Skip to main content

Tool Calling Patterns

Let LLMs decide which tools to use and with what parameters.

Basic Tool Selection​

enum Tool {
  search @description("Search the web for information")
  calculate @description("Perform mathematical calculations")
  lookup @description("Look up data in the database")
  none @description("No tool needed, respond directly")
}

class ToolCall {
  tool Tool
  query string @description("The input for the tool")
  reasoning string
}

function SelectTool(user_message: string) -> ToolCall {
  client GPT4
  prompt #"
    Given this user message, decide which tool to use:
    {{ user_message }}

    Available tools:
    - search: Search the web
    - calculate: Do math
    - lookup: Query database
    - none: Respond directly
  "#
}

Multi-Tool Orchestration​

class ToolPlan {
  steps ToolStep[]
  final_response_template string
}

class ToolStep {
  tool Tool
  input string
  output_variable string @description("Name to store result")
}

function PlanToolUse(task: string) -> ToolPlan {
  client GPT4
  prompt #"
    Plan how to accomplish this task using available tools:
    {{ task }}

    Create a step-by-step plan.
  "#
}

Executing Tool Calls​

const plan = await b.PlanToolUse(userTask);

const results = {};
for (const step of plan.steps) {
  switch (step.tool) {
    case 'search':
      results[step.output_variable] = await searchWeb(step.input);
      break;
    case 'calculate':
      results[step.output_variable] = await calculate(step.input);
      break;
    // ... other tools
  }
}