POML Standard Implementation

PomlSDK.jl implements the Prompt Orchestration Markup Language (POML) standard developed by Microsoft. This documentation explains how PomlSDK maps to the official POML specification.

What is POML?

POML (Prompt Orchestration Markup Language) is a standard for creating structured prompts for Large Language Models (LLMs). It provides a way to define complex prompt structures with metadata, examples, and multi-modal content, moving beyond simple string-based prompts.

PomlSDK Implementation

PomlSDK.jl provides a programmatic Julia API to build prompts according to the POML standard. The package maps directly to the POML XML structure but provides a more developer-friendly interface.

Core Elements Mapping

POML XML Element	PomlSDK.jl Function	Description
`<role>`	`role()`	Defines a role in the conversation (e.g., System, Assistant)
`<task>`	`task()`	Specifies the main task for the AI to perform
`<example>`	`example()`	Provides a single example for few-shot learning
`<example_set>`	`example_set()`	Groups multiple examples together
`<table>`	`table()`	Creates a markdown-formatted table
`<list>`	`list()`	Creates a list (ordered or unordered)
`<list_item>`	`list_item()`	Creates a list item
`<meta>`	`meta()`	Contains metadata about the prompt
`<meta_tag>`	`meta_tag()`	Adds a key-value metadata tag
`<meta_attribute>`	`meta_attribute()`	Adds a named metadata attribute
`<tool_request>`	`tool_request()`	Requests the use of a specific tool
`<tool_definition>`	`tool_definition()`	Defines a tool that can be used
`<image>`	`image()`	Embeds an image (base64 encoded)
`<document>`	`document()`	References an external document

Document Structure

The POML standard specifies that a document should have the following structure:

<poml>
  <role>...</role>
  <task>...</task>
  <example_set>...</example_set>
  <!-- Other elements -->
</poml>

PomlSDK.jl follows this structure exactly when serializing with dump_poml().

Advanced Features

PomlSDK implements several advanced POML features:

Tool Integration

POML supports tool definitions and requests, which PomlSDK implements through:

# Define a tool
tool_def = tool_definition(p, name="calculator", description="Performs basic math")

# Request a tool
tool_req = tool_request(p, name="calculator", parameters=Dict("operation" => "add", "a" => 5, "b" => 7))

PomlSDK supports embedding images and documents:

# Add an image
img_node = image(p, src="image/png;base64,...", alt="Description")

# Add a document reference
doc_node = document(p, src="data.pdf", parser="pdf")

Metadata

PomlSDK provides comprehensive metadata support:

# Create metadata section
meta_node = meta(p)

# Add tags
meta_tag(p, key="author", value="Your Name")
meta_tag(p, key="source", value="user_input")

# Add attributes
meta_attribute(p, name="version", value="1.0")

Differences from Official Implementation

While PomlSDK closely follows the POML standard, there are some implementation differences due to Julia's language characteristics:

Hierarchy Management: PomlSDK uses add_node!/pop_node! instead of context managers (like Python's with statements)
Attribute Handling: Attributes are handled through keyword arguments rather than separate method calls
Data Type Conversion: Automatic conversion of Julia types to appropriate string representations for XML

Validation

To validate that your generated POML conforms to the standard, you can:

Use an XML validator with the POML schema
Compare against the official POML examples
Process through a POML-compliant LLM system