OurCodesAreAllRight/FilesReadSystem
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

【智能助手增强】

Browse Source 
- 新增对话历史管理：MongoDB新增conversations集合，存储用户与AI的对话上下文，支持多轮对话意图延续
- 新增对话历史API（conversation.py）：GET/DELETE conversation历史、列出所有会话
- 意图解析增强：支持基于对话历史的意图识别，上下文理解更准确
- 字段提取优化：支持"提取文档中的医院数量"等自然语言模式，智能去除"文档中的"前缀
- 文档对比优化：从指令中提取文件名并精确匹配source_docs，支持"对比A和B两个文档"
- 文档摘要优化：使用LLM生成真实AI摘要而非返回原始文档预览

【Word模板填表核心功能】
- Word模板字段生成：空白Word上传后，自动从源文档（Excel/Word/TXT/MD）内容AI生成字段名
- Word模板填表（_fill_docx）：将提取数据写入Word模板表格，支持精确匹配、模糊匹配、追加新行
- 数据润色（_polish_word_filled_data）：LLM对多行Excel数据进行统计归纳（合计/平均/极值），转化为专业自然语言描述
- 段落格式输出：使用📌字段名+值段落+分隔线（灰色横线）格式，提升可读性
- 导出链打通：fill_template返回filled_file_path，export直接返回已填好的Word文件

【其他修复】
- 修复Word导出Windows文件锁问题：NamedTemporaryFile改为mkstemp+close
- 修复Word方框非法字符：扩展clean_text移除\uFFFD、□等Unicode替代符和零宽字符
- 修复文档对比"需要至少2个文档"：从指令提取具体文件名优先匹配而非取前2个
- 修复导出format硬编码：自动识别docx/xlsx格式
- Docx解析器增加备用解析方法和更完整的段落/表格/标题提取
- RAG服务新增MySQL数据源支持
This commit is contained in:
dj
2026-04-15 23:32:55 +08:00
parent 9e7f9df384
commit e5d4724e82
19 changed files with 2185 additions and 407 deletions
Download Patch File Download Diff File Expand all files Collapse all files

313
backend/app/instruction/executor.py
View File

@@ -5,9 +5,10 @@
"""
import logging
import json
import re
from typing import Any, Dict, List, Optional
from app.services.template_fill_service import template_fill_service
from app.services.template_fill_service import template_fill_service, TemplateField
from app.services.rag_service import rag_service
from app.services.markdown_ai_service import markdown_ai_service
from app.core.database import mongodb
@@ -15,6 +16,31 @@ from app.core.database import mongodb
logger = logging.getLogger(__name__)
def _extract_filenames_from_text(text: str) -> List[str]:
"""
从指令文本中提取文件名列表。
智能处理用'和'/'与'/'、分隔的多个文件名（尤其是带年号的统计公报）。
"""
# 先去掉"对比这两个文档"等引导语，只保留文件名部分
text = re.sub(r'^(?:对比|比较)这两个?文档[的差异]?[：:]?', '', text).strip()
text = re.sub(r'两个文档.*$', '', text).strip()
if not text:
return []
# 直接查找所有带扩展名的文件名模式
results = []
for m in re.finditer(r'[^\s，。！？、和与]+(?=\.(?:docx|xlsx|md|txt))', text):
start = m.start()
ext_match = re.search(r'\.(?:docx|xlsx|md|txt)', text[m.end():])
if ext_match:
fn = text[start:m.end() + ext_match.end()]
if fn:
results.append(fn)
return results
class InstructionExecutor:
"""指令执行器"""
@@ -41,9 +67,10 @@ class InstructionExecutor:
self.intent_parser = intent_parser
context = context or {}
context["instruction"] = instruction # 保存原始指令以便后续使用
# 解析意图
intent, params = await self.intent_parser.parse(instruction)
# 解析意图（传递对话历史上下文）
intent, params = await self.intent_parser.parse(instruction, context)
# 根据意图类型执行相应操作
if intent == "extract":
@@ -72,18 +99,48 @@ class InstructionExecutor:
async def _execute_extract(self, params: Dict[str, Any], context: Dict[str, Any]) -> Dict[str, Any]:
"""执行信息提取"""
try:
target_fields = params.get("field_refs", [])
# target_fields 来自意图解析，field_refs 来自引号/字段关键词匹配
target_fields = params.get("target_fields", []) or params.get("field_refs", [])
doc_ids = params.get("document_refs", [])
instruction_text = context.get("instruction", "")
# 如果没有指定文档，尝试按文件名精确搜索
if not doc_ids or "all_docs" in doc_ids:
if instruction_text:
import re
# 提取引号内的内容或文件名
filename_match = re.search(r'["""]([^"""]+)["""]', instruction_text)
if filename_match:
search_term = filename_match.group(1)
else:
match = re.search(r'([^\s]+\.(?:docx|xlsx|md|txt))', instruction_text)
search_term = match.group(1) if match else None
if search_term:
logger.info(f"提取时搜索文档: {search_term}")
searched_docs = await mongodb.search_documents(search_term, limit=5)
if searched_docs:
# 优先选择文件名完全匹配的文档
best_docs = [
d for d in searched_docs
if search_term.lower() in d.get("metadata", {}).get("original_filename", "").lower()
]
if not best_docs:
best_docs = [searched_docs[0]]
context["source_docs"] = best_docs
doc_ids = [doc.get("_id", "") for doc in best_docs]
logger.info(f"找到 {len(best_docs)} 个文档用于提取，最佳: {best_docs[0].get('metadata', {}).get('original_filename', '?')}")
if not target_fields:
return {
"success": False,
"intent": "extract",
"error": "未指定要提取的字段",
"message": "请明确说明要提取哪些字段，如：'提取医院数量和床位数'"
}
# 如果指定了文档，验证文档存在
if doc_ids and "all_docs" not in doc_ids:
# 如果指定了文档且还没有加载 source_docs，则验证并加载
if doc_ids and "all_docs" not in doc_ids and not context.get("source_docs"):
valid_docs = []
for doc_ref in doc_ids:
doc_id = doc_ref.replace("doc_", "")
@@ -93,20 +150,22 @@ class InstructionExecutor:
if not valid_docs:
return {
"success": False,
"intent": "extract",
"error": "指定的文档不存在",
"message": "请检查文档编号是否正确"
}
context["source_docs"] = valid_docs
# 构建字段列表
fields = []
for i, field_name in enumerate(target_fields):
fields.append({
"name": field_name,
"cell": f"A{i+1}",
"field_type": "text",
"required": False
})
# 构建字段列表（使用 TemplateField dataclass）
fields = [
TemplateField(
name=field_name,
cell=f"A{i+1}",
field_type="text",
required=False
)
for i, field_name in enumerate(target_fields)
]
# 调用填表服务
result = await template_fill_service.fill_template(
@@ -143,7 +202,7 @@ class InstructionExecutor:
}
# 获取源文档
source_docs = context.get("source_docs", [])
source_docs = context.get("source_docs", []) or []
source_doc_ids = [doc.get("_id") for doc in source_docs if doc.get("_id")]
# 获取字段
@@ -175,36 +234,103 @@ class InstructionExecutor:
}
async def _execute_summarize(self, params: Dict[str, Any], context: Dict[str, Any]) -> Dict[str, Any]:
"""执行摘要总结"""
"""执行摘要总结 - 使用 LLM 生成真实摘要"""
try:
docs = context.get("source_docs", [])
import re
docs = context.get("source_docs", []) or []
instruction_text = context.get("instruction", "")
# 从指令中提取文件名/关键词，优先搜索精确文档
search_term = None
if instruction_text:
filename_match = re.search(r'["""]([^"""]+)["""]', instruction_text)
if filename_match:
search_term = filename_match.group(1)
else:
file_match = re.search(r'([^\s，。！？,]+\.(?:docx|xlsx|md|txt))', instruction_text)
if file_match:
search_term = file_match.group(1)
# 如果没有文档或有更精确的搜索词，尝试重新搜索
if not docs or search_term:
if search_term:
logger.info(f"按关键词搜索文档: {search_term}")
searched_docs = await mongodb.search_documents(search_term, limit=5)
if searched_docs:
# 优先使用文件名最匹配的文档
docs = sorted(
searched_docs,
key=lambda d: 1 if search_term.lower() in d.get("metadata", {}).get("original_filename", "").lower() else 0,
reverse=True
)
logger.info(f"找到 {len(docs)} 个文档，最佳匹配: {docs[0].get('metadata', {}).get('original_filename', '?')}")
if not docs:
return {
"success": False,
"error": "没有可用的文档",
"message": "请先上传要总结的文档"
"success": True,
"intent": "summarize",
"action_needed": "provide_document",
"message": "我理解了，您想分析文档内容。",
"suggestion": "请提供已上传文档的名称（可以是文件名或部分名称），或者上传您想要分析的文档。\n\n支持的格式：docx、xlsx、md、txt\n\n例如：'分析2021年民政事业发展统计公报' 或 '总结卫生健康数据'"
}
summaries = []
for doc in docs[:5]: # 最多处理5个文档
content = doc.get("content", "")[:5000] # 限制内容长度
if content:
summaries.append({
"filename": doc.get("metadata", {}).get("original_filename", "未知"),
"content_preview": content[:500] + "..." if len(content) > 500 else content
})
# 对第一个（最佳匹配）文档生成 AI 摘要
primary_doc = docs[0]
content = primary_doc.get("content", "")
filename = primary_doc.get("metadata", {}).get("original_filename", "未知文档")
if not content:
return {
"success": False,
"intent": "summarize",
"error": "文档内容为空",
"message": f"文档 {filename} 没有可供分析的文本内容"
}
# 使用 LLM 生成摘要
content_for_summary = content[:12000] # 最多取前 12000 字
user_request = instruction_text or "请总结这份文档"
prompt = f"""请对以下文档进行全面、有条理的摘要分析。
文档名称：{filename}
用户要求：{user_request}
文档内容：
{content_for_summary}
请按以下格式输出摘要：
1. **文档概述**：简述文档主题和背景（2-3句）
2. **主要内容**：列出文档的核心数据和关键信息（用要点列出）
3. **重要数据**：提取文档中的重要数字、统计数据
4. **主要结论**：归纳文档的主要结论或趋势
要求：条理清晰，数据准确，不要遗漏关键信息。"""
from app.services.llm_service import llm_service
messages = [
{"role": "system", "content": "你是一个专业的文档分析助手，擅长提取关键信息并生成结构化摘要。"},
{"role": "user", "content": prompt}
]
response = await llm_service.chat(messages=messages, temperature=0.3, max_tokens=2000)
ai_summary = llm_service.extract_message_content(response)
return {
"success": True,
"intent": "summarize",
"summaries": summaries,
"message": f"找到 {len(summaries)} 个文档可供参考"
"ai_summary": ai_summary,
"filename": filename,
"doc_id": primary_doc.get("_id", ""),
"total_docs_found": len(docs),
"message": f"已生成文档摘要"
}
except Exception as e:
logger.error(f"摘要执行失败: {e}")
return {
"success": False,
"intent": "summarize",
"error": str(e),
"message": f"摘要生成失败: {str(e)}"
}
@@ -213,17 +339,39 @@ class InstructionExecutor:
"""执行问答"""
try:
question = params.get("question", "")
instruction_text = context.get("instruction", "")
if not question:
return {
"success": False,
"intent": "question",
"error": "未提供问题",
"message": "请输入要回答的问题"
}
# 使用 RAG 检索相关文档
docs = context.get("source_docs", [])
rag_results = []
docs = context.get("source_docs", []) or []
# 如果没有文档，尝试从指令中提取文件名搜索
if not docs:
filename_match = re.search(r'["""]([^"""]+\.(?:docx|xlsx|md|txt))["""]', instruction_text)
if not filename_match:
filename_match = re.search(r'([^\s，。！？]+\.(?:docx|xlsx|md|txt))', instruction_text)
if filename_match:
found = await mongodb.search_documents(filename_match.group(1), limit=5)
if found:
docs = found
if not docs:
return {
"success": True,
"intent": "question",
"question": question,
"answer": None,
"message": "请先上传文档，我才能回答您的问题"
}
# 使用 RAG 检索相关文档
rag_results = []
for doc in docs:
doc_id = doc.get("_id", "")
if doc_id:
@@ -241,12 +389,42 @@ class InstructionExecutor:
doc.get("content", "")[:3000] for doc in docs[:3] if doc.get("content")
])
if not context_text:
return {
"success": True,
"intent": "question",
"question": question,
"answer": None,
"message": "文档内容为空，无法回答问题"
}
# 使用 LLM 生成答案
filename = docs[0].get("metadata", {}).get("original_filename", "文档")
prompt = f"""基于以下文档内容，回答用户的问题。
文档名称：{filename}
用户问题：{question}
文档内容：
{context_text[:8000]}
请根据文档内容准确回答问题。如果文档中没有相关信息，请明确说明。"""
from app.services.llm_service import llm_service
messages = [
{"role": "system", "content": "你是一个专业的文档问答助手，根据提供的内容准确回答用户问题。"},
{"role": "user", "content": prompt}
]
response = await llm_service.chat(messages=messages, temperature=0.3, max_tokens=1500)
answer = llm_service.extract_message_content(response)
return {
"success": True,
"intent": "question",
"question": question,
"context_preview": context_text[:500] + "..." if len(context_text) > 500 else context_text,
"message": "已找到相关上下文，可进行问答"
"answer": answer,
"filename": filename,
"message": "已生成回答"
}
except Exception as e:
@@ -299,12 +477,53 @@ class InstructionExecutor:
async def _execute_compare(self, params: Dict[str, Any], context: Dict[str, Any]) -> Dict[str, Any]:
"""执行对比分析"""
try:
docs = context.get("source_docs", [])
docs = context.get("source_docs", []) or []
instruction_text = context.get("instruction", "")
# 优先从指令中提取具体的文件名
filenames = _extract_filenames_from_text(instruction_text)
if filenames:
# 只选择文件名匹配的那些文档
matched_docs = []
for doc in docs:
fname = doc.get("metadata", {}).get("original_filename", "").lower()
for fn in filenames:
if fn.lower() in fname or fname in fn.lower():
matched_docs.append(doc)
break
# 如果匹配到足够文档，用匹配的
if len(matched_docs) >= 2:
docs = matched_docs
else:
# 匹配不够，尝试按文件名搜索 MongoDB
all_found = []
for fn in filenames:
found = await mongodb.search_documents(fn, limit=5)
all_found.extend(found)
seen = set()
unique_docs = []
for d in all_found:
did = d.get("_id", "")
if did and did not in seen:
seen.add(did)
unique_docs.append(d)
if len(unique_docs) >= 2:
docs = unique_docs
elif len(unique_docs) == 1 and len(docs) >= 1:
# 找到一个指定的 + 用一个通用的
docs = unique_docs + docs[:1]
elif docs and len(filenames) == 1:
# 找到一个指定文件名但只有一个匹配，尝试补充
docs = unique_docs + [d for d in docs if d not in unique_docs]
docs = docs[:2]
if len(docs) < 2:
return {
"success": False,
"intent": "compare",
"error": "对比需要至少2个文档",
"message": "请上传至少2个文档进行对比"
"message": "请上传至少2个文档进行对比，或明确说出要对比的文档名称"
}
# 提取文档基本信息
@@ -329,6 +548,7 @@ class InstructionExecutor:
logger.error(f"对比执行失败: {e}")
return {
"success": False,
"intent": "compare",
"error": str(e),
"message": f"对比分析失败: {str(e)}"
}
@@ -336,10 +556,23 @@ class InstructionExecutor:
async def _execute_edit(self, params: Dict[str, Any], context: Dict[str, Any]) -> Dict[str, Any]:
"""执行文档编辑操作"""
try:
docs = context.get("source_docs", [])
docs = context.get("source_docs", []) or []
instruction_text = context.get("instruction", "")
# 如果没有文档，尝试从指令中提取文件名搜索
if not docs:
filename_match = re.search(r'["""]([^"""]+\.(?:docx|xlsx|md|txt))["""]', instruction_text)
if not filename_match:
filename_match = re.search(r'([^\s，。！？]+\.(?:docx|xlsx|md|txt))', instruction_text)
if filename_match:
found = await mongodb.search_documents(filename_match.group(1), limit=3)
if found:
docs = found
if not docs:
return {
"success": False,
"intent": "edit",
"error": "没有可用的文档",
"message": "请先上传要编辑的文档"
}
@@ -405,7 +638,7 @@ class InstructionExecutor:
- Word -> Markdown
"""
try:
docs = context.get("source_docs", [])
docs = context.get("source_docs", []) or []
if not docs:
return {
"success": False,