feat: 提示词和 AI 参数移至数据库管理
- AiSetting 模型新增 6 个提示词字段 + 3 个阶段温度字段 - main.py 迁移逻辑改为检测 ai_settings 表的新字段并填充默认值 - generate.py 从数据库读取生成提示词，去掉硬编码 - optimizer 从 API 读取各阶段提示词和温度，删除 prompts.py - crawler 从 API 读取提取/改写提示词和温度，删除 prompts.py - settings/active 端点去掉 token 认证（供爬虫/优化器使用） - 后台设置页新增提示词编辑区和温度调节控件 - 新增 _ensure_default_settings 自动创建默认配置
This commit is contained in:
bwstudio
+80
-20
@@ -1,35 +1,90 @@
|
||||
"""LLM 处理：调用 NVIDIA NIM（OpenAI 兼容 API）进行笑话提取、改写和分类。"""
|
||||
"""LLM 处理：调用 NVIDIA NIM（OpenAI 兼容 API）进行笑话提取、改写。"""
|
||||
|
||||
import json
|
||||
import os
|
||||
import httpx
|
||||
from openai import OpenAI
|
||||
|
||||
|
||||
class AiService:
|
||||
def __init__(self, api_base: str, api_key: str, model_name: str, temperature: float = 0.7, max_tokens: int = 2048):
|
||||
self.client = OpenAI(base_url=api_base, api_key=api_key)
|
||||
self.model = model_name
|
||||
self.temperature = temperature
|
||||
self.max_tokens = max_tokens
|
||||
def __init__(self, api_base: str, username: str, password: str):
|
||||
self.api_base = api_base.rstrip("/")
|
||||
self.username = username
|
||||
self.password = password
|
||||
self.token = None
|
||||
self.client = None
|
||||
self.model = ""
|
||||
self.ai_config = None
|
||||
|
||||
def _login(self) -> str:
|
||||
resp = httpx.post(
|
||||
f"{self.api_base}/api/auth/login",
|
||||
json={"username": self.username, "password": self.password},
|
||||
timeout=30,
|
||||
)
|
||||
resp.raise_for_status()
|
||||
return resp.json()["access_token"]
|
||||
|
||||
def _get(self, path: str) -> dict | list:
|
||||
resp = httpx.get(
|
||||
f"{self.api_base}{path}",
|
||||
headers={"Authorization": f"Bearer {self.token}"},
|
||||
timeout=30,
|
||||
)
|
||||
resp.raise_for_status()
|
||||
return resp.json()
|
||||
|
||||
def setup(self):
|
||||
"""登录并从 API 获取 AI 配置（含提示词和温度）"""
|
||||
self.token = self._login()
|
||||
self.ai_config = self._get("/api/admin/settings/active")
|
||||
self.client = OpenAI(
|
||||
base_url=self.ai_config["api_base"],
|
||||
api_key=self.ai_config["api_key"],
|
||||
)
|
||||
self.model = self.ai_config["model_name"]
|
||||
|
||||
def _get_prompt(self, key: str, default: str) -> str:
|
||||
if self.ai_config:
|
||||
val = self.ai_config.get(key)
|
||||
if val and val.strip():
|
||||
return val
|
||||
return default
|
||||
|
||||
def extract_jokes(self, page_content: str, known_types: list[str], known_crowds: list[str]) -> list[dict]:
|
||||
"""从页面内容中提取笑话，返回结构化数据。"""
|
||||
from crawler.prompts import EXTRACTION_SYSTEM_PROMPT, EXTRACTION_USER_PROMPT
|
||||
system_prompt = self._get_prompt("crawler_extract_prompt",
|
||||
"""你是一个笑话提取专家。从给定的网页文本中识别并提取所有笑话、幽默段子或有趣内容。
|
||||
要求：
|
||||
1. 只返回真正的笑话内容，不要提取普通文章或新闻
|
||||
2. 每条笑话需要包含：title（简短标题）、content（完整笑话内容）、type（类型）、crowd（人群）
|
||||
3. 如果网页中没有笑话，返回空数组 []
|
||||
4. 永远返回合法的 JSON 格式，根节点为数组或包含 jokes 键的对象""")
|
||||
|
||||
user_prompt = EXTRACTION_USER_PROMPT.format(
|
||||
page_content=page_content[:8000],
|
||||
known_types=", ".join(known_types),
|
||||
known_crowds=", ".join(known_crowds),
|
||||
)
|
||||
user_prompt = f"""网页内容：
|
||||
---
|
||||
{page_content[:8000]}
|
||||
---
|
||||
|
||||
已知笑话类型：{', '.join(known_types)}
|
||||
已知人群分类：{', '.join(known_crowds)}
|
||||
|
||||
请提取所有笑话，以 JSON 格式返回，示例：
|
||||
[
|
||||
{{"title": "程序员的幽默", "content": "程序员去相亲...", "types": ["谐音梗", "段子"], "crowds": ["职场", "大学生"]}},
|
||||
{{"title": "...", "content": "...", "types": ["..."], "crowds": ["..."]}}
|
||||
]
|
||||
|
||||
注意：types 和 crowds 是数组，可以填多个。
|
||||
只返回 JSON，不要其他文字。"""
|
||||
|
||||
response = self.client.chat.completions.create(
|
||||
model=self.model,
|
||||
messages=[
|
||||
{"role": "system", "content": EXTRACTION_SYSTEM_PROMPT},
|
||||
{"role": "system", "content": system_prompt},
|
||||
{"role": "user", "content": user_prompt},
|
||||
],
|
||||
temperature=self.temperature,
|
||||
max_tokens=self.max_tokens,
|
||||
temperature=self.ai_config.get("temperature", 0.7) if self.ai_config else 0.7,
|
||||
max_tokens=self.ai_config.get("max_tokens", 2048) if self.ai_config else 2048,
|
||||
)
|
||||
|
||||
raw = response.choices[0].message.content
|
||||
@@ -37,13 +92,19 @@ class AiService:
|
||||
|
||||
def rewrite_joke(self, content: str) -> str:
|
||||
"""润色单条笑话内容。"""
|
||||
from crawler.prompts import REWRITE_SYSTEM_PROMPT, REWRITE_USER_PROMPT
|
||||
system_prompt = self._get_prompt("crawler_rewrite_prompt",
|
||||
"""你是一个幽默作家，负责润色和改写笑话。
|
||||
要求：
|
||||
1. 保持笑话的核心笑点不变
|
||||
2. 语言更通顺、更幽默
|
||||
3. 字数控制在原内容的 80%-120% 之间
|
||||
4. 不要添加任何解释说明""")
|
||||
|
||||
response = self.client.chat.completions.create(
|
||||
model=self.model,
|
||||
messages=[
|
||||
{"role": "system", "content": REWRITE_SYSTEM_PROMPT},
|
||||
{"role": "user", "content": REWRITE_USER_PROMPT.format(content=content)},
|
||||
{"role": "system", "content": system_prompt},
|
||||
{"role": "user", "content": f"请润色以下笑话：\n\n{content}"},
|
||||
],
|
||||
temperature=0.8,
|
||||
max_tokens=500,
|
||||
@@ -60,7 +121,6 @@ class AiService:
|
||||
jokes = data
|
||||
return [j for j in jokes if isinstance(j, dict) and j.get("content")]
|
||||
except json.JSONDecodeError:
|
||||
# 尝试提取 markdown 代码块
|
||||
if "```json" in raw:
|
||||
raw = raw.split("```json")[1].split("```")[0]
|
||||
elif "```" in raw:
|
||||
|
||||
@@ -57,15 +57,13 @@ class Processor:
|
||||
self.token = self._login()
|
||||
print("[*] 登录成功")
|
||||
|
||||
ai_config = self._get("/api/admin/settings/active")
|
||||
self.ai = AiService(
|
||||
api_base=ai_config["api_base"],
|
||||
api_key=ai_config["api_key"],
|
||||
model_name=ai_config["model_name"],
|
||||
temperature=ai_config.get("temperature", 0.7),
|
||||
max_tokens=ai_config.get("max_tokens", 2048),
|
||||
api_base=self.api_base,
|
||||
username=self.username,
|
||||
password=self.password,
|
||||
)
|
||||
print(f"[*] AI 配置: {ai_config['model_name']}")
|
||||
self.ai.setup()
|
||||
print(f"[*] AI 配置: {self.ai.model}")
|
||||
|
||||
self.types = self._get("/api/categories/types")
|
||||
self.crowds = self._get("/api/categories/crowds")
|
||||
|
||||
@@ -1,41 +0,0 @@
|
||||
"""AI 提示词模板。"""
|
||||
|
||||
# ===== 笑话提取 =====
|
||||
EXTRACTION_SYSTEM_PROMPT = """你是一个笑话提取专家。从给定的网页文本中识别并提取所有笑话、幽默段子或有趣内容。
|
||||
要求：
|
||||
1. 只返回真正的笑话内容，不要提取普通文章或新闻
|
||||
2. 每条笑话需要包含：title（简短标题）、content（完整笑话内容）、type（类型）、crowd（人群）
|
||||
3. 如果网页中没有笑话，返回空数组 []
|
||||
4. 永远返回合法的 JSON 格式，根节点为数组或包含 jokes 键的对象"""
|
||||
|
||||
EXTRACTION_USER_PROMPT = """网页内容：
|
||||
---
|
||||
{page_content}
|
||||
---
|
||||
|
||||
已知笑话类型：{known_types}
|
||||
已知人群分类：{known_crowds}
|
||||
|
||||
请提取所有笑话，以 JSON 格式返回，示例：
|
||||
[
|
||||
{{"title": "程序员的幽默", "content": "程序员去相亲...", "types": ["谐音梗", "段子"], "crowds": ["职场", "大学生"]}},
|
||||
{{"title": "...", "content": "...", "types": ["..."], "crowds": ["..."]}}
|
||||
]
|
||||
|
||||
注意：types 和 crowds 是数组，可以填多个。
|
||||
只返回 JSON，不要其他文字。"""
|
||||
|
||||
|
||||
# ===== 笑话改写 =====
|
||||
REWRITE_SYSTEM_PROMPT = """你是一个幽默作家，负责润色和改写笑话。
|
||||
要求：
|
||||
1. 保持笑话的核心笑点不变
|
||||
2. 语言更通顺、更幽默
|
||||
3. 字数控制在原内容的 80%-120% 之间
|
||||
4. 不要添加任何解释说明"""
|
||||
|
||||
REWRITE_USER_PROMPT = """请润色以下笑话：
|
||||
|
||||
{content}
|
||||
|
||||
只返回润色后的笑话文字，不要其他内容。"""
|
||||
Reference in New Issue
Block a user