While those visions were a little blurry in the beginning, the focus has sharpened in recent weeks. Plane though we are still in the very early days of this technology, I’m downright giddy well-nigh the potential.
In this post, we’ll walk through the goals, the apps I’ve tested, and what precisely I’m doing in this current test. At the bottom, you can plane join the beta to test it out yourself.
Apps I Tested
This is the most troublemaking and frustrating part well-nigh this very moment in AI. It’s incredibly crowded, and every app is new and untested.
The result is that no app is truly a finished product. Very few plane have a reliable and stable reputation to know what we’re unquestionably getting.
I started by testing every app that had a self-ruling trial to requite me a sense of how they worked. I honestly couldn’t plane list all of the apps I tested here.
But the three most recent tests involved SiteGPT, AI Engine, and Chatbase.
I was unquestionably very tropical to moving forward with SiteGPT. AI Engine is a WordPress plugin, and I was running into too many bugs (could have been plugin conflicts) that created frustration.
Ultimately, I’m going with Chatbase. It seems to be the most polished while moreover offering everything I need in wing to future options like API integration.
Trained Content
To train the bot, we need to requite it content. Chatbase provides the options of files (PDFs or Word docs), text, website URLs, and Q&A (providing example questions and answers). Since my website has increasingly than 1,000 pages, my primary focus is on website URLs.
You have the option of psoriasis the website, submitting a sitemap, or inward in URLs manually.
Of course, I have to be shielding with this. My website was started in 2011. Since the bot (for now, at least) doesn’t understand recency and won’t know how to handle outdated and estranged information, it’s important that I feed it with the most relevant information possible.
This unquestionably helped motivate me to craft content for this bot. In the meantime, that content can moreover modernize usability of my website.
Here are examples of pieces of content that I created to help modernize the knowledge wiring of the bot:
In wing to the main “Philosophies” post, I created eight individual posts to expand on those philosophies.
This process moreover provided motivation to update my Ads FAQ and Glossary (or at least start) since those sections were getting outdated. But they could be perfect content for a useful bot.
I selected 141 pages of content in all to train this bot on. While some outdated information within these pages is inevitable, I prioritized content that was still relevant.
I should add that Chatbase doesn’t have a limit on the number of links that you can submit to train the bot (big benefit). It’s all based on characters, and that’s a limit that I obviously haven’t run into yet (the lowest level has a 2 Million weft limit that I have not reached).
Settings
I won’t imbricate every setting here, but I’ll imbricate those that are most important or impactful.
1. Wiring Prompt (system message)
This is where you requite the bot some vital instructions on how you want it to behave. This includes things like tone, role, and how it should write unrepealable questions.
This is where the learning lines is highest and there’s a lot of trial and error. I’m still sorting this out.
2. Temperature
Do you want to restrict the bot or requite it self-rule to be increasingly creative? “Creative” could midpoint increasingly interesting, but increasingly potential for making stuff up. For now, I’m going fully reserved, but I may loosen it up depending on results I see.
3. Rate Limiting
This is necessary due to the data limits applied. You don’t want one user pushing through all of your data limits. I have no context regarding what the rate limit should be or how difficult it will be to run into data limits. So we’ll see how this goes and retread accordingly.
4. Suggested Messages
This gives people something to select if they are unsure of how to use the bot. You can use the wiring prompt to establish how these questions should be answered, though it can be addressed in the Q&A section as well.
5. Customize Chat
You can customize the style, colors, and icons. I’ve washed-up very little of this, but here’s what my bot looks like…
Conversation Logs and Refinement
This is one of the most important sections of the admin side of the bot.
I’m worldly-wise to see a history of all conversations that are ongoing with the bot. That way, I can evaluate whether people are running into frustrations or if the bot is giving out false or inaccurate information.
An example of this I’ve seen with these bots is that they often reference wares and links that don’t exist. So, someone will ask for an vendible that addresses a specific topic, and the bot will make one up, well-constructed with a fake URL.
That’s not good! While I think I’ve mostly addressed the fake URL issue with the wiring prompt, I can still write answers that I don’t like by revising the answer.
Here’s an example…
The question is related to Facebook ads strategies. The wordplay starts out pretty good, but it finishes up with a scaling strategy that isn’t really relevant. It’s not that the wordplay is so much wrong here, it’s just awkward.
So, I can click that “revise answer” link and remove the part I don’t like.
This may seem tedious, but this is the third phase of how it learns. First, it learns based on my content. Second, it learns based on the instructions in the wiring prompt and settings. And finally, it learns when I revise its answers. It won’t (or shouldn’t?) reply that way again.
Keep in mind that the question doesn’t need to be identical for the corrected wordplay to be used next time. That’s what makes these language models so smart.
Watch the Bot in Action
You can watch the bot in whoopee below…
Join the AI Bot Beta Wait List!
I need to limit how many people have wangle to this beta in the early going. As we go and I have a largest idea of data usage, I’ll be worldly-wise to retread and unshut it increasingly widely.