Placeholder image

Exploring SAP RPT-1

| Amit Lal |

AI ABAP

YouTube

Spotify

Apple

Episode #276

Introduction

In episode 276 of our SAP on Azure video podcast we talk about RPT-1. On of the new foundational models from SAP.

One of the highlights at SAP TechEd last year was the announcement of two foundational new models, among them the RPT-1 or Rapid One. For me this was quite interesting, because SAP not only announce the model, but immediately made it available for testing on BTP and HuggingFace. Our amazing colleauge Amit Lal immediately jumped on it and explored the functionality. Today I am happy to have him with us again to talk about his experience with RPT-1.

Link to the repo on GitHub from Amit: SAP RPT-1-OSS | AI-Powered Enterprise ML Suite: https://github.com/amitlals/SAP-RPT-1-OSS-App-playground

Find all the links mentioned here: https://www.saponazurepodcast.de/episode276

Reach out to us for any feedback / questions:

#Microsoft #SAP #Azure #SAPonAzure #RPT1 #AI #Copilot

Summary created by AI

  • Introduction to SAP RPT-1 and Its Significance:
  • Holger, Goran, and Amit discussed the introduction of SAP’s RPT-1 (Rapid One), a foundational AI model for tabular data, highlighting its immediate availability for testing and its unique focus on SAP’s table-based enterprise data.
    • Background and Announcement: Holger introduced SAP RPT-1 as a significant announcement from SAP TechEd, emphasizing its immediate availability for testing on BTP AI Hub and Hugging Face, and noted Amit’s early exploration of the model.
    • Amit’s Role and Motivation: Amit described his position as a strategic technical lead at Microsoft, focusing on SAP workloads on Azure, and explained his motivation for experimenting with RPT-1 to bridge AI and SAP for enterprise customers.
    • Model Focus and Differentiation: Amit clarified that RPT-1 is SAP’s first foundation model specifically designed for tabular data, distinguishing it from traditional large language models that primarily handle unstructured text.
    • Open Source and Accessibility: Amit emphasized that his demonstrations and experiments use only the open-source version of RPT-1, making all related demos and code available on GitHub for partners and customers to replicate.
  • Technical Overview and Comparison with Traditional ML and LLMs:
  • Amit provided a technical deep dive into RPT-1, comparing its architecture, operational principles, and performance with traditional machine learning models and large language models, with Holger and Goran contributing to the discussion.
    • RPT-1 Architecture and Functionality: Amit explained that RPT-1 is a relational pre-trained transformer optimized for SAP’s tabular business data, operating natively on tables and supporting in-context learning without the need for model training or fine-tuning.
    • Limitations of Traditional ML and LLMs: Amit discussed the inefficiencies of traditional machine learning, which requires custom models for each prediction task, and highlighted the challenges LLMs face with tabular data, such as tokenizing numbers as text and hallucinating calculations.
    • Advantages of RPT-1: Amit outlined RPT-1’s benefits, including higher accuracy, lower resource consumption, instant predictions, and reduced burden on data science teams, as it processes numbers as relational values and does not memorize data.
    • Product Family and Use Cases: Amit described the different variants of RPT-1 (Small, Large, OSS), their intended use cases (real-time fraud detection, supply chain forecasting, local development), and the decision matrix for selecting the appropriate model.
  • Demonstrations of RPT-1 Applications and Use Cases:
  • Amit showcased several practical applications and demos of RPT-1, including financial dashboards, model comparison playgrounds, predictive integrity for SAP operations, and custom deployments, with Holger and Goran engaging in clarifying questions and observations.
    • SAP Playground and Financial Dashboard: Amit demonstrated the SAP RPT-1 playground, uploading synthetic sales and financial data to showcase instant predictions and discussed limitations such as row count restrictions in the playground environment.
    • Custom Open Source Applications: Amit presented four open-source applications: a financial dashboard, a forecast showdown comparing RPT-1 with other models, a predictive integrity tool for SAP job and interface failure prediction, and a custom deployment of RPT-1 OSS on Azure.
    • Model Comparison and Evaluation: Amit illustrated how RPT-1 outperformed other LLMs in tasks such as customer churn prediction, providing higher accuracy and more reliable labeling, and discussed the importance of validating predictions with real data.
    • Operational Risk Prediction: Amit showed how RPT-1 can be used to predict high-risk SAP jobs, transports, and interfaces, enabling proactive remediation and operational insights for enterprise environments.
    • Deployment Flexibility: Amit explained how RPT-1 OSS can be deployed on Hugging Face or Azure, allowing customers to run models in isolated environments and integrate predictions into their own applications.
  • Community Engagement and Feedback Invitation:
  • Holger, Goran, and Amit encouraged customers and partners to experiment with the open-source RPT-1 tools, provide feedback, and measure the model’s effectiveness in real-world scenarios.
    • Open Source Accessibility: Amit reiterated that all demos and code are available on GitHub, making it easy for users to clone, deploy, and test the solutions in their own environments.
    • Feedback Importance: Goran and Amit highlighted the value of user feedback for both Microsoft and SAP, noting that real-world testing and input will drive further adaptation and improvement of RPT-1.
  • Additional AI Tools and Innovations from Amit:
  • Amit briefly presented additional AI-driven tools developed in his ‘AI kitchen,’ including an agentic diagram generator for workshops and a cloud region latency advisor for SAP workloads, with Holger and Goran expressing interest.
    • Agentic Diagram Generator: Amit demonstrated a tool that uses AI agents to automatically generate architecture diagrams during workshops, saving significant time and allowing for real-time modifications.
    • Cloud Region Latency Advisor: Amit introduced a latency advisor tool that measures real-time network latency to various cloud regions (Azure, AWS, GCP), helping customers select optimal deployment locations for SAP workloads.

Links

Video index

Based on and using Hugo Cookbook Theme, Font Awesome Icons and Bulma CSS. | Thanks to Recipes4AJ to get me started. | Contact me via