyze the text before the user types anything. Is there a way to not prompt the user for input until the agent works on the content first, and then only prompts for input once that’s done? I’m just stuck with the initial prompt from the chat trigger, which frankly I never want the user to see.
/mike
Make sure to activate the workflow. Also, if you share the workflow, it will be much easier to help.
You can add the agent node (or a simple AI call) to the form workflow and pass whatever you want as parameters, for example..
I’ll need to prep some things before I can share the workflow, but will do so shortly.
Your suggestion to add the agent to the form workflow is exactly what I’m trying to do. Are you suggesting instead to split the agent into an analysis part and an interactive part? That might make sense to do, but I was hoping to test a single agent for this to see whether it was required, and I’ve never gotten it to work so that I could do this.
If I were to do that, the interactive agent needs access to the first agent’s analysis of what is missing. And frankly a 3rd agent might then be needed to rebuild the answers into the form structure. Can you suggest a workflow that would do all this? Note that my problems getting the interactive agent working seamlessly with the form entry would still exist.
/mike
Here’s my workflow below. Just so that I could test the agent part of the workflow easily, I’ve set the chat to “hosted” and have hardwired the select statement to retrieve a known form entry to feed into the agent. Here are my current questions/issues:
I know that’s lots of questions–I’ve tried asking LLMs these things and I don’t think any of them know n8n well enough to give truly valid answers.
/mike
Hi @trigrman thanks for sharing the workflow!
Let’s stick with one method “embedded mode” to keep things clear and avoid confusion with each other : )
I noticed some broken links in your workflow, it looks like you didn’t copy the last sample I uploaded.
As far as I know, this isn’t possible because the chat “Trigger” needs a user input to fire. Instead, you can control the initialMessages in the HTML, So you’ll probably need to add an AI node within the form submission save the AI output in the database then inject in again.
It’s fine this way since you’re just pulling form data based on the parameter and then load the data in the memory.
Yes, that happens because everything isn’t in a single flow, there’s the form and then the chat that depends on it. You’ll need to add another column in the database (for example AiOutput) and inject that both into the HTML start messages and into the memory, along with the system and user prompts, so it can be reused in the chat.
I think this will require some careful planning of the database and workflow design..
Initially, you can use the “Chat Memory Manager” to extract all responses and process them in another workflow that outputs a structured result in the format you need.
For now, I just edited the workflow so it works, even though it still needs a lot of optimization and probably better handling of memory/database,
Just to note: everything is running in production now, so that should remove any confusion with the links, BTW here’s the form you can test, I’ll keep this link active until you confirm that everything’s working for you..
Thanks for this–helpful to see your changes. I have tried your link and it fails for me with this error:
So, I also made my own copy and fixed a bunch of things, and put the URLs back to point to my own machine (production mode). My chat still does not work–I’m not getting a user entry field so I can respond to the chat.
Normally I’m a great troubleshooter, but I find tracing production runs of these workflows difficult–it doesn’t seem to record the progress of any entire run when the workflow has several disconnected flows inside it. I cannot see what’s happening at all in the chat portion of the workflow, for example–any tips on getting access to the logs there?
/mike
The model is overloaded. Please try again later
I think I reached the model limit, I switched to Flash 2.5, so it should work now.
That’s why I shared a live link, it’s actually working with the same workflow I shared above, After you submit the form, you’re automatically redirected to the chat like this
Yeah, this is really odd. I don’t get the same “redirect” message when I run it on my own machine, it just jumps into a chat, minus the “Type your question” prompt. I don’t even have that portion of the page showing up. I literally just copied your code and changed the URLs, I didn’t change anything else about the chat interaction, so not sure what’s causing the issue. Any ideas?
What do you get exactly?
Here is the chat from your last entry:
Can you try opening the link in an incognito window?
Oddly, I don’t get the “redirecting” message when running on my own box, it just goes straight from the form to the chat window. Perhaps that’s just due to network latency, which is non-existent in my case.
Anyway, my chat window looks like this when run using the production URLs all around (none are https in my case):
Yes, I updated some of the text fields, but I left the chat workflow you provided the same. When I check the Executions tab to look at the production run, it has logged the Form workflow, and the Webhook workflow, but NOT the chat workflow, so I can’t look at that to see what’s happening. I’ve double-checked the URLs and all is correct, so I’m starting to wonder if the lack of SSL is a contributing factor, as a bunch of the chat interface is simply not rendering.
/mike
