GitHub

Content Types

This guide explains how to add different types of content to your prompts using PomlSDK.jl.

Text Content

The most basic content type is plain text, added using add_text() which can only be done inside a tag:

p = Prompt()


# Add text within a role
role_node = role(p, caption="System")
add_node!(p, role_node)
add_text(p, "Follow these guidelines:")
add_text(p, "- Be concise")
add_text(p, "- Provide accurate information")
pop_node!(p)

Text Best Practices

  • Add spaces manually: add_text(p, "Hello, "); add_text(p, "world!") produces "Hello, world!"
  • Use line breaks: add_text(p, "Line 1\nLine 2") for multi-line content
  • Consider text length: Very long text segments may affect LLM performance

Tables

PomlSDK supports creating tables with the table() function:

p = Prompt()

# Create a table from 2D array data
sales_data = [
    ["Quarter", "Sales (USD)"],
    ["Q1", "1,200,000"],
    ["Q2", "1,500,000"],
    ["Q3", "1,100,000"],
    ["Q4", "1,800,000"]
]

table_node = table(p, data=sales_data)

Lists

Create ordered and unordered lists with list() and list_item():

p = Prompt()

# Unordered list (default)
list_node = list(p)
add_node!(p, list_node)

item1 = list_item(p)
add_node!(p, item1)
add_text(p, "First item")
pop_node!(p)

item2 = list_item(p)
add_node!(p, item2)
add_text(p, "Second item")
pop_node!(p)

pop_node!(p)

# Ordered list
ordered_list = list(p, style="ordered")
add_node!(p, ordered_list)
# Add items...
pop_node!(p)

Images

Embed images using base64 encoding with the image() function:

p = Prompt()

# Add an image from file
img_data = read("assets/chart.png", Vector{UInt8})
img_node = image(p, src=img_data, alt="Sales Chart", format="png")

# Or directly with base64 string 
img_node = image(p, src="image/png;base64,iVBORw0KG...", alt="Sales Chart")

Image Best Practices

  • Always provide alt text for accessibility
  • Compress images before embedding to reduce prompt size
  • Consider file format - PNG for charts, JPEG for photos
  • Use appropriate size - large images increase token count

Metadata

Add metadata to your prompts using the meta functions:

p = Prompt()

meta_node = meta(p)
add_node!(p, meta_node)

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

# Add metadata attributes
meta_attribute(p, name="version", value="1.0")
meta_attribute(p, name="created", value="2023-10-15")

pop_node!(p)

Metadata Use Cases

  • Source tracking: Where the prompt content originated
  • Versioning: Track prompt iterations
  • Context information: Additional information about the prompt's purpose
  • Provenance: Record who created or modified the prompt

Documents

Reference external documents with the document() function:

p = Prompt()

# Reference a PDF document
doc_node = document(p, src="reports/annual_summary.pdf", parser="pdf")

# Reference a text file
txt_node = document(p, src="data/notes.txt", parser="text")

Document Parameters

  • src: Path to the document
  • parser: Specifies how to process the document (pdf, text, etc.)
  • selectedPages: For PDFs, which pages to include
  • maxPages: Maximum number of pages to process

Combining Content Types

Complex prompts often combine multiple content types:

p = Prompt()

role_node = role(p, caption="System")
add_node!(p, role_node)
add_text(p, "You are a data analyst. Use the provided information to answer questions.")
pop_node!(p)

task_node = task(p, caption="Sales Analysis")
add_node!(p, task_node)
add_text(p, "Analyze the quarterly sales data below and identify trends.")

# Add sales table
sales_data = [
    ["Quarter", "Revenue", "Growth"],
    ["Q1", "$1.2M", "+5%"],
    ["Q2", "$1.5M", "+25%"],
    ["Q3", "$1.1M", "-27%"],
    ["Q4", "$1.8M", "+64%"]
]
table_node = table(p, data=sales_data)

# Add chart image
img_node = image(p, src="charts/sales_trend.png", alt="Quarterly sales trend")

# Add key points as a list
list_node = list(p, style="unordered")
add_node!(p, list_node)

item1 = list_item(p)
add_node!(p, item1)
add_text(p, "Q2 showed strongest growth at 25%")
pop_node!(p)

item2 = list_item(p)
add_node!(p, item2)
add_text(p, "Q3 experienced a significant dip")
pop_node!(p)

pop_node!(p)  # Pop list

pop_node!(p)  # Pop task

# Add metadata
meta_node = meta(p)
add_node!(p, meta_node)
meta_tag(p, key="dataset", value="2023_sales")
meta_attribute(p, name="analysis_date", value="2023-12-15")
pop_node!(p)

poml_string = dump_poml(p)

Advanced Content Handling

Handling Complex Data Types

PomlSDK automatically converts complex data types to appropriate string representations:

# Dictionary data
params = Dict("operation" => "sum", "values" => [1, 2, 3, 4])
tool_req = tool_request(p, name="calculator", parameters=params)

# Array data
items = ["apple", "banana", "cherry"]
list_node = list(p, items=items)

Custom Serialization

For specialized content, you can implement custom serialization:

function custom_serialize(p::Prompt, data::MyCustomType)
    # Convert to appropriate format
    json_str = JSON.json(data)
    
    # Add as text or metadata
    meta_attr = meta_attribute(p, name="custom_data", value=json_str)
    
    # Or add as a specialized tag
    custom_node = tag(p, "custom", data=json_str)
    add_node!(p, custom_node)
    # Add more structured content if needed
    pop_node!(p)
end

Best Practices for Content Types

  • Balance richness with token count: More complex content increases token usage
  • Use appropriate content type: Tables for structured data, lists for sequences
  • Provide context: Always explain what the content represents
  • Validate content: Ensure tables are properly formatted, images are valid
  • Consider accessibility: Provide alt text for images, clear captions for tables

By effectively using these content types, you can create rich, structured prompts that communicate more effectively with LLMs.