Gemini-powered business rule engine for SAP

This document describes a reference architecture for building a business rule engine powered by Large Language Models (LLMs) such as Gemini, to automate decision-making within your SAP business applications. It uses the various integration capabilities of ABAP SDK for Google Cloud.

The intended audience for this document includes ABAP developers, SAP solution architects, and cloud architects. The document assumes a basic understanding of ABAP SDK for Google Cloud, Vertex AI SDK for ABAP, and basic introduction to LLMs.

What is a business rule engine (BRE)?

A business rule engine (BRE) is a software system that lets you define, deploy, and execute business rules. These rules are essentially logic statements that dictate how your business operates. Instead of hardcoding the business rules into your applications, a BRE externalizes the logic statements, making it efficient to manage and modify the rules.

A Gemini-powered BRE uses the advanced language processing and reasoning capabilities of Gemini, Google's next-generation AI model, to automate decision-making based on a repository of engineered prompts that function as business rules. These prompts, crafted with specific instructions and expected output formats, guide Gemini in analyzing data, identifying patterns, and making decisions in a manner consistent with predefined business logic.

Architecture

The following diagram shows a reference architecture for creating a Gemini-powered BRE for SAP applications:

This reference architecture includes the following components:

Products used

This reference architecture uses the following Google Cloud products:

• ABAP SDK for Google Cloud: Helps you develop ABAP applications that connect your SAP systems to Google Cloud services, such as Cloud Storage.

• Vertex AI SDK for ABAP: Lets you use Google's Vertex AI within your SAP environment. This simplifies building AI-powered features in your SAP applications.

• Cloud Storage: A cloud service for storing and accessing any amount of data online.

• Vertex AI: Lets you build, deploy, and scale machine learning models for various applications.

Use case

Consider a scenario where your company receives inbound purchase orders (PO) as PDFs and uses an automated process to generate sales orders in SAP. You need to update this process to identify and mitigate risk-shifting clauses found in the delivery comments of incoming POs.

Risk-shifting clauses in a contract are terms that transfer risk from one party to another. Failing to identify and address these clauses can blindside your company with unexpected costs, legal battles, and reputational damage.

The following is an example of a risk-shifting clause written in the delivery comment of a PO document.

"Seller warrants that the products will be free from defects for a period of ten years from the date of delivery."

You need to write a rule for the following:

• Identify and categorize the risk-shifting clauses.
• Provide an explanation on why it is considered as a potential red flag.
• Provide recommendations on how to mitigate the risk.

You can build a BRE with Gemini to generate an intelligent JSON response, which enables your SAP business application to take automated decisions, trigger any process rejection, or initiate downstream human validations.

Sample risk-shifting clauses

The following table lists the sample risk-shifting clauses that can be found in inbound purchase orders:

Sample engineered prompt

The section describes a sample prompt engineered by using the RTF framework. The RTF framework in prompt engineering stands for Role, Task, and Format. It's an effective way to structure your prompts to get better and more consistent results from LLMs such as Gemini.

• Role: Persona that you want the LLM to adopt, such as "sales expert" or "legal advisor".
• Task: Specific action for the LLM, such as, "summarize this text" or "analyze something".
• Format: Desired output structure, such as list or JSON.

To learn about common prompting strategies that you can use to affect the model's responses, see Overview of prompting strategies

The following prompt acts as a defining rule to helps Gemini identify issues such as the risk-shifting clause:

Role

You are a legal advisor for my company.

Task

Analyze the below delivery instruction on a purchase order from my
customer and let me know if it could cause any issues to my company.

Format

Respond to me in plain text JSON format with 4 fields.

Field 1 should be "issue" with possible values Yes or No.

Field 2 should be "issueCategory" with possible values Indemnification
Overreach, Unilateral Termination Clause, Confidentiality Overextension,
Unreasonable Warranty, Ambiguous Governing Law, Unclear Dispute Resolution,
Force Majeure Limitation, Unbalanced Liability,
Intellectual Property Ambiguity, Compliance with Sanctions, Others or None.

Field 3 should be "explanation" - use this field to give a short
explanation for your response on Field 1.

Field 4 should be "recommendation" - use this field to give a short
recommendation of how to mitigate such issues.

Do not include backticks in the response.

Delivery Instruction: "Seller warrants that the products will be free
from defects for a period of ten years from the date of delivery."

{
  "issue": "Yes",
  "issueCategory": "Unreasonable Warranty",
  "explanation":   "A ten-year warranty is significantly longer than
  industry standard for most products. This could expose your company to
  substantial financial risk if issues arise with the products years
  down the line.",
  "recommendation": "Negotiate a shorter warranty period that aligns
  with industry standards and your product's expected lifespan.
  Offer extended warranty options at an additional cost if the
  customer requires longer coverage."
}

You can store this prompt in a repository such as Cloud Storage. The input data can be passed by the SAP transactions to the BRE module.

You then use the ABAP SDK for Google Cloud to retrieve the prompt from Cloud Storage and pass it to Gemini for execution, by using the Vertex AI SDK for ABAP.

The sample JSON response is generated by Gemini for this risk-shifting clause example.

The following sample code shows how to retrieve the rules and use it in your code. It assumes that you're implementing a BRE service as an ABAP class, such as lcl_bre.

 DATA(lo_bre) = NEW lcl_bre( iv_key_name  = 'CLIENT_KEY'
                             iv_model_key = 'MODEL_KEY' ).

 " Sample delivery text
 DATA(lv_delivery_text) = |Seller warrants that the products will be free from defects for| &&
                          |a period of ten years from the date of delivery.|.

 " Execute the rule for identifying Risk-Shifting clauses.
 DATA(lv_response) = lo_bre->execute_rule( iv_rule_id = 'RiskShiftingClause-Rule-001.prompt'
                                           iv_data    = lv_delivery_text ).
 " JSON response
 cl_demo_output=>display_json( lv_response ).

Replace the following:

• CLIENT_KEY: The client key configured for authentication to Vertex AI and Cloud Storage.
• MODEL_KEY: The model key name to access the LLM, which is configured in the model generation parameters.

The execute_rule method performs the following:

1. Reads the rule file from Cloud Storage by using the /GOOG/CL_STORAGE_V1 class.

2. Sends the processing rules to Gemini, which is set as instruction to the model, by using the /GOOG/CL_GENERATIVE_MODEL class. The model then generates a JSON response.

The following is the complete sample code for your reference:

REPORT zra_bre_example.

" Local class definition
CLASS lcl_bre DEFINITION FINAL.

  PUBLIC SECTION.
    METHODS constructor
      IMPORTING iv_key_name  TYPE /goog/keyname
                iv_model_key TYPE /goog/model_key
      RAISING   /goog/cx_sdk.

    METHODS convert_xstring_to_string
      IMPORTING iv_xstring       TYPE xstring
      RETURNING VALUE(iv_string) TYPE string.

    METHODS execute_rule
      IMPORTING iv_rule_id              TYPE string
                iv_data                 TYPE string
      RETURNING VALUE(rv_response_text) TYPE string
      RAISING   /goog/cx_sdk.

  PRIVATE SECTION.
    DATA mo_storage TYPE REF TO /goog/cl_storage_v1.
    DATA mo_model   TYPE REF TO /goog/cl_generative_model.
ENDCLASS.

" Local class implementation
CLASS lcl_bre IMPLEMENTATION.
  METHOD constructor.
    mo_storage = NEW /goog/cl_storage_v1( iv_key_name = iv_key_name ).
    mo_model = NEW /goog/cl_generative_model( iv_model_key = iv_model_key ).
  ENDMETHOD.

  METHOD convert_xstring_to_string.
DATA lv_file_length TYPE i.
    DATA lt_bin_data    TYPE STANDARD TABLE OF char1024.

    " Call function module to convert xstring to binary format
    CALL FUNCTION 'SCMS_XSTRING_TO_BINARY'
      EXPORTING buffer        = iv_xstring
      IMPORTING output_length = lv_file_length
      TABLES    binary_tab    = lt_bin_data.
    IF sy-subrc <> 0.
      " Handle error
    ENDIF.

    " Call function module to convert from binary to string format
    CALL FUNCTION 'SCMS_BINARY_TO_STRING'
      EXPORTING input_length = lv_file_length
      IMPORTING text_buffer  = iv_string
      TABLES    binary_tab   = lt_bin_data.
    IF sy-subrc <> 0.
      " Handle error
    ENDIF.
  ENDMETHOD.

  METHOD execute_rule.
    DATA lv_xdata       TYPE xstring.
    DATA lv_instruction TYPE string.

    " Get the rule file from rule repository(cloud storage)
    mo_storage->add_common_qparam( iv_name  = 'alt'
                                   iv_value = 'media' ).

    mo_storage->get_objects( EXPORTING iv_p_bucket = 'gemini-bre-repo'
                                       iv_p_object = iv_rule_id
                             IMPORTING es_output   = DATA(ls_output)
                                       ev_ret_code = DATA(lv_ret_code)
                                       ev_err_text = DATA(lv_err_text)
                                       es_err_resp = DATA(ls_err_resp)
                                       es_raw      = lv_xdata ).

    IF mo_storage->is_success( lv_ret_code ) = abap_true.
      " Set the instruction for Gemini as per the configured rule
      lv_instruction = convert_xstring_to_string( iv_xstring = lv_xdata ).
    ELSE.
      " Handle error
    ENDIF.

    " Set the text to be analyzed
    DATA(lv_prompt) = |Please analyze the Delivery Instruction:{ cl_abap_char_utilities=>newline }{ iv_data }|.

    " Call Gemini to process the rule and identify Risk-Shifting clauses
    rv_response_text = mo_model->set_system_instructions( iv_text = lv_instruction
                      )->add_safety_settings( iv_harm_category        = 'HARM_CATEGORY_DANGEROUS_CONTENT'
                                              iv_harm_block_threshold = 'BLOCK_NONE'
                      )->generate_content( iv_prompt_text = lv_prompt
                      )->get_text( ).
  ENDMETHOD.
ENDCLASS.

START-OF-SELECTION.
  TRY.
      DATA(lo_bre) = NEW lcl_bre( iv_key_name  = 'DEMO_AIPLATFORM'
                                  iv_model_key = 'gemini-flash' ).

      " Sample delivery text
 DATA(lv_delivery_text) = |Seller warrants that the products will be free from defects for| &&
                             |a period of ten years from the date of delivery.|.

      " Execute the rule for identifying Risk-Shifting clauses.
      DATA(lv_response) = lo_bre->execute_rule( iv_rule_id = 'RiskShiftingClause-Rule-001.prompt'
                                                iv_data    = lv_delivery_text ).
      " JSON response
      cl_demo_output=>display_json( lv_response ).

    CATCH /goog/cx_sdk INTO DATA(lo_exception).
      DATA(lv_msg) = lo_exception->get_text( ).
      " Handle error
  ENDTRY.

Design considerations

This section provides guidance to help you use this reference architecture to develop one or more architectures that help you to meet your specific requirements for security, privacy, compliance, cost, and performance.

Security, privacy, and compliance

Security and compliance are shared responsibilities. For detailed information, see Vertex AI shared responsibility.

For information about how Gemini uses your data, see How Gemini for Google Cloud uses your data.

Cost

For an estimate of the cost of the Google Cloud resources like Cloud Storage, use the Google Cloud Pricing Calculator.

For Vertex AI and Gemini API related pricing, see Vertex AI pricing.

Design alternative

While this document focuses on the on-premises or any cloud edition of ABAP SDK for Google Cloud, you can achieve similar results by using the SAP BTP edition of ABAP SDK for Google Cloud. The Vertex AI SDK for ABAP is not available for the SAP BTP environment, but you can use the code sample provided in this document for the on-premises edition of the SDK, adjust it, and build similar solutions within your SAP BTP environment.

What's next

• To learn about the Vertex AI SDK for ABAP, see Overview of the Vertex AI SDK for ABAP.
• To learn about the ABAP SDK for Google Cloud, see Overview of the ABAP SDK for Google Cloud.

• If you need help resolving problems with the ABAP SDK for Google Cloud, then do the following:

  • Refer to the ABAP SDK for Google Cloud troubleshooting guide.
  • Ask your questions and discuss the ABAP SDK for Google Cloud with the community on Cloud Forums.
  • Collect all available diagnostic information and contact Cloud Customer Care. For information about contacting Customer Care, see Getting support for SAP on Google Cloud.

Contributors

Authors: Ameya Suvarna | SAP Application Engineer Team Lead

Other contributors: Vikash Kumar | Technical Writer