Executing Functions and Returning Results
What you'll learn: How to actually run the function an LLM requested and feed the output back into the conversation for the model to continue reasoning.
The Execution Bridge
When an LLM generates a function call (lesson 1925), it's only describing what should happen—it's not actually executing code. Think of it like a customer ordering food: they tell you what they want, but someone still needs to prepare and deliver it.
Your application code must:
- Extract the function name and arguments from the LLM's response
- Execute the actual function in your environment
- Format the result as a new message to send back to the LLM
The Round-Trip Pattern
# Step 1: LLM requests function call
# {"name": "get_weather", "arguments": {"city": "Paris"}}
# Step 2: Your code executes the function
result = get_weather(city="Paris") # Returns: {"temp": 18, "condition": "cloudy"}
# Step 3: Format result as a "function" role message
messages.append({
"role": "function",
"name": "get_weather",
"content": json.dumps(result)
})
# Step 4: Send back to LLM for final answer
final_response = openai.ChatCompletion.create(
model="gpt-4",
messages=messages
)
Result Formatting Matters
The function result should be:
- JSON-serialized (even if it's just a string or number)
- Concise but complete (avoid dumping massive payloads)
- Error-aware (if the function fails, return a structured error message)
The LLM uses this result to generate its final user-facing response, like "The weather in Paris is currently 18°C and cloudy."
Key Takeaway: Function calling is a three-step dance: LLM requests → you execute → LLM synthesizes the result into a natural language answer.