Introduction

Enterprises running SAP S/4HANA are accustomed to delivering real‑time, transaction‑ready data to their users. Yet, when a business user asks a “natural‑language” question—“Which customers have not paid their invoices for more than 30 days?”—the system still forces them into a series of menu clicks, report selections, or custom ABAP queries.

Retrieval‑Augmented Generation (RAG) bridges that gap. By coupling a large language model (LLM) with a domain‑specific knowledge base, RAG can surface precise answers while grounding the response in the most recent SAP data. This blog post walks you through a hands‑on, end‑to‑end implementation that empowers S/4HANA users to ask conversational questions and receive accurate, actionable results directly within Fiori.

Who should read this?

• SAP Basis & Cloud Platform engineers who can provision AI services.
• ABAP developers who will build the front‑end integration.
• Business analysts interested in the technical feasibility of AI‑driven query assistance.

---

1. Why RAG Makes Sense for S/4HANA

Traditional Approach	RAG‑Enabled Approach
Pre‑defined reports, dashboards, and custom ABAP queries.	Users type natural‑language questions; the system retrieves relevant documents, runs a live query, and generates a concise answer.
Static data extracts or batch‑loaded knowledge graphs.	Real‑time vector embeddings from SAP HANA, guaranteeing up‑to‑the‑minute accuracy.
High learning curve for non‑technical users.	Conversational UI reduces training overhead and accelerates decision making.

Key benefits:

1. Contextual relevance – The LLM is “augmented” with actual SAP tables, CDS views, and configuration data, preventing hallucinations.
2. Scalability – Vector stores live in SAP HANA, leveraging the same in‑memory engine that powers S/4HANA.
3. Compliance – All data stays within the customer’s SAP Cloud Platform (SCP) tenant; no external data exfiltration.

---

2. High‑Level Architecture

graph TD
    A[User (Fiori)] --> B[ABAP OData Service]
    B --> C[AI Core RAG Service]
    C --> D[Vector Store (SAP HANA)]
    D --> E[Live S/4HANA Queries (ABAP CDS Views)]
    C --> F[LLM (e.g., SAP AI Foundation Model)]
    style A fill:#E3F2FD,stroke:#90A4AE,stroke-width:2px
    style B fill:#FFF3E0,stroke:#FFB74D,stroke-width:2px
    style C fill:#E8F5E9,stroke:#66BB6A,stroke-width:2px
    style D fill:#F3E5F5,stroke:#AB47BC,stroke-width:2px
    style E fill:#E0F7FA,stroke:#26C6DA,stroke-width:2px
    style F fill:#FFFDE7,stroke:#FDD835,stroke-width:2px

Components

Component	Role
Fiori UI	Front‑end where users type questions.
ABAP OData Service	Thin wrapper that forwards the request to the RAG endpoint and returns the LLM‑generated answer.
SAP AI Core (RAG Service)	Orchestrates retrieval from the vector store, runs the LLM, and merges results.
Vector Store (HANA)	Stores embeddings of SAP documentation, CDS view metadata, and optionally historic transactional snapshots.
Live S/4HANA Queries	Executed on‑the‑fly to fetch up‑to‑date data for the final answer.
Foundation Model	Pre‑trained LLM (e.g., SAP’s “BTP‑LLM‑4”) fine‑tuned for SAP terminology.

---

3. Prerequisites

Item	Minimum Version / Service
SAP S/4HANA	2022 SPS 04 or later (ABAP 7.55+)
SAP Business Technology Platform (BTP)	Subaccount with AI Core and AI Launchpad enabled
SAP HANA Cloud	Service instance with Vector Engine (available from HANA 2.0 SPS 05)
ABAP Development Tools (ADT)	Eclipse 2022‑12 or VS Code with ABAP extension
Fiori Elements	UI5 version ≥ 1.108
Optional: SAP AI Business Services	For content‑safety & moderation

Tip: Use the BTP CLI (btp) to provision services programmatically; the commands are listed in the “Provisioning” section below.

---

4. Step‑by‑Step Implementation

4.1. Provision AI Core and HANA Vector Store

# Log in to BTP CLI
btp login --url https://cpcli.cf.sap.hana.ondemand.com

# Create a subaccount (skip if you already have one)
btp create accounts/subaccount my-s4-rag --region eu10 --subdomain my-s4-rag

# Enable AI Core service
btp create service-instance ai-core standard my-ai-core -c '{ "plan":"standard" }'

# Enable HANA Cloud with Vector Engine
btp create service-instance hana-cloud hana-vector my-hana-vector -c '{
    "plan":"vector",
    "configuration": {
        "memorySize": "32GB",
        "region": "eu10"
    }
}'

After provisioning, note the service keys (btp get-service-key) for both AI Core and HANA. You will need the url, clientid, and clientsecret for authentication.

4.2. Create a Vector Index for SAP Knowledge

1. Collect source documents – Technical documentation, SAP Help PDFs, CDS view definitions, and a curated set of business questions.
2. Generate embeddings – Use the AI Core Embedding endpoint (/v2/embeddings).

import requests, json, base64, pathlib

# Load service credentials (example from service key)
creds = json.load(open('ai-core-key.json'))
token_url = f"{creds['url']}/oauth/token"
auth = (creds['clientid'], creds['clientsecret'])
token = requests.post(token_url, data={'grant_type':'client_credentials'}, auth=auth).json()['access_token']

def embed(text):
    payload = {
        "model": "text-embedding-ada-002",   # SAP‑provided embedding model
        "input": text
    }
    r = requests.post(f"{creds['url']}/v2/embeddings", headers={'Authorization': f"Bearer {token}"}, json=payload)
    return r.json()['data'][0]['embedding']

# Example: embed a CDS view description
cds_path = pathlib.Path('src/zcustomer_invoice.cds')
with cds_path.open() as f:
    source = f.read()
embedding = embed(source)

# Insert into HANA Vector Store (SQL)
sql = """
INSERT INTO VECTOR_STORE (DOC_ID, EMBEDDING, METADATA)
VALUES ('ZCUSTOMER_INVOICE', :embedding, :metadata);
"""
# Use pyhdb or sqlalchemy-hana to execute the SQL; omitted for brevity

Best practice: Batch‑process all documents nightly and keep the vector store incrementally updated. Use a timestamp column (LAST_MODIFIED) to identify new or changed assets.

4.3. Build the RAG Service on AI Core

Create a RAG pipeline that wires together three components:

Step	Component	Action
1	Retriever	vector-retriever – fetches top‑k similar documents from HANA.
2	Augmenter	sql-augmenter – runs a live CDS view based on retrieved metadata.
3	Generator	foundation-model – produces the final natural‑language answer.

Pipeline definition (JSON) – Save as rag-pipeline.json.

{
  "name": "s4hana-query-rag",
  "description": "RAG pipeline for conversational S/4HANA queries",
  "steps": [
    {
      "name": "retrieve_docs",
      "type": "vector-retriever",
      "configuration": {
        "vectorStore": {
          "type": "hana",
          "serviceKey": "hana-vector-key.json"
        },
        "topK": 5,
        "embeddingModel": "text-embedding-ada-002"
      }
    },
    {
      "name": "run_live_query",
      "type": "sql-augmenter",
      "configuration": {
        "datasource": {
          "type": "abap",
          "serviceKey": "s4hana-abap-key.json"
        },
        "queryTemplate": "SELECT * FROM {entity} WHERE {filter}"
      }
    },
    {
      "name": "generate_answer",
      "type": "foundation-model",
      "configuration": {
        "model": "sap-llm-4",
        "maxTokens": 256,
        "temperature": 0.2,
        "systemPrompt": "You are an SAP expert. Answer the user's question using only the retrieved data and live query results. Cite sources."
      }
    }
  ]
}

Deploy the pipeline:

btp create service-instance ai-core-rag rag-pipeline -c rag-pipeline.json

The AI Core UI will now expose an endpoint like:

POST https://ai-core.<region>.hana.ondemand.com/v2/pipelines/s4hana-query-rag/invoke

4.4. Expose a Lightweight ABAP OData Service

Create a service definition (Z_RAG_QUERY_SRV) and a service binding (Z_RAG_QUERY_SRV_OData). The service will have a single entity set Queries with an POST operation.

@AbapCatalog.sqlViewName: 'ZV_RAG_QRY'
@AbapCatalog.preserveKey: true
@EndUserText.label: 'RAG Query Input'
define view Z_RAG_QUERY_INPUT as select from dummy {
    key 'Q' as ID,
    @Semantics.text: true
    $session.user as USER,
    cast( $session.client as abap.int4 ) as CLIENT,
    @UI.lineItem: [{ position: 10 }]
    '' as QUESTION,   "filled at runtime
    '' as ANSWER      "filled after AI call
}

ABAP class to call AI Core (ZCL_RAG_QUERY_HANDLER):

CLASS zcl_rag_query_handler DEFINITION PUBLIC FINAL CREATE PUBLIC.
  PUBLIC SECTION.
    METHODS:
      invoke
        IMPORTING
          iv_question TYPE string
        RETURNING
          VALUE(rv_answer) TYPE string
        RAISING
          cx_http_communication_failure,
          cx_http_invalid_state.
ENDCLASS.

CLASS zcl_rag_query_handler IMPLEMENTATION.
  METHOD invoke.
    "--- 1. Get AI Core token -------------------------------------------------
    DATA(lo_http) = cl_http_client=>create_by_destination( 'AI_CORE_DEST' ).
    lo_http->request->set_method( if_http_request=>co_request_method_post ).
    lo_http->request->set_content_type( 'application/x-www-form-urlencoded' ).
    lo_http->request->set_cdata( |grant_type=client_credentials| ).
    lo_http->send( ).
    lo_http->receive( ).
    DATA(lv_body) = lo_http->response->get_cdata( ).
    DATA(ls_token) TYPE string.
    cl_json=>deserialize(
      EXPORTING json = lv_body
      RECEIVING  data = ls_token ).

    "--- 2. Build RAG payload -------------------------------------------------
    DATA(ls_payload) TYPE string.
    ls_payload = /ui2/cl_json=>serialize(
      data = VALUE #( question = iv_question )
      pretty_name = abap_false ).

    "--- 3. Call RAG pipeline -------------------------------------------------
    lo_http->reset( ).
    lo_http->request->set_method( if_http_request=>co_request_method_post ).
    lo_http->request->set_uri( '/v2/pipelines/s4hana-query-rag/invoke' ).
    lo_http->request->set_header_field( |Authorization|, |Bearer | && ls_token ).
    lo_http->request->set_content_type( 'application/json' ).
    lo_http->request->set_cdata( ls_payload ).
    lo_http->send( ).
    lo_http->receive( ).

    "--- 4. Extract answer ----------------------------------------------------
    DATA(lv_resp) = lo_http->response->get_cdata( ).
    DATA(ls_result) TYPE string.
    cl_json=>deserialize(
      EXPORTING json = lv_resp
      RECEIVING data = ls_result ).

    rv_answer = ls_result-answer. "Assumes JSON { "answer": "..." }
  ENDMETHOD.
ENDCLASS.

Expose via OData – Map the POST /Queries to call ZCL_RAG_QUERY_HANDLER=>INVOKE. The OData response contains:

{
  "question": "Which customers have overdue invoices?",
  "answer": "As of 2026‑02‑09, 42 customers have invoices overdue >30 days. The top three are: 100012 (EUR 12,340), 100045 (EUR 9,210), 100078 (EUR 8,975)."
}

4.5. Integrate with Fiori Elements

Create a Fiori Elements List Report (RAGQueryList) that shows a single input field and a result pane.

<!-- view/Query.view.xml -->
<Page id="page" title="AI‑Powered Query">
  <content>
    <VBox>
      <Input id="questionInput" placeholder="Ask a question…" liveChange="onLiveChange"/>
      <Button text="Ask" press="onAsk" enabled="{/isReady}"/>
      <ObjectStatus id="answerBox" text="{/answer}" state="Success"/>
    </VBox>
  </content>
</Page>

Controller (ES6) – calls the OData service:

onAsk() {
  const question = this.byId("questionInput").getValue();
  this.getView().setModel(new JSONModel({ isReady: false }));
  this.getModel().create("/Queries", { QUESTION: question }, {
    success: (data) => {
      this.byId("answerBox").setText(data.ANSWER);
      this.getView().getModel().setProperty("/isReady", true);
    },
    error: (err) => {
      MessageBox.error("AI service failed: " + err.message);
      this.getView().getModel().setProperty("/isReady", true);
    }
  });
}

Deploy the UI component to the SAP Launchpad; end users now have a single entry point for any analytical or transactional question.

---

5. Actionable Insights & Best Practices

5.1. Keep the Vector Store Fresh

• Incremental Updates – Schedule a nightly job (e.g., using SAP Cloud Platform Job Scheduler) that diffs source files and updates embeddings only for changed artifacts.
• Versioning – Store a VERSION column alongside each embedding; when you roll out a new data model, deprecate old vectors but keep them for auditability.

5.2. Guard Against Hallucinations

1. System Prompt Discipline – The system prompt in the pipeline (see systemPrompt above) explicitly instructs the model to cite sources.
2. Post‑Processing – In the ABAP handler, verify that the answer contains at least one reference to a known document ID (DOC_ID). If missing, return a fallback “I’m not sure; please refine the question.”

IF rv_answer CS 'DOC_ID:'.
  "OK
ELSE.
  rv_answer = |I could not locate a reliable source for that question.|
ENDIF.

5.3. Performance Tuning

Bottleneck	Mitigation
Vector retrieval latency	Ensure HANA vector index uses IVF‑PQ or HNSW for sub‑millisecond ANN search.
LLM inference time	Choose a smaller, instruction‑tuned model for simple lookup questions; fall back to a larger model only when the query is ambiguous.
ABAP‑to‑AI round‑trip	Enable HTTP/2 on the destination and reuse the OAuth token for the life of the user session.

5.4. Security & Compliance

• Scope‑limited OAuth – Grant the AI Core destination only the ai_core.rag.invoke scope.
• Data Masking – If the query may expose PII (e.g., customer names), add a masking step in the run_live_query augmenter to replace sensitive fields before they reach the LLM.

SELECT CUSTOMER_ID,
       CASE WHEN :MASK_PII = 'X' THEN '*****' ELSE NAME END AS NAME,
       AMOUNT
FROM Z_INVOICE
WHERE DUE_DATE < CURRENT_DATE - 30

---

6. Key Takeaways

• RAG adds factual grounding to LLM‑driven assistants, eliminating the “hallucination” problem for SAP data.
• SAP HANA’s vector engine provides ultra‑low‑latency similarity search, keeping the retrieval step on‑par with S/4HANA’s performance expectations.
• A lightweight ABAP OData façade isolates the AI Core integration, allowing you to reuse existing security and transport mechanisms.
• Fiori Elements can surface the entire pipeline with minimal UI code, delivering a seamless conversational experience.