Cleaning Data Returned by Report Engine's LLM
This commit is contained in:
马一丁
@@ -14,6 +14,7 @@ from ..prompts import (
|
||||
SYSTEM_PROMPT_DOCUMENT_LAYOUT,
|
||||
build_document_layout_prompt,
|
||||
)
|
||||
from ..utils.json_parser import RobustJSONParser, JSONParseError
|
||||
from .base_node import BaseNode
|
||||
|
||||
|
||||
@@ -27,6 +28,12 @@ class DocumentLayoutNode(BaseNode):
|
||||
def __init__(self, llm_client):
|
||||
"""记录LLM客户端并设置节点名字，供BaseNode日志使用"""
|
||||
super().__init__(llm_client, "DocumentLayoutNode")
|
||||
# 初始化鲁棒JSON解析器，启用所有修复策略
|
||||
self.json_parser = RobustJSONParser(
|
||||
enable_json_repair=True,
|
||||
enable_llm_repair=False, # 可以根据需要启用LLM修复
|
||||
max_repair_attempts=3,
|
||||
)
|
||||
|
||||
def run(
|
||||
self,
|
||||
@@ -82,8 +89,14 @@ class DocumentLayoutNode(BaseNode):
|
||||
"""
|
||||
解析LLM返回的JSON文本，若失败则抛出友好错误。
|
||||
|
||||
使用鲁棒JSON解析器进行多重修复尝试：
|
||||
1. 清理markdown标记和思考内容
|
||||
2. 本地语法修复（括号平衡、逗号补全、控制字符转义等）
|
||||
3. 使用json_repair库进行高级修复
|
||||
4. 可选的LLM辅助修复
|
||||
|
||||
参数:
|
||||
raw: LLM原始返回字符串，允许带```包裹。
|
||||
raw: LLM原始返回字符串，允许带```包裹、思考内容等。
|
||||
|
||||
返回:
|
||||
dict: 结构化的设计稿。
|
||||
@@ -91,19 +104,25 @@ class DocumentLayoutNode(BaseNode):
|
||||
异常:
|
||||
ValueError: 当响应为空或JSON解析失败时抛出。
|
||||
"""
|
||||
cleaned = raw.strip()
|
||||
if cleaned.startswith("```json"):
|
||||
cleaned = cleaned[7:]
|
||||
if cleaned.startswith("```"):
|
||||
cleaned = cleaned[3:]
|
||||
if cleaned.endswith("```"):
|
||||
cleaned = cleaned[:-3]
|
||||
cleaned = cleaned.strip()
|
||||
if not cleaned:
|
||||
raise ValueError("文档设计LLM返回空内容")
|
||||
try:
|
||||
return json.loads(cleaned)
|
||||
except json.JSONDecodeError as exc:
|
||||
result = self.json_parser.parse(
|
||||
raw,
|
||||
context_name="文档设计",
|
||||
expected_keys=["title", "toc", "hero"],
|
||||
)
|
||||
# 验证关键字段的类型
|
||||
if not isinstance(result.get("title"), str):
|
||||
logger.warning("文档设计缺少title字段或类型错误，使用默认值")
|
||||
result.setdefault("title", "未命名报告")
|
||||
if not isinstance(result.get("toc"), (list, dict)):
|
||||
logger.warning("文档设计缺少toc字段或类型错误，使用空列表")
|
||||
result.setdefault("toc", [])
|
||||
if not isinstance(result.get("hero"), dict):
|
||||
logger.warning("文档设计缺少hero字段或类型错误，使用空对象")
|
||||
result.setdefault("hero", {})
|
||||
return result
|
||||
except JSONParseError as exc:
|
||||
# 转换为原有的异常类型以保持向后兼容
|
||||
raise ValueError(f"文档设计JSON解析失败: {exc}") from exc
|
||||
|
||||
|
||||
|
||||
@@ -12,6 +12,7 @@ from loguru import logger
|
||||
|
||||
from .base_node import BaseNode
|
||||
from ..prompts import SYSTEM_PROMPT_TEMPLATE_SELECTION
|
||||
from ..utils.json_parser import RobustJSONParser, JSONParseError
|
||||
|
||||
|
||||
class TemplateSelectionNode(BaseNode):
|
||||
@@ -25,13 +26,19 @@ class TemplateSelectionNode(BaseNode):
|
||||
def __init__(self, llm_client, template_dir: str = "ReportEngine/report_template"):
|
||||
"""
|
||||
初始化模板选择节点
|
||||
|
||||
|
||||
Args:
|
||||
llm_client: LLM客户端
|
||||
template_dir: 模板目录路径
|
||||
"""
|
||||
super().__init__(llm_client, "TemplateSelectionNode")
|
||||
self.template_dir = template_dir
|
||||
# 初始化鲁棒JSON解析器，启用所有修复策略
|
||||
self.json_parser = RobustJSONParser(
|
||||
enable_json_repair=True,
|
||||
enable_llm_repair=False,
|
||||
max_repair_attempts=3,
|
||||
)
|
||||
|
||||
def run(self, input_data: Dict[str, Any], **kwargs) -> Dict[str, Any]:
|
||||
"""
|
||||
@@ -137,20 +144,22 @@ class TemplateSelectionNode(BaseNode):
|
||||
|
||||
# 调用LLM
|
||||
response = self.llm_client.stream_invoke_to_string(SYSTEM_PROMPT_TEMPLATE_SELECTION, user_message)
|
||||
|
||||
|
||||
# 检查响应是否为空
|
||||
if not response or not response.strip():
|
||||
logger.error("LLM返回空响应")
|
||||
return None
|
||||
|
||||
|
||||
logger.info(f"LLM原始响应: {response}")
|
||||
|
||||
# 尝试解析JSON响应
|
||||
|
||||
# 尝试解析JSON响应，使用鲁棒解析器
|
||||
try:
|
||||
# 清理响应文本
|
||||
cleaned_response = self._clean_llm_response(response)
|
||||
result = json.loads(cleaned_response)
|
||||
|
||||
result = self.json_parser.parse(
|
||||
response,
|
||||
context_name="模板选择",
|
||||
expected_keys=["template_name", "selection_reason"],
|
||||
)
|
||||
|
||||
# 验证选择的模板是否存在
|
||||
selected_template_name = result.get('template_name', '')
|
||||
for template in available_templates:
|
||||
@@ -161,38 +170,16 @@ class TemplateSelectionNode(BaseNode):
|
||||
'template_content': template['content'],
|
||||
'selection_reason': result.get('selection_reason', 'LLM智能选择')
|
||||
}
|
||||
|
||||
|
||||
logger.error(f"LLM选择的模板不存在: {selected_template_name}")
|
||||
return None
|
||||
|
||||
except json.JSONDecodeError as e:
|
||||
|
||||
except JSONParseError as e:
|
||||
logger.error(f"JSON解析失败: {str(e)}")
|
||||
# 尝试从文本响应中提取模板信息
|
||||
return self._extract_template_from_text(response, available_templates)
|
||||
|
||||
def _clean_llm_response(self, response: str) -> str:
|
||||
"""
|
||||
清理LLM响应。
|
||||
|
||||
去掉 ```json``` 包裹以及前后空白，方便 `json.loads`。
|
||||
|
||||
参数:
|
||||
response: LLM原始响应。
|
||||
|
||||
返回:
|
||||
str: 适合直接做JSON解析的纯文本。
|
||||
"""
|
||||
# 移除可能的markdown代码块标记
|
||||
if '```json' in response:
|
||||
response = response.split('```json')[1].split('```')[0]
|
||||
elif '```' in response:
|
||||
response = response.split('```')[1].split('```')[0]
|
||||
|
||||
# 移除前后空白
|
||||
response = response.strip()
|
||||
|
||||
return response
|
||||
|
||||
def _extract_template_from_text(self, response: str, available_templates: List[Dict[str, Any]]) -> Optional[Dict[str, Any]]:
|
||||
"""
|
||||
从文本响应中提取模板信息。
|
||||
|
||||
@@ -14,6 +14,7 @@ from ..prompts import (
|
||||
SYSTEM_PROMPT_WORD_BUDGET,
|
||||
build_word_budget_prompt,
|
||||
)
|
||||
from ..utils.json_parser import RobustJSONParser, JSONParseError
|
||||
from .base_node import BaseNode
|
||||
|
||||
|
||||
@@ -27,6 +28,12 @@ class WordBudgetNode(BaseNode):
|
||||
def __init__(self, llm_client):
|
||||
"""仅记录LLM客户端引用，方便run阶段发起请求"""
|
||||
super().__init__(llm_client, "WordBudgetNode")
|
||||
# 初始化鲁棒JSON解析器，启用所有修复策略
|
||||
self.json_parser = RobustJSONParser(
|
||||
enable_json_repair=True,
|
||||
enable_llm_repair=False, # 可以根据需要启用LLM修复
|
||||
max_repair_attempts=3,
|
||||
)
|
||||
|
||||
def run(
|
||||
self,
|
||||
@@ -79,8 +86,14 @@ class WordBudgetNode(BaseNode):
|
||||
"""
|
||||
将LLM输出的JSON文本转为字典，失败时提示规划异常。
|
||||
|
||||
使用鲁棒JSON解析器进行多重修复尝试：
|
||||
1. 清理markdown标记和思考内容
|
||||
2. 本地语法修复（括号平衡、逗号补全、控制字符转义等）
|
||||
3. 使用json_repair库进行高级修复
|
||||
4. 可选的LLM辅助修复
|
||||
|
||||
参数:
|
||||
raw: LLM返回值，可能包含```包裹。
|
||||
raw: LLM返回值，可能包含```包裹、思考内容等。
|
||||
|
||||
返回:
|
||||
dict: 合法的篇幅规划JSON。
|
||||
@@ -88,19 +101,25 @@ class WordBudgetNode(BaseNode):
|
||||
异常:
|
||||
ValueError: 当响应为空或JSON解析失败时抛出。
|
||||
"""
|
||||
cleaned = raw.strip()
|
||||
if cleaned.startswith("```json"):
|
||||
cleaned = cleaned[7:]
|
||||
if cleaned.startswith("```"):
|
||||
cleaned = cleaned[3:]
|
||||
if cleaned.endswith("```"):
|
||||
cleaned = cleaned[:-3]
|
||||
cleaned = cleaned.strip()
|
||||
if not cleaned:
|
||||
raise ValueError("篇幅规划LLM返回空内容")
|
||||
try:
|
||||
return json.loads(cleaned)
|
||||
except json.JSONDecodeError as exc:
|
||||
result = self.json_parser.parse(
|
||||
raw,
|
||||
context_name="篇幅规划",
|
||||
expected_keys=["totalWords", "globalGuidelines", "chapters"],
|
||||
)
|
||||
# 验证关键字段的类型
|
||||
if not isinstance(result.get("totalWords"), (int, float)):
|
||||
logger.warning("篇幅规划缺少totalWords字段或类型错误，使用默认值")
|
||||
result.setdefault("totalWords", 10000)
|
||||
if not isinstance(result.get("globalGuidelines"), list):
|
||||
logger.warning("篇幅规划缺少globalGuidelines字段或类型错误，使用空列表")
|
||||
result.setdefault("globalGuidelines", [])
|
||||
if not isinstance(result.get("chapters"), (list, dict)):
|
||||
logger.warning("篇幅规划缺少chapters字段或类型错误，使用空列表")
|
||||
result.setdefault("chapters", [])
|
||||
return result
|
||||
except JSONParseError as exc:
|
||||
# 转换为原有的异常类型以保持向后兼容
|
||||
raise ValueError(f"篇幅规划JSON解析失败: {exc}") from exc
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user