Tool definitions, parallel calls, structured extraction, agentic loops.

Function Calling — Connecting LLMs to External Systems

Function calling (tool use) allows LLMs to request calls to external functions/APIs by generating structured invocations. The model selects which tool to use and with what arguments — the application executes the function and returns results. This transforms LLMs from text generators into capable agents. Key capabilities: multi-step reasoning with real data, parallel function execution, structured data extraction (no regex), and workflow automation. Supported by all major providers (OpenAI, Anthropic, Google).

Interactive Pipeline Flow

UnderstandImplementOptimise

📋 Define — JSON schema tools

Define tools as JSON Schema with name, description, and parameters. The description is critical — the model uses it to decide when to invoke the tool. Be specific: "Search the product database. Returns {id, name, price, stock}."

Architecture Diagrams

GPT-4o executing multiple tool calls simultaneously.

Loading diagram...

Key Design Decisions

📋Tool descriptions are more important than the schema

The model reads the description to decide when and how to use a tool. 'Query the database' is poor. 'Search the products table by name, category, or price range. Returns product_id, name, price, stock_count.' enables much better decisions.

🔁Parallel tool calling 3x reduces latency

When the model needs data from multiple sources, it can request all calls simultaneously — eliminating serial round-trip latency. Comparing 3 stock prices sequentially: 3 × 500ms = 1.5s. In parallel: 500ms total.

🔍Use function calling for structured extraction, not regex

Instead of prompting 'Return a JSON object with fields x, y, z' (fragile), define a tool schema and force tool_choice='required'. The model guarantees valid JSON matching your schema — production-grade extraction.

Discussion

Function Calling — Connecting LLMs to External Systems

Key Design Decisions

📋Tool descriptions are more important than the schema

🔁Parallel tool calling 3x reduces latency

🔍Use function calling for structured extraction, not regex