/

Building Apps

/

Personas allow you to train LLM models on your blog to follow your tone and style.

Before getting started with this feature, you should be aware of the following:

  • Personas are currently only available for Team & Business plans.

  • You must have a blog with at least 50 posts with 400+ words each.

  • Training a Persona costs around 2000-3000 credits, you will get the exact amount before you start training.

Creating your Persona

Head to the Personas page and click on the +Add button, you will first be prompted to provide a sitemap url to your blog. You can also filter the extracted sitemap URLs by the slug. For example, if your blogs are all hosted under the /blog/{title} slug you should use /blog/ as the filter.

Once the sitemap is validated, you'll be able to name your Persona, and select the base model for training (Currently only GPT 4o Mini is available). Additionally, you'll see the list of urls extracted from your sitemap, we randomly sample 50 pages but you can cherry pick the pages you want to include. For example, pages such as a privacy policy, contact and any non-blog post pages should be excluded. You must have exactly 50 pages selected.

After that, click on the Prepare Training Data to create your persona (costs 100 credits). This does not yet start the training process but instead it will start generating the training data required which takes less than 10 minutes.

Once the data preparation process is complete. You'll be able to start the training process which usually takes less than 30 minutes. Additionally, you'll be able to accurately see how many credits the training will consume.

Finally when the training process in done, your persona will be ready to be tested and used in your app workflows. We recommend testing your persona before using it, to do so, click on the test button. This will open a panel where you can provide a prompt and see how the response of your persona compares with the base model response.

Additionally, from the pages you've added, we reserve 5 for testing, so you can click on any of the tests under the prompt box, this will load the post contents on the left, and it will generate the prompt for writing it. Once you run this test prompt, you'll be able to quantitatively assess your personas performance. You'll see a Similarity Score for each the base and the persona, this is telling you basically how similar your personas response is to the actual post and how similar the base model response is to the actual post.

In this test we did for the Semrush blog, the finetuned persona got 91% similarity while the base model got 79%.

Using it in your Apps

Your trained personas will show up in the Chat Model, model field options. You can also click on the Create App using Persona button on the persona item to create a starter workflow using the persona.