Function calling in LLMs in Agentic AI - Model Pipeline Trace

Model Pipeline - Function calling in LLMs

This pipeline shows how a large language model (LLM) uses function calling to improve responses by invoking external functions during text generation.

Data Flow - 4 Stages

1User Input

1 text string→Receive user question or command→1 text string

"What is the weather in Paris today?"

↓

2LLM Processing

1 text string→Analyze input and decide if function call is needed→1 text string or function call request

"Call weather_api(city='Paris')"

↓

3Function Call Execution

Function call request with parameters→Invoke external function (e.g., API) and get result→Function output data

{"temperature": "18°C", "condition": "Cloudy"}

↓

4LLM Response Generation

Function output data→Incorporate function output into final text response→1 text string

"The weather in Paris today is 18°C and cloudy."

Training Trace - Epoch by Epoch

Loss
1.0 |\
0.9 | \
0.8 |  \
0.7 |   \
0.6 |    \
0.5 |     \
0.4 |      \
0.3 |       \
0.2 |        \
0.1 |         \
0.0 +----------
      1 2 3 4 5
      Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	0.85	0.60	Model starts learning to detect when to call functions.
2	0.65	0.72	Improved accuracy in predicting function calls.
3	0.50	0.81	Better integration of function outputs in responses.
4	0.38	0.88	Model effectively calls functions and generates accurate answers.
5	0.30	0.92	Training converges with high accuracy and low loss.

Prediction Trace - 4 Layers

Layer 1: Input Text

Layer 2: Function Call Decision

Layer 3: Function Execution

Layer 4: Response Generation

Model Quiz - 3 Questions

What does the LLM do after receiving the user input?

A Decides if a function call is needed

B Immediately calls an external function

C Returns a random answer

D Ignores the input

Key Insight

Function calling lets LLMs get real-time data or perform tasks by invoking external functions, making their answers more accurate and useful.