In a recent post, Implementing a Custom GPT at WebWorks, I introduced WebWorks Wizard – a custom GPT model built on ChatGPT, trained on ePublisher's documentation. It was an exciting experiment that showed how AI could help users find information faster. But as our team and customers started using WebWorks Wizard in real scenarios, we bumped into a few limitations. This follow-up post is a candid look at what didn't work so well and how we're fixing it by baking an AI chat assistant directly into ePublisher's Reverb 2.0 output.
What We Learned from WebWorks Wizard
WebWorks Wizard proved that an AI assistant for our docs can be incredibly useful, but it isn't perfect. After some hands-on time, we've discovered a few key issues:
-
Friction from login: To use the custom GPT, users have to log in with an OpenAI account. This extra step creates unwanted friction. It's like having to sign in to a separate tool just to get help – not ideal when you want quick answers.
-
Context disconnect: The GPT chat lives on OpenAI's site, outside of our documentation environment. Users have to leave our help system to use it, which can feel disconnected and clunky. The whole point is to assist users within our docs, and this approach misses that mark.
-
Static knowledge base: Perhaps the biggest limitation is how the GPT uses our documentation. It references our uploaded knowledge base only at the start of a conversation. If the conversation drifts to a new topic that isn't part of the original query, the GPT stops pulling in fresh info from ours docs. Instead, it relies on whatever it remembers or its general training. You can guess the outcome: sometimes it goes off-track or gives less accurate answers because it isn't doing a new search in the docs mid-chat.
Realizing these pain points has been a bit of a reality check. The custom GPT was a great first step, but we know we can do better for our users. We want an AI helper that feels like a natural part of the documentation, and that stays accurate even when the conversation shifts.
Our Solution - AI Chat Built into Reverb 2.0
To solve these issues, we're working on bringing the AI assistant home into our own product. Instead of sending users to ChatGPT's website, we're developing an integrated AI chat experience that will be built directly into ePublisher's Reverb 2.0 output (Reverb 2.0 is our modern HTML5 help format). In other words, when you publish your documentation with ePublisher in the future, your readers will get a built-in AI chat box right alongside the content. They'll be able to ask it a question, and it will reply using your documentation as the knowledge source – all without leaving the help site or dealing with separate logins.
This integrated chat approach we're developing tackles the earlier limitations head-on. First, there will be no separate login hurdle; once the feature is set up, any user of your published help will be able to start chatting with the assistant immediately. Second, the AI will be in-context – it will "live" on the documentation page, feeling like a seamless extension of the help system rather than a detached Q&A bot. And most importantly, the assistant won't be stuck with a one-time snapshot of your knowledge base. It will search the entire documentation fresh for each query. If a user switches to a new topic or asks an unrelated follow-up, the AI will look up the relevant info from the docs on the fly. The result: far more accurate answers that stay grounded in the correct context, no matter where the conversation goes.
How It Will Work for You
We're designing this feature so that adding an AI chat to your documentation will be as simple as possible. Here's the basic process we're planning for ePublisher users:
- Upload your docs to our new online platform – you'll ensure your source content is in an AI-friendly format like PDF or Markdown++. This content will be used to generate the knowledge base for the AI chat assistant.
- Write a simple prompt for the assistant – a one-time step to give the AI a bit of context or personality. This could be as straightforward as "You are the WebWorks Wizard, an AI assistant that helps users find information in the ePublisher User Guide." (If you don't want to customize this, the default behavior will work out-of-the-box.)
- Paste in your Assistant ID – you'll drop your Assistant ID into a new Target setting in ePublisher. This ID will let the published output call the AI model behind the scenes to generate answers (securely and only for your documentation).
Once implemented, you'll generate your Reverb 2.0 output with these settings, and your published help site will come with an AI chat interface ready to assist. Users browsing your documentation will be able to ask questions in natural language, and the AI will respond by pulling answers from the very content you've published. No logging into external sites, no context-switching – everything will happen right within your help system.
Markdown++ - Structured Docs for Smarter AI
You might be wondering how to make your content as AI-friendly as possible. One thing we've noticed is that structured, clean content helps the AI find answers more effectively. This is where Markdown++ comes into play. If you haven't heard of it, Markdown++ is WebWorks' own extended flavor of Markdown – it's like Markdown on steroids, giving technical writers a lot more structure and richness while staying clean and simple in syntax.
Why is that a big deal for AI? Well, documentation written in Markdown++ tends to be very well-organized (headings, lists, clearly marked sections, etc.), which makes it easier for an AI to parse and retrieve specific info. There's less noise from overly complex styling or hidden metadata – the content is nicely structured text. Feeding the AI with Markdown++-authored docs is almost like giving it a neatly categorized reference manual. We've found this can improve the relevance of the AI's answers because the important details are easier for it to pinpoint.
If you have a bunch of legacy content in other formats, ePublisher even provides tools to convert Word, FrameMaker, or DITA docs into Markdown++ without losing your styling or structure. By moving to a Markdown++ workflow, you're not just simplifying collaboration and publishing (which is a whole benefit on its own), but you're also making your documentation more digestible for AI assistance. It's a small change that can have a big impact on how well the AI helps your users.
Conclusion: A Better AI Help Experience Coming Soon
The journey from our first custom GPT to this integrated AI chat has been enlightening. We started with an out-of-the-box solution (literally on ChatGPT's site) and learned its limits the hard way. Now we're taking those lessons and developing something much more user-friendly and powerful for our customers. By embedding the AI assistant directly into ePublisher's Reverb 2.0 output, we'll remove the extra steps and uncertainty, making the experience feel like a natural part of the documentation. It will keep users in the zone – your help portal – and give them accurate answers backed by the full breadth of your content.
We're pretty excited about how this is shaping up. As of now, our team is deep into development and testing of the integrated chat feature, which we plan to release later this year. If you loved the idea of WebWorks Wizard but felt the friction, we think you'll love this new approach a whole lot more. It's all about making documentation smarter and smoother to use.
Stay tuned for more updates as we continue development. In the meantime, feel free to share your thoughts or questions about what you'd like to see in this feature. (Our WebWorks Community Discord is always open for these conversations.) Technical documentation is evolving with AI, and we're working hard to make sure you're equipped with the best tools to ride that wave. Here's to a help experience that will soon be smarter, faster, and truly hassle-free!