mirror of
https://github.com/snowy-yang/xiaonuan.git
synced 2026-10-21 08:04:48 +00:00
snowy-yang
4494987e54
- FastAPI + WebSocket 服务，Textual TUI 终端客户端 - PostgreSQL 存储专用连接层（asyncpg 连接池，主程序零 SQL） - 无限且永久的记忆：语义检索注入 + 每夜巩固 + 身世档案（lore） - 作息推断（猜上下学时间）、定时事件（叫早/每日提醒） - 情绪回访 / 承诺追踪 / 慢信 / 回忆杀 / 她的生活 / 连续陪伴天数 - 原生 function calling 工具：天气 / 新闻 / 搜索（免 Key） - 极小模型快速路由（AI_SMALL_*）+ 夜间批量通道（AI_BATCH_*） - 小学徒小满：夜间琐事整理，小暖晨间检查转述 - 离线测试：路由器 / 工具循环 / 杂项 / 向量 / TUI 冒烟
325 lines
14 KiB
Python
325 lines
14 KiB
Python
"""AI 连接器：大模型能力的统一入口（OpenAI 兼容接口）。
|
||
|
||
- 主模型：流式对话已由「原生工具调用循环」取代（run_model_tools）
|
||
- 小模型（AI_SMALL_*，缺省复用主模型）：每条消息一次快速路由判定
|
||
- 夜间通道（AI_BATCH_*，缺省复用主模型）：小学徒夜间琐事用
|
||
- 回声模式：什么模型都没配时，诚实回声，不做任何假装
|
||
"""
|
||
|
||
import asyncio
|
||
import json
|
||
import os
|
||
import re
|
||
from datetime import datetime, timedelta
|
||
|
||
import httpx
|
||
from dotenv import load_dotenv
|
||
|
||
load_dotenv()
|
||
|
||
API_KEY = os.getenv("AI_API_KEY", "")
|
||
BASE_URL = os.getenv("AI_BASE_URL", "https://api.openai.com/v1").rstrip("/")
|
||
MODEL = os.getenv("AI_MODEL", "gpt-4o-mini")
|
||
EMBED_MODEL = os.getenv("AI_EMBED_MODEL", "text-embedding-3-small")
|
||
EMBED_DIM = 256 # 回声模式哈希 embedding 的维度
|
||
|
||
# 夜间/批量通道：给小学徒干琐事用的便宜通道，不配则回落主模型
|
||
BATCH_KEY = os.getenv("AI_BATCH_KEY", "") or API_KEY
|
||
BATCH_BASE = (os.getenv("AI_BATCH_BASE_URL", "") or BASE_URL).rstrip("/")
|
||
BATCH_MODEL = os.getenv("AI_BATCH_MODEL", "") or MODEL
|
||
BATCH_CONFIGURED = bool(os.getenv("AI_BATCH_MODEL")
|
||
or os.getenv("AI_BATCH_BASE_URL"))
|
||
|
||
# 极小模型快速路由：不配则复用主模型
|
||
SMALL_KEY = os.getenv("AI_SMALL_KEY", "") or API_KEY
|
||
SMALL_BASE = (os.getenv("AI_SMALL_BASE_URL", "") or BASE_URL).rstrip("/")
|
||
SMALL_MODEL = os.getenv("AI_SMALL_MODEL", "") or MODEL
|
||
|
||
# 小暖的小学徒：夜间琐事都是她干的
|
||
ASSISTANT_NAME = os.getenv("AI_ASSISTANT_NAME", "小满")
|
||
ASSISTANT_PERSONA = os.getenv(
|
||
"AI_ASSISTANT_PERSONA",
|
||
"你是一位小学徒，勤快、话少、偶尔偷懒。你负责夜间的琐事：整理回忆、"
|
||
"写回顾草稿。你写的东西都会经小暖检查后才交给用户。",
|
||
)
|
||
|
||
|
||
# ---- 回声模式 ----
|
||
|
||
async def _echo_stream(text: str):
|
||
"""回声模式：把用户的话切片吐回去"""
|
||
reply = f"（回声模式，未配置 AI_API_KEY）你说的是：{text}"
|
||
for ch in reply:
|
||
yield ch
|
||
await asyncio.sleep(0.015)
|
||
|
||
|
||
# ---- 向量 ----
|
||
|
||
def hash_embed(text: str) -> list[float]:
|
||
"""无 embedding 接口时的本地替身：字符二元组哈希到 256 维并归一化"""
|
||
import hashlib
|
||
vec = [0.0] * EMBED_DIM
|
||
t = text.strip().lower()
|
||
grams = [t[i:i + 2] for i in range(len(t) - 1)] or [t]
|
||
for g in grams:
|
||
h = int(hashlib.md5(g.encode("utf-8")).hexdigest(), 16)
|
||
vec[h % EMBED_DIM] += 1.0
|
||
norm = sum(v * v for v in vec) ** 0.5 or 1.0
|
||
return [v / norm for v in vec]
|
||
|
||
|
||
async def embed(texts: list[str]) -> list[list[float]]:
|
||
"""批量把文本变成向量。没配 embedding 接口时用本地哈希"""
|
||
if not API_KEY:
|
||
return [hash_embed(t) for t in texts]
|
||
async with httpx.AsyncClient(timeout=60) as client:
|
||
res = await client.post(
|
||
f"{BASE_URL}/embeddings",
|
||
headers={"Authorization": f"Bearer {API_KEY}"},
|
||
json={"model": EMBED_MODEL, "input": texts},
|
||
)
|
||
res.raise_for_status()
|
||
return [d["embedding"] for d in res.json()["data"]]
|
||
|
||
|
||
# ---- 基础调用 ----
|
||
|
||
_TEST_TRANSPORT = None # 测试注入点
|
||
|
||
|
||
def _make_client() -> httpx.AsyncClient:
|
||
if _TEST_TRANSPORT is not None:
|
||
return httpx.AsyncClient(timeout=120, transport=_TEST_TRANSPORT)
|
||
return httpx.AsyncClient(timeout=120)
|
||
|
||
|
||
async def _complete(base: str, key: str, model: str,
|
||
messages: list[dict], temperature: float = 0) -> str:
|
||
async with _make_client() as client:
|
||
res = await client.post(
|
||
f"{base}/chat/completions",
|
||
headers={"Authorization": f"Bearer {key}"},
|
||
json={"model": model, "messages": messages,
|
||
"temperature": temperature},
|
||
)
|
||
res.raise_for_status()
|
||
return res.json()["choices"][0]["message"]["content"]
|
||
|
||
|
||
async def complete(messages: list[dict], temperature: float = 0) -> str:
|
||
"""主模型非流式调用：问候/日记/关怀这类一次性生成用"""
|
||
return await _complete(BASE_URL, API_KEY, MODEL, messages, temperature)
|
||
|
||
|
||
async def complete_night(messages: list[dict], temperature: float = 0.3) -> str:
|
||
"""夜间通道：配置了 AI_BATCH_* 走便宜通道，否则回落主模型"""
|
||
return await _complete(BATCH_BASE, BATCH_KEY, BATCH_MODEL,
|
||
messages, temperature)
|
||
|
||
|
||
# ---- 原生工具调用（function calling）----
|
||
|
||
TOOLS_SCHEMA = [
|
||
{"type": "function", "function": {
|
||
"name": "get_weather",
|
||
"description": "查询某地现在的天气和今日温度、降水概率。"
|
||
"用户问到天气、要不要带伞、冷不冷，或你在写和出门"
|
||
"有关的问候需要天气时使用。",
|
||
"parameters": {"type": "object", "properties": {
|
||
"city": {"type": "string",
|
||
"description": "城市名。用户没提到就省略，"
|
||
"系统会尝试从记忆里推断。"}},
|
||
"required": []}}},
|
||
{"type": "function", "function": {
|
||
"name": "get_news",
|
||
"description": "看新闻。用户问新闻、热点、最近发生什么事时使用。",
|
||
"parameters": {"type": "object", "properties": {
|
||
"query": {"type": "string",
|
||
"description": "新闻主题关键词，如 科技；省略则给头条"}},
|
||
"required": []}}},
|
||
{"type": "function", "function": {
|
||
"name": "web_search",
|
||
"description": "联网搜索你不知道的、或时效性强的事实。",
|
||
"parameters": {"type": "object", "properties": {
|
||
"query": {"type": "string", "description": "搜索词"}},
|
||
"required": ["query"]}}},
|
||
]
|
||
|
||
|
||
async def run_model_tools(messages: list[dict], executor,
|
||
max_rounds: int = 3) -> str:
|
||
"""带原生工具调用的对话循环：模型自己决定调什么工具，
|
||
executor(name, args) 负责执行并返回文本结果。返回最终回复文本。"""
|
||
msgs = list(messages)
|
||
async with _make_client() as client:
|
||
for _ in range(max_rounds):
|
||
res = await client.post(
|
||
f"{BASE_URL}/chat/completions",
|
||
headers={"Authorization": f"Bearer {API_KEY}"},
|
||
json={"model": MODEL, "messages": msgs,
|
||
"tools": TOOLS_SCHEMA},
|
||
)
|
||
res.raise_for_status()
|
||
m = res.json()["choices"][0]["message"]
|
||
calls = m.get("tool_calls")
|
||
if not calls:
|
||
return m.get("content") or ""
|
||
msgs.append({"role": "assistant",
|
||
"content": m.get("content") or "",
|
||
"tool_calls": calls})
|
||
for tc in calls:
|
||
fn = tc.get("function") or {}
|
||
try:
|
||
args = json.loads(fn.get("arguments") or "{}")
|
||
except json.JSONDecodeError:
|
||
args = {}
|
||
result = await executor(fn.get("name"), args)
|
||
msgs.append({"role": "tool",
|
||
"tool_call_id": tc.get("id") or "",
|
||
"content": str(result)})
|
||
# 超出轮次上限：不带工具强制收尾，宁可平淡也不空转
|
||
res = await client.post(
|
||
f"{BASE_URL}/chat/completions",
|
||
headers={"Authorization": f"Bearer {API_KEY}"},
|
||
json={"model": MODEL, "messages": msgs},
|
||
)
|
||
res.raise_for_status()
|
||
return res.json()["choices"][0]["message"].get("content") or ""
|
||
|
||
|
||
# ---- 极小模型快速路由：每条消息一次调用，判完所有快事 ----
|
||
|
||
_ROUTER_VIEWS = {"memories", "timetable", "diary", "digest",
|
||
"assistant", "status", "lore"}
|
||
_ROUTER_LABELS = {"出门上学", "到校", "放学", "到家", "起床", "睡觉"}
|
||
|
||
|
||
async def route(text: str, now: datetime, existing: list[str]) -> dict:
|
||
"""小模型一次判定：查看档案 / 慢信 / 情绪 / 城市 / 各类提取。
|
||
所有字段已清洗；模型不可用或解析失败时返回空判定。"""
|
||
empty = {"view": None, "letter": None, "negative_mood": False,
|
||
"city_hint": "", "facts": [], "events": [], "routines": [],
|
||
"promises": [], "lore": []}
|
||
if not (SMALL_KEY and SMALL_MODEL):
|
||
return empty
|
||
|
||
user = (
|
||
f"现在是 {now:%Y-%m-%d %H:%M}（本地时间）。对下面的用户发言做判定，"
|
||
'只输出一个 JSON 对象：\n'
|
||
'{"view": null或"memories|timetable|diary|digest|assistant|status|lore",'
|
||
' "letter": null或{"days":7,"body":"信的内容"}或{"ask":true},'
|
||
' "negative_mood": false, "city_hint": "",'
|
||
' "facts": ["用户叫小明"],'
|
||
' "events": [{"kind":"alarm或reminder","fire_at":"ISO时间",'
|
||
'"text":"起床","repeat":"once或daily"}],'
|
||
' "routines": [{"label":"出门上学|到校|放学|到家|起床|睡觉 之一",'
|
||
'"time":"HH:MM","kind":"explicit或inferred"}],'
|
||
' "promises": [{"note":"面试","due_at":"ISO时间或null"}],'
|
||
' "lore": [{"topic":"诞生","body":"小暖是用户亲手造的"}]}\n'
|
||
"字段含义：\n"
|
||
"- view：用户想看她的某份档案——记忆/时间表/日记/昨天回顾/学徒/状态/身世；"
|
||
"普通聊天就是 null。\n"
|
||
"- letter：写给未来自己的慢信。days=天数、body=内容；"
|
||
'想写但没说完整给 {"ask":true}；不是写信就是 null。\n'
|
||
"- negative_mood：是否在倾诉负面情绪。\n"
|
||
'- city_hint：这句话里提到的用户所在城市（"我在杭州，明天冷不冷"→"杭州"），'
|
||
"没有为空串。\n"
|
||
"- facts：值得长期记住的关于用户的信息，以「用户」开头，"
|
||
"已有的不要重复，最多5条。\n"
|
||
"- events：定时任务（叫我起床/提醒我/每天XX点…）；"
|
||
"说「每天」repeat=daily 否则 once；fire_at 推算成最近一个将来的时间。\n"
|
||
"- routines：作息规律(explicit，如「每天7点半出门上学」)或"
|
||
"当下日常活动(inferred，如「到学校啦」，time 用当前时间)。\n"
|
||
"- promises：用户提到将来要做/会发生、值得之后回访的事；"
|
||
"due_at 说不清给 null。\n"
|
||
"- lore：用户在定义小暖自己的身世、身份、来历"
|
||
"（「你是XX造的」「你的生日是X」「你叫小暖因为X」）。\n"
|
||
"哪类都没有就 null / 空数组 / false。\n"
|
||
f"已有记忆：{json.dumps(existing, ensure_ascii=False)}\n"
|
||
f"用户发言：{text}"
|
||
)
|
||
try:
|
||
raw = await _complete(
|
||
SMALL_BASE, SMALL_KEY, SMALL_MODEL,
|
||
[{"role": "system",
|
||
"content": "你是意图路由器，只输出 JSON。"},
|
||
{"role": "user", "content": user}],
|
||
temperature=0)
|
||
m = re.search(r"\{.*\}", raw, re.S)
|
||
if not m:
|
||
return empty
|
||
data = json.loads(m.group())
|
||
except Exception:
|
||
return empty
|
||
|
||
out = dict(empty)
|
||
if data.get("view") in _ROUTER_VIEWS:
|
||
out["view"] = data["view"]
|
||
|
||
letter = data.get("letter")
|
||
if isinstance(letter, dict):
|
||
if letter.get("ask"):
|
||
out["letter"] = ("ask", "")
|
||
else:
|
||
try:
|
||
days = int(letter.get("days"))
|
||
body = str(letter.get("body") or "").strip()
|
||
if 1 <= days <= 365 and body:
|
||
out["letter"] = (days, body[:2000])
|
||
except (TypeError, ValueError):
|
||
pass
|
||
|
||
out["negative_mood"] = bool(data.get("negative_mood"))
|
||
out["city_hint"] = str(data.get("city_hint") or "").strip()[:20]
|
||
|
||
if isinstance(data.get("facts"), list):
|
||
for f in data["facts"][:5]:
|
||
s = str(f).strip()
|
||
if s and s not in existing and s not in out["facts"]:
|
||
out["facts"].append(s)
|
||
|
||
fallback = (now + timedelta(days=1)).replace(hour=18, minute=0,
|
||
second=0, microsecond=0)
|
||
for item in (data.get("events") or [])[:3]:
|
||
try:
|
||
fire = datetime.fromisoformat(str(item["fire_at"]))
|
||
except (KeyError, ValueError, TypeError):
|
||
continue
|
||
if fire <= now:
|
||
continue
|
||
out["events"].append({
|
||
"kind": "alarm" if item.get("kind") == "alarm" else "reminder",
|
||
"fire_at": fire,
|
||
"text": str(item.get("text") or "提醒").strip(),
|
||
"repeat": "daily" if item.get("repeat") == "daily" else "once"})
|
||
|
||
for item in (data.get("routines") or [])[:2]:
|
||
label = str(item.get("label") or "")
|
||
mm = re.match(r"^(\d{1,2}):(\d{2})$", str(item.get("time") or ""))
|
||
if label in _ROUTER_LABELS and mm:
|
||
out["routines"].append({
|
||
"label": label,
|
||
"minute": int(mm.group(1)) * 60 + int(mm.group(2)),
|
||
"kind": "explicit" if item.get("kind") == "explicit"
|
||
else "inferred"})
|
||
|
||
for item in (data.get("promises") or [])[:3]:
|
||
note = str(item.get("note") or "").strip()
|
||
if not note:
|
||
continue
|
||
try:
|
||
due = datetime.fromisoformat(str(item.get("due_at")))
|
||
if due <= now:
|
||
due = fallback
|
||
except (ValueError, TypeError):
|
||
due = fallback
|
||
out["promises"].append({"note": note, "due_at": due})
|
||
|
||
for item in (data.get("lore") or [])[:3]:
|
||
topic = str(item.get("topic") or "").strip()[:12]
|
||
body = str(item.get("body") or "").strip()[:200]
|
||
if topic and body:
|
||
out["lore"].append({"topic": topic, "body": body})
|
||
return out
|