Tool Calling And Function Calling
Topic
Name: Tool calling and function calling
Why it matters for FDE roles: Many AI workflows need to search records, query APIs, create tickets, update statuses, or trigger business processes.
Plain-English Definition
Tool calling is an LLM application pattern where the model selects a defined function or tool, passes structured arguments, and receives a result that the application can use in the next step.
Where It Shows Up
- Job listing signal: agents, tool use, function calling, API automation, AI workflows.
- Portfolio project connection: Ops Knowledge Copilot can expose tools for search, record lookup, recommendation drafting, and status updates.
- Real customer scenario: An AI assistant checks a customer record, retrieves relevant docs, drafts a ticket update, and waits for approval before writing back.
Core Concepts
- Tool schema: name, description, arguments, and return shape.
- Tool selection: deciding when a tool is needed.
- Argument validation: checking model-proposed inputs before execution.
- Permissions: limiting which tools and records are available.
- Idempotency: making repeated calls safe when retries happen.
- Auditability: recording tool calls, inputs, outputs, and user approvals.
Failure Modes
- Vague tool descriptions that make the model choose the wrong action.
- Tool arguments accepted without validation.
- Tools that modify customer systems without human review.
- Missing logs, so failures cannot be reconstructed.
- Overusing tools when retrieval or a normal API call would be simpler.
Tiny Practice Task
Define three tools for Ops Knowledge Copilot: search_records, get_record_by_id, and draft_recommendation.
Interview Language
One sentence I could say in an interview:
I treat tool calling as an integration boundary: the model can propose an action, but code should validate arguments, enforce permissions, log the call, and require approval for risky writes.
Relevant work experience for this topic.