mirror of
https://github.com/handsomezhuzhu/QQuiz.git
synced 2026-02-20 20:10:14 +00:00
feat: 添加 Gemini 支持和 AI 参考答案生成功能
主要功能： - 🎯 新增 Google Gemini AI 提供商支持 - 原生 PDF 理解能力（最多1000页） - 完整保留图片、表格、公式等内容 - 支持自定义 Base URL（用于代理/中转服务） - 🤖 实现 AI 参考答案自动生成 - 当题目缺少答案时自动调用 AI 生成参考答案 - 支持单选、多选、判断、简答等所有题型 - 答案标记为"AI参考答案："便于识别 - 🔧 优化文档解析功能 - 改进中文 Prompt 提高识别准确度 - 自动修复 JSON 中的控制字符（换行符等） - 智能题目类型验证和自动转换（proof→short等） - 增加超时时间和重试机制 - 🎨 完善管理后台配置界面 - 新增 Gemini 配置区域 - 突出显示 PDF 原生支持特性 - 为其他提供商添加"仅文本"警告 - 支持 Gemini Base URL 自定义 技术改进： - 添加 google-genai 依赖 - 实现异步 API 调用适配 - 完善错误处理和日志输出 - 统一配置管理和数据库存储 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
@@ -6,6 +6,8 @@ import json
|
||||
from typing import List, Dict, Any, Optional
|
||||
from openai import AsyncOpenAI
|
||||
from anthropic import AsyncAnthropic
|
||||
from google import genai
|
||||
from google.genai import types
|
||||
import httpx
|
||||
|
||||
from models import QuestionType
|
||||
@@ -36,9 +38,17 @@ class LLMService:
|
||||
|
||||
self.client = AsyncOpenAI(
|
||||
api_key=api_key,
|
||||
base_url=base_url
|
||||
base_url=base_url,
|
||||
timeout=120.0, # 增加超时时间到 120 秒
|
||||
max_retries=3 # 自动重试 3 次
|
||||
)
|
||||
|
||||
# Log configuration for debugging
|
||||
print(f"[LLM Config] Provider: OpenAI", flush=True)
|
||||
print(f"[LLM Config] Base URL: {base_url}", flush=True)
|
||||
print(f"[LLM Config] Model: {self.model}", flush=True)
|
||||
print(f"[LLM Config] API Key: {api_key[:10]}...{api_key[-4:] if len(api_key) > 14 else 'xxxx'}", flush=True)
|
||||
|
||||
elif self.provider == "anthropic":
|
||||
api_key = (config or {}).get("anthropic_api_key") or os.getenv("ANTHROPIC_API_KEY")
|
||||
self.model = (config or {}).get("anthropic_model") or os.getenv("ANTHROPIC_MODEL", "claude-3-haiku-20240307")
|
||||
@@ -60,9 +70,58 @@ class LLMService:
|
||||
|
||||
self.client = AsyncOpenAI(
|
||||
api_key=api_key,
|
||||
base_url=base_url
|
||||
base_url=base_url,
|
||||
timeout=120.0, # 增加超时时间到 120 秒
|
||||
max_retries=3 # 自动重试 3 次
|
||||
)
|
||||
|
||||
elif self.provider == "gemini":
|
||||
api_key = (config or {}).get("gemini_api_key") or os.getenv("GEMINI_API_KEY")
|
||||
base_url = (config or {}).get("gemini_base_url") or os.getenv("GEMINI_BASE_URL")
|
||||
self.model = (config or {}).get("gemini_model") or os.getenv("GEMINI_MODEL", "gemini-2.0-flash-exp")
|
||||
|
||||
if not api_key:
|
||||
raise ValueError("Gemini API key not configured")
|
||||
|
||||
# Create client with optional custom base URL
|
||||
if base_url:
|
||||
# Use custom base URL (for proxy/relay services)
|
||||
print(f"[LLM Config] Using custom Gemini base URL: {base_url}", flush=True)
|
||||
|
||||
# Try different methods to set custom base URL
|
||||
try:
|
||||
# Method 1: Try http_options parameter
|
||||
self.client = genai.Client(
|
||||
api_key=api_key,
|
||||
http_options={'api_endpoint': base_url}
|
||||
)
|
||||
print(f"[LLM Config] ✓ Set base URL via http_options", flush=True)
|
||||
except TypeError:
|
||||
try:
|
||||
# Method 2: Try vertexai parameter (some versions)
|
||||
self.client = genai.Client(
|
||||
api_key=api_key,
|
||||
vertexai=False,
|
||||
client_options={'api_endpoint': base_url}
|
||||
)
|
||||
print(f"[LLM Config] ✓ Set base URL via client_options", flush=True)
|
||||
except:
|
||||
# Method 3: Set environment variable and create client
|
||||
print(f"[LLM Config] ⚠️ SDK doesn't support custom URL parameter, using environment variable", flush=True)
|
||||
os.environ['GOOGLE_API_BASE'] = base_url
|
||||
self.client = genai.Client(api_key=api_key)
|
||||
print(f"[LLM Config] ✓ Set base URL via environment variable", flush=True)
|
||||
else:
|
||||
# Use default Google API
|
||||
self.client = genai.Client(api_key=api_key)
|
||||
|
||||
# Log configuration for debugging
|
||||
print(f"[LLM Config] Provider: Gemini", flush=True)
|
||||
print(f"[LLM Config] Model: {self.model}", flush=True)
|
||||
if base_url:
|
||||
print(f"[LLM Config] Base URL: {base_url}", flush=True)
|
||||
print(f"[LLM Config] API Key: {api_key[:10]}...{api_key[-4:] if len(api_key) > 14 else 'xxxx'}", flush=True)
|
||||
|
||||
else:
|
||||
raise ValueError(f"Unsupported AI provider: {self.provider}")
|
||||
|
||||
@@ -82,33 +141,49 @@ class LLMService:
|
||||
...
|
||||
]
|
||||
"""
|
||||
prompt = """You are a professional question parser. Parse the given document and extract all questions.
|
||||
prompt = """你是一个专业的试题解析专家。请仔细分析下面的文档内容，提取其中的所有试题。
|
||||
|
||||
For each question, identify:
|
||||
1. Question content (the question text)
|
||||
2. Question type: single (单选), multiple (多选), judge (判断), short (简答)
|
||||
3. Options (for choice questions only, format: ["A. Option1", "B. Option2", ...])
|
||||
4. Correct answer
|
||||
5. Analysis/Explanation (if available)
|
||||
请注意：
|
||||
- 文档中可能包含中文或英文题目
|
||||
- 题目可能有多种格式，请灵活识别
|
||||
- 即使格式不标准，也请尽量提取题目内容
|
||||
- 如果文档只是普通文章而没有题目，请返回空数组 []
|
||||
|
||||
Return ONLY a JSON array of questions, with no additional text:
|
||||
对于每道题目，请识别：
|
||||
1. 题目内容 (完整的题目文字)
|
||||
2. 题目类型（**只能**使用以下4种类型之一）：
|
||||
- single：单选题
|
||||
- multiple：多选题
|
||||
- judge：判断题
|
||||
- short：简答题（包括问答题、计算题、证明题、填空题等所有非选择题）
|
||||
3. 选项 (仅针对选择题，格式: ["A. 选项1", "B. 选项2", ...])
|
||||
4. 正确答案 (请仔细查找文档中的答案。如果确实没有答案，可以填 null)
|
||||
5. 解析/说明 (如果有的话)
|
||||
|
||||
**重要**：题目类型必须是 single、multiple、judge、short 之一，不要使用其他类型名称！
|
||||
|
||||
返回格式：请**只返回** JSON 数组，不要有任何其他文字或 markdown 代码块：
|
||||
[
|
||||
{
|
||||
"content": "question text",
|
||||
{{
|
||||
"content": "题目内容",
|
||||
"type": "single",
|
||||
"options": ["A. Option1", "B. Option2", "C. Option3", "D. Option4"],
|
||||
"options": ["A. 选项1", "B. 选项2", "C. 选项3", "D. 选项4"],
|
||||
"answer": "A",
|
||||
"analysis": "explanation"
|
||||
},
|
||||
"analysis": "解析说明"
|
||||
}},
|
||||
...
|
||||
]
|
||||
|
||||
Document content:
|
||||
文档内容：
|
||||
---
|
||||
{content}
|
||||
---
|
||||
|
||||
IMPORTANT: Return ONLY the JSON array, no markdown code blocks or explanations."""
|
||||
重要提示：
|
||||
- 仔细阅读文档内容
|
||||
- 识别所有看起来像试题的内容
|
||||
- 如果文档中没有题目（比如只是普通文章），返回 []
|
||||
- **只返回 JSON 数组**，不要包含 ```json 或其他标记"""
|
||||
|
||||
try:
|
||||
if self.provider == "anthropic":
|
||||
@@ -120,6 +195,20 @@ IMPORTANT: Return ONLY the JSON array, no markdown code blocks or explanations."
|
||||
]
|
||||
)
|
||||
result = response.content[0].text
|
||||
elif self.provider == "gemini":
|
||||
# Gemini uses different API
|
||||
import asyncio
|
||||
|
||||
def _generate_content():
|
||||
return self.client.models.generate_content(
|
||||
model=self.model,
|
||||
contents=prompt.format(content=content)
|
||||
)
|
||||
|
||||
print(f"[Gemini Text] Calling Gemini API with model: {self.model}", flush=True)
|
||||
response = await asyncio.to_thread(_generate_content)
|
||||
result = response.text
|
||||
print(f"[Gemini Text] API call completed", flush=True)
|
||||
else: # OpenAI or Qwen
|
||||
response = await self.client.chat.completions.create(
|
||||
model=self.model,
|
||||
@@ -131,28 +220,398 @@ IMPORTANT: Return ONLY the JSON array, no markdown code blocks or explanations."
|
||||
)
|
||||
result = response.choices[0].message.content
|
||||
|
||||
# Log original response for debugging
|
||||
import sys
|
||||
print(f"[LLM Raw Response] Length: {len(result)} chars", flush=True)
|
||||
print(f"[LLM Raw Response] First 300 chars:\n{result[:300]}", flush=True)
|
||||
print(f"[LLM Raw Response] Last 200 chars:\n{result[-200:]}", flush=True)
|
||||
sys.stdout.flush()
|
||||
|
||||
# Clean result and parse JSON
|
||||
result = result.strip()
|
||||
|
||||
# Remove markdown code blocks
|
||||
if result.startswith("```json"):
|
||||
result = result[7:]
|
||||
if result.startswith("```"):
|
||||
elif result.startswith("```"):
|
||||
result = result[3:]
|
||||
|
||||
if result.endswith("```"):
|
||||
result = result[:-3]
|
||||
|
||||
result = result.strip()
|
||||
|
||||
questions = json.loads(result)
|
||||
# Try to find JSON array if there's extra text
|
||||
if not result.startswith('['):
|
||||
# Find the first '[' character
|
||||
start_idx = result.find('[')
|
||||
if start_idx != -1:
|
||||
print(f"[JSON Cleanup] Found '[' at position {start_idx}, extracting array...")
|
||||
result = result[start_idx:]
|
||||
else:
|
||||
print(f"[JSON Error] No '[' found in response!")
|
||||
raise Exception("LLM response does not contain a JSON array")
|
||||
|
||||
# Add content hash to each question
|
||||
if not result.endswith(']'):
|
||||
# Find the last ']' character
|
||||
end_idx = result.rfind(']')
|
||||
if end_idx != -1:
|
||||
print(f"[JSON Cleanup] Found last ']' at position {end_idx}")
|
||||
result = result[:end_idx + 1]
|
||||
|
||||
result = result.strip()
|
||||
|
||||
# Log the cleaned result for debugging
|
||||
print(f"[LLM Cleaned JSON] Length: {len(result)} chars")
|
||||
print(f"[LLM Cleaned JSON] First 300 chars:\n{result[:300]}")
|
||||
|
||||
try:
|
||||
questions = json.loads(result)
|
||||
except json.JSONDecodeError as je:
|
||||
print(f"[JSON Error] Failed to parse JSON at line {je.lineno}, column {je.colno}")
|
||||
print(f"[JSON Error] Error: {je.msg}")
|
||||
|
||||
# If error is about control characters, try to fix them
|
||||
if "control character" in je.msg.lower() or "invalid \\escape" in je.msg.lower():
|
||||
print(f"[JSON Cleanup] Attempting to fix control characters...", flush=True)
|
||||
|
||||
# Fix unescaped control characters in JSON string values
|
||||
import re
|
||||
|
||||
def fix_string_value(match):
|
||||
"""Fix control characters inside a JSON string value"""
|
||||
string_content = match.group(1)
|
||||
# Escape control characters
|
||||
string_content = string_content.replace('\n', '\\n')
|
||||
string_content = string_content.replace('\r', '\\r')
|
||||
string_content = string_content.replace('\t', '\\t')
|
||||
string_content = string_content.replace('\b', '\\b')
|
||||
string_content = string_content.replace('\f', '\\f')
|
||||
return f'"{string_content}"'
|
||||
|
||||
# Match string values in JSON
|
||||
# Pattern matches: "..." (handles escaped quotes and backslashes)
|
||||
# (?:[^"\\]|\\.)* means: either non-quote-non-backslash OR backslash-followed-by-anything, repeated
|
||||
fixed_result = re.sub(r'"((?:[^"\\]|\\.)*)"', fix_string_value, result)
|
||||
|
||||
print(f"[JSON Cleanup] Retrying with fixed control characters...", flush=True)
|
||||
try:
|
||||
questions = json.loads(fixed_result)
|
||||
print(f"[JSON Cleanup] ✅ Successfully parsed after fixing control characters!", flush=True)
|
||||
except json.JSONDecodeError as je2:
|
||||
print(f"[JSON Error] Still failed after fix: {je2.msg}", flush=True)
|
||||
# Print context around the error
|
||||
lines = result.split('\n')
|
||||
if je.lineno <= len(lines):
|
||||
start = max(0, je.lineno - 3)
|
||||
end = min(len(lines), je.lineno + 2)
|
||||
print(f"[JSON Error] Context (lines {start+1}-{end}):")
|
||||
for i in range(start, end):
|
||||
marker = " >>> " if i == je.lineno - 1 else " "
|
||||
print(f"{marker}{i+1}: {lines[i]}")
|
||||
raise Exception(f"Invalid JSON format from LLM: {je.msg} at line {je.lineno}")
|
||||
else:
|
||||
# Print context around the error
|
||||
lines = result.split('\n')
|
||||
if je.lineno <= len(lines):
|
||||
start = max(0, je.lineno - 3)
|
||||
end = min(len(lines), je.lineno + 2)
|
||||
print(f"[JSON Error] Context (lines {start+1}-{end}):")
|
||||
for i in range(start, end):
|
||||
marker = " >>> " if i == je.lineno - 1 else " "
|
||||
print(f"{marker}{i+1}: {lines[i]}")
|
||||
raise Exception(f"Invalid JSON format from LLM: {je.msg} at line {je.lineno}")
|
||||
|
||||
# Validate that we got a list
|
||||
if not isinstance(questions, list):
|
||||
raise Exception(f"Expected a list of questions, got {type(questions)}")
|
||||
|
||||
if len(questions) == 0:
|
||||
raise Exception("No questions found in the parsed result")
|
||||
|
||||
# Validate and fix question types
|
||||
valid_types = {"single", "multiple", "judge", "short"}
|
||||
type_mapping = {
|
||||
"proof": "short",
|
||||
"essay": "short",
|
||||
"calculation": "short",
|
||||
"fill": "short",
|
||||
"填空": "short",
|
||||
"证明": "short",
|
||||
"计算": "short",
|
||||
"问答": "short",
|
||||
"单选": "single",
|
||||
"多选": "multiple",
|
||||
"判断": "judge",
|
||||
"简答": "short"
|
||||
}
|
||||
|
||||
# Add content hash and validate types
|
||||
for q in questions:
|
||||
if "content" not in q:
|
||||
print(f"[Warning] Question missing 'content' field: {q}")
|
||||
continue
|
||||
|
||||
# Validate and fix question type
|
||||
q_type = q.get("type", "short")
|
||||
if isinstance(q_type, str):
|
||||
q_type_lower = q_type.lower()
|
||||
if q_type_lower not in valid_types:
|
||||
# Try to map to valid type
|
||||
if q_type_lower in type_mapping:
|
||||
old_type = q_type
|
||||
q["type"] = type_mapping[q_type_lower]
|
||||
print(f"[Type Fix] Changed '{old_type}' to '{q['type']}' for question: {q['content'][:50]}...", flush=True)
|
||||
else:
|
||||
# Default to short answer
|
||||
print(f"[Type Fix] Unknown type '{q_type}', defaulting to 'short' for question: {q['content'][:50]}...", flush=True)
|
||||
q["type"] = "short"
|
||||
else:
|
||||
q["type"] = q_type_lower
|
||||
else:
|
||||
q["type"] = "short"
|
||||
|
||||
q["content_hash"] = calculate_content_hash(q["content"])
|
||||
|
||||
return questions
|
||||
|
||||
except Exception as e:
|
||||
print(f"Error parsing document: {e}")
|
||||
print(f"[Error] Document parsing failed: {str(e)}")
|
||||
raise Exception(f"Failed to parse document: {str(e)}")
|
||||
|
||||
async def parse_document_with_pdf(self, pdf_bytes: bytes, filename: str) -> List[Dict[str, Any]]:
|
||||
"""
|
||||
Parse PDF document using Gemini's native PDF understanding.
|
||||
Only works with Gemini provider.
|
||||
|
||||
Args:
|
||||
pdf_bytes: PDF file content as bytes
|
||||
filename: Original filename for logging
|
||||
|
||||
Returns:
|
||||
List of question dictionaries
|
||||
"""
|
||||
if self.provider != "gemini":
|
||||
raise ValueError("PDF parsing is only supported with Gemini provider")
|
||||
|
||||
prompt = """你是一个专业的试题解析专家。请仔细分析这个 PDF 文档，提取其中的所有试题。
|
||||
|
||||
请注意：
|
||||
- PDF 中可能包含中文或英文题目
|
||||
- 题目可能有多种格式，请灵活识别
|
||||
- 即使格式不标准，也请尽量提取题目内容
|
||||
- 题目内容如果包含代码或换行，请将换行符替换为空格或\\n
|
||||
|
||||
对于每道题目，请识别：
|
||||
1. 题目内容 (完整的题目文字，如果有代码请保持在一行或用\\n表示换行)
|
||||
2. 题目类型（**只能**使用以下4种类型之一）：
|
||||
- single：单选题
|
||||
- multiple：多选题
|
||||
- judge：判断题
|
||||
- short：简答题（包括问答题、计算题、证明题、填空题等所有非选择题）
|
||||
3. 选项 (仅针对选择题，格式: ["A. 选项1", "B. 选项2", ...])
|
||||
4. 正确答案 (请仔细查找文档中的答案。如果确实没有答案，可以填 null)
|
||||
5. 解析/说明 (如果有的话)
|
||||
|
||||
**重要**：题目类型必须是 single、multiple、judge、short 之一，不要使用其他类型名称！
|
||||
|
||||
返回格式要求：
|
||||
1. **必须**返回一个完整的 JSON 数组（以 [ 开始，以 ] 结束）
|
||||
2. **不要**返回 JSONL 格式（每行一个 JSON 对象）
|
||||
3. **不要**包含 markdown 代码块标记（```json 或 ```）
|
||||
4. **不要**包含任何解释性文字
|
||||
|
||||
正确的格式示例：
|
||||
[
|
||||
{{
|
||||
"content": "题目内容",
|
||||
"type": "single",
|
||||
"options": ["A. 选项1", "B. 选项2", "C. 选项3", "D. 选项4"],
|
||||
"answer": "A",
|
||||
"analysis": "解析说明"
|
||||
}},
|
||||
{{
|
||||
"content": "第二道题",
|
||||
"type": "judge",
|
||||
"options": [],
|
||||
"answer": "对",
|
||||
"analysis": null
|
||||
}}
|
||||
]
|
||||
|
||||
重要提示：
|
||||
- 请仔细查看 PDF 的每一页
|
||||
- 识别所有看起来像试题的内容
|
||||
- 如果找不到明确的选项，可以根据上下文推断题目类型
|
||||
- 题目内容中的换行请用\\n或空格替换，确保 JSON 格式正确
|
||||
- **只返回一个 JSON 数组**，不要包含其他任何内容"""
|
||||
|
||||
try:
|
||||
print(f"[Gemini PDF] Processing PDF: {filename}", flush=True)
|
||||
print(f"[Gemini PDF] File size: {len(pdf_bytes)} bytes", flush=True)
|
||||
|
||||
# Use Gemini's native PDF processing
|
||||
# Run sync API in thread pool to avoid blocking
|
||||
import asyncio
|
||||
|
||||
def _generate_content():
|
||||
return self.client.models.generate_content(
|
||||
model=self.model,
|
||||
contents=[
|
||||
types.Part.from_bytes(
|
||||
data=pdf_bytes,
|
||||
mime_type='application/pdf',
|
||||
),
|
||||
prompt
|
||||
]
|
||||
)
|
||||
|
||||
print(f"[Gemini PDF] Calling Gemini API with model: {self.model}", flush=True)
|
||||
response = await asyncio.to_thread(_generate_content)
|
||||
print(f"[Gemini PDF] API call completed", flush=True)
|
||||
|
||||
result = response.text
|
||||
print(f"[Gemini PDF] Response retrieved, checking content...", flush=True)
|
||||
|
||||
# Log original response for debugging
|
||||
import sys
|
||||
print(f"[LLM Raw Response] Length: {len(result)} chars", flush=True)
|
||||
print(f"[LLM Raw Response] First 300 chars:\n{result[:300]}", flush=True)
|
||||
print(f"[LLM Raw Response] Last 200 chars:\n{result[-200:]}", flush=True)
|
||||
sys.stdout.flush()
|
||||
|
||||
# Clean result and parse JSON (same as text method)
|
||||
result = result.strip()
|
||||
|
||||
# Remove markdown code blocks
|
||||
if result.startswith("```json"):
|
||||
result = result[7:]
|
||||
elif result.startswith("```"):
|
||||
result = result[3:]
|
||||
|
||||
if result.endswith("```"):
|
||||
result = result[:-3]
|
||||
|
||||
result = result.strip()
|
||||
|
||||
# Try to find JSON array if there's extra text
|
||||
if not result.startswith('['):
|
||||
start_idx = result.find('[')
|
||||
if start_idx != -1:
|
||||
print(f"[JSON Cleanup] Found '[' at position {start_idx}, extracting array...", flush=True)
|
||||
result = result[start_idx:]
|
||||
else:
|
||||
print(f"[JSON Error] No '[' found in response!", flush=True)
|
||||
raise Exception("LLM response does not contain a JSON array")
|
||||
|
||||
if not result.endswith(']'):
|
||||
end_idx = result.rfind(']')
|
||||
if end_idx != -1:
|
||||
print(f"[JSON Cleanup] Found last ']' at position {end_idx}", flush=True)
|
||||
result = result[:end_idx + 1]
|
||||
|
||||
result = result.strip()
|
||||
|
||||
# Log the cleaned result for debugging
|
||||
print(f"[LLM Cleaned JSON] Length: {len(result)} chars", flush=True)
|
||||
print(f"[LLM Cleaned JSON] First 300 chars:\n{result[:300]}", flush=True)
|
||||
|
||||
try:
|
||||
questions = json.loads(result)
|
||||
except json.JSONDecodeError as je:
|
||||
print(f"[JSON Error] Failed to parse JSON at line {je.lineno}, column {je.colno}", flush=True)
|
||||
print(f"[JSON Error] Error: {je.msg}", flush=True)
|
||||
# Print context around the error
|
||||
lines = result.split('\n')
|
||||
if je.lineno <= len(lines):
|
||||
start = max(0, je.lineno - 3)
|
||||
end = min(len(lines), je.lineno + 2)
|
||||
print(f"[JSON Error] Context (lines {start+1}-{end}):", flush=True)
|
||||
for i in range(start, end):
|
||||
marker = " >>> " if i == je.lineno - 1 else " "
|
||||
print(f"{marker}{i+1}: {lines[i]}", flush=True)
|
||||
raise Exception(f"Invalid JSON format from LLM: {je.msg} at line {je.lineno}")
|
||||
|
||||
# Validate that we got a list
|
||||
if not isinstance(questions, list):
|
||||
raise Exception(f"Expected a list of questions, got {type(questions)}")
|
||||
|
||||
if len(questions) == 0:
|
||||
# Provide more helpful error message
|
||||
print(f"[Gemini PDF] ⚠️ Gemini returned empty array - PDF may not contain recognizable questions", flush=True)
|
||||
print(f"[Gemini PDF] 💡 Trying to get Gemini's explanation...", flush=True)
|
||||
|
||||
# Ask Gemini what it saw in the PDF
|
||||
def _ask_what_gemini_sees():
|
||||
return self.client.models.generate_content(
|
||||
model=self.model,
|
||||
contents=[
|
||||
types.Part.from_bytes(
|
||||
data=pdf_bytes,
|
||||
mime_type='application/pdf',
|
||||
),
|
||||
"Please describe what you see in this PDF document. What is the main content? Are there any questions, exercises, or test items? Respond in Chinese."
|
||||
]
|
||||
)
|
||||
|
||||
import asyncio
|
||||
explanation_response = await asyncio.to_thread(_ask_what_gemini_sees)
|
||||
explanation = explanation_response.text
|
||||
print(f"[Gemini PDF] 📄 Gemini sees: {explanation[:500]}...", flush=True)
|
||||
|
||||
raise Exception(f"No questions found in PDF. Gemini's description: {explanation[:200]}...")
|
||||
|
||||
# Validate and fix question types
|
||||
valid_types = {"single", "multiple", "judge", "short"}
|
||||
type_mapping = {
|
||||
"proof": "short",
|
||||
"essay": "short",
|
||||
"calculation": "short",
|
||||
"fill": "short",
|
||||
"填空": "short",
|
||||
"证明": "short",
|
||||
"计算": "short",
|
||||
"问答": "short",
|
||||
"单选": "single",
|
||||
"多选": "multiple",
|
||||
"判断": "judge",
|
||||
"简答": "short"
|
||||
}
|
||||
|
||||
# Add content hash and validate types
|
||||
for q in questions:
|
||||
if "content" not in q:
|
||||
print(f"[Warning] Question missing 'content' field: {q}", flush=True)
|
||||
continue
|
||||
|
||||
# Validate and fix question type
|
||||
q_type = q.get("type", "short")
|
||||
if isinstance(q_type, str):
|
||||
q_type_lower = q_type.lower()
|
||||
if q_type_lower not in valid_types:
|
||||
# Try to map to valid type
|
||||
if q_type_lower in type_mapping:
|
||||
old_type = q_type
|
||||
q["type"] = type_mapping[q_type_lower]
|
||||
print(f"[Type Fix] Changed '{old_type}' to '{q['type']}' for question: {q['content'][:50]}...", flush=True)
|
||||
else:
|
||||
# Default to short answer
|
||||
print(f"[Type Fix] Unknown type '{q_type}', defaulting to 'short' for question: {q['content'][:50]}...", flush=True)
|
||||
q["type"] = "short"
|
||||
else:
|
||||
q["type"] = q_type_lower
|
||||
else:
|
||||
q["type"] = "short"
|
||||
|
||||
q["content_hash"] = calculate_content_hash(q["content"])
|
||||
|
||||
print(f"[Gemini PDF] Successfully extracted {len(questions)} questions", flush=True)
|
||||
return questions
|
||||
|
||||
except Exception as e:
|
||||
print(f"[Error] PDF parsing failed: {str(e)}", flush=True)
|
||||
raise Exception(f"Failed to parse PDF document: {str(e)}")
|
||||
|
||||
async def grade_short_answer(
|
||||
self,
|
||||
question: str,
|
||||
@@ -201,6 +660,18 @@ Return ONLY the JSON object, no markdown or explanations."""
|
||||
]
|
||||
)
|
||||
result = response.content[0].text
|
||||
elif self.provider == "gemini":
|
||||
# Gemini uses different API
|
||||
import asyncio
|
||||
|
||||
def _generate_content():
|
||||
return self.client.models.generate_content(
|
||||
model=self.model,
|
||||
contents=prompt
|
||||
)
|
||||
|
||||
response = await asyncio.to_thread(_generate_content)
|
||||
result = response.text
|
||||
else: # OpenAI or Qwen
|
||||
response = await self.client.chat.completions.create(
|
||||
model=self.model,
|
||||
|
||||
Reference in New Issue
Block a user