Tool Configurations | NVIDIA NeMo Data Designer

Tool configurations define how LLM columns access external tools during generation. Each ToolConfig specifies which MCP providers to use, which tools are allowed, and operational limits.

Overview

A ToolConfig connects LLM columns to MCP providers. When you create column configurations (like LLMTextColumnConfig or LLMCodeColumnConfig), you reference a tool configuration by its alias. Data Designer uses the tool configuration to determine which tools are available and how to manage tool calls.

ToolConfig Structure

The ToolConfig class has the following fields:

Field	Type	Required	Description
`tool_alias`	`str`	Yes	Unique identifier for this tool configuration (referenced by columns)
`providers`	`list[str]`	Yes	List of MCP provider names to use (can reference multiple providers)
`allow_tools`	`list[str]`	No	Restrict to specific tools (`None` = allow all tools from providers)
`max_tool_call_turns`	`int`	No	Maximum tool-calling iterations (default: 5)
`timeout_sec`	`float`	No	Per-call timeout in seconds (default: 60.0)

Examples

Basic Tool Configuration

1 import data_designer.config as dd
2 
3 tool_config = dd.ToolConfig(
4     tool_alias="my-tools",
5     providers=["demo-mcp"],
6 )

Restricting Allowed Tools

1 import data_designer.config as dd
2 
3 tool_config = dd.ToolConfig(
4     tool_alias="search-only",
5     providers=["demo-mcp"],
6     allow_tools=["search_docs", "list_docs"],  # Only these tools allowed
7 )

Using Multiple Providers

A single ToolConfig can reference multiple MCP providers, allowing tools to be drawn from different sources:

1 import data_designer.config as dd
2 
3 tool_config = dd.ToolConfig(
4     tool_alias="multi-search",
5     providers=["doc-search-mcp", "web-search-mcp"],
6     allow_tools=["search_docs", "search_web", "list_docs"],
7     max_tool_call_turns=10,
8 )

When the model requests a tool call, Data Designer automatically finds which provider hosts that tool and routes the call appropriately.

Setting Operational Limits

1 import data_designer.config as dd
2 
3 tool_config = dd.ToolConfig(
4     tool_alias="limited-tools",
5     providers=["demo-mcp"],
6     max_tool_call_turns=3,   # Maximum 3 tool-calling iterations
7     timeout_sec=30.0,        # 30 seconds per tool call
8 )

Adding to Config Builder

Tool configurations can be added to the config builder in two ways:

1 import data_designer.config as dd
2 
3 tool_config = dd.ToolConfig(
4     tool_alias="my-tools",
5     providers=["demo-mcp"],
6 )
7 
8 # Method 1: Pass at initialization
9 builder = dd.DataDesignerConfigBuilder(tool_configs=[tool_config])
10 
11 # Method 2: Add later
12 builder = dd.DataDesignerConfigBuilder()
13 builder.add_tool_config(tool_config)

Understanding Turn-Based Limiting

The max_tool_call_turns parameter limits how many tool-calling iterations (turns) are permitted, not the total number of individual tool calls.

Turn-based vs call-based counting A turn is one iteration where the LLM requests tool calls. With parallel tool calling, a single turn may execute multiple tools simultaneously.

For example, if the model requests 3 tools in parallel, that counts as 1 turn, not 3. This gives models flexibility to use parallel calling efficiently while still bounding total iterations.

When the turn limit is reached, Data Designer gracefully refuses additional tool calls rather than failing abruptly. The model receives feedback explaining the limit was reached and can produce a final response based on the tools it already called.