Four Game-Changing NAN Updates for Building Better Chatbots

I just wrapped up a video showing off some cool new NAN chatbot updates that honestly blew my mind. There are four major features that'll change how we build these things.

First up, there's this updated node for web search abilities. It's been around for a bit, but now it lets you send multiple messages in one conversation or add conditional workflow logic. Pretty neat, right? This means you can split conversations into different branches on your canvas. I've found this super helpful when designing more complex interactions.

Another big one is direct human-in-the-loop access as a tool call. Besides Gmail, you can use other nodes like WhatsApp. So when your chatbot gets stuck, it can quickly grab info from a human and keep the conversation flowing. No awkward pauses!

I also showed two derivative nodes that work as tool calls. These let the agent send dynamic messages during execution. So if the agent realizes it needs more context, instead of starting over, it just asks for what it needs and continues.

To show how this all works together, I ran a demo using a simple "hi there" message. The agent's response goes through the "respond to chat" node. You gotta bring this node into your canvas and tweak the settings to use respond mode.

The OpenAI node now has built-in tools like web search, file search, and code interpreter. For my demo, I restricted web searches to my domain, which lets the agent pull real-time info about products or sales without complex RAG systems. But heads up - while GPT 4.1 handles this filtered search fine, the mini version doesn't.

When I asked "What's the weather in Melbourne?", the agent first sent a message saying it was checking, then gave me the actual weather. This dynamic chatbot integration makes conversations feel way more natural.

These NAN chatbot features are so new there's no documentation yet! You'll need the latest version to access them. I walked through adding the "respond to chat" tool, showing options like static messages, AI-generated ones, and the "Wait for user reply" setting. This last one's crucial for conditional workflow logic - it pauses until the user provides context.

The human-in-the-loop chatbot integration is probably my favorite. I set it up to trigger when someone asks for a discount. When I requested 10%, the agent told me it was checking with management, then sent a Gmail request that I manually approved. The agent then confirmed my discount. This human-in-the-loop approach makes chatbots way more useful in real situations.

So what do you think? Would you use these dynamic chatbot features in your projects? Let me know in the comments if you want me to build out more complex demos with conditional routes. I'm happy to dive deeper if there's interest!