The journey behind launching the HelpKit AI Notion chatbot

Hey folks!

Today I want to talk with you a little bit about the journey behind launching HelpKit AI. HelpKit is a solo founded and fully bootstrapped tool that turns your Notion docs into a professional help center and documentation site.

Since HelpKit is already hosting hundreds of customers Notion knowledge bases it only made sense to supercharge them with a custom AI trained chatbot. So far the results have been amazing.

The public beta is out since two weeks and we have onboarded more than 15 customers on the add-on and the AI bot held 2000+ successful conversations. And now, we just finally launched on Product Hunt as well!

The launch was planned two months ago but some things stopped me from releasing it. If you are curious about the background feel free to read on.

Now, let me explain why the journey to the launch was a bit rougher than expected.

The UX Hurdle

First and foremost, I was stuck with developing a great user experience. I have only come across a handful of products that were sort of implementing the same concept and I really did not like their UX implementation. Everything felt clunky and just did not fit into the concept of an AI enabled knowledge base system.

You see our AI chatbot needs to be integrated into HelpKit’s two layouts: Help Center and Documentation and also work well on the widget and mobile.

Quite a tricky task to get right everywhere! I brainstormed for a few days, tried out some implementations but none of them felt natural for a user in my opinion. I even spent a whole week building an interface only to then throw away the entire feature code branch and start from scratch.

What I needed was a break. So I took a weekend break, spent some time at the beach and came back refreshed with some new ideas (that strangely all came to me during my Sunday evening shower).

I realised that the most intuitive way to present the option to chat with an AI powered chatbot assistant is when the user is already searching for something. So this way, should our normal search not be able to deliver an instant search result, users can be “catched” and still have their problem resolved by the AI.

Once you click into the search bar the Ask AI… button allows you to start chatting 👇

In terms of the UI, I decided to go with a slick right sidebar interface for the chatbot on the help center layout and a nicely integrated chat window within the search command bar of the documentation layout.

Within the HelpKit embeddable widget as well as on mobile phones, HelpKit AI will show a full-screen chatbot interface to the user.

Overall, the HelpKit public beta has been running for 2 weeks and users as well as customers are really satisfied with the great user experience. Which… also obviously makes me super happy ツ

Building on top of AI

HelpKit AI is my first real touchpoint developing a product with Large Language Models (LLMs). After doing some simple warm-up Hello World AI tests, everything seemed so simple.

Unfortunately I had to soon find out that I sincerely underestimated the complexity involved in developing a conversational AI chatbot that closely mirrors human interaction.

Following a considerable period of experimentation with a plethora of existing tools, it's become apparent that the majority merely repurpose around 10 lines of Langchain code, presenting it as a fully-fledged product.

The outcomes of these efforts are, at their best are "meh". Some OK, but not great.

I nearly fell into the same trap, having developed a feature that was 90% complete before acknowledging the subpar and cluttered results, as well as my limited comprehension of this new technology.

Engaging with LLMs presents an entirely unique and novel paradigm.

Consequently, I revisited the initial plans and invested a lot of more time immersing myself in the intricacies of LLMs, AI User Experience, and a slew of way too complex research papers, all in an effort to better understand how to construct tools around LLMs.

Interestingly, due to the novelty of this field, there's no definitive roadmap to a successful implementation. The process is characterized by a great deal of experimentation. You really need a flexible approach and swift adaptation to the frequent and rapid enhancements being made to these tools. As an engineer being used to having deterministic outcomes I just couldn’t believe the quirkiness of testing out bunch of prompts to find out what works the best. In the end everything finally clicked and I know exactly what I was doing and how I can make HelpKit AI as best as possible.

Wrapping up

So that’s my little background story of building HelpKit AI and some of the struggles that I had. The biggest takeaway lesson from me here was to realize that sometimes you need to stop being so harsh on yourself and accept that some things are more difficult than others. You might need a little bit more time than you thought. The beautiful thing for us Indiehackers is that we don’t have any pressuring project manager sitting behind our back pressuring us to release something despite the quality but somehow we still tend pressure ourselves even harder into it. Take a step back, breath, learn more and then come back stronger.

In case you are interested you can find the HelpKit AI landing page here and I also recorded a little Youtube video demoing HelpKit AI you can check out.

That's it! Hope this was useful to some of you reading this and also thinking about embarking on the journey of building a big new feature for your product.