Message roles (system, user, assistant) in Prompt Engineering / GenAI - Model Pipeline Trace

Model Pipeline - Message roles (system, user, assistant)

This pipeline shows how a generative AI model processes different message roles: system, user, and assistant. Each role guides the model to understand context, user input, and generate helpful responses.

Data Flow - 4 Stages

1Input messages

3 messages with role and content→Collect messages with roles: system, user, assistant→3 messages with role and content

[{"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "What is AI?"}, {"role": "assistant", "content": "AI means artificial intelligence."}]

↓

2Preprocessing

3 messages with role and content→Convert messages into tokens and add role embeddings→Token sequences with role embeddings

Tokens for 'What is AI?' plus role info for 'user'

↓

3Model encoding

Token sequences with role embeddings→Process tokens through transformer layers to understand context→Contextual embeddings representing message meaning

Vector representations capturing 'What is AI?' in user context

↓

4Response generation

Contextual embeddings→Generate next tokens as assistant reply→Token sequence for assistant response

Tokens forming 'AI means artificial intelligence.'

Training Trace - Epoch by Epoch

Loss: 2.3 |\
       1.8 | \
       1.3 |  \
       0.9 |   \
       0.6 |    \__
         +----------------
          1  2  3  4  5 Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	2.3	0.25	Model starts learning basic message patterns
2	1.8	0.40	Model improves understanding of roles
3	1.3	0.55	Better context handling between system and user
4	0.9	0.70	Assistant responses become more relevant
5	0.6	0.80	Model converges with good role-based replies

Prediction Trace - 3 Layers

Layer 1: Input message encoding

Layer 2: Transformer encoding

Layer 3: Response generation

Model Quiz - 3 Questions

Which message role sets the overall behavior of the AI model?

A System

B User

C Assistant

D null

Key Insight

Understanding message roles helps the AI model respond appropriately by distinguishing instructions (system), questions (user), and answers (assistant). This role-based context improves conversation quality.