KNOWLEDGE WIKI  (LOCAL-MD-001)MODE: READ_ONLYSYS_TIME: --:--:--
SECTION:fdePAGES:40CURRENT:study-topics/tool-calling-and-function-calling.md
FDE-039study-topics/tool-calling-and-function-calling.mdUPDATED: 06/18/2026

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.