As Galal implied, foundation models enable organizations to interface with software in natural language. “You just type in what you want,” Galal explained. This fact alone greatly democratizes the use of applications for constructing pro- cesses, automating them, and making them accessible to almost anyone. This same boon extends to systems that use LLM techniques for question-answering, search, and query generation. For query generation, users type a prompt in natural language, which the model translates into the language of the specified database or application. Ensuring accurate results and minimizing what Christine Browning, head of product experience at SAS R&D, termed “hallucinations” entails these steps:
♦ Prompt engineering: According to Galal, prompt engineering is required to give foundation models “boundaries like, ‘Generate these types of things and not these types of things.’ It constrains the model to things that are plausible, lawful, and private.” Vendors can limit model responses to vendors’ internal systems through prompt engineering. One approach to doing so is to “use an ontology to constrain the queries that are generated,” Martin revealed.
♦ Fine-tuning: It’s also necessary to fine-tune models to generate the most appropriate content, action, and results. “Fine-tuning is a step beyond prompt engineering in terms of giving [the model] an example of questions and answers,” Galal indicated. “You give it like, ‘Here’s a question,’ and ‘Here’s a right answer,’ and give it a couple of samples of those so it knows this is the range of possibilities.”
♦ Chain of thought: This precept typifies the sophistication and complexity of foundation models for tasks such as process automation. It’s what allows organizations to “have a dialogue, if you will, while the large language model still remembers that you started a process with two steps; you’ve added a third step; you’ve modified the second step; you’ve added a conditional step,” Galal commented. “That’s a chain of thought because it remembers the original concept of ‘I want an insurance claim approval process.’”
Process automation
The natural language interfaces LLMs can support are invaluable for process automation. Models can generate recommendations for germane steps in processes and domain- or task-specific data models. They can also generate data—known as synthetic data—to test applications built with low-code process automation techniques. This functionality is instrumental for expediting low-code application development to automate any number of processes, from human resources to healthcare services. The greater merit derived from the generative prowess of foundation models stems from blending these textual applications with visual ones involving images.
Galal described a process wherein users can upload a document as a PDF “and we can make the LLM generate a comparable digital form of that document that is understood throughout our platform.” The instantaneous creation of this digital rendering has implications for developing forms for any use case or industry, particularly when considered at enterprise scale. Eventually, users will be able to simply upload what Galal termed “napkin drawings,” and the system will render a process, workflow, or application to match.