第一次提交

2026-08-01 16:50:55 +08:00
commit c8f9e39039
68 changed files with 6016 additions and 0 deletions
@@ -0,0 +1,111 @@
+"""AI 任务：判分、讲解、知识点总结。统一记录到 ai_generations。"""
+import json
+import logging
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.models.practice import AiGeneration
+from app.models.question import Question
+from app.models.user import UserSettings
+from app.services.llm.client import LLMClient
+from app.services.llm.prompts import (
+    EXPLAIN_SYSTEM,
+    EXPLAIN_USER_TEMPLATE,
+    JUDGE_SYSTEM,
+    JUDGE_USER_TEMPLATE,
+    TAGS_SYSTEM,
+    TAGS_USER_TEMPLATE,
+)
+
+logger = logging.getLogger(__name__)
+
+
+async def load_llm(db: AsyncSession, user_id: str) -> LLMClient:
+    """加载用户的 LLMClient（未配置会抛 LLMConfigError）。"""
+    row = await db.scalar(select(UserSettings).where(UserSettings.user_id == user_id))
+    return LLMClient(row)
+
+
+def _parse_json(raw: str) -> dict:
+    text = raw.strip()
+    if text.startswith("```"):
+        text = text.split("```", 2)[1]
+        if text.startswith("json"):
+            text = text[4:]
+    return json.loads(text.strip())
+
+
+async def _record(
+    db: AsyncSession, user_id: str, question_id: str | None, task: str, result, content: str
+):
+    db.add(
+        AiGeneration(
+            user_id=user_id,
+            question_id=question_id,
+            task=task,
+            model=result.model,
+            request_tokens=result.prompt_tokens,
+            response_tokens=result.completion_tokens,
+            content_markdown=content,
+        )
+    )
+
+
+async def judge_answer(
+    db: AsyncSession, llm: LLMClient, user_id: str, q: Question, user_answer: list[str]
+) -> dict:
+    """主观题判分。返回 {is_correct, feedback_markdown, key_points_missed}。"""
+    answer = "、".join(q.correct_answer) if q.correct_answer else "（未提供标准答案）"
+    result = await llm.chat_text(
+        JUDGE_SYSTEM,
+        JUDGE_USER_TEMPLATE.format(
+            stem=q.stem_markdown, answer=answer, user_answer="\n".join(user_answer)
+        ),
+        json_mode=True,
+    )
+    await _record(db, user_id, q.id, "judge", result, result.content)
+    try:
+        data = _parse_json(result.content)
+        return {
+            "is_correct": bool(data.get("is_correct")),
+            "feedback_markdown": data.get("feedback_markdown", ""),
+            "key_points_missed": data.get("key_points_missed", []),
+        }
+    except (json.JSONDecodeError, KeyError) as e:
+        logger.warning("judge 解析失败：%s", e)
+        return {
+            "is_correct": False,
+            "feedback_markdown": result.content,
+            "key_points_missed": [],
+        }
+
+
+async def explain_question(
+    db: AsyncSession, llm: LLMClient, user_id: str, q: Question
+) -> str:
+    answer_hint = ""
+    if q.correct_answer:
+        answer_hint = f"参考答案：{'、'.join(q.correct_answer)}\n\n"
+    result = await llm.chat_text(
+        EXPLAIN_SYSTEM,
+        EXPLAIN_USER_TEMPLATE.format(stem=q.stem_markdown, answer_hint=answer_hint),
+    )
+    await _record(db, user_id, q.id, "explain", result, result.content)
+    return result.content
+
+
+async def summarize_tags(
+    db: AsyncSession, llm: LLMClient, user_id: str, q: Question
+) -> list[str]:
+    result = await llm.chat_text(
+        TAGS_SYSTEM, TAGS_USER_TEMPLATE.format(stem=q.stem_markdown), json_mode=True
+    )
+    await _record(db, user_id, q.id, "summarize", result, result.content)
+    try:
+        data = _parse_json(result.content)
+        tags = data.get("tags", [])
+        return [str(t).strip() for t in tags if str(t).strip()][:4]
+    except (json.JSONDecodeError, KeyError) as e:
+        logger.warning("tags 解析失败：%s", e)
+        return []
@@ -0,0 +1,103 @@
+"""统一的 OpenAI 兼容 LLM client：文本 + 多模态。
+
+配置来自用户的 UserSettings（base_url / api_key / 模型名）。api_key 在库里
+是 Fernet 加密的，这里解密后使用。
+"""
+import base64
+from dataclasses import dataclass
+
+from openai import AsyncOpenAI
+
+from app.models.user import UserSettings
+from app.services.crypto import decrypt
+
+
+class LLMConfigError(RuntimeError):
+    """用户尚未正确配置 AI 接口。"""
+
+
+@dataclass
+class LLMResult:
+    content: str
+    model: str | None
+    prompt_tokens: int | None
+    completion_tokens: int | None
+
+
+class LLMClient:
+    def __init__(self, settings: UserSettings):
+        if not settings or not settings.llm_base_url or not settings.llm_api_key_encrypted:
+            raise LLMConfigError("请先在设置里配置 AI 接口（base_url 与 api_key）")
+        api_key = decrypt(settings.llm_api_key_encrypted)
+        if not api_key:
+            raise LLMConfigError("api_key 解密失败，请在设置里重新填写")
+        self._client = AsyncOpenAI(base_url=settings.llm_base_url, api_key=api_key)
+        self._text_model = settings.llm_text_model
+        self._vision_model = settings.llm_vision_model or settings.llm_text_model
+
+    async def chat_text(
+        self,
+        system: str,
+        user: str,
+        *,
+        json_mode: bool = False,
+        model: str | None = None,
+    ) -> LLMResult:
+        m = model or self._text_model
+        if not m:
+            raise LLMConfigError("未配置文本模型名")
+        kwargs = {}
+        if json_mode:
+            kwargs["response_format"] = {"type": "json_object"}
+        resp = await self._client.chat.completions.create(
+            model=m,
+            messages=[
+                {"role": "system", "content": system},
+                {"role": "user", "content": user},
+            ],
+            **kwargs,
+        )
+        return self._to_result(resp, m)
+
+    async def chat_vision(
+        self,
+        system: str,
+        user_text: str,
+        image_bytes: bytes,
+        mime: str = "image/jpeg",
+        *,
+        json_mode: bool = False,
+        model: str | None = None,
+    ) -> LLMResult:
+        m = model or self._vision_model
+        if not m:
+            raise LLMConfigError("未配置多模态模型名")
+        data_url = f"data:{mime};base64,{base64.b64encode(image_bytes).decode()}"
+        kwargs = {}
+        if json_mode:
+            kwargs["response_format"] = {"type": "json_object"}
+        resp = await self._client.chat.completions.create(
+            model=m,
+            messages=[
+                {"role": "system", "content": system},
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "text", "text": user_text},
+                        {"type": "image_url", "image_url": {"url": data_url}},
+                    ],
+                },
+            ],
+            **kwargs,
+        )
+        return self._to_result(resp, m)
+
+    @staticmethod
+    def _to_result(resp, model: str) -> LLMResult:
+        usage = getattr(resp, "usage", None)
+        return LLMResult(
+            content=resp.choices[0].message.content or "",
+            model=model,
+            prompt_tokens=getattr(usage, "prompt_tokens", None),
+            completion_tokens=getattr(usage, "completion_tokens", None),
+        )
@@ -0,0 +1,75 @@
+"""各 AI 任务的 prompt 模板。"""
+
+# ---- OCR：多模态直接看图，转成含 LaTeX 的 Markdown（不作答）----
+VLM_OCR_SYSTEM = (
+    "你是一个精准的题目识别助手。用户会给你一张包含题目的图片，"
+    "请把图片中的题目内容原样转写为 Markdown 文本。要求：\n"
+    "1. 数学公式用 LaTeX 表示，行内用 $...$，独立公式用 $$...$$。\n"
+    "2. 保留题目的选项（如 A/B/C/D）、题号、结构。\n"
+    "3. 只转写题目本身，不要作答、不要解释、不要添加任何额外内容。\n"
+    "4. 如果有多道题，全部转写。"
+)
+VLM_OCR_USER = "请转写这张图片里的题目。"
+
+# ---- 结构化提取：把 OCR/图片内容解析成结构化题目 JSON ----
+EXTRACT_SYSTEM = """你是一个题目结构化助手。用户会给你一段题目文本（Markdown，可能含 LaTeX 公式），
+其中可能包含一道或多道题。请把**每一道题**解析成一个对象，输出 JSON：
+{
+  "questions": [
+    {
+      "type": "unclassified | single_choice | multiple_choice | true_false | fill_blank | short_answer",
+      "stem_markdown": "题干（保留原始 LaTeX）",
+      "options": [{"key": "A", "text_markdown": "..."}]  // 仅选择/判断题；否则为 null,
+      "correct_answer": ["A"] 或 ["文本答案"] 或 null,  // 原文未给出答案则为 null
+      "difficulty": null 或 1-5
+    }
+  ]
+}
+题型判断规则：
+- 有 ABCD 等多个选项且只选一个 → single_choice；可多选 → multiple_choice
+- 判断对错（正确/错误、对/错、True/False）→ true_false，options 用 [{"key":"T","text_markdown":"正确"},{"key":"F","text_markdown":"错误"}]
+- 有下划线、括号、填空 → fill_blank
+- 其余开放性问答 → short_answer
+- 判断不了、或者这段内容不像一道完整的题 → unclassified（这是可以接受的，不要硬猜）
+
+重要：
+- 有几道题就输出几个对象，不要合并成一条，也不要漏掉。
+- 只有一道题时也要放在 questions 数组里。
+- 原文没给答案就填 null，不要自己解题填答案。
+只输出 JSON，不要额外文字。"""
+
+EXTRACT_USER_TEMPLATE = "题目文本：\n\n{content}"
+
+# ---- 判分（主观题）----
+JUDGE_SYSTEM = """你是一个阅卷助手。根据题目、标准答案（可能没有）和学生答案，判断对错并给出反馈。
+输出 JSON：
+{
+  "is_correct": true/false,
+  "feedback_markdown": "简短反馈，指出对错原因（可含 LaTeX）",
+  "key_points_missed": ["遗漏的要点", ...]
+}
+只输出 JSON。"""
+
+JUDGE_USER_TEMPLATE = """题目：
+{stem}
+
+标准答案：{answer}
+
+学生答案：
+{user_answer}"""
+
+# ---- 讲解 ----
+EXPLAIN_SYSTEM = (
+    "你是一位耐心的老师。请针对给定题目给出面向自学者的分步讲解，"
+    "解释解题思路而非只给结论。使用 Markdown，公式用 LaTeX（$...$ / $$...$$）。"
+)
+EXPLAIN_USER_TEMPLATE = """题目：
+{stem}
+
+{answer_hint}请给出详细讲解。"""
+
+# ---- 知识点标签 ----
+TAGS_SYSTEM = """你是一个知识点归类助手。根据题目内容，输出 1-4 个简短的知识点标签。
+输出 JSON：{"tags": ["标签1", "标签2"]}
+标签要简洁（2-8 字），是学科知识点，不要句子。只输出 JSON。"""
+TAGS_USER_TEMPLATE = "题目：\n{stem}"