llm_xml_parser

XML Parser for Structured LLM Output

General Description

This parser implements a stack-based algorithm to analyze XML hierarchies, tracking parent-child relationships and structure depth. It offers flexible configurations for extracting specific tags (both single and lists) at defined depths, with a rigorous validation system. The implementation is optimized for small XML files (typically a few KB) with simple hierarchical structures (maximum 3-4 levels of depth).

Core Features

The parser handles:

• Critical Errors: missing tags, unauthorized duplicates, empty lists
• Non-Blocking Warnings: nested tags in unexpected structures, mixed XML/text content
• Untagged Text Handling: preserves text not enclosed in XML tags
• Case-Insensitive Tag Search: tag names are matched regardless of case.
• Original Text Capitalization: output text retains the exact capitalization from the input.

Key Technical Features

• Exclusive use of Python standard libraries (no external dependencies)
• Case-insensitive only for tag names during search
• Preservation of original text, including capitalization and formatting
• Integrated logging system for warnings and errors

Use Cases and Behaviors

Important Note: The ParseResult structure is flat. All configured tags, including those defined as "children" in the configuration, are accessible as first-level attributes of the result object. There is no nested hierarchical structure.

✅ Correctly Handled Cases

Elements extracted without warnings or errors.

1. Single Element at Depth 0

XML:

<thinking>This is a test</thinking>

Config: {'thinking': 'single'} Output:

result.thinking → "This is a test"

2. List with 2+ Elements

XML:

<step>A</step>
<step>B</step>
<step>C</step>

Config: {'step': 'list'} Output:

result.step → ["A", "B", "C"]

3. Explicitly Configured Hierarchy

XML:

<exercise>
    <question>What is 2+2?</question>
    <answer>The answer is 4</answer>
</exercise>

Config:

config = {
    'exercise': {
        'type': 'single',
        'children': {
            'question': 'single',
            'answer': 'single'
        }
    }
}

Output:

result.exercise → "<question>What is 2+2?</question><answer>The answer is 4</answer>" 
result.question → "What is 2+2?" 
result.answer → "The answer is 4"

4. Correct Untagged Text Extraction

LLM Input:

<thinking>Analysis</thinking>
Answer: 42

After Pre-Processing:

<root>
    <thinking>Analysis</thinking>
    Answer: 42
</root>

Note: Preprocessing always adds an artificial root tag, since LLM output often does not include a valid root tag. The root tag is shown here only to illustrate the handling of unlabeled text.

Output:

result.thinking → "Analysis"
result.untagged → "Answer: 42"

⚠️ Warnings (Extraction with Notification)

Elements extracted with warnings in logs and accessible via ParseResult.warnings

1. Nested Tags in Single Element

XML:

<thinking>Use <formula>E=mc²</formula></thinking>

Config: {'thinking': 'single'} Output:

result.thinking → "Use <formula>E=mc²</formula>"
# Log and warnings: "Unconfigured tag <formula> found inside 'thinking'"

2. Unconfigured Tag in List

XML:

<step>
    A<detail>some info</detail>
</step>
<step>B</step>

Config: {'step': 'list'} Output:

result.step → ["A<detail>some info</detail>", "B"]
# Log and warnings: "Unconfigured tag <detail> found inside 'step'"

3. List with 1 Element

XML:

<step>Single step</step>

Config: {'step': 'list'} Output:

result.step → ["Single step"]
# Log and warnings: "List <step> contains only 1 element."

❌ Blocking Errors

Interrupt execution by raising specific exceptions.

1. Missing Single Element

XML: <answer>42</answer> Config: {'thinking': 'single'} Error:

XMLStructureError: Tag <thinking> not found (single required).

2. Too Many Single Elements

XML:

<answer>42</answer>
<answer>43</answer>

Config: {'answer': 'single'} Error:

XMLStructureError: Multiple <answer> found, but 'single' is required.

3. Empty List

XML: <answer>stuff</answer> Config: {'steps': 'list'} Error:

XMLStructureError: List <steps> is empty (1+ elements required).

4. Malformed XML

XML:

<thinking>Test<thinking>

or

<a><b></a></b>

Error:

XMLFormatError: Unclosed tags remain. Malformed XML structure.

or

XMLFormatError: Mismatched tags: opened <b>, but closed </a>

Unhandled Cases

The parser is designed to be minimalist and efficiently handle simple XML output from LLMs. The following cases are not supported:

• Namespaces

  • ❌ Invalid: <ns:thinking>Test</ns:thinking>

• Tags with Special Characters

  • ✅ Valid: <distro_linux>Test</distro_linux> (underscore supported)
  • ❌ Invalid: <foo-bar>Test</foo-bar>
  • ❌ Invalid: <foo bar>Test</foo bar>

• CDATA and Complex Content

  • CDATA sections not interpreted
  • XML entities not processed
  • ❌ Not handled: <thinking><![CDATA[<test>]]></thinking>

• XML Attributes

  • Tag attributes ignored
  • ❌ Not handled: <tag attr="val">Test</tag>

Implementation Details

Execution Modes

• Normal Mode (Default, strict_mode=False)

  • Warnings logged but non-blocking
  • Only errors generate exceptions
  • Output preserved even with warnings

• Strict Mode (strict_mode=True)

  • All warnings become blocking errors
  • Maximum validation rigor

Input Handling

• Size: Optimized for small XML files (10-20 KB)
• Depth: Efficient support up to 3-4 levels
• Case Sensitivity:
  • Tags: case-insensitive during search
  • Content: preserved exactly as in input

Pre-Processing

• Automatic removal of XML comments
• Automatic root tag addition
• Basic input normalization

Technical Requirements

• Python: Version 3.7 or higher
• Dependencies: Python standard library only
• Encoding: UTF-8 for input/output