Agentforce

Chapter 10What Is RAG, and Why Should I Care?

RAG is an acronym that stands for retrieval-augmented generation – as in generative AI using retrievers that augment the prompt. More or less.

We've seen that a lot of Agentforce and Salesforce AI more generally exists for the enterprise-ification of AI, the making-it-useful to big companies. But as my father told me once, our strengths are our weaknesses. GenAI and LLMs are resolutely nondeterministic: it is literally impossible to predict exactly what will come out as a response. Why? Because LLMs do not follow rules but do their own form of reasoning. It's what makes them more exciting than paint-by-numbers chatbots or tree-based service bots, no matter how complicated.

But nondeterminism is also a weakness. Imagine this statement coming from marketing to the legal team: “Hey, we're not exactly sure what the email is going to say, but we know it will be super-relevant to each customer.” Or from customer service: “We're not sure what our agents will say, but customers will love us!”

Thus, we need the Trust Layer and auditing and governance. Don't let your agents loose without it.

Likewise, RAG is used to fill another enterprise-sized gap in GenAI's résumé: making sure the model uses the company's own data somehow. Without incorporating the company's data, the so-called agentic loop would consist of third-party models telling companies what to say to their customers without knowing anything about them – maybe beyond what they find ...