PomlSDK.jl This package implements the Prompt Orchestration Markup Language (POML) standard developed by Microsoft Research : POML specification.

This is a Julia package for creating structured prompts for Large Language Models (LLMs) using a Prompt Object Model Language (POML)-like approach. It provides a programmatic API to build complex, hierarchical prompt structures with metadata, examples, and multi-modal content, moving beyond simple string-based prompts.

You can install it directly from GitHub:

using Pkg
Pkg.add(url="https://github.com/imohag9/PomlSDK.jl")

PomlSDK offers a fluent API to construct prompts as structured object models. This allows for better organization, reusability, and programmatic manipulation of prompts before they are serialized into a textual format suitable for LLMs.

Key features include:

• Hierarchical Structure: Build prompts using a tree of nodes (roles, tasks, examples, etc.).
• Rich Content Types: Support for text, tables (markdown), lists, images (base64), and metadata.
• Tool Integration: Define and request tools/functions with structured parameters.
• Metadata Support: Attach source information, attributes, and annotations.
• Serialization: Convert the structured prompt object into a POML-like string format.

The core workflow involves creating a Prompt object and then using specialized functions to add tags and content to it. The add_node! and pop_node! functions manage the current position within the prompt hierarchy.

using PomlSDK

# 1. Create a new Prompt
p = Prompt()


# 2 Add a structured component, like a role
role_node = role(p, caption="System")
add_node!(p, role_node) # Make 'role_node' the current parent
add_text(p, "Adhere to the highest standards of accuracy.")
pop_node!(p) # Move back to the previous parent (the root)

# 3. Add a task
task_node = task(p, priority="high", caption="Data Summary")
add_node!(p, task_node)
add_text(p, "Summarize the provided sales data.")

# 4. Add an example set for the task
example_set_node = example_set(p)
add_node!(p, example_set_node)

example_node = example(p)
add_node!(p, example_node)
# Add input and output parts (simplified - often nested tags)
input_node = tag(p, "input") # Using generic tag
add_node!(p, input_node)
add_text(p, "Q1 Sales: $1.2M, Q2 Sales: $1.5M")
pop_node!(p)

output_node = tag(p, "output")
add_node!(p, output_node)
add_text(p, "Sales increased by 25% from Q1 to Q2.")
pop_node!(p)
pop_node!(p) # Pop example

# Add another example if needed...

pop_node!(p) # Pop example_set
pop_node!(p) # Pop task

# 5. Add a table with data
sales_data = [
    ["Quarter", "Sales (USD)"],
    ["Q1", "1200000"],
    ["Q2", "1500000"],
    ["Q3", "1100000"],
    ["Q4", "1800000"]
]
table_node = table(p, data=sales_data)
# table() function handles adding the node and its content

# 6. Serialize the prompt to POML format
poml_string = dump_poml(p)
println(poml_string)

• Prompt(): Creates a new prompt object.
• tag(p, name, attrs...): Creates a generic XML tag node. Most other specific functions use this internally.
• role(p, attrs...), task(p, attrs...), example(p, attrs...), example_set(p, attrs...), table(p, data=...), list(p, attrs...), list_item(p, attrs...), image(p, src=..., attrs...), tool_request(p, name=..., attrs...), tool_definition(p, name=..., description=..., attrs...): Create specific semantic tags.
• meta(p, attrs...), meta_tag(p, key=..., value=...), meta_attribute(p, name=..., value=...): Handle metadata.
• add_text(p, content): Adds text content to the current node in the hierarchy.
• add_node!(p, node): Makes node the new parent for subsequent additions.
• pop_node!(p): Moves the current parent back up one level in the hierarchy.
• dump_poml(p): Serializes the prompt object structure into a POML-formatted string.

Contributions are welcome! Please feel free to submit issues or pull requests.

This project is licensed under the MIT License - see the LICENSE file for details.

This package draws inspiration from the concepts of structured prompting and markup languages like XML, adapting them for the Julia ecosystem.