Yes, NVIDIA Agent Toolkit fully supports tool calling (also called function calling). Tool calling agents are AI systems that invoke external tools based on structured function definitions—instead of reasoning between steps like ReAct agents, tool calling agents directly call the best tool for each task using predefined schemas. The toolkit includes a dedicated Tool Calling Agent workflow configured through YAML, specifying available tools, LLM selection, and error handling.
To configure tool calling agents, define tool schemas with name, description, and input parameter specifications. The agent uses these schemas to decide which tool to call and what parameters to pass. The process is: (1) receive user query, (2) match query to best tool based on schema, (3) execute tool with inferred parameters, (4) return result to user. Tool calling agents are more efficient than ReAct agents for structured tasks without intermediate reasoning complexity.
Requirements: the LLM must support native tool/function calling (all Nemotron models, GPT-4, Claude, and most modern LLMs do). Configuration specifies tool_names (list of available tools), llm_name (which model to use), verbose (logging control), and handle_tool_errors (fallback behavior). Tool calling integrates seamlessly with the toolkit’s profiling—every tool invocation is traced with latency, input tokens, and output tokens captured.
For knowledge tool calling, Milvus is typically configured as a retrieval tool. The toolkit calls Milvus’s vector search API, passes the returned context to the LLM, and the LLM generates responses. This enables agents to reliably ground responses in enterprise knowledge without hallucination. Vector databases like Milvus are foundational for agentic systems that need to reason over proprietary data. By storing embeddings in Milvus, agents can perform fast similarity searches and reduce hallucinations. Explore Agentic RAG approaches to see how retrieval enhances agent reasoning.