A team spent two weeks engineering prompts for a customer churn prediction model. The inputs were structured ERP records - sales history, service logs, payment patterns. The outputs were inconsistent. The team kept adjusting the system prompt, trying different LLMs, running new tests. Eventually someone asked: why are we using a language model for a structured prediction problem? An RPT model would have taken two hours and required no prompt engineering at all.
Not every task needs a large language model. This lesson covers how to use the Model Library to compare options, and when to choose SAP RPT models instead of an LLM.
The Model Library
The Model Library in SAP Generative AI Hub is where you compare and choose models. You can filter by task type, modality, provider, context window, and cost. A leaderboard with benchmark data lets you compare performance side by side. Model cards give you capability details, input types, and pricing. You can deploy a model or open it directly in Chat or Prompt Editor from the library.
LLM or SAP RPT Model: choosing the right tool
Choose an LLM when the task involves language (summarizing, answering questions, drafting, conversation) or when your input is unstructured text (documents, emails, support tickets, chat messages).
Choose SAP RPT models when your input is structured rows and columns from a business system (ERP records, sales tables, service logs, HR data) and the task is a business prediction (delivery delays, payment defaults, upsell likelihood, churn, technician routing). No prompt engineering or model training required. You provide labeled context rows. The model predicts.
Hint
Quick test: if you can describe your task as "predict [outcome] from [rows of business data]" - that is an RPT use case. If the input is free text or the task involves generating language - that is an LLM.
What is new: SAP RPT-1.5
SAP RPT-1.5 reached general availability in July 2026. Available in Small and Large variants in the Generative AI Hub model library alongside LLMs.
Enhanced explainability — factor-level attribution shows exactly which input features drove the prediction
Improved large-scale performance — better accuracy on high-volume ERP datasets with many unique values
Tabular Orchestration — automated context engineering for production workflows, now generally available
Note
Don't make this mistake
Using an LLM for structured business prediction tasks because it is more familiar. LLMs are not trained on your ERP data and were not designed for row-level structured prediction. You will spend days prompt engineering around a fundamental model mismatch. If your data lives in rows and columns and the task is a business prediction - use an RPT model. It will be faster to configure and more accurate on your data.