Introducing the SAP-RPT Model

Objective

After completing this lesson, you will be able to introducing the SAP-RPT Model

The SAP Relational Pre-Trained Transformer (SAP-RPT)

What Is SAP-RPT?

SAP-RPT stands for Relational Pre-trained Transformer (officially pronounced "rapid"). It is SAP's first tabular foundation model, a pre-trained AI model built specifically for structured, relational business data : tables, rows, columns, the kind of data that lives in every SAP system.

Hint

As you will see below, you can think of SAP-RPT as a kind of GPT for ABAP or non-ABAP tables – except it predicts missing or future values in a tabular structure, not the next word in a sentence.

Quick facts

SAP-RPT has been released for the first time (SAP-RPT-1) in Q4 2025. Since then, two new versions have been released successfully:

  • SAP-RPT-1.5 in July 2026
  • SAP-RPT-1.6 in September 2026.

Caution

SAP-RPT-1 will be retired by December 31, 2026. Before then, you should start working with SAP-RPT-1.6, the most performant SAP-RPT model to date.

SAP-RPT mostly does Classification and Regression on tabular data. It is not a Large Language Model (LLM), because it operates on tabular data, not text-based data (words, sentences, paragraphs, and so on).

SAP-RPT comes with two different variants:

  • SAP-RPT-1.6 for medium-complexity scenarios where low latency and high throughput are the priority
  • SAP-RPT-1.6-large when highest prediction quality is the primary objective, with support for significantly larger context and more target classes.

What are the Key Benefits of SAP-RPT?

Before RPT: the Classical Machine Learning Workflow

Getting a predictive model into production in an SAP landscape traditionally means:

  1. Identify a business question (for example, "will this invoice be paid late?" )
  2. Export data, clean it, engineer features
  3. Choose an algorithm, train a model on historical data
  4. Evaluate, tune, retrain
  5. Deploy and monitor; retrain when the world changes
  6. Repeat for every new prediction task

This process requires data science expertise, can take a lot of time, and produces a narrow model that works for one task only. Most enterprises end up with a graveyard of one-trick models, each needing its own maintenance cycle.

What RPT changes

The following table shows how SAP-RPT compares to traditional approaches.

Pain pointBefore RPTWith RPT
Time to first predictionWeeks to monthsHours to days
Training data requiredLarge labelled datasetAs few as 2–3 example rows
ML expertise neededRequiredNot required
Feature engineeringManual, case-by-caseNone
Model retrainingRequired when data driftsNo retraining: context changes with data
One model per taskYesNo: one model, many tasks
Incomplete/changing dataOften breaks modelsIs handled natively

How SAP-RPT Works: In-Context Learning

SAP-RPT uses tabular in-context learning, the same general idea as prompting an LLM, but applied to tables instead of text.

Instead of training on your data, you feed the model with examples at inference time. The model reads them, identifies the pattern, makes the prediction, and then discards your data. Nothing is stored from your specific rows: learning happens on-the-fly.

Anatomy of an RPT request

Every request on a "relatively" small table or two, containing two kinds of rows:

  • Context rows

    Labelled historical records that teach the model the pattern.

  • Prediction row(s)

    Record(s) where the target column carries the [PREDICT] marker. Several columns can be predicted at the same time.

A diagram showing the prediction workflow from SAP HANA Cloud, to SAP AI Core and back to SAP HANA Cloud, centered around an SQLScript Stored Procedure that brings together context data, predict table, and the call to the SAP-RPT model.

When executed, the model returns a predicted value plus a confidence score (for classification) or a confidence interval (for regression).

Example Table for an SAP-RPT Request: Late Delivery Prediction

order_idcustomer_tiertotal_amountdelivery_dayslate_delivery
1001gold25020 (context row)
1002silver12061 (context row)
1003gold3003[PREDICT]

The execution of the RPT workflow will predict, for the third row (order_id 1003) marked with "[PREDICT]", a classification ("will the delivery happen with delay or not?"), and also generate a confidence score for this prediction.

RPT-1.6 Requirements and Limits

  • At least 2 complete context rows to establish a pattern.
  • Better results can be achieved with more context rows (up to the context window limit).
  • RPT-1.6 supports up to 512 prediction rows and 10 target columns per API call.

Variants of SAP-RPT-1.6 Models

The table below describes which variants of SAP-RPT-1.6 are available, so that you can identify which one is the best fit for your requirements.

SAP-RPT-1.6 Variants

ValueSAP-RPT-1.6SAP-RPT-1.6-large
Max. accepted context length2,048 rows65,536 rows
Max. columns100 columns256 columns
Max. prediction rows per API call512512
Recommended context length (based on SAP experimentation)500–2,000 rows4,000–36,000 rows
Recommended max. no. of target classes (classification only)~250~5,000

Auto-Context (Tabular Context Registry) Option

The auto-context with Tabular Context Registry (TCR) option delegates the choice of context rows to the model (so, there is no need to determine context rows yourself); it allows an unlimited set of context rows. This is useful when datasets exceed context window limitations.

Note

The auto-context with TCR option can be applied on top or either SAP-RPT-1.6 or SAP-RPT-1.6-large.

Leveraging the Deep context mode with SAP-RPT-1.6-large

The SAP-RPT-1.6-large variant also comes with a Context mode option. You can keep this option to Default, or choose Deep when the number of context rows is greater than 8,000 and you need more accuracy for predicted rows.

What SAP-RPT Can (and Cannot) Do

Supported Prediction Scenarios

Task typeDescriptionExample
ClassificationPredicts a category label + confidence score (0 to 1)Payment risk tier (Low / Medium / High)
RegressionPredicts a numeric value + confidence intervalEstimated days late

As of SAP-RPT-1.6, the model does not support other/more elaborate scenarios, such as clustering, time series, and so on.

When to Use RPT vs. the Alternatives

CriterionSAP-RPTTraditional Machine LearningLarge Language Model
Data typeStructured / tabularStructured / tabularUnstructured text
OutputLabel or numberLabel, number, scoreText
Training neededNoYesNo (prompting)
Setup effortLowHighMedium
Cold-start (little data)StrongWeakN/A
Feature engineeringNoneManualN/A
Hallucination riskVery lowNoneHigh
Time series / clusteringNot supportedSupportedNot supported
Latency target < 200 msNot recommendedSupportedNot supported
Max accuracy (large data)Good (91–96% of tuned Gradient Boosted Decision Trees)MaximumPoor on tables

Hint

In the SAP Architecture Center, the SAP's AI Golden Path (Build and Deliver→Classic ML Scenarios) suggests to start with SAP-RPT for classification or regression on tabular data, and to fall back to traditional Machine Learning only if you need a specialized algorithm, time-series processing, or sub-200 ms latency.

How to Go First for Quick Experiment

SAP offers a free SAP-RPT-1.6 experience with its SAP-RPT Playground, at https://rpt.cloud.sap/data-tablehttps://rpt.cloud.sap/. You can work with a data sample automatically generated for you, or bring you own data sample (CSV file) to experiment on your data.

In addition to the User Interface, the SAP-RPT Playground also allows you to call the model with API calls, using your own scripts. Use the API documentation available here: https://rpt.cloud.sap/docs/api/overview.

A screenshot of the SAP-RPT Playground User Interface showing a table with some rows to predict, as well as a chat panel commenting on the scenario and its data, and providing guidance.

Note

The SAP-RPT Playground is not for productive use. It works on a limited version of RPT. In particular, for the UI flavor: you can have a maximum of 25 context rows, 4 predict columns and 50 classes. Still, the public API supports up to 1,000 requests per hour. The SAP AI Core Trial offers higher limits, but still a limitation in terms of SLA, support, resource groups or concurrent deployments.

With the 3-row example table shown in a previous section, a typical call to the SAP-RPT Playground API would look as follows:

Code Snippet
123456789101112131415161718192021222324252627282930313233
POST https://rpt.cloud.sap/api/predict Authorization: Bearer <YOUR_PLAYGROUND_TOKEN> Content-Type: application/json { "rows": [ { "order_id": "1001", "customer_tier": "gold", "total_amount": 250, "delivery_days": 2, "late_delivery": 0 }, { "order_id": "1002", "customer_tier": "silver", "total_amount": 120, "delivery_days": 6, "late_delivery": 1 }, { "order_id": "1003", "customer_tier": "gold", "total_amount": 300, "delivery_days": 3, "late_delivery": "[PREDICT]" } ], "index_column": "order_id", "explanations": { "top_column_scores": 3, "top_relevant_context_rows": 2 } }

Note

The "index_column": "order_id" entry is here to specify an index to keep for the returned rows. Additionally, it is essential to remove the order_id column from the context. If it were kept, that would become a feature column and could hamper the prediction process.

Technical Access Paths

There are different ways to call SAP-RPT from your application or database. Let's consider two of them:

How to Include SAP-RPT in your Business Scenario

 Path A – SAP AI CorePath B – HANA Cloud SQL
Primary useSAP Business AI Platform apps, CAP services, any REST clientData-centric scenarios, data analysts
InterfaceREST API ( /v2/inference/deployments/{id}/predict )SQL stored procedure / PREDICT function
Requires AI Core instanceYesYes (HANA Cloud connects via X.509 service key)
Best forApplication integration, agentic workflowsIn-database analytics, no data movement

Key Limitations to Know

  • RPT supports classification and regression only — not time series, clustering, anomaly detection, or recommendations (use PAL/APL for those).
  • It is NOT suitable for real-time scenarios with latency requirements below 200 ms (GPU-based inference).
  • The context window can go up to 2,048 rows (RPT-1.6) to 65,536 rows (RPT-1.6 large). In addition, using the auto-context mode with Tabular Context Registry (TCR) allows you to use more rows if the scenario requires it.
  • SAP-RPT-1 is retired on December 31, 2026. Use RPT-1.6 for all new projects.

Where to Go Next

Below, you find a few resources to go further in your SAP-RPT discovery journey

ResourceLink
SAP-RPT product pagesap.com/products/artificial-intelligence/sap-rpt
Free playground (no account needed)rpt.cloud.sap
Developer tutorial (AI Core step-by-step)Strategic Prediction with SAP RPT-1 — developers.sap.com
AI Golden Path (architecture guidance)Predictive Tabular AI — architecture.learning.sap.com
Hello World with AI Core (community blog)community.sap.com
SAP-RPT-1 Open-source model (HuggingFace)SAP/sap-rpt-1-oss
SAP-RPT-1.6 model cardSAP-RPT-1.6 Model Card on sap.com