Skip to content

Class ZCL_LLM_TEMPLATE_PARSER

AI Generated documentation.

Overview

The ZCL_LLM_TEMPLATE_PARSER class implements a template engine inspired by limited Jinja2 syntax, supporting variable substitution, control structures (conditionals, loops), filters, and error handling. Key functionalities include:

  • Variable resolution for flat/nested structures and tables.
  • Control structures: if/elif/else, for loops with metadata (e.g., loop.last).
  • Filters for string transformations (upper, lower, default).
  • Error detection for invalid paths, unclosed tokens, and type mismatches.

Public Methods

Method Purpose
ADD_TEMPLATE Registers templates with optional replacement; tokenizes content.
RENDER Generates output by parsing tokens against a data context.

Dependencies

  • Exception class ZCX_LLM_TEMPLATE_PARSER for error reporting.
  • ABAP Runtime Type Services (RTTS) for dynamic data structure resolution.

Details

Tokenization & Parsing Flow

flowchart LR  
    A[Template Input] --> B(Tokenize)  
    B --> C{Token Type?}  
    C -->|Text| D[Add to Output]  
    C -->|Variable| E[Resolve from Context]  
    C -->|Control| F[Update Control Stack]  
    F -->|Conditionals| G[Evaluate Logic]  
    F -->|Loops| H[Iterate Collection]  
    G/H --> I[Generate Output]  
    I --> J[Rendered Result]

Key Implementation Aspects

  1. Tokenization:
  2. Splits templates into TEXT, VARIABLE, CONTROL, and removes comments.

  3. Handles escape sequences (e.g., \{, \}) and multiline text.

  4. Control Stack:

  5. Manages nested IF/FOR structures with state tracking: 

    TYPES: BEGIN OF control_stack_type,  
         type              TYPE string,  " 'IF' or 'FOR'  
         condition_met     TYPE abap_bool,  " For conditionals  
         loop_var          TYPE string,     " Loop variable name  
         collection        TYPE REF TO data," Iterated data  
         loop_tokens       TYPE tokens_type," Nested loop content  
       END OF control_stack_type.

  6. Variable Resolution:

  7. Supports dot notation (address.street) and table indexing (items[1]).
  8. Applies filters (e.g., {{ name | upper }}) during resolution.

  9. Uses RTTS to dynamically navigate structures/tables.

  10. Loop Processing:

  11. Creates modified contexts for iterations, injecting:
    • Loop variable (e.g., item).
    • Metadata (LOOP.index, LOOP.first).

  12. Handles nested loops via token accumulation in loop_tokens.

  13. Error Handling:

  14. Detects unclosed control structures during final stack checks.
  15. Validates table indices and variable paths using RTTS.

Type Resolution Logic

For structured data:

  1. Resolve initial context reference.
  2. Traverse path segments (e.g., user.address.city):
  3. For structures: Access component via ASSIGN COMPONENT.
  4. For tables: Validate index bounds (1-based).
  5. Format tables as comma-separated values or key-value pairs.

Performance Considerations

  • Tokenization occurs once per template (cached in tokens field).
  • Context cloning during loops may impact performance for large datasets.