🙏 Help me please...

[swift echo]

Hello everyone! I hope you're doing well. I've been building voice agents using Retell and Make.com (optionally n8n).

Yesterday, I was assigned a new task to build an email-based bot.

After an in-depth discussion, we decided to use n8n for the email-based bot.

However, I only have limited experience with n8n, and I'm not sure how to approach this. Could you please advise? 🙏

[prisma trout]

Hi @swift echo

[balmy temple]

@swift echo

You’re on the right track using n8n
A typical setup would be Email trigger (IMAP/Gmail) → parsing & IF logic → AI (OpenAI) → optional memory/DB → generate reply → send response.

If you want, I can help you design the workflow end-to-end or share a working example to get you moving faster.

[runic moth]

For n8n, a good approach is email → intent → action → reply.

Use an IMAP/Gmail trigger, clean the email, send it to an LLM for intent classification, then route with IF/Switch nodes and reply via SMTP/Gmail.

Biggest difference from voice bots is state + threading, so storing context in static data or a DB helps a lot.

If you want, I’m happy to walk you through the setup or help you get it done.

[sly ibex]

sounds interesting
if you want i can guide you with my experience and also can show what i have made

[swift echo]

Thank you for your messages, 🙏

The main problem is one: I used KB(knowledge base - website or files) when building voice agent with retell. But I'd like to know if I can use the similar feature like KB on n8n.

Actually, this bot should be able to response based off the content of website or some files.

[cosmic lichen]

Hi! You’re on the right track choosing n8n for this 👍
If you’ve already built voice agents with Retell + Make, the mental model transfers quite well.

A good way to approach an email-based bot in n8n is to break it into a few clear stages:

1. Ingestion – Use an IMAP or Gmail trigger to capture incoming emails and normalize the data (sender, subject, body, thread ID).

2. Filtering & classification – Add lightweight checks (spam, auto-replies, “real question” detection) and classify intent.

3. Context & data lookup – Pull relevant customer data (CRM, Shopify, Stripe, internal DB) and fetch any knowledge needed for the reply.

4. Reasoning layer – This is where shared prompts + RAG work well. Keep prompts modular and retrieve templates or policies dynamically instead of hardcoding.

5. Response & actions – Generate the reply, store results, and send the email via a worker to avoid blocking the main flow.

6. Escalation & logging – Flag low-confidence cases and log everything for review and iteration.

My main advice: keep n8n as the orchestrator, not the place where all logic lives. Heavy reasoning, long prompts, and stateful memory are usually better handled outside n8n and called via HTTP nodes.
If you want, I’m happy to walk you through a reference architecture or review your first draft flow and suggest improvements.

[swift echo]

Thanks for your message, @cosmic lichen Really useful, man!

I've been built a few workflows using n8n before, so I could guess a simple flow for email based bot agent.

When I am buildig voice agents on retell, I used Knowledge Base containing website contents and readable files (pdf, txt, md). So the agent could answer to the caller based on that knowledge base.

I think the email based bot works like that, so I think that n8n workflow have to integrate something like the knowledge base of retell.

That is what I wanna know.

[cosmic lichen]

thank you for your feedback.

I want to discuss more detail.

please contact me

[peak olive]

It really depends on the size of the knowledge base.

If the knowledge base is large or expected to grow over time, then using a RAG setup is the better approach. It keeps things scalable, efficient, and more accurate because the system retrieves only the relevant chunks instead of loading everything into the prompt.

If the knowledge base is small and not too complex, then we can keep it simple. In that case, we can just ingest the PDF or content directly into the AI agent as context or attach it as a tool, without setting up a full RAG pipeline.

So basically, small and static → simple agent setup is fine.

[stone surge]

ChatGPT and Claude can also help in desiging your approach if share the requirements in detail. Ask it to create a workflow to import in n8n. Once imported, you can test and customize.

[swift echo]

Thank you for your message. 💟

[swift echo]

Thank you.

[wooden wigeon]

Hi kevin i have recently completed my 3rd model using n8n and retell ai
I specialize in architecting high-conversion voice agents, with a proven track record of successful deployments across Miami and Canada. Given the requirements, I am confident my solutions will deliver the specific results you're looking for.

When are you available this week for a brief call to discuss how we can scale these results for your project

[kind saffron]

When building complex workflows in N8N - I use Claude Opus 4.5 and N8N AI (which is the Claude 3.5 LLM) using both in tandem to check design and troubleshoot errors. Opus 4.5 has saved me hours of troubleshooting