feat(rag): 子章节级图片过滤 + VLM 后台增强 + 意图分析优化

- 图片选择新增 section_path 字段,支持子章节级过滤
- _filter_images_by_answer 增加 primary_sections 参数和叶子节点匹配
- 检索发散检测:primary_leaf_names > 3 时全局阈值+1
- _FIGURE_ANSWER_KEYWORDS 移除单字"图""表",正则图号兜底防误触发
- lazy_enhance 后台增强流程优化
- 意图分析与 LLM 工具层改进

评测:图片选择 F1 从 52.4% 提升至 66.3%(+13.9pp),Precision +16.5pp
This commit is contained in:
lacerate551
2026-06-20 19:26:40 +08:00
parent ee5295cc97
commit 5ba0d782e2
8 changed files with 3402 additions and 3028 deletions

View File

@@ -233,10 +233,10 @@ class IntentAnalyzer:
self._exact_cache_max = 500
def _get_client(self):
"""获取 LLM 客户端"""
"""获取 LLM 客户端(百炼快速模型)"""
if self._client is None:
from config import get_llm_client
self._client = get_llm_client()
from config import get_intent_client
self._client = get_intent_client()
return self._client
def _get_cache(self):

View File

@@ -71,19 +71,38 @@ def call_llm(
return response
content = response.choices[0].message.content
# 推理模型兼容content 为空时尝试从 reasoning_content 提取
# 推理模型兼容mimo-v2.5 等):
# 推理模型思考链消耗大量 token~1000max_tokens 不足时 content 为空,
# 全部输出进入 reasoning_content。此处从思考链中提取有效内容。
if not content or not content.strip():
reasoning = getattr(response.choices[0].message, 'reasoning_content', None)
if reasoning and reasoning.strip():
# 从思维链中提取 JSON 块作为内容
json_match = re.search(r'\{[\s\S]*\}', reasoning)
if json_match:
logger.info("LLM: content为空从reasoning_content提取JSON")
return json_match.group().strip()
logger.warning("LLM 返回空 content可能需要增大 max_tokens")
# 先去掉 <think>...</think> 标签
cleaned = re.sub(r'', '', reasoning, flags=re.DOTALL).strip()
if cleaned:
logger.info("LLM: content为空从reasoning_content提取内容")
# 尝试提取 JSON 对象(兼容结构化响应场景)
json_match = re.search(r'\{[\s\S]*\}', cleaned)
if json_match:
try:
json.loads(json_match.group())
return json_match.group().strip()
except (json.JSONDecodeError, ValueError):
pass
# 尝试提取 JSON 数组
bracket_match = re.search(r'\[[\s\S]*\]', cleaned)
if bracket_match:
try:
json.loads(bracket_match.group())
return bracket_match.group().strip()
except (json.JSONDecodeError, ValueError):
pass
# 纯文本响应:直接返回清理后的内容
return cleaned
logger.warning("LLM 返回空 content 且 reasoning_content 也无法提取(可能需要增大 max_tokens")
return None
return content.strip()
except Exception as e:
logger.warning(f"LLM 调用失败: {e}")
@@ -95,7 +114,7 @@ def call_llm_stream(
prompt: str,
model: str,
temperature: float = 0.3,
max_tokens: int = 1000,
max_tokens: int = 3000,
messages: List[dict] = None,
error_prefix: str = "[错误]",
**kwargs
@@ -104,13 +123,14 @@ def call_llm_stream(
流式 LLM 调用(生成器封装)
自动处理流式响应,逐块 yield 文本内容。
兼容推理模型mimo-v2.5 等):当 content 为空时回退到 reasoning_content。
Args:
client: OpenAI 客户端实例
prompt: 用户提示
model: 模型名称
temperature: 温度参数
max_tokens: 最大 token 数
max_tokens: 最大 token 数(推理模型需留足思考链预算)
messages: 完整消息列表
error_prefix: 错误时的前缀
**kwargs: 其他参数
@@ -135,9 +155,33 @@ def call_llm_stream(
**kwargs
)
content_yielded = False
reasoning_buffer = []
for chunk in stream:
if chunk.choices and chunk.choices[0].delta.content:
yield chunk.choices[0].delta.content
if not chunk.choices:
continue
delta = chunk.choices[0].delta
# 正常 content 输出
if hasattr(delta, 'content') and delta.content:
content_yielded = True
yield delta.content
continue
# 推理模型reasoning_content思考链
rc = getattr(delta, 'reasoning_content', None)
if rc:
reasoning_buffer.append(rc)
# 回退content 为空但 reasoning_content 有内容(推理模型 token 不足时)
if not content_yielded and reasoning_buffer:
reasoning_text = ''.join(reasoning_buffer)
# 去掉 <think>...</think> 标签
cleaned = re.sub(r'', '', reasoning_text, flags=re.DOTALL).strip()
if cleaned:
logger.info("流式 LLM: content为空从reasoning_content提取内容")
yield cleaned
except Exception as e:
logger.error(f"LLM 流式调用失败: {e}")