主要更新: - 新增 transform 意图:支持 Word/Excel/Markdown 格式互转 - 新增 edit 意图:使用 LLM 润色编辑文档内容 - 智能指令接口增加异步执行模式(async_execute 参数) - 修复 Word 模板导出文档损坏问题(改用临时文件方式) - 优化 intent_parser 增加 transform/edit 关键词识别 新增文件: - app/api/endpoints/instruction.py: 智能指令 API 端点 - app/services/multi_doc_reasoning_service.py: 多文档推理服务 其他优化: - RAG 服务混合搜索(BM25 + 向量)融合 - 模板填充服务表头匹配增强 - Word AI 解析服务返回结构完善 - 前端 InstructionChat 组件对接真实 API
640 lines
22 KiB
Python
640 lines
22 KiB
Python
"""
|
||
Word 文档 AI 解析服务
|
||
|
||
使用 LLM (GLM) 对 Word 文档进行深度理解,提取结构化数据
|
||
"""
|
||
import logging
|
||
from typing import Dict, Any, List, Optional
|
||
import json
|
||
|
||
from app.services.llm_service import llm_service
|
||
from app.core.document_parser.docx_parser import DocxParser
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
|
||
class WordAIService:
|
||
"""Word 文档 AI 解析服务"""
|
||
|
||
def __init__(self):
|
||
self.llm = llm_service
|
||
self.parser = DocxParser()
|
||
|
||
async def parse_word_with_ai(
|
||
self,
|
||
file_path: str,
|
||
user_hint: str = ""
|
||
) -> Dict[str, Any]:
|
||
"""
|
||
使用 AI 解析 Word 文档,提取结构化数据
|
||
|
||
适用于从非结构化的 Word 文档中提取表格数据、键值对等信息
|
||
|
||
Args:
|
||
file_path: Word 文件路径
|
||
user_hint: 用户提示词,指定要提取的内容类型
|
||
|
||
Returns:
|
||
Dict: 包含结构化数据的解析结果
|
||
"""
|
||
try:
|
||
# 1. 先用基础解析器提取原始内容
|
||
parse_result = self.parser.parse(file_path)
|
||
|
||
if not parse_result.success:
|
||
return {
|
||
"success": False,
|
||
"error": parse_result.error,
|
||
"structured_data": None
|
||
}
|
||
|
||
# 2. 获取原始数据
|
||
raw_data = parse_result.data
|
||
paragraphs = raw_data.get("paragraphs", [])
|
||
paragraphs_with_style = raw_data.get("paragraphs_with_style", [])
|
||
tables = raw_data.get("tables", [])
|
||
content = raw_data.get("content", "")
|
||
images_info = raw_data.get("images", {})
|
||
metadata = parse_result.metadata or {}
|
||
|
||
image_count = images_info.get("image_count", 0)
|
||
image_descriptions = images_info.get("descriptions", [])
|
||
|
||
logger.info(f"Word 基础解析完成: {len(paragraphs)} 个段落, {len(tables)} 个表格, {image_count} 张图片")
|
||
|
||
# 3. 提取图片数据(用于视觉分析)
|
||
images_base64 = []
|
||
if image_count > 0:
|
||
try:
|
||
images_base64 = self.parser.extract_images_as_base64(file_path)
|
||
logger.info(f"提取到 {len(images_base64)} 张图片的 base64 数据")
|
||
except Exception as e:
|
||
logger.warning(f"提取图片 base64 失败: {str(e)}")
|
||
|
||
# 4. 根据内容类型选择 AI 解析策略
|
||
# 如果有图片,先分析图片
|
||
image_analysis = ""
|
||
if images_base64:
|
||
image_analysis = await self._analyze_images_with_ai(images_base64, user_hint)
|
||
logger.info(f"图片 AI 分析完成: {len(image_analysis)} 字符")
|
||
|
||
# 优先处理:表格 > (表格+文本) > 纯文本
|
||
if tables and len(tables) > 0:
|
||
structured_data = await self._extract_tables_with_ai(
|
||
tables, paragraphs, image_count, user_hint, metadata, image_analysis
|
||
)
|
||
elif paragraphs and len(paragraphs) > 0:
|
||
structured_data = await self._extract_from_text_with_ai(
|
||
paragraphs, content, image_count, image_descriptions, user_hint, image_analysis
|
||
)
|
||
else:
|
||
structured_data = {
|
||
"success": True,
|
||
"type": "empty",
|
||
"message": "文档内容为空"
|
||
}
|
||
|
||
# 添加图片分析结果
|
||
if image_analysis:
|
||
structured_data["image_analysis"] = image_analysis
|
||
|
||
return structured_data
|
||
|
||
except Exception as e:
|
||
logger.error(f"AI 解析 Word 文档失败: {str(e)}")
|
||
return {
|
||
"success": False,
|
||
"error": str(e),
|
||
"structured_data": None
|
||
}
|
||
|
||
async def _extract_tables_with_ai(
|
||
self,
|
||
tables: List[Dict],
|
||
paragraphs: List[str],
|
||
image_count: int,
|
||
user_hint: str,
|
||
metadata: Dict,
|
||
image_analysis: str = ""
|
||
) -> Dict[str, Any]:
|
||
"""
|
||
使用 AI 从 Word 表格和文本中提取结构化数据
|
||
|
||
Args:
|
||
tables: 表格列表
|
||
paragraphs: 段落列表
|
||
image_count: 图片数量
|
||
user_hint: 用户提示
|
||
metadata: 文档元数据
|
||
image_analysis: 图片 AI 分析结果
|
||
|
||
Returns:
|
||
结构化数据
|
||
"""
|
||
try:
|
||
# 构建表格文本描述
|
||
tables_text = self._build_tables_description(tables)
|
||
|
||
# 构建段落描述
|
||
paragraphs_text = "\n".join(paragraphs[:50]) if paragraphs else "(无正文文本)"
|
||
if len(paragraphs) > 50:
|
||
paragraphs_text += f"\n...(共 {len(paragraphs)} 个段落,仅显示前50个)"
|
||
|
||
# 图片提示
|
||
image_hint = f"注意:此文档包含 {image_count} 张图片/图表。" if image_count > 0 else ""
|
||
|
||
prompt = f"""你是一个专业的数据提取专家。请从以下 Word 文档的完整内容中提取结构化数据。
|
||
|
||
【用户需求】
|
||
{user_hint if user_hint else "请提取文档中的所有结构化数据,包括表格数据、键值对、列表项等。"}
|
||
|
||
【文档正文(段落)】
|
||
{paragraphs_text}
|
||
|
||
【文档表格】
|
||
{tables_text}
|
||
|
||
【文档图片信息】
|
||
{image_hint}
|
||
|
||
请按照以下 JSON 格式输出:
|
||
{{
|
||
"type": "table_data",
|
||
"headers": ["列1", "列2", ...],
|
||
"rows": [["行1列1", "行1列2", ...], ["行2列1", "行2列2", ...], ...],
|
||
"key_values": {{"键1": "值1", "键2": "值2", ...}},
|
||
"list_items": ["项1", "项2", ...],
|
||
"description": "文档内容描述"
|
||
}}
|
||
|
||
重点:
|
||
- 优先从表格中提取结构化数据
|
||
- 如果表格中有表头,headers 是表头,rows 是数据行
|
||
- 如果文档中有键值对(如 名称: 张三),提取到 key_values 中
|
||
- 如果文档中有列表项,提取到 list_items 中
|
||
- 图片内容无法直接提取,但请在 description 中说明图片的大致主题(如"包含流程图"、"包含数据图表"等)
|
||
"""
|
||
|
||
messages = [
|
||
{"role": "system", "content": "你是一个专业的数据提取助手。请严格按JSON格式输出。"},
|
||
{"role": "user", "content": prompt}
|
||
]
|
||
|
||
response = await self.llm.chat(
|
||
messages=messages,
|
||
temperature=0.1,
|
||
max_tokens=50000
|
||
)
|
||
|
||
content = self.llm.extract_message_content(response)
|
||
|
||
# 解析 JSON
|
||
result = self._parse_json_response(content)
|
||
|
||
if result:
|
||
logger.info(f"AI 表格提取成功: {len(result.get('rows', []))} 行数据, key_values={len(result.get('key_values', {}))}, list_items={len(result.get('list_items', []))}")
|
||
return {
|
||
"success": True,
|
||
"type": "table_data",
|
||
"headers": result.get("headers", []),
|
||
"rows": result.get("rows", []),
|
||
"description": result.get("description", ""),
|
||
"key_values": result.get("key_values", {}),
|
||
"list_items": result.get("list_items", [])
|
||
}
|
||
else:
|
||
# 如果 AI 返回格式不对,尝试直接解析表格
|
||
return self._fallback_table_parse(tables)
|
||
|
||
except Exception as e:
|
||
logger.error(f"AI 表格提取失败: {str(e)}")
|
||
return self._fallback_table_parse(tables)
|
||
|
||
async def _extract_from_text_with_ai(
|
||
self,
|
||
paragraphs: List[str],
|
||
full_text: str,
|
||
image_count: int,
|
||
image_descriptions: List[str],
|
||
user_hint: str,
|
||
image_analysis: str = ""
|
||
) -> Dict[str, Any]:
|
||
"""
|
||
使用 AI 从 Word 纯文本中提取结构化数据
|
||
|
||
Args:
|
||
paragraphs: 段落列表
|
||
full_text: 完整文本
|
||
image_count: 图片数量
|
||
image_descriptions: 图片描述列表
|
||
user_hint: 用户提示
|
||
image_analysis: 图片 AI 分析结果
|
||
|
||
Returns:
|
||
结构化数据
|
||
"""
|
||
try:
|
||
# 限制文本长度
|
||
text_preview = full_text[:8000] if len(full_text) > 8000 else full_text
|
||
|
||
# 图片提示
|
||
image_hint = f"\n【文档图片】此文档包含 {image_count} 张图片/图表。" if image_count > 0 else ""
|
||
if image_descriptions:
|
||
image_hint += "\n" + "\n".join(image_descriptions)
|
||
|
||
prompt = f"""你是一个专业的数据提取专家。请从以下 Word 文档的完整内容中提取结构化数据。
|
||
|
||
【用户需求】
|
||
{user_hint if user_hint else "请识别并提取文档中的关键信息,包括:表格数据、键值对、列表项等。"}
|
||
|
||
【文档正文】{image_hint}
|
||
{text_preview}
|
||
|
||
请按照以下 JSON 格式输出:
|
||
{{
|
||
"type": "structured_text",
|
||
"tables": [{{"headers": [...], "rows": [...]}}],
|
||
"key_values": {{"键1": "值1", "键2": "值2", ...}},
|
||
"list_items": ["项1", "项2", ...],
|
||
"summary": "文档内容摘要"
|
||
}}
|
||
|
||
重点:
|
||
- 如果文档包含表格数据,提取到 tables 中
|
||
- 如果文档包含键值对(如 名称: 张三),提取到 key_values 中
|
||
- 如果文档包含列表项,提取到 list_items 中
|
||
- 如果文档包含图片,请根据上下文推断图片内容(如"流程图"、"数据折线图"等)并在 description 中说明
|
||
- 如果无法提取到结构化数据,至少提供一个详细的摘要
|
||
"""
|
||
|
||
messages = [
|
||
{"role": "system", "content": "你是一个专业的数据提取助手。请严格按JSON格式输出。"},
|
||
{"role": "user", "content": prompt}
|
||
]
|
||
|
||
response = await self.llm.chat(
|
||
messages=messages,
|
||
temperature=0.1,
|
||
max_tokens=50000
|
||
)
|
||
|
||
content = self.llm.extract_message_content(response)
|
||
|
||
result = self._parse_json_response(content)
|
||
|
||
if result:
|
||
logger.info(f"AI 文本提取成功: type={result.get('type')}")
|
||
return {
|
||
"success": True,
|
||
"type": result.get("type", "structured_text"),
|
||
"tables": result.get("tables", []),
|
||
"key_values": result.get("key_values", {}),
|
||
"list_items": result.get("list_items", []),
|
||
"summary": result.get("summary", ""),
|
||
"raw_text_preview": text_preview[:500]
|
||
}
|
||
else:
|
||
return {
|
||
"success": True,
|
||
"type": "text",
|
||
"summary": text_preview[:500],
|
||
"raw_text_preview": text_preview[:500]
|
||
}
|
||
|
||
except Exception as e:
|
||
logger.error(f"AI 文本提取失败: {str(e)}")
|
||
return {
|
||
"success": False,
|
||
"error": str(e)
|
||
}
|
||
|
||
async def _analyze_images_with_ai(
|
||
self,
|
||
images: List[Dict[str, str]],
|
||
user_hint: str = ""
|
||
) -> str:
|
||
"""
|
||
使用视觉模型分析 Word 文档中的图片
|
||
|
||
Args:
|
||
images: 图片列表,每项包含 base64 和 mime_type
|
||
user_hint: 用户提示
|
||
|
||
Returns:
|
||
图片分析结果文本
|
||
"""
|
||
try:
|
||
# 调用 LLM 的视觉分析功能
|
||
result = await self.llm.analyze_images(
|
||
images=images,
|
||
user_prompt=user_hint or "请详细描述图片内容,提取所有文字和数据信息。"
|
||
)
|
||
|
||
if result.get("success"):
|
||
analysis = result.get("analysis", {})
|
||
if isinstance(analysis, dict):
|
||
description = analysis.get("description", "")
|
||
text_content = analysis.get("text_content", "")
|
||
data_extracted = analysis.get("data_extracted", {})
|
||
|
||
result_text = f"【图片分析结果】\n{description}"
|
||
if text_content:
|
||
result_text += f"\n\n【图片中的文字】\n{text_content}"
|
||
if data_extracted:
|
||
result_text += f"\n\n【提取的数据】\n{json.dumps(data_extracted, ensure_ascii=False)}"
|
||
return result_text
|
||
else:
|
||
return str(analysis)
|
||
else:
|
||
logger.warning(f"图片 AI 分析失败: {result.get('error')}")
|
||
return ""
|
||
|
||
except Exception as e:
|
||
logger.error(f"图片 AI 分析异常: {str(e)}")
|
||
return ""
|
||
|
||
def _build_tables_description(self, tables: List[Dict]) -> str:
|
||
"""构建表格的文本描述"""
|
||
result = []
|
||
|
||
for idx, table in enumerate(tables):
|
||
rows = table.get("rows", [])
|
||
if not rows:
|
||
continue
|
||
|
||
result.append(f"\n--- 表格 {idx + 1} ---")
|
||
|
||
for row_idx, row in enumerate(rows[:50]): # 限制每表格最多50行
|
||
if isinstance(row, list):
|
||
result.append(" | ".join(str(cell).strip() for cell in row))
|
||
elif isinstance(row, dict):
|
||
result.append(str(row))
|
||
|
||
if len(rows) > 50:
|
||
result.append(f"...(共 {len(rows)} 行,仅显示前50行)")
|
||
|
||
return "\n".join(result) if result else "(无表格内容)"
|
||
|
||
def _parse_json_response(self, content: str) -> Optional[Dict]:
|
||
"""解析 JSON 响应,处理各种格式问题"""
|
||
import re
|
||
|
||
# 清理 markdown 标记
|
||
cleaned = content.strip()
|
||
cleaned = re.sub(r'^```json\s*', '', cleaned, flags=re.MULTILINE)
|
||
cleaned = re.sub(r'^```\s*', '', cleaned, flags=re.MULTILINE)
|
||
cleaned = cleaned.strip()
|
||
|
||
# 找到 JSON 开始位置
|
||
json_start = -1
|
||
for i, c in enumerate(cleaned):
|
||
if c == '{':
|
||
json_start = i
|
||
break
|
||
|
||
if json_start == -1:
|
||
logger.warning("无法找到 JSON 开始位置")
|
||
return None
|
||
|
||
json_text = cleaned[json_start:]
|
||
|
||
# 尝试直接解析
|
||
try:
|
||
return json.loads(json_text)
|
||
except json.JSONDecodeError:
|
||
pass
|
||
|
||
# 尝试修复并解析
|
||
try:
|
||
# 找到闭合括号
|
||
depth = 0
|
||
end_pos = -1
|
||
for i, c in enumerate(json_text):
|
||
if c == '{':
|
||
depth += 1
|
||
elif c == '}':
|
||
depth -= 1
|
||
if depth == 0:
|
||
end_pos = i + 1
|
||
break
|
||
|
||
if end_pos > 0:
|
||
fixed = json_text[:end_pos]
|
||
# 移除末尾逗号
|
||
fixed = re.sub(r',\s*([}]])', r'\1', fixed)
|
||
return json.loads(fixed)
|
||
except Exception as e:
|
||
logger.warning(f"JSON 修复失败: {e}")
|
||
|
||
return None
|
||
|
||
def _fallback_table_parse(self, tables: List[Dict]) -> Dict[str, Any]:
|
||
"""当 AI 解析失败时,直接解析表格"""
|
||
if not tables:
|
||
return {
|
||
"success": True,
|
||
"type": "empty",
|
||
"data": {},
|
||
"message": "无表格内容"
|
||
}
|
||
|
||
all_rows = []
|
||
all_headers = None
|
||
|
||
for table in tables:
|
||
rows = table.get("rows", [])
|
||
if not rows:
|
||
continue
|
||
|
||
# 查找真正的表头行(跳过标题行)
|
||
header_row_idx = 0
|
||
for idx, row in enumerate(rows[:5]): # 只检查前5行
|
||
if not isinstance(row, list):
|
||
continue
|
||
# 如果某一行包含"表"字开头且单元格内容很长,这可能是标题行
|
||
first_cell = str(row[0]) if row else ""
|
||
if first_cell.startswith("表") and len(first_cell) > 15:
|
||
header_row_idx = idx + 1
|
||
continue
|
||
# 如果某一行有超过3个空单元格,可能是无效行
|
||
empty_count = sum(1 for cell in row if not str(cell).strip())
|
||
if empty_count > 3:
|
||
header_row_idx = idx + 1
|
||
continue
|
||
# 找到第一行看起来像表头的行(短单元格,大部分有内容)
|
||
avg_len = sum(len(str(c)) for c in row) / len(row) if row else 0
|
||
if avg_len < 20: # 表头通常比数据行短
|
||
header_row_idx = idx
|
||
break
|
||
|
||
if header_row_idx >= len(rows):
|
||
continue
|
||
|
||
# 使用找到的表头行
|
||
if rows and isinstance(rows[header_row_idx], list):
|
||
headers = rows[header_row_idx]
|
||
if all_headers is None:
|
||
all_headers = headers
|
||
|
||
# 数据行(从表头之后开始)
|
||
for row in rows[header_row_idx + 1:]:
|
||
if isinstance(row, list) and len(row) == len(headers):
|
||
all_rows.append(row)
|
||
|
||
if all_headers and all_rows:
|
||
return {
|
||
"success": True,
|
||
"type": "table_data",
|
||
"headers": all_headers,
|
||
"rows": all_rows,
|
||
"description": "直接从 Word 表格提取"
|
||
}
|
||
|
||
return {
|
||
"success": True,
|
||
"type": "raw",
|
||
"tables": tables,
|
||
"message": "表格数据(未AI处理)"
|
||
}
|
||
|
||
async def fill_template_with_ai(
|
||
self,
|
||
file_path: str,
|
||
template_fields: List[Dict[str, Any]],
|
||
user_hint: str = ""
|
||
) -> Dict[str, Any]:
|
||
"""
|
||
使用 AI 解析 Word 文档并填写模板
|
||
|
||
这是主要入口函数,前端调用此函数即可完成:
|
||
1. AI 解析 Word 文档
|
||
2. 根据模板字段提取数据
|
||
3. 返回填写结果
|
||
|
||
Args:
|
||
file_path: Word 文件路径
|
||
template_fields: 模板字段列表 [{"name": "字段名", "hint": "提示词"}, ...]
|
||
user_hint: 用户提示
|
||
|
||
Returns:
|
||
填写结果
|
||
"""
|
||
try:
|
||
# 1. AI 解析文档
|
||
parse_result = await self.parse_word_with_ai(file_path, user_hint)
|
||
|
||
if not parse_result.get("success"):
|
||
return {
|
||
"success": False,
|
||
"error": parse_result.get("error", "解析失败"),
|
||
"filled_data": {},
|
||
"source": "ai_parse_failed"
|
||
}
|
||
|
||
# 2. 根据字段类型提取数据
|
||
filled_data = {}
|
||
extract_details = []
|
||
|
||
parse_type = parse_result.get("type", "")
|
||
|
||
if parse_type == "table_data":
|
||
# 表格数据:直接匹配列名
|
||
headers = parse_result.get("headers", [])
|
||
rows = parse_result.get("rows", [])
|
||
|
||
for field in template_fields:
|
||
field_name = field.get("name", "")
|
||
values = self._extract_field_from_table(headers, rows, field_name)
|
||
filled_data[field_name] = values
|
||
extract_details.append({
|
||
"field": field_name,
|
||
"values": values,
|
||
"source": "ai_table_extraction",
|
||
"confidence": 0.9 if values else 0.0
|
||
})
|
||
|
||
elif parse_type == "structured_text":
|
||
# 结构化文本:尝试从 key_values 和 list_items 提取
|
||
key_values = parse_result.get("key_values", {})
|
||
list_items = parse_result.get("list_items", [])
|
||
|
||
for field in template_fields:
|
||
field_name = field.get("name", "")
|
||
value = key_values.get(field_name, "")
|
||
if not value and list_items:
|
||
value = list_items[0] if list_items else ""
|
||
filled_data[field_name] = [value] if value else []
|
||
extract_details.append({
|
||
"field": field_name,
|
||
"values": [value] if value else [],
|
||
"source": "ai_text_extraction",
|
||
"confidence": 0.7 if value else 0.0
|
||
})
|
||
|
||
else:
|
||
# 其他类型:返回原始解析结果供后续处理
|
||
for field in template_fields:
|
||
field_name = field.get("name", "")
|
||
filled_data[field_name] = []
|
||
extract_details.append({
|
||
"field": field_name,
|
||
"values": [],
|
||
"source": "no_ai_data",
|
||
"confidence": 0.0
|
||
})
|
||
|
||
# 3. 返回结果
|
||
max_rows = max(len(v) for v in filled_data.values()) if filled_data else 1
|
||
|
||
return {
|
||
"success": True,
|
||
"filled_data": filled_data,
|
||
"fill_details": extract_details,
|
||
"ai_parse_result": {
|
||
"type": parse_type,
|
||
"description": parse_result.get("description", "")
|
||
},
|
||
"source_doc_count": 1,
|
||
"max_rows": max_rows
|
||
}
|
||
|
||
except Exception as e:
|
||
logger.error(f"AI 填表失败: {str(e)}")
|
||
return {
|
||
"success": False,
|
||
"error": str(e),
|
||
"filled_data": {},
|
||
"fill_details": []
|
||
}
|
||
|
||
def _extract_field_from_table(
|
||
self,
|
||
headers: List[str],
|
||
rows: List[List],
|
||
field_name: str
|
||
) -> List[str]:
|
||
"""从表格中提取指定字段的值"""
|
||
# 查找匹配的列
|
||
target_col_idx = None
|
||
for col_idx, header in enumerate(headers):
|
||
if field_name.lower() in str(header).lower() or str(header).lower() in field_name.lower():
|
||
target_col_idx = col_idx
|
||
break
|
||
|
||
if target_col_idx is None:
|
||
return []
|
||
|
||
# 提取该列所有值
|
||
values = []
|
||
for row in rows:
|
||
if isinstance(row, list) and target_col_idx < len(row):
|
||
val = str(row[target_col_idx]).strip()
|
||
if val:
|
||
values.append(val)
|
||
|
||
return values
|
||
|
||
|
||
# 全局单例
|
||
word_ai_service = WordAIService()
|