hz/ebiz-dify-ai
1
0
Fork 0
mirror of http://112.124.100.131/huang.ze/ebiz-dify-ai.git synced 2025-12-16 06:16:53 +08:00
Code Issues Packages Projects Releases Wiki Activity

refactor advanced prompt core. (#1350)

Browse Source
This commit is contained in:
Garfield Dai
2023-10-18 20:02:52 +08:00
committed by GitHub
parent 52ebffa857
commit fe14130b3c
9 changed files with 405 additions and 272 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
api/core/model_providers/models/llm/baichuan_model.py
View File

@@ -37,12 +37,6 @@ class BaichuanModel(BaseLLM):
prompts = self._get_prompt_from_messages(messages)
return self._client.generate([prompts], stop, callbacks)
def prompt_file_name(self, mode: str) -> str:
if mode == 'completion':
return 'baichuan_completion'
else:
return 'baichuan_chat'
def get_num_tokens(self, messages: List[PromptMessage]) -> int:
"""
get num tokens of prompt messages.

218
api/core/model_providers/models/llm/base.py
View File

@@ -1,28 +1,18 @@
import json
import os
import re
import time
from abc import abstractmethod
from typing import List, Optional, Any, Union, Tuple
from typing import List, Optional, Any, Union
import decimal
import logging
from langchain.callbacks.manager import Callbacks
from langchain.memory.chat_memory import BaseChatMemory
from langchain.schema import LLMResult, SystemMessage, AIMessage, HumanMessage, BaseMessage, ChatGeneration
from langchain.schema import LLMResult, BaseMessage, ChatGeneration
from core.callback_handler.std_out_callback_handler import DifyStreamingStdOutCallbackHandler, DifyStdOutCallbackHandler
from core.helper import moderation
from core.model_providers.models.base import BaseProviderModel
from core.model_providers.models.entity.message import PromptMessage, MessageType, LLMRunResult, to_prompt_messages, \
to_lc_messages
from core.model_providers.models.entity.message import PromptMessage, MessageType, LLMRunResult, to_lc_messages
from core.model_providers.models.entity.model_params import ModelType, ModelKwargs, ModelMode, ModelKwargsRules
from core.model_providers.providers.base import BaseModelProvider
from core.prompt.prompt_builder import PromptBuilder
from core.prompt.prompt_template import PromptTemplateParser
from core.third_party.langchain.llms.fake import FakeLLM
import logging
from extensions.ext_database import db
logger = logging.getLogger(__name__)
@@ -320,206 +310,8 @@ class BaseLLM(BaseProviderModel):
def support_streaming(self):
return False
def get_prompt(self, mode: str,
pre_prompt: str, inputs: dict,
query: str,
context: Optional[str],
memory: Optional[BaseChatMemory]) -> \
Tuple[List[PromptMessage], Optional[List[str]]]:
prompt_rules = self._read_prompt_rules_from_file(self.prompt_file_name(mode))
prompt, stops = self._get_prompt_and_stop(prompt_rules, pre_prompt, inputs, query, context, memory)
return [PromptMessage(content=prompt)], stops
def get_advanced_prompt(self, app_mode: str,
app_model_config: str, inputs: dict,
query: str,
context: Optional[str],
memory: Optional[BaseChatMemory]) -> List[PromptMessage]:
model_mode = app_model_config.model_dict['mode']
conversation_histories_role = {}
raw_prompt_list = []
prompt_messages = []
if app_mode == 'chat' and model_mode == ModelMode.COMPLETION.value:
prompt_text = app_model_config.completion_prompt_config_dict['prompt']['text']
raw_prompt_list = [{
'role': MessageType.USER.value,
'text': prompt_text
}]
conversation_histories_role = app_model_config.completion_prompt_config_dict['conversation_histories_role']
elif app_mode == 'chat' and model_mode == ModelMode.CHAT.value:
raw_prompt_list = app_model_config.chat_prompt_config_dict['prompt']
elif app_mode == 'completion' and model_mode == ModelMode.CHAT.value:
raw_prompt_list = app_model_config.chat_prompt_config_dict['prompt']
elif app_mode == 'completion' and model_mode == ModelMode.COMPLETION.value:
prompt_text = app_model_config.completion_prompt_config_dict['prompt']['text']
raw_prompt_list = [{
'role': MessageType.USER.value,
'text': prompt_text
}]
else:
raise Exception("app_mode or model_mode not support")
for prompt_item in raw_prompt_list:
prompt = prompt_item['text']
# set prompt template variables
prompt_template = PromptTemplateParser(template=prompt)
prompt_inputs = {k: inputs[k] for k in prompt_template.variable_keys if k in inputs}
if '#context#' in prompt:
if context:
prompt_inputs['#context#'] = context
else:
prompt_inputs['#context#'] = ''
if '#query#' in prompt:
if query:
prompt_inputs['#query#'] = query
else:
prompt_inputs['#query#'] = ''
if '#histories#' in prompt:
if memory and app_mode == 'chat' and model_mode == ModelMode.COMPLETION.value:
memory.human_prefix = conversation_histories_role['user_prefix']
memory.ai_prefix = conversation_histories_role['assistant_prefix']
histories = self._get_history_messages_from_memory(memory, 2000)
prompt_inputs['#histories#'] = histories
else:
prompt_inputs['#histories#'] = ''
prompt = prompt_template.format(
prompt_inputs
)
prompt = re.sub(r'<\|.*?\|>', '', prompt)
prompt_messages.append(PromptMessage(type = MessageType(prompt_item['role']) ,content=prompt))
if memory and app_mode == 'chat' and model_mode == ModelMode.CHAT.value:
memory.human_prefix = MessageType.USER.value
memory.ai_prefix = MessageType.ASSISTANT.value
histories = self._get_history_messages_list_from_memory(memory, 2000)
prompt_messages.extend(histories)
if app_mode == 'chat' and model_mode == ModelMode.CHAT.value:
prompt_messages.append(PromptMessage(type = MessageType.USER ,content=query))
return prompt_messages
def prompt_file_name(self, mode: str) -> str:
if mode == 'completion':
return 'common_completion'
else:
return 'common_chat'
def _get_prompt_and_stop(self, prompt_rules: dict, pre_prompt: str, inputs: dict,
query: str,
context: Optional[str],
memory: Optional[BaseChatMemory]) -> Tuple[str, Optional[list]]:
context_prompt_content = ''
if context and 'context_prompt' in prompt_rules:
prompt_template = PromptTemplateParser(template=prompt_rules['context_prompt'])
context_prompt_content = prompt_template.format(
{'context': context}
)
pre_prompt_content = ''
if pre_prompt:
prompt_template = PromptTemplateParser(template=pre_prompt)
prompt_inputs = {k: inputs[k] for k in prompt_template.variable_keys if k in inputs}
pre_prompt_content = prompt_template.format(
prompt_inputs
)
prompt = ''
for order in prompt_rules['system_prompt_orders']:
if order == 'context_prompt':
prompt += context_prompt_content
elif order == 'pre_prompt':
prompt += pre_prompt_content
query_prompt = prompt_rules['query_prompt'] if 'query_prompt' in prompt_rules else '{{query}}'
if memory and 'histories_prompt' in prompt_rules:
# append chat histories
tmp_human_message = PromptBuilder.to_human_message(
prompt_content=prompt + query_prompt,
inputs={
'query': query
}
)
if self.model_rules.max_tokens.max:
curr_message_tokens = self.get_num_tokens(to_prompt_messages([tmp_human_message]))
max_tokens = self.model_kwargs.max_tokens
rest_tokens = self.model_rules.max_tokens.max - max_tokens - curr_message_tokens
rest_tokens = max(rest_tokens, 0)
else:
rest_tokens = 2000
memory.human_prefix = prompt_rules['human_prefix'] if 'human_prefix' in prompt_rules else 'Human'
memory.ai_prefix = prompt_rules['assistant_prefix'] if 'assistant_prefix' in prompt_rules else 'Assistant'
histories = self._get_history_messages_from_memory(memory, rest_tokens)
prompt_template = PromptTemplateParser(template=prompt_rules['histories_prompt'])
histories_prompt_content = prompt_template.format({'histories': histories})
prompt = ''
for order in prompt_rules['system_prompt_orders']:
if order == 'context_prompt':
prompt += context_prompt_content
elif order == 'pre_prompt':
prompt += (pre_prompt_content + '\n') if pre_prompt_content else ''
elif order == 'histories_prompt':
prompt += histories_prompt_content
prompt_template = PromptTemplateParser(template=query_prompt)
query_prompt_content = prompt_template.format({'query': query})
prompt += query_prompt_content
prompt = re.sub(r'<\|.*?\|>', '', prompt)
stops = prompt_rules.get('stops')
if stops is not None and len(stops) == 0:
stops = None
return prompt, stops
def _read_prompt_rules_from_file(self, prompt_name: str) -> dict:
# Get the absolute path of the subdirectory
prompt_path = os.path.join(
os.path.dirname(os.path.dirname(os.path.dirname(os.path.dirname(os.path.realpath(__file__))))),
'prompt/generate_prompts')
json_file_path = os.path.join(prompt_path, f'{prompt_name}.json')
# Open the JSON file and read its content
with open(json_file_path, 'r') as json_file:
return json.load(json_file)
def _get_history_messages_from_memory(self, memory: BaseChatMemory,
max_token_limit: int) -> str:
"""Get memory messages."""
memory.max_token_limit = max_token_limit
memory_key = memory.memory_variables[0]
external_context = memory.load_memory_variables({})
return external_context[memory_key]
def _get_history_messages_list_from_memory(self, memory: BaseChatMemory,
max_token_limit: int) -> List[PromptMessage]:
"""Get memory messages."""
memory.max_token_limit = max_token_limit
memory.return_messages = True
memory_key = memory.memory_variables[0]
external_context = memory.load_memory_variables({})
memory.return_messages = False
return to_prompt_messages(external_context[memory_key])
def _get_prompt_from_messages(self, messages: List[PromptMessage],
model_mode: Optional[ModelMode] = None) -> Union[str | List[BaseMessage]]:
model_mode: Optional[ModelMode] = None) -> Union[str , List[BaseMessage]]:
if not model_mode:
model_mode = self.model_mode

9
api/core/model_providers/models/llm/huggingface_hub_model.py
View File

@@ -66,15 +66,6 @@ class HuggingfaceHubModel(BaseLLM):
prompts = self._get_prompt_from_messages(messages)
return self._client.get_num_tokens(prompts)
def prompt_file_name(self, mode: str) -> str:
if 'baichuan' in self.name.lower():
if mode == 'completion':
return 'baichuan_completion'
else:
return 'baichuan_chat'
else:
return super().prompt_file_name(mode)
def _set_model_kwargs(self, model_kwargs: ModelKwargs):
provider_model_kwargs = self._to_model_kwargs_input(self.model_rules, model_kwargs)
self.client.model_kwargs = provider_model_kwargs

9
api/core/model_providers/models/llm/openllm_model.py
View File

@@ -49,15 +49,6 @@ class OpenLLMModel(BaseLLM):
prompts = self._get_prompt_from_messages(messages)
return max(self._client.get_num_tokens(prompts), 0)
def prompt_file_name(self, mode: str) -> str:
if 'baichuan' in self.name.lower():
if mode == 'completion':
return 'baichuan_completion'
else:
return 'baichuan_chat'
else:
return super().prompt_file_name(mode)
def _set_model_kwargs(self, model_kwargs: ModelKwargs):
pass

9
api/core/model_providers/models/llm/xinference_model.py
View File

@@ -59,15 +59,6 @@ class XinferenceModel(BaseLLM):
prompts = self._get_prompt_from_messages(messages)
return max(self._client.get_num_tokens(prompts), 0)
def prompt_file_name(self, mode: str) -> str:
if 'baichuan' in self.name.lower():
if mode == 'completion':
return 'baichuan_completion'
else:
return 'baichuan_chat'
else:
return super().prompt_file_name(mode)
def _set_model_kwargs(self, model_kwargs: ModelKwargs):
pass