[feat][add intent mudule]

Author: lightislost

muagent/base_configs/prompts/intention_template_prompt.py

@@ -1,3 +1,6 @@
+from typing import Union, Optional
+
+
 RECOGNIZE_INTENTION_PROMPT = """你是一个任务决策助手，能够将理解用户意图并决策采取最合适的行动，尽可能地以有帮助和准确的方式回应人类，
 使用 JSON Blob 来指定一个返回的内容，提供一个 action（行动）。
 有效的 'action' 值为：'planning'(需要先进行拆解计划) or 'only_answer' （不需要拆解问题即可直接回答问题）or "tool_using" (使用工具来回答问题) or 'coding'(生成可执行的代码)。
@@ -59,3 +62,77 @@
 
 ### Ouput Response
 '''
+
+
+def get_intention_prompt(
+    background: str, intentions: Union[list, tuple], examples: Optional[dict]=None
+) -> str:
+    nums_zh = ('一', '两', '三', '四', '五', '六', '七', '八', '九', '十')
+
+    intention_num = len(intentions)
+    num_zh = nums_zh[intention_num - 1] if intention_num <= 10 else intention_num
+    prompt = f'##背景##\n{background}\n\n##任务##\n辨别用户的询问意图，包括以下{num_zh}类：\n'
+
+    for i, val in enumerate(intentions):
+        if isinstance(val, (list, tuple)):
+            k, v = val
+            cur_intention = f'{i + 1}. {k}：{v}\n'
+        else:
+            cur_intention = f'{i + 1}. {val}\n'
+        prompt += cur_intention
+    
+    prompt += '\n##注意事项##\n'
+    num_range_str = '、'.join(map(str, range(1, intention_num + 1)))
+    prompt += f'回答：数字{num_range_str}中的一个来表示用户的询问意图，对应上述{num_zh}种分类。避免提供额外解释或其他信息。\n\n'
+
+    if examples:
+        prompt += '##示例##\n'
+        intention_idx_map = {k[0]: idx + 1 for idx, k in enumerate(intentions)}
+        for query, ans in examples.items():
+            ans = intention_idx_map[ans]
+            prompt += f'询问：{query}\n回答：{ans}\n\n'
+    
+    prompt += '##用户询问##\n询问：{query}\n回答：'
+    return prompt
+
+
+INTENTIONS_CONSULT_WHICH = [
+    ('整体计划查询', '用户询问关于某个解决方案的完整流程或步骤，包含但不限于“整个流程”、“步骤”、“流程图”等词汇或概念。'),
+    ('下一步任务查询', '用户询问在某个解决方案的特定步骤中应如何操作或处理，通常会提及“下一步”、“具体操作”、“如何做”等，且明确指向解决方案中的某个特定环节。'),
+    ('闲聊', '用户询问的内容与当前的技术问题或解决方案无关，更多是出于兴趣或社交性质的交流。')
+]
+CONSULT_WHICH_PROMPT = get_intention_prompt(
+    '作为运维领域的客服，您的职责是根据用户询问的内容，精准判断其背后的意图，以便提供最恰当的服务和支持。',
+    INTENTIONS_CONSULT_WHICH,
+    {
+        '系统升级的整个流程是怎样的？': '整体计划查询',
+        '听说你们采用了新工具，能讲讲它的特点吗？': '闲聊'
+    }
+)
+
+INTENTIONS_WHETHER_EXEC = [
+    ('执行', '当用户声明自己在使用某平台、服务、产品或功能时遇到具体问题，且明确表示不知道如何解决时，其意图应被分类为“执行”。'),
+    ('询问', '当用户明确询问某些解决方案的背景、流程、方式方法等信息，或只是出于好奇想要了解更多信息，或只是简单闲聊时，其意图应被分类为“询问”。')
+]
+WHETHER_EXECUTE_PROMPT = get_intention_prompt(
+    '作为运维领域的客服，您需要根据用户询问判断其主要意图，以确定接下来的运维流程。',
+    INTENTIONS_WHETHER_EXEC,
+    {
+        '为什么我的优惠券使用失败？': '执行',
+        '我想知道如何才能更好地优化我的服务器性能，你们有什么建议吗？': '询问'
+    }
+)
+
+DIRECT_CHAT_PROMPT = """##背景##
+作为运维领域的客服，您的职责是根据自身专业知识回答用户询问，以便提供最恰当的服务和支持。
+
+##任务##
+基于您所掌握的领域知识，对用户的提问进行回答。
+
+##注意事项##
+请尽量从客观的角度来回答问题，内容符合事实、有理有据。
+
+##用户询问##
+询问：{query}
+回答：
+"""

muagent/db_handler/vector_db_handler/tbase_handler.py

@@ -101,7 +101,7 @@ def search(self, query, index_name: str = None, query_params: dict = {}, limit=1
         res = index.search(query, query_params=query_params)
         return res
 
-    def vector_search(self, base_query: str, index_name: str = None, query_params: dict={}):
+    def vector_search(self, base_query: str, index_name: str = None, query_params: dict={}, limit=10):
         '''
         vector_search
         :param base_query:
@@ -114,7 +114,7 @@ def vector_search(self, base_query: str, index_name: str = None, query_params: d
                 .sort_by('distance')
                 .dialect(2)
         )
-        r = self.search(query, index_name, query_params)
+        r = self.search(query, index_name, query_params, limit=limit)
         return r
 
     def delete(self, content):

muagent/schemas/common/auto_extract_graph_schema.py

@@ -33,6 +33,9 @@ class GNode(BaseModel):
     type: str
     attributes: Dict
 
+    def __getattr__(self, name: str):
+        return self.attributes.get(name)
+
 
 class GEdge(BaseModel):
     start_id: str

muagent/service/ekg_construct/ekg_construct_base.py

@@ -28,6 +28,7 @@
 
 from muagent.base_configs.env_config import KB_ROOT_PATH
 
+from muagent.service.ekg_inference.intention_router import IntentionRouter
 from muagent.llm_models.get_embedding import get_embedding
 from muagent.utils.common_utils import getCurrentDatetime, getCurrentTimestap
 from muagent.utils.common_utils import double_hashing
@@ -53,6 +54,7 @@ def __init__(
             gb_config: GBConfig = None,
             tb_config: TBConfig = None,
             sls_config: SLSConfig = None,
+            intention_router: IntentionRouter = None,
             do_init: bool = False,
             kb_root_path: str = KB_ROOT_PATH,
         ):
@@ -71,6 +73,7 @@ def __init__(
 
         # get llm model
         self.model = getChatModelFromConfig(self.llm_config)
+        self.intention_router = IntentionRouter(self.model, embed_config=self.embed_config)
         # init db handler
         self.init_handler()
 
@@ -519,16 +522,17 @@ def create_ekg(
             text: str, 
             teamid: str, 
             service_name: str, 
+            rootid: str,
             intent_text: str = None, 
             intent_nodes: List[str] = [], 
             all_intent_list: List=[],
-            do_save: bool = False
+            do_save: bool = False,
         ):
 
         if intent_nodes:
             ancestor_list = intent_nodes
         elif intent_text or text:
-            ancestor_list, all_intent_list = self.get_intents(intent_text or text)
+            ancestor_list, all_intent_list = self.get_intents(rootid, intent_text or text)
         else:
             raise Exception(f"must have intent infomation")
 
@@ -542,6 +546,9 @@ def create_ekg(
         graph = self.write2kg(result["sls_graph"], result["tbase_graph"], teamid, do_save=do_save)
         result["graph"] = graph
         return result
+    
+    def dsl2graph(self, ):
+        pass
 
     def text2graph(self, text: str, intents: List[str], all_intent_list: List[str], teamid: str) -> dict:
         # generate graph by llm
@@ -600,34 +607,17 @@ def returndsl(self, graph_datas_by_path: dict, intents: List[str], ) -> dict:
         res["graph"] = Graph(nodes=merge_gbase_nodes, edges=merge_gbase_edges, paths=[])
         return res
 
-    def get_intent_by_alarm(self, ):
-        pass 
-
-    def get_intents(self, alarm_list: list[dict], ) -> EKGIntentResp:
+    def get_intents(self, rootid, text: str):
         '''according contents search intents'''
-        ancestor_list = set()
-        all_intent_list = []
-        for alarm in alarm_list:
-            ancestor, all_intent = self.get_intent_by_alarm(alarm)
-            if ancestor is None:
-                continue
-            ancestor_list.add(ancestor)
-            all_intent_list.append(all_intent)
-
-        return list(ancestor_list), all_intent_list
-
-    def get_intents_by_alarms(self, alarm_list: list[dict], ) -> EKGIntentResp:
-        '''according contents search intents'''
-        ancestor_list = set()
-        all_intent_list = []
-        for alarm in alarm_list:
-            ancestor, all_intent = self.get_intent_by_alarm(alarm)
-            if ancestor is None:
-                continue
-            ancestor_list.add(ancestor)
-            all_intent_list.append(all_intent)
-
-        return list(ancestor_list), all_intent_list
+        result = self.intention_router.get_intention_by_node_info_nlp(
+            root_node_id=rootid,
+            query=text,
+            start_from_root=True,
+            gb_handler=self.gb,
+            tb_handler=self.tb,
+            agent=self.model
+        )
+        return result.get("node_id", ), []
 
     def get_graph_by_text(self, text: str) -> EKGSlsData:
         '''according text generate ekg's raw datas'''

muagent/service/ekg_inference/intention_match_rule.py

@@ -0,0 +1,41 @@
+import re
+import Levenshtein
+from muagent.schemas.common import GNode
+
+
+class MatchRule:
+    @classmethod
+    def edit_distance(cls, node: GNode, pattern=None, **kwargs):
+        if len(kwargs) == 0:
+            return -float('inf')
+
+        s = list(kwargs.values())[0]
+        desc: str = node.attributes.get('description', '')
+
+        if pattern is None:
+            return -Levenshtein.distance(desc, s)
+
+        desc_list = re.findall(pattern, desc)
+        if not desc_list:
+            return -float('inf')
+
+        return max([-Levenshtein.distance(x, s) for x in desc_list])
+    
+    @classmethod
+    def edit_distance_integer(cls, node: GNode, **kwargs):
+        return cls.edit_distance(node, pattern='\d+', **kwargs)
+
+
+class RuleDict(dict):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+    def save(self):
+        """save the rules"""
+        raise NotImplementedError
+
+    def load(self, **kwargs):
+        """load the rules"""
+        raise NotImplementedError
+
+rule_dict = RuleDict()