Safety and Limits

This guide covers the safety controls available for tool use, including allowlists, turn budgets, and timeouts. These controls help prevent runaway loops and ensure predictable generation behavior.

Overview

When LLM columns use tools, the model can make multiple tool calls in a loop until it produces a final answer. Without limits, this could lead to:

• Excessive API calls and costs
• Long generation times
• Infinite loops if the model keeps requesting tools

Data Designer provides three types of controls:

Tool Allowlists

Restrict which tools are available using allow_tools:

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

Behavior

Tools not in the allowlist won’t be included in the schemas sent to the model, so the model won’t know they exist.

Turn Budgets

Limit the number of tool-calling iterations using max_tool_call_turns:

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=5,  # Maximum 5 iterations (default)
7
)

Understanding Turns

A turn is one iteration where the LLM requests tool calls. With parallel tool calling, a single turn may execute multiple tools simultaneously.

This approach gives models flexibility to use parallel calling efficiently while still bounding total iterations.

Graceful Budget Exhaustion

When the turn limit is reached, Data Designer doesn’t abruptly stop generation. Instead:

1. The model’s tool call request is recorded in the conversation
2. Tool “results” are returned with a refusal message explaining the limit was reached
3. The model receives this feedback and can produce a final response

This ensures the model can still provide a useful answer based on the tools it already called, rather than failing silently.

Timeouts

Limit how long each tool call can take using timeout_sec:

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

Timeout Behavior

When a timeout occurs:

1. The tool call is terminated
2. An error message is returned to the model
3. The model can attempt recovery (retry, skip, or answer without the result)

1
# Example error in trace when timeout occurs
2
{
3
    "role": "tool",
4
    "content": "Error: Tool 'search_docs' failed: Connection timeout after 30s",
5
    "tool_call_id": "call_abc123"
6
}

Default Timeout

The default timeout is 60 seconds. Adjust based on your tools:

Combining Controls

You can use all controls together for defense in depth:

1
import data_designer.config as dd
2
3
tool_config = dd.ToolConfig(
4
    tool_alias="secure-tools",
5
    providers=["demo-mcp"],
6
    allow_tools=["search_docs", "get_fact"],  # Restricted tools
7
    max_tool_call_turns=3,                     # Limited iterations
8
    timeout_sec=15.0,                          # Fast timeout
9
)

See Also

• Tool Configurations: Complete ToolConfig reference
• Message Traces: Monitor tool usage patterns