Transform Data

Next, let's imagine that you only had access to initial customer inquiries, but not the support agent responses. In this scenario, you would have only imported a single column in the previous step. In order to fine-tune a model, you need to pair the inquiries with high-quality responses.

The Entry Point transforms feature lets you apply prompts to your data to produce missing or complimentary fields.

In this section, we'll use Entry Point's transform functionality to:

• Craft a prompt that gives us a desired AI agent response to a customer inquiry

• Run that prompt against every inbound request in our data set

• Save the responses as a new field so they can be used to fine-tune a model

Create a New Transform

Select Transforms from the left-hand sidebar. Then press "+" in the top-right corner.

Enter your desired system message. In this example, we used the prompt below.

You are a friendly and helpful support agent. Your job is to do the initial intake of support events. You will gather additional information from the customer before we connect them with a human agent. 

Please start the conversation with the customer by: 
- Greet them warmly - thank them for reaching out, apologize for the inconvenience they are experiencing, assure them that we'll get them taken care
- Tell them you're an AI agent and that you'll get them connected with the support team after they answer a few very brief questions

Then, ask the customer the questions below: 

- How long have they been experiencing the issue? 
- Have they taken any steps to try to resolve it already? If so, ask them to explain briefly. 
- Are they able to replicate the issue? Does it happen all the time, or only intermittently? 

Format the list of questions as bullet points to make it easier to read. Also, if the initial client outreach already contains an answer to any of the three questions, please do not ask them to provide that information again.

Stick to the basics and keep things concise. Use a casual tone; there's no need to be super formal. Do not use emojis.

In the user message field, select "Inbound Request".

Press "Save" in the lower left. Your transform will automatically be given a name.

Test Transform

Next, we can test the prompt we wrote in the previous section against an example from our dataset to see how it performs.

Scroll down until you see the Preview section. Here, you'll see a preview prompt containing your System and User messages. Note that you can click on the icon with two arrows in the top-right to randomly select a different customer inquiry to test with.

Click on the settings icon in the Test Transform section. In this example, we're going to

• Leave the Model as Open AI

• Set the Model to GPT-4 Omni

• Set Temperature to 0.2

Press "Generate". Review the response. Adjust your system and user messages as needed and repeat the process until you are happy with the results.

Create a Transform Job

Now that we like the results we're getting, we're ready to start our transform job. In this step, we'll take the prompt that we crafted in the previous steps and run it against all of the examples in our dataset.

Start by scrolling down and click the "+" button in the Jobs section.

Next, we need to name the "New Destination Field". This is the Field in which the responses will be stored.

In this example, we've entered "AI Agent Response" for the Name. The Reference will automatically generate.

Press "Next".

You'll now be asked to select your Platform and Model. We suggest using the same settings you used in the Test Transform section above.

Press "Next".

Enter your desired temperature setting. Again, we recommend using the same setting you used to test the prompt in the section above.

Entry Point will provide an estimated cost to run the transform job. If this looks high, set a reasonable limit on max tokens for a better estimate. Once you're ready, press "Start". Depending on the size of the job, it may take a few minutes to run.

Once the job is complete, you'll see the new field has been created and responses have been added to all of your examples!

Next, we'll use Entry Point's Synthesis function to easily add more examples to our fine-tuning dataset.