Define Properties and Data Types
What Are Properties?
Properties are the building blocks of your agent’s output.
Each one defines a field of data you want to extract, classify, summarize, or generate.
Think of them as your columns in a table or fields in a form — each property has a defined type, an associated tool, and a prompt that tells the AI what to do.
Standard Property Types
Below is a breakdown of the core property types used when building agents:
| Property Type | Best For | Notes |
|---|---|---|
| File | Uploading files for reference. | Upload and manage files from PDFs to CSVs and more. |
| Text | Free-form answers, summaries, notes | Enables AI Citations; use when flexibility is needed |
| Number | Prices, interest rates, percentages, counts | Enables AI Citations; formats automatically (e.g. currency, decimals) |
| Single Select | Choosing one option from a fixed list | Constrains outputs; supports routing logic and sub-workflows |
| Multi-Select | Selecting multiple discrete options from a fixed list. | Useful for inspections, risks, features, etc., any time you there is a defined list of outputs with multiple options. |
| Collection | Tables or sub-rows within a document/workflow | Enables complex, structured outputs (e.g., line items, photo logs) |
| JSON | Structured multi-value extraction in one pass | Best for grouped fields, improved accuracy, token efficiency |
| Reference | Connect other agents to this one to enable more structured workflows. | Best when you need to share data across agent entities, and connect workflows together. |
| URL | Document links, sources, web references | Typically populated by manual input, or web search. |
| Page Splitter | Split files into a collection, with each page becoming its own entity/row. | Best when dealing with very large PDF or Excel files, in order to make their extraction and QA more manageable. |
What Is Grounding?
Grounding ensures that AI-generated outputs are backed by specific evidence from the source document.
In V7 Go, grounding appears as AI citations — clickable highlights that show exactly where in the file the AI found its answer.
This is essential for:
- Trust & auditability
- Human-in-the-loop QA
- Legal, compliance, and financial workflows
Where AI Citations Work Today
Currently, AI citations are supported for the following property types:
| Property Type | Grounding Support | Notes |
|---|---|---|
| Text | ✅ Yes | Most common use case |
| Number | ✅ Yes | Works well for prices, and numeric values. |
| JSON | 🔶 Partial | Supported if you cite in-text JSON values; better control using Python downstream |
| Collections | 🔶 Partial | Citations are available when pointing to a Collection as an Input (e.g. File Bundle) but do not currently work when used as a Collection output. |
| Selects | ❌ Not Yet | On the roadmap — currently no citations |
| Multi-Select | ❌ Not Yet | Same as above |
What’s Coming: “Grounding Everywhere”
- Citations for Collections (input bundles) are actively being developed
- Single/Multi Select grounding is on the roadmap
- Goal: Grounding available everywhere, across all tools and document types
Tips for Grounding-First Prompts
If grounding is important for a field:
- Use Text or Number property types
- Write clear, narrow prompts that focus on specific terms
- Avoid combining too many concepts in one prompt
General Tips
- Start simple: Most POCs begin with Text, Number, Single Select, Multi-Select and Collection properties for core fields.
- Add complexity as needed: Use Selects and Collections once field options or document structures are better understood.
- Use JSON to avoid conflating similar fields (e.g. Start Date vs. End Date).
Real-world Use Case: Logistics - Purchase Order Processing
Matching Property to Output
Use this mental checklist:
- Is the output free-form or structured? → Use Text/Select/Number/Collection
- Is there only one correct answer? Or multiple? → Use Single Select or Multi-Select
- Do I want multiple related values in one field? → Use JSON
- Am I dealing with tabular or repeating data inside a doc? → Use Collection
Key Takeaway
Choose property types that reflect your final output format and improve downstream usability.
The more structured your properties, the more scalable and automated your workflow becomes.