Pinja is a Jinja-subset-compatible template engine for Object Pascal. It's primarily designed for usage with PasLLM for Large Language Model applications, since LLM models have often Jinja templates built-in for prompt processing. But it can be used in any Object Pascal application for other purposes as well, where a lightweight Jinja-like template engine is needed.

A lot of Jinja features are supported, including:

• Expression Evaluation: Full expression support with proper operator precedence

  • Arithmetic operators: +, -, *, /, // (floor division), % (modulo), ** (power)
  • Comparison operators: ==, !=, <, <=, >, >=
  • Logical operators: and, or, not
  • Bitwise operators: &, ^, bor, <<, >>, ~
  • String concatenation: ~
  • Membership operators: in, not in
  • Identity tests: is, is not
  • Ternary expressions: value if condition else alternative

• Control Flow Structures

  • Conditional statements: {% if %}, {% elif %}, {% else %}, {% endif %}
  • For loops: {% for item in items %} with optional filter conditions
  • Loop control: {% break %}, {% continue %}
  • Filter blocks: {% filter name %}...{% endfilter %}
  • Generation blocks: {% generation %}...{% endgeneration %}

• Variable Management

  • Variable assignment: {% set name = value %}
  • Set blocks: {% set name %}...{% endset %}
  • Attribute access: object.attribute
  • Array/object indexing: array[index]
  • Slicing support: array[start:stop:step]

• Macros and Callable Functions

  • Macro definitions: {% macro name(params) %}...{% endmacro %}
  • Positional and keyword arguments
  • Call blocks: {% call macro_name(args) %}...{% endcall %}
  • Call parameters: {% call(params) macro_name(args) %}...{% endcall %}

• Comprehensive Built-in Filters (40+ filters)

  • String manipulation: upper, lower, trim, strip, capitalize, title, replace, center, truncate
  • String tests: startswith, endswith, count
  • List operations: first, last, join, sort, reverse, unique, slice, batch, random
  • Formatting: format, indent, wordcount, striptags, urlize, pprint
  • Type conversions: int, str/string, list, to_json
  • Data manipulation: default, length, dictsort, groupby
  • Escaping: escape/e, safe, xmlattr
  • Math: round

• Built-in Functions (20+ functions)

  • Type conversions: int(), float(), str(), bool(), list()
  • Math operations: abs(), min(), max(), sum(), round()
  • Sequence operations: len(), range(), keys(), values(), items()
  • Filtering: select(), reject(), selectattr(), rejectattr(), map()
  • Utilities: namespace(), joiner(), tojson(), equalto(), in(), raise_exception()

• Data Types

  • Null values
  • Booleans
  • Integers (64-bit)
  • Floating-point numbers
  • Strings (UTF-8 encoded)
  • Arrays (dynamic lists)
  • Objects (key-value dictionaries)
  • Callable objects (functions and macros)

• Safety and Escaping

  • HTML escaping for safe output
  • JSON serialization support
  • Safe filter to bypass auto-escaping

• LLM Integration

  • Compatible with Jinja2/3 chat template formats used by Large Language Models
  • Separate PinjaChatTemplate unit for LLM-specific functionality
  • Designed for prompt processing in AI applications

• Object Pascal Features

  • Compatible with Delphi ≥11.2 and FreePascal ≥3.3.1
  • No external dependencies except PasJSON and PasDblStrUtils
  • Cross-platform support (32-bit and 64-bit)
  • zlib license for flexible usage

Missing features compared to full Jinja2:

• No support for asynchronous templates
• Limited support for custom filters and tests
• Some advanced features like template inheritance are not fully implemented
• Inclusion of external templates is not supported
• And some more... but for LLM prompt processing, this is usually sufficient and already more than needed.

It's licensed under the zlib license. See the source code headers for details.