Concept card

SAP RPT-1 / RPT-1.5 — Relational Foundation Model Architecture

As of 2026-07-24T14:00:00Z

Edited by Cédric-Olivier Mâry, sole editor and publication director — 27 years in the SAP ecosystem. How these pages are produced, and where automation is used.

What is SAP RPT-1 / RPT-1.5 — Relational Foundation Model Architecture?

SAP RPT-1.5 pretrains directly over relational schema structure instead of flattening rows into text, and its RAP mechanism collapses model customization from weeks of ML pipeline work to seconds of in-context assembly.

What It Is

SAP RPT-1, short for Relational Pretrained Transformer, is the first foundation model architected natively for tabular, structured, relational business data rather than for unstructured text. Standard large language models treat a database row as a flattened string of text, which throws away column semantics, foreign-key relationships, and schema structure — the very things that make a relational table useful. RPT was pretrained instead over millions of relational schemas, learning patterns at the row, column, and foreign-key level directly from table structure. It was announced in 2025 and upgraded to RPT-1.5 at Sapphire 2026.

Why It Matters

RPT targets a specific, well-documented failure mode that stalled a large share of enterprise AI pilots from 2023 onward: SAP transactional data lives in normalized relational schemas, and forcing that data through a text-token large language model requires expensive extract-transform-load work plus prompt engineering that, in practice, still lands accuracy below classical machine-learning baselines such as XGBoost on standard tabular benchmarks. RPT closes that gap by working with the table structure directly instead of translating it into prose first.

How It Works

Why it matters in practice

Forcing SAP transactional data through text-token LLMs requires expensive ETL plus prompt-engineering that still drops accuracy below XGBoost and classical ML baselines on tabular benchmarks.
RAP lets the model ingest arbitrarily large relational tables at inference time and return typed predictions — classification, regression, forecast, anomaly score, imputation — respecting foreign-key joins, with no retraining.
Three variants target different needs: sap-rpt-1-small for interactive prototyping, sap-rpt-1-large for harder tasks like churn scoring or fraud detection, and sap-rpt-1-oss published open-source on HuggingFace/GitHub at Sapphire 2026.

Key points

First foundation model architected natively for tabular relational data — pretrains over millions of relational schemas, learns row + column + foreign-key patterns directly from table structure, not text tokens.
RPT-1.5 adds RAP (Retrieval-Augmented Prediction) — ingest arbitrarily large structured datasets at inference time, in-context typed predictions (classification, regression, anomaly, imputation) with no retraining; RAG-pattern for tabular data.
Three variants: sap-rpt-1-small (fast inference, playground prototyping) · sap-rpt-1-large (higher accuracy on hard tasks: multi-step forecast, fraud, churn) · sap-rpt-1-oss (open-source on HuggingFace + GitHub for external benchmarking).
Solves the 2023-2025 pilot-plateau problem — text-LLMs consistently underperform XGBoost on SAP transactional data; RPT removes the ETL-to-tokens bottleneck and is competitive with or exceeds classical baselines on tabular leaderboards.
Natural-language chat assistant in the playground — analysts query tabular datasets in plain English; collapses model-customisation cycle from weeks to seconds via RAP context assembly.
Typed prediction outputs — RPT returns predictions that respect column data types and foreign-key constraints, not free-text strings that require post-processing parsing; a material accuracy advantage over text-LLM workarounds.
Benchmark surface — sap-rpt-1-oss allows direct comparison against TabPFN (prior SOTA for tabular foundation models), XGBoost, LightGBM, and TabTransformer on standard UCI + Kaggle tabular leaderboards.
Deployment via SAP BTP — the playground and serving infrastructure are BTP-native; the correct deployment path is SAP AI Core + AI Launchpad, not a standalone model server.
When not to use — genuinely unstructured data (free-text, documents, images), no BTP footprint, or when the ML task is already solved with high accuracy by a classical pipeline that is cheaper to maintain.
SAP ecosystem integration — RPT predictions feed downstream into SAP Datasphere Analytic Models, SAC Planning, and IBP demand signals via the SAP AI Core API, closing the loop from foundation-model prediction to business decision.

Terms used on this page

RPT (Relational Pretrained Transformer): SAP's foundation-model architecture pretrained over millions of relational database schemas; the first transformer designed natively for tabular structured business data, learning column semantics and foreign-key patterns rather than text tokens.
RAP (Retrieval-Augmented Prediction): The central architectural addition in RPT-1.5; analogous to RAG but for tabular data — the model retrieves relevant rows and columns from arbitrarily large structured datasets at inference time using a learned retrieval head, then produces a typed prediction without retraining.
sap-rpt-1-oss: The open-source variant of RPT published on HuggingFace and GitHub at Sapphire 2026; enables external researchers to benchmark RPT against TabPFN, TabTransformer, XGBoost, and LightGBM on standard tabular leaderboards, building community trust.
Tabular foundation model: A foundation model architecture specifically designed for structured tabular data with schemas, foreign keys, and column-level data types — distinct from text-token LLMs; RPT is SAP's production bid, TabPFN (Hollmann et al., 2022) is the prior open-research SOTA.
TabPFN: Prior-Data Fitted Networks — open-source tabular foundation model from Hollmann et al. (NeurIPS 2022/2023) that RPT-1.5 benchmarks against; achieves strong performance on small tabular datasets (<10k rows) but lacks the relational schema pretraining that RPT targets for SAP data.
Pilot-plateau problem: The 2023-2025 pattern where enterprise GenAI pilots using text-LLMs on SAP transactional data consistently underperform classical ML (XGBoost, LightGBM) on numeric prediction tasks — the architectural gap RPT is designed to close.
SAP AI Core: The BTP service that hosts, serves, and scales custom AI models including RPT; the required deployment target for production RPT serving in SAP customer landscapes.
Foreign-key-aware prediction: RPT's ability to produce predictions that respect referential integrity across related tables — e.g. anomaly scores that account for vendor master data when scoring purchase-order lines — a capability text-LLMs cannot replicate without explicit ETL.

Sources

Full card available to members. What the full card adds: the full decision framework · the SAP vs Snowflake / Databricks / Fabric comparison · the common pitfalls and their fix · the cheat sheet · the architecture schemas · the code blocks · the facts worth quoting.

Open in the app →