Turn Your Data Into Custom AI Models With Einstein Studio

Out of the box, Agentforce provides Salesforce Admins with a plethora of AI models to work with from leading vendors like Anthropic, OpenAI, and Google. When using tools like Prompt Builder, admins can select these models specifically to see if different models achieve different results.  

While they typically produce fairly similar results, newer or larger models will run slower but more accurately, and smaller, more lightweight models may be faster but less powerful. It’s never a bad idea to swap out models to test out new outcomes.

What if there was a more precise way to fine-tune and control the behavior of models? Tools like Prompt Builder have a wide set of potential use cases. The generative nature of large language models (LLMs) means there’s always a bit of randomness at play, which might be fine if you want a bit of creativity and flourish, perhaps for marketing-related content. Other times, you might want to keep the model from straying too far from the prompt, like for HR material.  

How can an admin fine-tune these models to behave more specifically? To be more or less random? Or even have a model focused on a specific dataset? Here’s where Einstein Studio and its handy companion Model Builder come into play. 

Let’s explore how Einstein Studio can configure existing models to fit your needs.

Configure models with Einstein Studio

From Data 360, open the Einstein Studio tab. From there, you’ll see the ability to list models you’ve generated with Model Builder, the related retrievers, and the Model Library. Select the Model Library and you’ll see a list of the foundational models Salesforce has associated with Agentforce. Select one of the GPT models and click New Configuration. From there, you’ll see an interface like this:

Let’s focus on the Advanced Settings in the lower-right corner. Depending on the model, you may only see some of these options, but let’s break down the three that are possible:

• Temperature controls the randomness or creativity of an AI model’s responses. When the temperature is set low, the model becomes very deterministic, almost always choosing the most likely next word. This makes its responses focused and predictable, which is ideal for tasks that require accuracy or consistency, such as coding or factual answers. Increasing the temperature makes the model more creative and varied, allowing it to take risks with word choice and produce more surprising, original output. This is useful for brainstorming, creative writing, and generating diverse ideas.
• Frequency Penalty influences how often the model repeats the same words or phrases. A low frequency penalty allows the model to reuse words freely if they are relevant, which can make the text more consistent but also more repetitive. Raising the frequency penalty discourages the model from reusing words too often, encouraging more varied vocabulary. This is especially helpful when generating longer outputs or lists where repetitive language could become distracting.
• Presence Penalty affects how likely the model is to introduce new words or topics that it hasn’t mentioned yet. With a low presence penalty, the model tends to stay on the same subject and expand on it, offering more depth rather than breadth. Increasing the presence penalty pushes the model to bring in new ideas or concepts, reducing its tendency to circle around the same content. This setting is useful if you want the model to explore a wider range of possibilities or perspectives.

At a high level:

• Low temperature + low penalties: Best for factual, technical, or code tasks
  
• Medium temperature + medium penalties: Best for general writing or chat
  
• High temperature + high penalties: Best for brainstorming, idea generation, or creative writing

Note that models will generally be configured with medium-level settings. To see this in action, I’ve generally started with a simple version of the task I’d be trying in a tool like Prompt Builder with shorter answers to see how the variations play out. For example, “Write a mission statement for a small startup software consultancy” is more of a brainstorming project where we might want to see a lot of variation, so a high temperature + high penalty configuration will give more options.

Convert your data into a custom model with Model Builder

What if you could create a model that was laser focused on your own data to make predictions? If you could convert historical data into something to help guess the future? With Einstein Studio’s Model Builder, you can do just that—all with no clicks and no code required. At a high level, you would:

1. Create a data model object (DMO) based on the dataset you want to predict. 
2. Filter that dataset down to what you want to leverage for the prediction (that is, for leads maybe if the status is Closed).
3. Determine related fields that factor into prediction (that is, the size and shape of the lead).
4. Create actionable variables that might help change the prediction (that is, a discount offer for a lead).

Einstein Studio will train and evaluate your model. Here’s an example of a healthy predictive model.

Once activated, you can use Flow with predictive outputs to:

• Use a screen flow to enrich records with predictions.
  Build a screen flow that retrieves records from Salesforce objects and displays prediction results on a Lightning page. For example, you can show a lead conversion prediction score directly on a lead record to help users make informed decisions.
• Use a record-triggered flow to automate actions when a prediction changes.
  Set up a record-triggered flow to automatically run actions when records are created, updated, or deleted. For instance, you can trigger actions when a case’s prediction score is updated and exceeds a defined threshold, signaling a potential escalation.

Make the precise model you need with Einstein Studio

Finding that the out-of-the-box foundational models aren’t giving you the experience you want? With Einstein Studio, you can configure those models or generate your own to create the outcomes you need, whether that’s a reliable response with little change, a creative response with a lot of different outcomes, or predicting outcomes based on the dataset of your choice. All of this being done declaratively with clicks, not code. Meanwhile, with Model Builder, you can tailor your experience to the data you want. Together, these tools give you much more powerful control over the experience you give your users with Agentforce.

See a walkthrough of Einstein Studio here.

Looking to learn more? Want free training for Agentforce? Sign up for an Agentforce NOW workshop.

Resources

• Trailhead: Get to Know Einstein Studio
• Trailhead: Build AI Models in Einstein Studio
• Trailhead: Create and Evaluate an AI Model