feat: 简易应用页面参数优化 (#1182)

Author: shaohuzhang1

apps/application/chat_pipeline/step/generate_human_message_step/i_generate_human_message_step.py

@@ -37,6 +37,8 @@ class InstanceSerializer(serializers.Serializer):
             "最大携带知识库段落长度"))
         # 模板
         prompt = serializers.CharField(required=True, error_messages=ErrMessage.char("提示词"))
+        system = serializers.CharField(required=False, allow_null=True, allow_blank=True,
+                                       error_messages=ErrMessage.char("系统提示词(角色)"))
         # 补齐问题
         padding_problem_text = serializers.CharField(required=False, error_messages=ErrMessage.char("补齐问题"))
         # 未查询到引用分段
@@ -59,6 +61,7 @@ def execute(self,
                 prompt: str,
                 padding_problem_text: str = None,
                 no_references_setting=None,
+                system=None,
                 **kwargs) -> List[BaseMessage]:
         """
 
@@ -71,6 +74,7 @@ def execute(self,
         :param padding_problem_text        用户修改文本
         :param kwargs:                     其他参数
         :param no_references_setting:     无引用分段设置
+        :param system                     系统提示称
         :return:
         """
         pass

apps/application/chat_pipeline/step/generate_human_message_step/impl/base_generate_human_message_step.py

@@ -9,6 +9,7 @@
 from typing import List, Dict
 
 from langchain.schema import BaseMessage, HumanMessage
+from langchain_core.messages import SystemMessage
 
 from application.chat_pipeline.I_base_chat_pipeline import ParagraphPipelineModel
 from application.chat_pipeline.step.generate_human_message_step.i_generate_human_message_step import \
@@ -27,6 +28,7 @@ def execute(self, problem_text: str,
                 prompt: str,
                 padding_problem_text: str = None,
                 no_references_setting=None,
+                system=None,
                 **kwargs) -> List[BaseMessage]:
         prompt = prompt if (paragraph_list is not None and len(paragraph_list) > 0) else no_references_setting.get(
             'value')
@@ -35,6 +37,11 @@ def execute(self, problem_text: str,
         history_message = [[history_chat_record[index].get_human_message(), history_chat_record[index].get_ai_message()]
                            for index in
                            range(start_index if start_index > 0 else 0, len(history_chat_record))]
+        if system is not None and len(system) > 0:
+            return [SystemMessage(system), *flat_map(history_message),
+                    self.to_human_message(prompt, exec_problem_text, max_paragraph_char_number, paragraph_list,
+                                          no_references_setting)]
+
         return [*flat_map(history_message),
                 self.to_human_message(prompt, exec_problem_text, max_paragraph_char_number, paragraph_list,
                                       no_references_setting)]

apps/application/chat_pipeline/step/reset_problem_step/i_reset_problem_step.py

@@ -29,6 +29,8 @@ class InstanceSerializer(serializers.Serializer):
                                                     error_messages=ErrMessage.list("历史对答"))
         # 大语言模型
         chat_model = ModelField(required=False, allow_null=True, error_messages=ErrMessage.base("大语言模型"))
+        problem_optimization_prompt = serializers.CharField(required=False, max_length=102400,
+                                                            error_messages=ErrMessage.char("问题补全提示词"))
 
     def get_step_serializer(self, manage: PipelineManage) -> Type[serializers.Serializer]:
         return self.InstanceSerializer
@@ -47,5 +49,6 @@ def _run(self, manage: PipelineManage):
 
     @abstractmethod
     def execute(self, problem_text: str, history_chat_record: List[ChatRecord] = None, chat_model: BaseChatModel = None,
+                problem_optimization_prompt=None,
                 **kwargs):
         pass

apps/application/chat_pipeline/step/reset_problem_step/impl/base_reset_problem_step.py

@@ -21,6 +21,7 @@
 
 class BaseResetProblemStep(IResetProblemStep):
     def execute(self, problem_text: str, history_chat_record: List[ChatRecord] = None, chat_model: BaseChatModel = None,
+                problem_optimization_prompt=None,
                 **kwargs) -> str:
         if chat_model is None:
             self.context['message_tokens'] = 0
@@ -30,15 +31,19 @@ def execute(self, problem_text: str, history_chat_record: List[ChatRecord] = Non
         history_message = [[history_chat_record[index].get_human_message(), history_chat_record[index].get_ai_message()]
                            for index in
                            range(start_index if start_index > 0 else 0, len(history_chat_record))]
+        reset_prompt = problem_optimization_prompt if problem_optimization_prompt else prompt
         message_list = [*flat_map(history_message),
-                        HumanMessage(content=prompt.format(**{'question': problem_text}))]
+                        HumanMessage(content=reset_prompt.replace('{question}', problem_text))]
         response = chat_model.invoke(message_list)
         padding_problem = problem_text
         if response.content.__contains__("<data>") and response.content.__contains__('</data>'):
             padding_problem_data = response.content[
                                    response.content.index('<data>') + 6:response.content.index('</data>')]
             if padding_problem_data is not None and len(padding_problem_data.strip()) > 0:
                 padding_problem = padding_problem_data
+        elif len(response.content) > 0:
+            padding_problem = response.content
+
         try:
             request_token = chat_model.get_num_tokens_from_messages(message_list)
             response_token = chat_model.get_num_tokens(padding_problem)

apps/application/migrations/0014_application_problem_optimization_prompt.py

@@ -0,0 +1,18 @@
+# Generated by Django 4.2.15 on 2024-09-13 18:57
+
+from django.db import migrations, models
+
+
+class Migration(migrations.Migration):
+
+    dependencies = [
+        ('application', '0013_application_tts_type'),
+    ]
+
+    operations = [
+        migrations.AddField(
+            model_name='application',
+            name='problem_optimization_prompt',
+            field=models.CharField(blank=True, default='()里面是用户问题,根据上下文回答揣测用户问题({question}) 要求: 输出一个补全问题,并且放在<data></data>标签中', max_length=102400, null=True, verbose_name='问题优化提示词'),
+        ),
+    ]

apps/application/models/application.py

@@ -35,7 +35,7 @@ def get_dataset_setting_dict():
 
 
 def get_model_setting_dict():
-    return {'prompt': Application.get_default_model_prompt()}
+    return {'prompt': Application.get_default_model_prompt(), 'no_references_prompt': '{question}'}
 
 
 class Application(AppModelMixin):
@@ -54,8 +54,13 @@ class Application(AppModelMixin):
     work_flow = models.JSONField(verbose_name="工作流数据", default=dict)
     type = models.CharField(verbose_name="应用类型", choices=ApplicationTypeChoices.choices,
                             default=ApplicationTypeChoices.SIMPLE, max_length=256)
-    tts_model = models.ForeignKey(Model, related_name='tts_model_id', on_delete=models.SET_NULL, db_constraint=False, blank=True, null=True)
-    stt_model = models.ForeignKey(Model, related_name='stt_model_id', on_delete=models.SET_NULL, db_constraint=False, blank=True, null=True)
+    problem_optimization_prompt = models.CharField(verbose_name="问题优化提示词", max_length=102400, blank=True,
+                                                   null=True,
+                                                   default="()里面是用户问题,根据上下文回答揣测用户问题({question}) 要求: 输出一个补全问题,并且放在<data></data>标签中")
+    tts_model = models.ForeignKey(Model, related_name='tts_model_id', on_delete=models.SET_NULL, db_constraint=False,
+                                  blank=True, null=True)
+    stt_model = models.ForeignKey(Model, related_name='stt_model_id', on_delete=models.SET_NULL, db_constraint=False,
+                                  blank=True, null=True)
     tts_model_enable = models.BooleanField(verbose_name="语音合成模型是否启用", default=False)
     stt_model_enable = models.BooleanField(verbose_name="语音识别模型是否启用", default=False)
     tts_type = models.CharField(verbose_name="语音播放类型", max_length=20, default="BROWSER")

apps/application/serializers/application_serializers.py

@@ -120,7 +120,12 @@ class DatasetSettingSerializer(serializers.Serializer):
 
 
 class ModelSettingSerializer(serializers.Serializer):
-    prompt = serializers.CharField(required=True, max_length=2048, error_messages=ErrMessage.char("提示词"))
+    prompt = serializers.CharField(required=False, allow_null=True, allow_blank=True, max_length=102400,
+                                   error_messages=ErrMessage.char("提示词"))
+    system = serializers.CharField(required=False, allow_null=True, allow_blank=True, max_length=102400,
+                                   error_messages=ErrMessage.char("角色提示词"))
+    no_references_prompt = serializers.CharField(required=True, max_length=102400, allow_null=True, allow_blank=True,
+                                                 error_messages=ErrMessage.char("无引用分段提示词"))
 
 
 class ApplicationWorkflowSerializer(serializers.Serializer):
@@ -174,7 +179,7 @@ class ApplicationSerializer(serializers.Serializer):
                                  error_messages=ErrMessage.char("应用描述"))
     model_id = serializers.CharField(required=False, allow_null=True, allow_blank=True,
                                      error_messages=ErrMessage.char("模型"))
-    multiple_rounds_dialogue = serializers.BooleanField(required=True, error_messages=ErrMessage.char("多轮对话"))
+    dialogue_number = serializers.BooleanField(required=True, error_messages=ErrMessage.char("会话次数"))
     prologue = serializers.CharField(required=False, allow_null=True, allow_blank=True, max_length=4096,
                                      error_messages=ErrMessage.char("开场白"))
     dataset_id_list = serializers.ListSerializer(required=False, child=serializers.UUIDField(required=True),
@@ -185,6 +190,8 @@ class ApplicationSerializer(serializers.Serializer):
     model_setting = ModelSettingSerializer(required=True)
     # 问题补全
     problem_optimization = serializers.BooleanField(required=True, error_messages=ErrMessage.boolean("问题补全"))
+    problem_optimization_prompt = serializers.CharField(required=False, max_length=102400,
+                                                        error_messages=ErrMessage.char("问题补全提示词"))
     # 应用类型
     type = serializers.CharField(required=True, error_messages=ErrMessage.char("应用类型"),
                                  validators=[
@@ -364,8 +371,8 @@ class Edit(serializers.Serializer):
                                      error_messages=ErrMessage.char("应用描述"))
         model_id = serializers.CharField(required=False, allow_blank=True, allow_null=True,
                                          error_messages=ErrMessage.char("模型"))
-        multiple_rounds_dialogue = serializers.BooleanField(required=False,
-                                                            error_messages=ErrMessage.boolean("多轮会话"))
+        dialogue_number = serializers.IntegerField(required=False,
+                                                   error_messages=ErrMessage.boolean("多轮会话"))
         prologue = serializers.CharField(required=False, allow_null=True, allow_blank=True, max_length=4096,
                                          error_messages=ErrMessage.char("开场白"))
         dataset_id_list = serializers.ListSerializer(required=False, child=serializers.UUIDField(required=True),
@@ -430,13 +437,14 @@ def insert_simple(self, application: Dict):
         def to_application_model(user_id: str, application: Dict):
             return Application(id=uuid.uuid1(), name=application.get('name'), desc=application.get('desc'),
                                prologue=application.get('prologue'),
-                               dialogue_number=3 if application.get('multiple_rounds_dialogue') else 0,
+                               dialogue_number=application.get('dialogue_number', 0),
                                user_id=user_id, model_id=application.get('model_id'),
                                dataset_setting=application.get('dataset_setting'),
                                model_setting=application.get('model_setting'),
                                problem_optimization=application.get('problem_optimization'),
                                type=ApplicationTypeChoices.SIMPLE,
                                model_params_setting=application.get('model_params_setting', {}),
+                               problem_optimization_prompt=application.get('problem_optimization_prompt', None),
                                work_flow={}
                                )
 

apps/application/serializers/chat_message_serializers.py

@@ -60,6 +60,17 @@ def __init__(self,
         self.chat_record_list: List[ChatRecord] = []
         self.work_flow_version = work_flow_version
 
+    @staticmethod
+    def get_no_references_setting(dataset_setting, model_setting):
+        no_references_setting = dataset_setting.get(
+            'no_references_setting', {
+                'status': 'ai_questioning',
+                'value': '{question}'})
+        if no_references_setting.get('status') == 'ai_questioning':
+            no_references_prompt = model_setting.get('no_references_prompt', '{question}')
+            no_references_setting['value'] = no_references_prompt if len(no_references_prompt) > 0 else "{question}"
+        return no_references_setting
+
     def to_base_pipeline_manage_params(self):
         dataset_setting = self.application.dataset_setting
         model_setting = self.application.model_setting
@@ -80,20 +91,21 @@ def to_base_pipeline_manage_params(self):
             'history_chat_record': self.chat_record_list,
             'chat_id': self.chat_id,
             'dialogue_number': self.application.dialogue_number,
+            'problem_optimization_prompt': self.application.problem_optimization_prompt if self.application.problem_optimization_prompt is not None and len(
+                self.application.problem_optimization_prompt) > 0 else '()里面是用户问题,根据上下文回答揣测用户问题({question}) 要求: 输出一个补全问题,并且放在<data></data>标签中',
             'prompt': model_setting.get(
-                'prompt') if 'prompt' in model_setting else Application.get_default_model_prompt(),
+                'prompt') if 'prompt' in model_setting and len(model_setting.get(
+                'prompt')) > 0 else Application.get_default_model_prompt(),
+            'system': model_setting.get(
+                'system', None),
             'model_id': model_id,
             'problem_optimization': self.application.problem_optimization,
             'stream': True,
             'model_params_setting': model_params_setting if self.application.model_params_setting is None or len(
                 self.application.model_params_setting.keys()) == 0 else self.application.model_params_setting,
             'search_mode': self.application.dataset_setting.get(
                 'search_mode') if 'search_mode' in self.application.dataset_setting else 'embedding',
-            'no_references_setting': self.application.dataset_setting.get(
-                'no_references_setting') if 'no_references_setting' in self.application.dataset_setting else {
-                'status': 'ai_questioning',
-                'value': '{question}',
-            },
+            'no_references_setting': self.get_no_references_setting(self.application.dataset_setting, model_setting),
             'user_id': self.application.user_id
         }