ChatGPT took the world by storm. However, many prompts provide generic responses that are forgettable and unengaging.
This is why we went on a quest to see what would happen if we were to train an AI on our expertise. In January 2023, we launched “Ask Flow SEO”: A chatbot that answers your SEO questions based on our own library of content.
Shoutout to Lukas Fingerle who created this for us and shared the exact process at SEO Reborn Summit.
In this blog post, we’ll explore how businesses can leverage AI technology to create chatbot experiences that leverage a company’s unique expertise.
We delve into repurposing content for conversational interactions, fine-tuning AI models, and creating engaging chatbot experiences.
Want to learn more? Check out this recording to understand how we trained GPT-3.5 on Flow SEO knowledge to create a unique SEO chatbot.
4 Steps of How We Built Ask Flow SEO with GPT-3.5 and WordPress:
1. Defining Our Goals and Tech
What do you want to achieve with it? What problems do you want it to solve? This will help guide the design and development process and ensure that it meets the needs of your users.
For example, when creating Ask Flow SEO, our goal was to provide website users with an improved search experience and easily surface the knowledge we already had on the site in a different way.
We are working with:
- The WordPress API to pull our existing content
- GPT-3.5 as a language model
- Vercel for a small frontend
We got inspired by “Ask my Book” from Sahil Lavingia, founder of Gumroad which he published in December 2022. Check his video “AI in a Day: GPT-3, Stable Diffusion, Embeddings, Fine-Tuning” for context.
2. Customizing AI for particular needs
There are two main ways to train GPT-3.5:
- Fine-tuning involves training the AI model on specific data to improve performance on a particular task.
- Embeddings involve creating a content landscape and mapping out different topics and subtopics, which helps to provide context to the AI and design effective prompts that deliver relevant responses.
You can do both independently from one another, but you can also combine them very well for even more accurate results.
What we used for Ask Flow SEO were embeddings. We utilized the WordPress API off the Flow SEO blog, scraped the blog content, created smaller content pieces from it of a couple of 100 words each, and then created the embeddings.
This way, when a user asks a question, the system identifies the most relevant content by finding the closest match in terms of embeddings. This content is then fed into the AI prompt, instructing GPT-3 to answer the question based on the given context.
Although the process may seem technical, it’s important to understand the main concept: using embeddings to provide context to the AI, enabling it to generate more accurate and relevant responses based on the content provided
3. Prompt Design with GPT-3.5
When it comes to prompt design, we found that making small tweaks could significantly improve the quality of the AI’s responses.
For Ask Flow, we created a specific format that begins with a prefix, instructing the AI to act as an SEO consultant and answer questions truthfully. By doing so, we were able to guide the AI to generate responses that align with our expertise.
4. Temperature Setting with GPT-3.5
Another key thing from the customizing process is the temperature API setting, which tells the AI how creative it can be and how much randomness it can put into the answers.
At a temperature of zero, there is no room for interpretation; it takes clear facts from the context you gave. In this case, the chatbot sounds very much like Flow SEO. However, our data set of about 50.000 words was very limited – so the chatbot is also very brief,
So, the higher you go into temperature, the more in-depth its answer, but also the more likely it mixes random information from its base model into its reply – the same outdated SEO advice you can find on Google, not what we wanted.
You want to make sure that you find a sweet spot between too low and too high to ensure that the AI remains closely aligned with the actual content provided.
The Future of Chatbots: AI-Powered Experiences
Chatbots have become essential for businesses to enhance customer service and engagement. While there are still challenges to overcome, the potential benefits are worth the effort. So, are you ready to create chatbots that wow your users and set your business apart?