Create Embeddings

Embeddings are great for uploading tons of text so your AI can respond to your users with the correct content. But there is some structure to it! To help you get the best responses possible. So, in this module we'll take you through the best way to structure them.

Where to find embeddings

First navigate to the main dashboard and select the Integrations tab to the left hand side task bar. Next select the OpenAI tab and the OpenAI window will display. At the top on the screen select the Embeddings tab.

The next step is to click the +New Embedding button to the top right side. This will now open the embeddings editing display.

Let's breakdown the 3 input boxes shown in the image:

• Type: The type box is where you'll insert a title for the embedding. Although it's actually more than just a title. The Type input is how OpenAI searches all of your embeddings, so naming the title needs to be unique and not duplicated in other embeds.

• Heading: Just like the type input, OpenAI also will search through the Heading context to 1, filter the search and also allow you to use conditions later on in your workflows. So again, keep the name here unique.

• Text: You can insert up to 1000 characters in this window. This is where you content/information will go.

Structure the text

This is optional. After many attempts, testing and user feed back we have found that structuring your text can help with providing a better response.

If you just want to copy and paste your existing text then that's fine too. We personally find that OpenAI will read the text better when adding a little structure to it.

So how do we structure the text input. As you'll see in the image below, that each section is broken down in to Headings and Sub-headings. For example:

Instead of just typing in the whole information about the pricing and plans, we broke it up in to smaller chunks denoted by the words Heading and Sub-heading.

Be sure to test it yourself in your own embeddings, try it without headings/sub-headings and test it with them. Gauge if you see any differences on the output response.

Side-note:

You an also inject prompting in to the text box to if required. We also found that this method works really well also, not only in providing a better response, but also reducing hallucinated replies. Just like we added in our integrations embedding. Labelled: Additional