Principle:Run llama Llama index Tool Definition

Overview

Tool Definition is the foundational concept in agentic AI that enables LLM agents to interact with external functions, data sources, and services. In LlamaIndex, tools are the bridge between an agent's reasoning capabilities and real-world actions -- they wrap arbitrary Python functions, query engines, and retrievers into a standardized interface that the LLM can discover, understand, and invoke through function calling.

The tool abstraction consists of three key elements: the callable logic (what the tool does), the metadata (how the LLM understands the tool), and the schema (what parameters the tool accepts). Together, these allow the agent to make informed decisions about which tool to use and how to call it correctly.

AI Agents Function Calling Tool Use LlamaIndex

The Tool Abstraction

Every tool in LlamaIndex implements a common interface defined by BaseTool and AsyncBaseTool. This interface requires:

• A metadata property returning a ToolMetadata object (name, description, schema, return_direct flag)
• A call method for synchronous execution
• An acall method for asynchronous execution

The ToolMetadata dataclass is the critical piece that enables function calling:

@dataclass
class ToolMetadata:
    description: str                              # Natural language description for the LLM
    name: Optional[str] = None                    # Tool identifier used in function calls
    fn_schema: Optional[Type[BaseModel]] = DefaultToolFnSchema  # Pydantic schema for parameters
    return_direct: bool = False                   # If True, tool output is returned directly without further LLM processing

The description is particularly important -- it is what the LLM reads to decide whether a tool is relevant for a given task. A well-written description dramatically improves agent accuracy.

Types of Tools

LlamaIndex provides three primary tool classes, each wrapping a different kind of callable:

Tool Class	Wraps	Use Case
FunctionTool	Any Python function (sync or async)	General-purpose tool creation from arbitrary functions
QueryEngineTool	A BaseQueryEngine	Natural language Q&A over indexed documents
RetrieverTool	A BaseRetriever	Document retrieval from a knowledge base

FunctionTool

The most flexible tool type. It wraps any Python callable and automatically generates the parameter schema by inspecting the function signature and docstring. It supports:

• Sync and async functions -- provide either or both; the missing variant is auto-generated
• Callbacks -- post-processing hooks that can override the tool output
• Partial parameters -- pre-bind certain arguments so they are excluded from the LLM's schema
• Workflow context injection -- functions annotated with Context automatically receive the workflow context

QueryEngineTool

Wraps a query engine (e.g., a VectorStoreIndex or a KnowledgeGraphIndex) as a tool. The LLM provides a natural language query string, and the tool returns the query engine's response. This is the primary way to give an agent access to RAG (Retrieval-Augmented Generation) pipelines.

RetrieverTool

Similar to QueryEngineTool but operates at the retriever level, returning raw document nodes rather than synthesized responses. This gives the agent access to the retrieved documents themselves, which is useful when the agent needs to reason over multiple source passages.

Schema Generation

When creating a FunctionTool, the parameter schema is automatically derived from the function's type annotations using create_schema_from_function. This generates a Pydantic model that is then converted to a JSON Schema for the LLM's function calling interface.

Docstring parameter descriptions (in Sphinx, Google, or Javadoc style) are automatically extracted and added to the schema fields, giving the LLM richer context about each parameter:

def search_documents(query: str, max_results: int = 10) -> str:
    """Search the document database.

    Args:
        query (str): The search query to find relevant documents.
        max_results (int): Maximum number of results to return.
    """
    ...

The generated schema will include the descriptions "The search query to find relevant documents." and "Maximum number of results to return." for the respective parameters.

Tool Output

All tools return a ToolOutput object containing:

• content -- the string representation of the output (what the LLM sees)
• blocks -- a list of ContentBlock objects (text, images, audio)
• tool_name -- which tool produced the output
• raw_input -- the original arguments passed to the tool
• raw_output -- the unprocessed return value of the underlying function
• is_error -- whether the tool call resulted in an error

Knowledge Sources

LlamaIndex Tools Documentation ReAct Paper - Yao et al. 2022 LlamaIndex GitHub Repository

Implementation

Implementation:Run_llama_Llama_index_FunctionTool_From_Defaults