AI 原生开放教材

外观

Ed Donner：AI Agents 实战课系统课程·英文原文 ↗原件 ↓仓库 ↗

API Response & Conversation History ​

1. API Response is not a string ​

When we call the model:

python
response = client_which_is_gemini.chat.completions.create(
    model="gemini-2.5-flash-lite",
    messages=messages
)

you can also use other models, gemini 3.5 has even higher limits (reason:With every new generation of AI, Google optimizes the underlying architecture and TPU (Tensor Processing Unit) hardware.

Gemini 3.1 Flash Lite is a highly compressed, ultra-efficient model specifically engineered for high-volume, low-latency tasks. 1)
Because it is so highly optimized, it requires significantly less computing power and cost for Google to run per request compared to the older Gemini 2.5 architecture. Google can easily afford to give you 500 free requests of a highly optimized model because it barely dents their server capacity.2. Encouraging Migration to the New PlatformGoogle wants developers building on their newest, most advanced platform. Gemini 3.1 comes with a completely redesigned native reasoning engine (letting you control "thinking levels" from minimal to high) and improved tool integrations.By offering generous free limits on Gemini 3.1, Google incentivizes developers to migrate their projects to the newer version.By contrast, restricting older "legacy" models like Gemini 2.5 to 20 daily requests gently pushes developers away from outdated tech.

The model does not return just the generated text.

Instead, it returns a response object containing several pieces of information, such as:

  • Metadata (e.g., response ID)
  • Token usage
  • Finish reason
  • Choices
  • Assistant message

A simplified response looks like:

python
response = {
    "id": "...",
    "choices": [
        {
            "message": {
                "role": "assistant",
                "content": "If a clockmaker creates..."
            }
        }
    ]
}

To access only the generated text, use:

python
print(response.choices[0].message.content)

Roles in Chat APIs ​

Every message sent to the model has a role that identifies who is speaking.

User ​

Represents a message sent by the human.

python
{"role": "user", "content": "How are you?"}

Assistant ​

Represents a message generated by the AI.

python
{"role": "assistant", "content": "I'm doing well!"}

System ​

Provides instructions that guide the model's behavior throughout the conversation.

python
{"role": "system", "content": "You are a helpful Python tutor."}

Conversation History (messages) ​

The model maintains context using a list called messages.

Each new message is appended (added to the end) of this list.

python
messages = []

messages.append({"role": "user", "content": "How are you?"})

messages.append({"role": "assistant", "content": "I'm doing well!"})

messages.append({"role": "user", "content": "Tell me a joke."})

The conversation history now looks like:

python
messages = [
    {"role": "user", "content": "How are you?"},
    {"role": "assistant", "content": "I'm doing well!"},
    {"role": "user", "content": "Tell me a joke."}
]

When this list is sent back to the model, it can understand the entire conversation and respond with the correct context.

Key Takeaways ​

  • response is an object, not a string.

  • The generated text is accessed using:

    python
    response.choices[0].message.content

  • messages stores the entire conversation history.

  • role indicates who spoke the message (user, assistant, or system).

  • Every new message is appended to the messages list, allowing the model to maintain context across the conversation.