jyapayne/dify
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

feat: universal chat in explore (#649)

Browse source 
Co-authored-by: StyleZhang <jasonapring2015@outlook.com>
This commit is contained in:
John Wang 2023-07-27 13:08:57 +08:00 • committed by GitHub
commit 4fdb37771a
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
64 changed files with 3186 additions and 858 deletions
Download patch file Download diff file Expand all files Collapse all files

71
api/services/app_model_config_service.py
View file

@ -1,6 +1,7 @@
import re
import uuid
from core.agent.agent_executor import PlanningStrategy
from core.constant import llm_constant
from models.account import Account
from services.dataset_service import DatasetService
@ -31,6 +32,16 @@ MODELS_BY_APP_MODE = {
]
}
SUPPORT_AGENT_MODELS = [
"gpt-4",
"gpt-4-32k",
"gpt-3.5-turbo",
"gpt-3.5-turbo-16k",
]
SUPPORT_TOOLS = ["dataset", "google_search", "web_reader", "wikipedia"]
class AppModelConfigService:
@staticmethod
def is_dataset_exists(account: Account, dataset_id: str) -> bool:
@ -58,7 +69,8 @@ class AppModelConfigService:
if not isinstance(cp["max_tokens"], int) or cp["max_tokens"] <= 0 or cp["max_tokens"] > \
llm_constant.max_context_token_length[model_name]:
raise ValueError(
"max_tokens must be an integer greater than 0 and not exceeding the maximum value of the corresponding model")
"max_tokens must be an integer greater than 0 "
"and not exceeding the maximum value of the corresponding model")
# temperature
if 'temperature' not in cp:
@ -148,11 +160,6 @@ class AppModelConfigService:
if not isinstance(config["speech_to_text"]["enabled"], bool):
raise ValueError("enabled in speech_to_text must be of boolean type")
provider_name = LLMBuilder.get_default_provider(account.current_tenant_id, 'whisper-1')
if config["speech_to_text"]["enabled"] and provider_name != 'openai':
raise ValueError("provider not support speech to text")
# more_like_this
if 'more_like_this' not in config or not config["more_like_this"]:
@ -169,6 +176,33 @@ class AppModelConfigService:
if not isinstance(config["more_like_this"]["enabled"], bool):
raise ValueError("enabled in more_like_this must be of boolean type")
# sensitive_word_avoidance
if 'sensitive_word_avoidance' not in config or not config["sensitive_word_avoidance"]:
config["sensitive_word_avoidance"] = {
"enabled": False
}
if not isinstance(config["sensitive_word_avoidance"], dict):
raise ValueError("sensitive_word_avoidance must be of dict type")
if "enabled" not in config["sensitive_word_avoidance"] or not config["sensitive_word_avoidance"]["enabled"]:
config["sensitive_word_avoidance"]["enabled"] = False
if not isinstance(config["sensitive_word_avoidance"]["enabled"], bool):
raise ValueError("enabled in sensitive_word_avoidance must be of boolean type")
if "words" not in config["sensitive_word_avoidance"] or not config["sensitive_word_avoidance"]["words"]:
config["sensitive_word_avoidance"]["words"] = ""
if not isinstance(config["sensitive_word_avoidance"]["words"], str):
raise ValueError("words in sensitive_word_avoidance must be of string type")
if "canned_response" not in config["sensitive_word_avoidance"] or not config["sensitive_word_avoidance"]["canned_response"]:
config["sensitive_word_avoidance"]["canned_response"] = ""
if not isinstance(config["sensitive_word_avoidance"]["canned_response"], str):
raise ValueError("canned_response in sensitive_word_avoidance must be of string type")
# model
if 'model' not in config:
raise ValueError("model is required")
@ -274,6 +308,12 @@ class AppModelConfigService:
if not isinstance(config["agent_mode"]["enabled"], bool):
raise ValueError("enabled in agent_mode must be of boolean type")
if "strategy" not in config["agent_mode"] or not config["agent_mode"]["strategy"]:
config["agent_mode"]["strategy"] = PlanningStrategy.ROUTER.value
if config["agent_mode"]["strategy"] not in [member.value for member in list(PlanningStrategy.__members__.values())]:
raise ValueError("strategy in agent_mode must be in the specified strategy list")
if "tools" not in config["agent_mode"] or not config["agent_mode"]["tools"]:
config["agent_mode"]["tools"] = []
@ -282,8 +322,8 @@ class AppModelConfigService:
for tool in config["agent_mode"]["tools"]:
key = list(tool.keys())[0]
if key not in ["sensitive-word-avoidance", "dataset"]:
raise ValueError("Keys in agent_mode.tools list can only be 'sensitive-word-avoidance' or 'dataset'")
if key not in SUPPORT_TOOLS:
raise ValueError("Keys in agent_mode.tools must be in the specified tool list")
tool_item = tool[key]
@ -293,19 +333,7 @@ class AppModelConfigService:
if not isinstance(tool_item["enabled"], bool):
raise ValueError("enabled in agent_mode.tools must be of boolean type")
if key == "sensitive-word-avoidance":
if "words" not in tool_item or not tool_item["words"]:
tool_item["words"] = ""
if not isinstance(tool_item["words"], str):
raise ValueError("words in sensitive-word-avoidance must be of string type")
if "canned_response" not in tool_item or not tool_item["canned_response"]:
tool_item["canned_response"] = ""
if not isinstance(tool_item["canned_response"], str):
raise ValueError("canned_response in sensitive-word-avoidance must be of string type")
elif key == "dataset":
if key == "dataset":
if 'id' not in tool_item:
raise ValueError("id is required in dataset")
@ -324,6 +352,7 @@ class AppModelConfigService:
"suggested_questions_after_answer": config["suggested_questions_after_answer"],
"speech_to_text": config["speech_to_text"],
"more_like_this": config["more_like_this"],
"sensitive_word_avoidance": config["sensitive_word_avoidance"],
"model": {
"provider": config["model"]["provider"],
"name": config["model"]["name"],

28
api/services/completion_service.py
View file

@ -37,6 +37,8 @@ class CompletionService:
if not query:
raise ValueError('query is required')
query = query.replace('\x00', '')
conversation_id = args['conversation_id'] if 'conversation_id' in args else None
conversation = None
@ -140,6 +142,7 @@ class CompletionService:
suggested_questions=json.dumps(model_config['suggested_questions']),
suggested_questions_after_answer=json.dumps(model_config['suggested_questions_after_answer']),
more_like_this=json.dumps(model_config['more_like_this']),
sensitive_word_avoidance=json.dumps(model_config['sensitive_word_avoidance']),
model=json.dumps(model_config['model']),
user_input_form=json.dumps(model_config['user_input_form']),
pre_prompt=model_config['pre_prompt'],
@ -171,7 +174,7 @@ class CompletionService:
generate_worker_thread.start()
# wait for 5 minutes to close the thread
# wait for 10 minutes to close the thread
cls.countdown_and_close(generate_worker_thread, pubsub, user, generate_task_id)
return cls.compact_response(pubsub, streaming)
@ -179,9 +182,9 @@ class CompletionService:
@classmethod
def get_real_user_instead_of_proxy_obj(cls, user: Union[Account, EndUser]):
if isinstance(user, Account):
user = db.session.query(Account).get(user.id)
user = db.session.query(Account).filter(Account.id == user.id).first()
elif isinstance(user, EndUser):
user = db.session.query(EndUser).get(user.id)
user = db.session.query(EndUser).filter(EndUser.id == user.id).first()
else:
raise Exception("Unknown user type")
@ -226,12 +229,15 @@ class CompletionService:
@classmethod
def countdown_and_close(cls, worker_thread, pubsub, user, generate_task_id) -> threading.Thread:
# wait for 5 minutes to close the thread
timeout = 300
# wait for 10 minutes to close the thread
timeout = 600
def close_pubsub():
sleep_iterations = 0
while sleep_iterations < timeout and worker_thread.is_alive():
if sleep_iterations > 0 and sleep_iterations % 10 == 0:
PubHandler.ping(user, generate_task_id)
time.sleep(1)
sleep_iterations += 1
@ -369,7 +375,7 @@ class CompletionService:
if len(value) > max_length:
raise ValueError(f'{variable} in input form must be less than {max_length} characters')
filtered_inputs[variable] = value
filtered_inputs[variable] = value.replace('\x00', '') if value else None
return filtered_inputs
@ -418,6 +424,10 @@ class CompletionService:
yield "data: " + json.dumps(cls.get_chain_response_data(result.get('data'))) + "\n\n"
elif event == 'agent_thought':
yield "data: " + json.dumps(cls.get_agent_thought_response_data(result.get('data'))) + "\n\n"
elif event == 'ping':
yield "event: ping\n\n"
else:
yield "data: " + json.dumps(result) + "\n\n"
except ValueError as e:
if e.args[0] != "I/O operation on closed file.": # ignore this error
logging.exception(e)
@ -467,16 +477,14 @@ class CompletionService:
def get_agent_thought_response_data(cls, data: dict):
response_data = {
'event': 'agent_thought',
'id': data.get('agent_thought_id'),
'id': data.get('id'),
'chain_id': data.get('chain_id'),
'task_id': data.get('task_id'),
'message_id': data.get('message_id'),
'position': data.get('position'),
'thought': data.get('thought'),
'tool': data.get('tool'), # todo use real dataset obj replace it
'tool': data.get('tool'),
'tool_input': data.get('tool_input'),
'observation': data.get('observation'),
'answer': data.get('answer') if not data.get('thought') else '',
'created_at': int(time.time())
}