Input Formats

Agent Actions accepts multiple input formats. Place files in agent_io/staging/ before running your workflow.

Supported Formats

Format	Extension	Use Case
JSON	.json	Structured data, API responses
CSV	.csv	Tabular data, spreadsheets
TSV	.tsv	Tab-separated tabular data
Excel	.xlsx	Spreadsheet data
XML	.xml	Structured markup data
Text	.txt	Plain text documents
Markdown	.md	Documentation, formatted text
HTML	.html	Web content
PDF	.pdf	PDF documents
Word	.docx	Word documents

JSON Input

A JSON file holds a list of records — one array, one object per record. Each element becomes one record.

json

JSON

[
  {
    "page_content": "The text content to process...",
    "url": "https://source.example.com/doc",
    "metadata": {
      "author": "John Doe",
      "date": "2024-01-15"
    }
  }
]

A file holding one record is that same list with one element in it. A lone object at the top level is refused at staging, and the fix is to wrap it: {...} becomes [{...}], which stages the record it always did.

A record is always an object, so a bare value is refused whether it is the whole document or a row inside one — wrapping it changes nothing. The refusal names the row it stopped on when there is one.

CSV/Tabular Input

Each row becomes a separate record. Header row defines field names.

csv

CSV

id,title,content,category
1,First Doc,Content here...,technical
2,Second Doc,More content...,general

Excel Input

The same rule reaches .xlsx through the sheet rather than the file: a sheet is a list of records, each row becomes one record, and the header row names the fields. A one-row sheet is one record, not a record of its own shape.

Accessing Source Data

Reference source fields in prompts using {{ source.field }}:

yaml

YAML

prompt: |
  Analyze this content: {{ source.page_content }}
  From: {{ source.url }}
  Author: {{ source.metadata.author }}

Nested Fields

Access nested objects with dot notation:

yaml

YAML

prompt: |
  Author: {{ source.metadata.author }}
  Date: {{ source.metadata.date }}

Iteration

Loop over arrays in source data:

yaml

YAML

prompt: |
  Process these items:
  {% for item in source.items %}
  - {{ item.name }}: {{ item.value }}
  {% endfor %}

Best Practices

1. Use consistent structure: All input files should have the same field structure
2. Include metadata: Add tracking fields like id, source, timestamp
3. Analyze schemas: Use agac schema -a my_workflow to check field dependencies before running

See Also

• Output Format — How outputs are structured
• Field References — Reference syntax details
• Chunking — Split large documents