用大模型分析 120 万条聊天记录：我是个糟糕的朋友吗？

import json
import os
from collections import defaultdict
from datetime import datetime
from openai import OpenAI

# ---------- 配置 ----------
API_KEY = "sk-你的key"
TELEGRAM_EXPORT_PATH = "telegram_export/result.json"
MODEL = "gpt-4o-mini"  # 用便宜模型做摘要即可

client = OpenAI(api_key=API_KEY)

# ---------- 第一步：加载并按联系人分桶 ----------
with open(TELEGRAM_EXPORT_PATH, "r", encoding="utf-8") as f:
    data = json.load(f)

buckets = defaultdict(list)  # key: 联系人名, value: 消息列表

for chat in data.get("chats", {}).get("list", []):
    name = chat.get("name", "unknown")
    for msg in chat.get("messages", []):
        if msg.get("type") != "message":
            continue
        buckets[name].append({
            "date": msg.get("date", ""),
            "from": msg.get("from", ""),
            "text": msg.get("text", "") if isinstance(msg.get("text"), str) else "",
        })

print(f"共 {len(buckets)} 个联系人，总消息 {sum(len(v) for v in buckets.values())} 条")

# ---------- 第二步：按年切片 + 统计基础指标 ----------
def yearly_stats(messages):
    stats = defaultdict(lambda: {
        "total": 0, "sent": 0, "received": 0,
        "avg_delay_min": 0, "texts_sample": []
    })
    prev_date = None
    for msg in messages:
        try:
            dt = datetime.strptime(msg["date"][:10], "%Y-%m-%d")
        except ValueError:
            continue
        year = dt.year
        s = stats[year]
        s["total"] += 1
        if msg["from"] == "我自己的名字":  # ← 替换成你在 Telegram 中的显示名
            s["sent"] += 1
        else:
            s["received"] += 1
        # 取每年第 10 条消息做抽样，避免摘要过长
        if s["total"] % 10 == 0 and msg["text"]:
            s["texts_sample"].append(msg["text"][:80])
    return dict(stats)

# ---------- 第三步：用大模型生成友谊摘要 ----------
def summarize_friend(name, stats_dict):
    prompt_parts = []
    for year, s in sorted(stats_dict.items()):
        prompt_parts.append(
            f"{year}年: 共{s['total']}条, 我发{s['sent']}条/收到{s['received']}条, "
            f"抽样内容: {'; '.join(s['texts_sample'][:5])}"
        )
    prompt = (
        f"以下是我和朋友「{name}」的逐年聊天统计摘要:\n"
        + "\n".join(prompt_parts)
        + "\n\n请从以下维度评价这段友谊:\n"
        "1. 主动发起聊天的比例（我主动 vs 对方主动）\n"
        "2. 关系热度变化趋势（升温/平稳/降温）\n"
        "3. 情绪支持度（对方倾诉时我的回应质量推测）\n"
        "4. 总体评价：我是个好朋友吗？给出具体理由。\n"
        "用中文回答，简洁有力。"
    )
    resp = client.chat.completions.create(
        model=MODEL,
        messages=[{"role": "user", "content": prompt}],
        max_tokens=500,
    )
    return resp.choices[0].message.content

# ---------- 执行 ----------
for name, msgs in list(buckets.items())[:5]:  # 先分析前 5 个联系人
    if len(msgs) < 50:  # 太少的跳过
        continue
    stats = yearly_stats(msgs)
    summary = summarize_friend(name, stats)
    print(f"\n{'='*40}")
    print(f"朋友: {name} | 消息数: {len(msgs)}")
    print(summary)