liuguancen
846380879b
三个独立修复 / 排查:
1. R1 思考过程不显示
- 根因: chat_test.py 等 <think> 开标签出现才进思考态，但 R1
流式输出本来就在 reasoning 态启动，永远不出 <think>，所有
reasoning 全部当 text 走到答案区
- 修法: 引入 r1_thinking_done 状态机，默认在思考态，
看到 </think> 切换；R1-70B 直连本地代理 deepseek-r1
（官方 deepseek-reasoner 把 reasoning 放独立字段，旧版
callback 取不到）
- 结果验证: "1+1" → 269 think + 40 text，思考与答案正确分流
2. 选题推荐场景拒答 + chat 模板标记泄漏
- 根因: prompt 写死了 "你只能回答有关选题推荐的问题"
+ 直接嵌入 <|im_start|>/<|im_end|> Qwen chat 标记
- 修法: 改写 Topic Recommend Assistant prompt，删 chat 标记，
行为准则改为"沾边查询用工具回答"；agent_v2 增加 strip
防守层
- 结果验证: "钢铁行业研究重点方向" → agent 调工具，不再拒答
3. 知识库召回 0 排查（数据问题，未根治）
- 根因: kb_config.py 把所有 KB（政策库/钢铁库/报告库等）
都映射到 t_policy_total_bge_new_v2，但 Milvus 里根本没有
这个 collection（实际只有 11 个 p_* 个人库 + 1 个
t_journal_article_bge_v1）
- 临时改: search_tool.py 加诊断日志 [RAG诊断] 输出每个 KB
召回数；rag_search 内 for 循环里首个 KB 空就 return 的
bug 改 continue
- 待决策: kb_config 是否把默认 KB 映射到唯一存在的
t_journal_article_bge_v1，或重建对应集合
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
336 lines
16 KiB
Python
336 lines
16 KiB
Python
import asyncio
|
||
import concurrent
|
||
from datetime import datetime
|
||
import json
|
||
import logging
|
||
import re
|
||
from typing import List
|
||
|
||
from fastapi import logger
|
||
|
||
from configs import kb_config
|
||
from configs.model_config import LLM_MODELS
|
||
from server.agent.tools import duckduckgo_search
|
||
from server.agent.tools.duckduckgo_search import duckduckgo_search_iter
|
||
from server.agent.tools.knowledgebase_kgo_search import knowledgebase_kgo_iter
|
||
from server.agent.tools.rag_search import rag_search1
|
||
from server.chat import utils
|
||
from server.chat.policy_fun_iast import get_llm_model_response
|
||
from server.chat.utils import doc_to_list, get_similar_documents1, solve_knowledge_map, solve_mental_data
|
||
from server.knowledge_base.kb_doc_api import search_docs
|
||
|
||
|
||
|
||
def rag_search(query: str,uid):
|
||
"""
|
||
根据用户输入的query，返回rag搜索结果
|
||
"""
|
||
source_docs = []
|
||
try:
|
||
search = json.loads(query)
|
||
logging.info(f'模型输入: {search["query"]}')
|
||
original_query = search["query"]
|
||
search_query = get_llm_model_response(
|
||
strategy_name="rag_search_rewrite",
|
||
llm_model_name=LLM_MODELS[0],
|
||
template_prompt_name="rag_search_rewrite",
|
||
prompt_param_dict={"input": search["query"], "year": datetime.now().strftime("%Y")},
|
||
temperature=0.3,
|
||
max_tokens=512
|
||
)
|
||
logging.info(f'模型改写: {search_query}')
|
||
search_keywords = []
|
||
search_text = f"{search_query}"
|
||
# if type(search["keywords"]) == list:
|
||
# search_keywords = search["keywords"]
|
||
# for keyword in search_keywords:
|
||
# search_text += f" {keyword}"
|
||
# else:
|
||
# search_keywords = search["keywords"].split(",")
|
||
# for keyword in search_keywords:
|
||
# search_text += f" {keyword}"
|
||
self_database = utils.get_shared_variable(uid)
|
||
result = []
|
||
|
||
knownledge_name = []
|
||
if type(search["knowledge_name"]) == list:
|
||
knownledge_name=search["knowledge_name"]
|
||
else:
|
||
knownledge_name=search["knowledge_name"].split(",")
|
||
if "美术专业知识库" in knownledge_name:
|
||
knownledge_name.remove("美术专业知识库")
|
||
if "database" in self_database:
|
||
self_database["database"]= self_database["database"].append("p_cafa0101011")
|
||
else:
|
||
self_database["database"] = ["p_cafa0101011"]
|
||
# 添加个人知识库
|
||
if "database" in self_database:
|
||
knownledge_name.extend(self_database["database"])
|
||
knownledge_name = [knownledge for knownledge in knownledge_name
|
||
if (knownledge in kb_config.CH_BASE_NAME
|
||
or knownledge in kb_config.EN_BASE_NAME
|
||
or knownledge in getattr(kb_config, "YJ_BASE_NAME", [])
|
||
or kb_config.SELF_KNOWLEDGE_BASE.match(knownledge)
|
||
or knownledge == "coding")]
|
||
if len(knownledge_name)==0:
|
||
#result.append(f"没有找到匹配的知识库,请必须更换联网思索搜索更多知识库内容")
|
||
return result,source_docs
|
||
# knownledge_name=kb_config.CH_BASE_NAME
|
||
|
||
knownledge_name=solve_knowledge_map(knownledge_name)
|
||
#knownledge_name = ["p_c88859a3d06e4265bd01d816ef2650d1"]
|
||
num = 0
|
||
temp=utils.get_shared_variable(uid)
|
||
for knownledge in knownledge_name:
|
||
seen_docs = set()
|
||
duplicate_indices = []
|
||
# 针对中国钢铁行业动态库增加日期范围过滤
|
||
expr_param = ""
|
||
if knownledge == kb_config.STEEL_KB:
|
||
time_today = datetime.now().strftime("%Y-%m-%d")
|
||
# 调用LLM生成日期表达式，模板沿用 get_policy_time
|
||
try:
|
||
expr_candidate = get_llm_model_response(
|
||
strategy_name="get steel time",
|
||
llm_model_name=LLM_MODELS[0],
|
||
template_prompt_name="get_steel_time",
|
||
prompt_param_dict={"query": original_query, "time": time_today},
|
||
temperature=0.01,
|
||
max_tokens=512
|
||
).replace("None", "").strip()
|
||
expr_param = expr_candidate if expr_candidate else ""
|
||
except Exception as _:
|
||
expr_param = ""
|
||
|
||
doc_list = search_docs(
|
||
usr_query=original_query,
|
||
fileName=[],
|
||
top_k=20,
|
||
score_threshold=1.0,
|
||
query=search_text,
|
||
knowledge_base_name=knownledge,
|
||
expr=expr_param
|
||
)
|
||
logging.info(f"[RAG诊断] kb={knownledge!r} expr={expr_param!r} 召回 {len(doc_list)} docs")
|
||
|
||
if len(doc_list)==0:
|
||
# 修 bug: 原代码 return 导致首个 KB 空就放弃全部 KB；改 continue 继续尝试下一个
|
||
continue
|
||
titles = temp["title"]
|
||
doc_list,title = utils.remove_docs1(titles,doc_list)
|
||
titles.extend(title)
|
||
for inum,doc in enumerate(doc_list):
|
||
solve_mental_data(knownledge,doc_list,doc=doc,seen_docs=seen_docs,duplicate_indices=duplicate_indices,knowledge=result,inum=inum)
|
||
|
||
# 从policydocs中删除重复的文档（从后往前删除以防止索引错位）
|
||
for index in sorted(duplicate_indices, reverse=True):
|
||
del doc_list[index]
|
||
# 处理原文来源进入数组。使用开关语句明确各个条件分支
|
||
match knownledge:
|
||
# 属于政策库分支，入参为中文政策库名称
|
||
case kb_config.DEFAULT_POLICY_BASE:
|
||
doc_to_list(num,kb_config.DEFAULT_POLICY_BASE_NAME,doc_list,source_docs)
|
||
# 属于期刊论文库分支，入参为期刊论文库的中文名称
|
||
case kb_config.DEFAULT_JOURNAL_BASE:
|
||
doc_to_list(num,kb_config.DEFAULT_JOURNAL_BASE_NAME,doc_list,source_docs)
|
||
# 属于报告库分支，入参为报告库中文名称
|
||
case kb_config.DEFAULT_REPORT_BASE1:
|
||
doc_to_list(num,kb_config.DEFAULT_REPORT_BASE_NAME,doc_list,source_docs)
|
||
# 属于冶金行业新闻库分支，入参为冶金行业新闻库中文名称
|
||
case kb_config.GY_NEWS_BASE:
|
||
doc_to_list(num,kb_config.GY_NEWS_BASE_NAME,doc_list,source_docs)
|
||
# 属于冶金行业报告库分支，入参为冶金行业报告库中文名称
|
||
case kb_config.GY_REPORT_BASE:
|
||
doc_to_list(num,kb_config.GY_REPORT_BASE_NAME,doc_list,source_docs)
|
||
# 属于冶金专业知识库分支，入参为冶金专业知识库中文名称
|
||
case kb_config.GY_JOURNAL_BASE:
|
||
doc_to_list(num,kb_config.GY_JOURNAL_BASE_NAME,doc_list,source_docs)
|
||
# 新增冶金新闻库（2024年以及之前）
|
||
case kb_config.YJ_NEWS_BASE:
|
||
doc_to_list(num,kb_config.YJ_NEWS_BASE_NAME,doc_list,source_docs)
|
||
# 新增冶金中文期刊库
|
||
case kb_config.YJ_CH_JOURNAL_BASE:
|
||
doc_to_list(num,kb_config.YJ_CH_JOURNAL_BASE_NAME,doc_list,source_docs)
|
||
# 新增冶金外文期刊库
|
||
case kb_config.YJ_FOR_JOURNAL_BASE:
|
||
doc_to_list(num,kb_config.YJ_FOR_JOURNAL_BASE_NAME,doc_list,source_docs)
|
||
# 新增冶金OA期刊库
|
||
case kb_config.YJ_OA_JOURNAL_BASE:
|
||
doc_to_list(num,kb_config.YJ_OA_JOURNAL_BASE_NAME,doc_list,source_docs)
|
||
# 新增冶金政策库
|
||
case kb_config.YJ_POLICYS_BASE:
|
||
doc_to_list(num,kb_config.YJ_POLICYS_BASE_NAME,doc_list,source_docs)
|
||
# 新增中国钢铁行业动态库
|
||
case kb_config.STEEL_KB:
|
||
doc_to_list(num,kb_config.STEEL_KB_NAME,doc_list,source_docs)
|
||
# 属于个人知识库分支
|
||
case _ if kb_config.SELF_KNOWLEDGE_BASE.match(knownledge) or knownledge == "coding":
|
||
doc_to_list(num,knownledge,doc_list,source_docs)
|
||
case _:
|
||
print(f"输入了没有的知识库名称")
|
||
return "输入了没有的知识库名称",source_docs
|
||
# num += len(source_docs[knownledge])
|
||
# 构建缓存对象用于h'per_query_cache'用于传递给其他方法使用uuid获取本轮对话的文献来源
|
||
# del num
|
||
# source = utils.get_shared_variable(uid)
|
||
# print(utils.get_shared_variable(uid))
|
||
# source["source_docs"]=source_docs
|
||
# utils.set_shared_variable(uid,source)
|
||
except Exception as e:
|
||
logging.error(f"Error in rag_search: {e}")
|
||
# return "入参格式需为{\"knowledge_name\":[\"XXX\",\"XXX\"],\"query\":\"XXX\",\"keywords\":[\"XXX\", \"XXX\", \"XXX\", \"XXX\"]}检查输入参数如果没有缺少必要值，当前工具异常请换其他工具"
|
||
return "当前工具异常！请换其他工具",source_docs
|
||
|
||
return result,source_docs
|
||
|
||
|
||
|
||
|
||
def knowledgebase_kgo_search(query: str, uid) -> List[str]:
|
||
try:
|
||
res = knowledgebase_kgo_iter(query,uid)
|
||
try:
|
||
if type(res[0])==list and type(res[1])==list:
|
||
return res
|
||
elif type(res[1])==list and len(res[1])>0:
|
||
res[0]=[]
|
||
return res
|
||
else:
|
||
temp = []
|
||
temp[0]=[]
|
||
temp[1]=[]
|
||
return temp
|
||
except Exception as e:
|
||
temp = {}
|
||
logging.error(f"No docs: {e}")
|
||
temp[0]=[]
|
||
temp[1]=[]
|
||
return temp
|
||
except json.JSONDecodeError:
|
||
# 如果JSON解码失败，则返回错误消息
|
||
logging.error("Invalid JSON format in query.")
|
||
return "<关键指令>不需要再调用该工具了</关键指令>"
|
||
except KeyError:
|
||
# 如果解析的JSON对象中缺少必要的键，则返回错误消息
|
||
return "请尝试调用其他工具"
|
||
except Exception as e:
|
||
# 捕获其他所有异常，并返回通用错误消息
|
||
return f"发生错误：{str(e)}，请尝试调用其他工具"
|
||
|
||
|
||
def inner_duckduckgo_search(query: str, uuid:str,) :
|
||
logging.info(f"模型输入: {query}")
|
||
combined_result = asyncio.run(duckduckgo_search_iter(query, uuid, "y","default" ))
|
||
# 以标准json格式输出
|
||
logging.info("返回JSON格式的结果给到模型...")
|
||
return combined_result
|
||
|
||
|
||
def search_tool(query: str):
|
||
"""获取到uid并拆分query"""
|
||
if "<param>"in query:
|
||
query = query.replace("<param>","").replace("</param>","")
|
||
matches = re.findall(r'\{.*?\}', query)
|
||
if len(matches)>=2:
|
||
query = matches[0]
|
||
else:
|
||
return "<关键指令>当前工具不需要再调用</关键指令>"
|
||
time_based_uuid = json.loads(matches[1])["uuid"]
|
||
search = json.loads(query)
|
||
if type(search["query"])==list and len(search["query"])>0:
|
||
searches = search["query"][0]
|
||
elif type(search["query"])==list and len(search["query"]) == 0:
|
||
searches = "无"
|
||
else:
|
||
searches = search["query"]
|
||
"""
|
||
根据用户输入的query，返回rag搜索结果
|
||
"""
|
||
|
||
try:
|
||
with concurrent.futures.ThreadPoolExecutor() as executor:
|
||
# 提交任务并发执行
|
||
|
||
test = {}
|
||
test["num"]=0
|
||
test["source_docs"]=[]
|
||
test["END"] = ""
|
||
test["title"] = []
|
||
utils.set_shared_variable(time_based_uuid+"￥",test)
|
||
# future2 = executor.submit(knowledgebase_kgo_search,search["query"],time_based_uuid+"q")
|
||
future1 = executor.submit(rag_search,query,time_based_uuid)
|
||
# if not "type" in utils.get_shared_variable(time_based_uuid):
|
||
# future2 = executor.submit(knowledgebase_kgo_search,searches,time_based_uuid+"￥")
|
||
if not "type" in utils.get_shared_variable(time_based_uuid):
|
||
future2 = executor.submit(knowledgebase_kgo_search,searches,time_based_uuid+"￥")
|
||
result3 = []
|
||
# 获取结果
|
||
result1,sourcedocs = future1.result()
|
||
# 诊断：看 rag_search 实际召回多少
|
||
logging.info(f"[RAG诊断] rag_search 返回 result1={len(result1) if isinstance(result1, list) else type(result1).__name__}, sourcedocs={len(sourcedocs) if isinstance(sourcedocs, list) else type(sourcedocs).__name__}, kb={search.get('knowledge_name')}, query={search.get('query')!r}")
|
||
result2 = {}
|
||
if "type" in utils.get_shared_variable(time_based_uuid):
|
||
result2[0] =[]
|
||
result2[1] = []
|
||
else:
|
||
result2 = future2.result()
|
||
logging.info(f"[RAG诊断] zhipu 返回 result2[0]={len(result2[0]) if isinstance(result2[0], list) else type(result2[0]).__name__}, result2[1]={len(result2[1]) if isinstance(result2[1], list) else type(result2[1]).__name__}")
|
||
# if "type" in utils.get_shared_variable(time_based_uuid):
|
||
# result2[0] =[]
|
||
# result2[1] = []
|
||
# else:
|
||
# result2 = future2.result()
|
||
# result2[0] = []
|
||
# result2[1] = []
|
||
utils.remove_shared_variable(time_based_uuid+"q")
|
||
if type(result2[1]) == list:
|
||
if type(sourcedocs) == list:
|
||
sourcedocs.extend(result2[1])
|
||
else:
|
||
sourcedocs = []
|
||
if type(result1) == list:
|
||
result1.extend(result2[0])
|
||
result3 = result1
|
||
else:
|
||
result3 = result2[0]
|
||
|
||
logging.info(f"result2:{result2[1]}")
|
||
source = []
|
||
res=[]
|
||
sources = utils.get_shared_variable(time_based_uuid)
|
||
i = sources["num"]
|
||
num = sources["num"]
|
||
for result in sourcedocs:
|
||
try:
|
||
i+=1
|
||
res3 = re.sub(r'\[\d+\]', f"[{i}]", result.replace("\n",""), count=1)
|
||
if res3:
|
||
source.append(re.sub(r'\[\d+\]', f"[{i}]", result.replace("\n",""), count=1))
|
||
else:
|
||
i -= 1
|
||
except Exception as e:
|
||
i -= 1
|
||
pass
|
||
# internet_search_res = f"参考资料[{len(result1)+1}-{len(source)}]:{result2[0]}"
|
||
# internet_search_res = f"参考资料:{result2[0]}"
|
||
j = sources["num"]
|
||
for result in result3:
|
||
j+=1
|
||
res.append(re.sub(r'\[\d+\]', f"[{j}]", result, count=1))
|
||
|
||
print(utils.get_shared_variable(time_based_uuid))
|
||
# sources["source_docs"]=source
|
||
sources["source_docs"].extend(source)
|
||
sources["num"]=i
|
||
# sources["END"] = "ok"
|
||
utils.set_shared_variable(time_based_uuid,sources)
|
||
logging.info(f"result1:{result1},sourcedocs:{sourcedocs}")
|
||
logging.info(f"result2:{result2}")
|
||
logging.info(f"{res}")
|
||
if len(res) ==0 and len(source)==0:
|
||
return f"尝试调整入参重新调用知识库联想工具（同一个问题调用超过三次就不要再使用知识库联想工具了，浪费时间）"
|
||
return f"<关键指令>如果你在写文章禁止在非规定位置输出参考资料</关键指令>资料：{res}\n资料来源为：{source}\n 注意：如果你在根据大纲撰写文章，撰写中间部分章节禁止输出综上所述之类的影响文风的话，撰写中间部分禁止输出附录引用文献等！！！"
|
||
except Exception as e:
|
||
logging.error(f"Error occurred during search_tool execution.{e}")
|
||
return "同一个问题调用知识库联想工具超过5次就不要再调用知识库联想"
|
||
|