# 简化示意：历史太长时，让模型自己把"旧历史"总结成一段摘要
def compress(history, keep_recent=4):
    if len(history) <= keep_recent + 2:
        return history                       # 还不长，不用压
    old = history[1:-keep_recent]            # 除了 system 和最近几轮，都算"旧的"
    text = "\n".join(f"{m['role']}: {m['content']}" for m in old)
    summary = client.chat.completions.create(  # 让模型把旧对话总结成要点
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": "把以下对话压成要点摘要：\n" + text}]
    ).choices[0].message.content
    # 重组：system + 一条摘要 + 最近几轮原话
    return [history[0],
            {"role": "system", "content": "早前对话摘要：" + summary}
            ] + history[-keep_recent:]

def build_messages(system_prompt, retrieved_docs, history, user_question):
    msgs = [{"role": "system", "content": system_prompt}]        # ① 人设放最前
    if retrieved_docs:                                            # ② 相关资料（只放相关的！）
        docs = "\n---\n".join(retrieved_docs)
        msgs.append({"role": "system", "content": f"参考资料：\n{docs}"})
    msgs += history[-6:]                                          # ③ 只留最近几轮（更早的应已压缩）
    msgs.append({"role": "user", "content":                      # ④ 当前问题放最后（结尾=高注意力）
                 f"{user_question}\n\n请只依据上面的参考资料回答。"}) # 关键指令在结尾重申
    return msgs
# 之后照常 client.chat.completions.create(model=..., messages=build_messages(...))

✋ 动手 10 分钟：给聊天机器人加"历史压缩"

在 Day11/12 的环境里，给多轮机器人加一个"历史过长就摘要"的机制，直观感受上下文管理：

from openai import OpenAI
from dotenv import load_dotenv
load_dotenv()
client = OpenAI()
history = [{"role": "system", "content": "你是助手，回答简洁"}]

def summarize(msgs):
    text = "\n".join(f"{m['role']}: {m['content']}" for m in msgs)
    r = client.chat.completions.create(model="gpt-4o-mini",
        messages=[{"role": "user", "content": "用要点压缩以下对话：\n" + text}])
    return r.choices[0].message.content

while True:
    q = input("\n你：")
    if q == "quit": break
    history.append({"role": "user", "content": q})

    # 简单预算：超过 10 条就把中间压成摘要，只留 system + 摘要 + 最近 4 条
    if len(history) > 10:
        s = summarize(history[1:-4])
        history[:] = [history[0], {"role":"system","content":"早前摘要："+s}] + history[-4:]
        print("（已压缩历史，节省上下文）")

    r = client.chat.completions.create(model="gpt-4o-mini", messages=history)
    a = r.choices[0].message.content
    history.append({"role": "assistant", "content": a})
    print("AI：", a, f"  [当前历史 {len(history)} 条]")

多聊几轮，观察历史条数在触发压缩后如何被"缩回去"，机器人却还记得关键信息。