Chat vs. Completion: The Language of Messages

In the early days of LLMs (GPT-3), we used "Completion" models. You gave them a string of text, and they "finished" it. Today, we use "Chat" models. You give them a List of Messages, and they "Respond" to them.

1. Completion Models (The Old Way)

• Input: "The capital of France is"
• Output: " Paris."
• LangChain class: OpenAI() (Legacy)

---

2. Chat Models (The Modern Way)

• Input: A structured list of identified speakers.
• Output: An AIMessage object.
• LangChain class: ChatOpenAI()

---

3. The 3 Core Message Types

In LangChain, every conversation is built using these three types:

Type	Role	Purpose
SystemMessage	The Instructions	Setting the "Identity" (e.g., "You are a helpful pirate").
HumanMessage	The User	The specific query from the person chatting.
AIMessage	The Model	The response generated by the AI.

---

4. Code Example: Structured Conversation

from langchain_core.messages import HumanMessage, SystemMessage

messages = [
    SystemMessage(content="You are a professional chef. Answer briefly."),
    HumanMessage(content="How do I cook a perfect steak?")
]

response = model.invoke(messages)
print(response.content)

---

5. Why Roles Matter

By separating the System instruction from the Human input, the model can better differentiate between what it counts as "The Truth" (System) and "A Query" (Human). This is the first level of defense against prompt injection (Module 13).

---

Key Takeaways

• Chat Models are the industry standard for reasoning and interaction.
• System Messages set the ground rules and persona.
• Human Messages carry the user's specific request.
• Always use ChatOpenAI (or equivalent) for modern agentic applications.