Tool Calling

​

What is Tool Calling?

• Structured Outputs: Tool arguments are always in JSON format, enforced by JSONSchema at the API level. See our Structured Outputs documentation for more details.
• Efficient Communication: Tool calling is a concept built into the model, reducing token usage and improving understanding.
• Model Routing: Facilitates setting up modular prompts with specific responsibilities.
• Prompt Injection Protection: Strict schema definitions at the model level make it harder to “jailbreak” the model.

​

Creating Custom Tools

​

Creating Visually

​

Publishing Programmatically

​

Tool Variables

• Different customers or tenants need different sets of tools
• Your available tools change based on runtime context (e.g., user permissions, feature flags)
• You want to manage tool definitions outside of your prompt template

​

Adding a Tool Variable

1. Open the Tool & Output Editor in the Prompt Registry
2. Click the dropdown arrow on the Add Tool button and select Tool Variable
3. Enter a variable name (e.g., dynamic_tools) and click Add
4. The variable will appear in your Input Variable Sets alongside other template variables

​

Passing Tools at Runtime

response = pl.run(
    prompt_name="my-agent",
    input_variables={
        "dynamic_tools": [
            {
                "name": "get_knowledge",
                "description": "Search the knowledge base",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "query": {"type": "string", "description": "Search query"}
                    },
                    "required": ["query"]
                }
            },
            {
                "name": "create_ticket",
                "description": "Create a support ticket",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "title": {"type": "string"},
                        "priority": {"type": "string", "enum": ["low", "medium", "high"]}
                    },
                    "required": ["title"]
                }
            }
        ]
    }
)

const response = await pl.run({
    promptName: "my-agent",
    inputVariables: {
        dynamic_tools: [
            {
                name: "get_knowledge",
                description: "Search the knowledge base",
                parameters: {
                    type: "object",
                    properties: {
                        query: { type: "string", description: "Search query" }
                    },
                    required: ["query"]
                }
            },
            {
                name: "create_ticket",
                description: "Create a support ticket",
                parameters: {
                    type: "object",
                    properties: {
                        title: { type: "string" },
                        priority: { type: "string", enum: ["low", "medium", "high"] }
                    },
                    required: ["title"]
                }
            }
        ]
    }
});

​

Built-in Tools

• OpenAI / Azure OpenAI (Responses API)
• Anthropic
• Google (Gemini)
• Vertex AI

​

How to Add Built-in Tools

1. Open your prompt in the Prompt Registry and navigate to the prompt editor
2. Select your LLM provider in the provider settings at the bottom of the editor
3. Open the Function & Output Schema Editor by clicking the Functions & Output button
4. Click the Built-in tools button (on the right side) to browse available tools for your selected provider
5. Click Add Tool for the tool you want to use — it will appear in your function definitions list
6. Configure tool_choice (optional) — set to auto to let the model decide when to use the tool
7. Save and run your prompt — the model will use the built-in tools when appropriate

​

OpenAI and Azure OpenAI (Responses API)

​

Available Tools

​

Using File Search with Vector Stores

Setting Up File Search

1. Enable File Search by following the steps above to add it as a built-in tool
2. Create and configure a Vector Store:
   • Click Manage Vector Stores in the File Search configuration
   • Click Create to make a new vector store with a custom name
   • Upload files via drag-and-drop or file selection (single or multiple files)
   • View storage usage, file counts, and manage attached files
3. Attach Vector Stores to your prompt:
   • Select one or more vector stores using checkboxes
   • Click Save Selection to attach them
   • The vector store IDs are added to your tool configuration
4. Run your prompt:
   • The LLM will automatically search vector stores when relevant
   • Retrieved context is used to generate informed responses
   • Sources can be traced back to specific documents

​

Using Code Interpreter

​

Using Image Generation

• A collapsible revised prompt showing the optimized text the model used for generation
• Generation parameters such as size, quality, background, and output format
• The generated image displayed in a rich card format

​

Learn More

• Image Generation Guide
• OpenAI Web Search Guide
• OpenAI File Search Guide
• OpenAI Code Interpreter Guide
• OpenAI Image Generation Guide
• OpenAI Responses API Announcement
• Deep Research API Cookbook

​

Anthropic

​

Available Tools

​

Learn More

• Anthropic Tool Use Documentation
• Anthropic Computer Use

​

Google (Gemini)

​

Available Tools

​

Using File Search with Google

1. Add the File Search tool from the built-in tools menu
2. Configure a file search store — specify the store name(s) that contain your indexed documents
3. Run your prompt — Gemini will automatically search through indexed documents when relevant context is needed

​

Learn More

• Google Gemini API Tools Documentation
• Grounding with Google Search

​

Vertex AI

​

For Gemini Models on Vertex AI

​

For Claude Models on Vertex AI

​

Learn More

• Vertex AI Gemini Documentation
• Claude on Vertex AI