youngkingdom/dify
1
0
Fork 0
mirror of https://github.com/langgenius/dify.git synced 2026-05-06 10:28:10 +08:00
Code Issues Packages Projects Releases Wiki Activity

feat: improve follow-up settings (#35442)

Browse Source 
Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>
This commit is contained in:
非法操作
2026-04-22 16:55:16 +08:00
committed by GitHub
parent 491061b8f4
commit afec528f51
42 changed files with 1086 additions and 349 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
api/core/app/app_config/features/suggested_questions_after_answer/manager.py
View File

@ -1,5 +1,7 @@
from typing import Any
CUSTOM_FOLLOW_UP_PROMPT_MAX_LENGTH = 1000
class SuggestedQuestionsAfterAnswerConfigManager:
@classmethod
@ -20,7 +22,11 @@ class SuggestedQuestionsAfterAnswerConfigManager:
@classmethod
def validate_and_set_defaults(cls, config: dict[str, Any]) -> tuple[dict[str, Any], list[str]]:
"""
Validate and set defaults for suggested questions feature
Validate and set defaults for suggested questions feature.
Optional fields:
- prompt: custom instruction prompt.
- model: provider/model configuration for suggested question generation.
:param config: app model config args
"""
@ -39,4 +45,27 @@ class SuggestedQuestionsAfterAnswerConfigManager:
if not isinstance(config["suggested_questions_after_answer"]["enabled"], bool):
raise ValueError("enabled in suggested_questions_after_answer must be of boolean type")
prompt = config["suggested_questions_after_answer"].get("prompt")
if prompt is not None and not isinstance(prompt, str):
raise ValueError("prompt in suggested_questions_after_answer must be of string type")
if isinstance(prompt, str) and len(prompt) > CUSTOM_FOLLOW_UP_PROMPT_MAX_LENGTH:
raise ValueError(
f"prompt in suggested_questions_after_answer must be less than or equal to "
f"{CUSTOM_FOLLOW_UP_PROMPT_MAX_LENGTH} characters"
)
if "model" in config["suggested_questions_after_answer"]:
model_config = config["suggested_questions_after_answer"]["model"]
if not isinstance(model_config, dict):
raise ValueError("model in suggested_questions_after_answer must be of object type")
if "provider" not in model_config or not isinstance(model_config["provider"], str):
raise ValueError("provider in suggested_questions_after_answer.model must be of string type")
if "name" not in model_config or not isinstance(model_config["name"], str):
raise ValueError("name in suggested_questions_after_answer.model must be of string type")
if "completion_params" in model_config and not isinstance(model_config["completion_params"], dict):
raise ValueError("completion_params in suggested_questions_after_answer.model must be of object type")
return config, ["suggested_questions_after_answer"]

103
api/core/llm_generator/llm_generator.py
View File

@ -2,7 +2,7 @@ import json
import logging
import re
from collections.abc import Sequence
from typing import Any, Protocol, TypedDict, cast
from typing import Any, NotRequired, Protocol, TypedDict, cast
import json_repair
from sqlalchemy import select
@ -13,13 +13,13 @@ from core.llm_generator.output_parser.rule_config_generator import RuleConfigGen
from core.llm_generator.output_parser.suggested_questions_after_answer import SuggestedQuestionsAfterAnswerOutputParser
from core.llm_generator.prompts import (
CONVERSATION_TITLE_PROMPT,
DEFAULT_SUGGESTED_QUESTIONS_MAX_TOKENS,
DEFAULT_SUGGESTED_QUESTIONS_TEMPERATURE,
GENERATOR_QA_PROMPT,
JAVASCRIPT_CODE_GENERATOR_PROMPT_TEMPLATE,
LLM_MODIFY_CODE_SYSTEM,
LLM_MODIFY_PROMPT_SYSTEM,
PYTHON_CODE_GENERATOR_PROMPT_TEMPLATE,
SUGGESTED_QUESTIONS_MAX_TOKENS,
SUGGESTED_QUESTIONS_TEMPERATURE,
SYSTEM_STRUCTURED_OUTPUT_GENERATE,
WORKFLOW_RULE_CONFIG_PROMPT_GENERATE_TEMPLATE,
)
@ -41,6 +41,36 @@ from models.workflow import Workflow
logger = logging.getLogger(__name__)
class SuggestedQuestionsModelConfig(TypedDict):
provider: str
name: str
completion_params: NotRequired[dict[str, object]]
def _normalize_completion_params(completion_params: dict[str, object]) -> tuple[dict[str, object], list[str]]:
"""
Normalize raw completion params into invocation parameters and stop sequences.
This mirrors the app-model access path by separating ``stop`` from provider
parameters before invocation, then drops non-positive token limits because
some plugin-backed models reject ``0`` after mapping ``max_tokens`` to their
provider-specific output-token field.
"""
normalized_parameters = dict(completion_params)
stop_value = normalized_parameters.pop("stop", [])
if isinstance(stop_value, list) and all(isinstance(item, str) for item in stop_value):
stop = stop_value
else:
stop = []
for token_limit_key in ("max_tokens", "max_output_tokens"):
token_limit = normalized_parameters.get(token_limit_key)
if isinstance(token_limit, int | float) and token_limit <= 0:
normalized_parameters.pop(token_limit_key, None)
return normalized_parameters, stop
class WorkflowServiceInterface(Protocol):
def get_draft_workflow(self, app_model: App, workflow_id: str | None = None) -> Workflow | None:
pass
@ -123,8 +153,15 @@ class LLMGenerator:
return name
@classmethod
def generate_suggested_questions_after_answer(cls, tenant_id: str, histories: str) -> Sequence[str]:
output_parser = SuggestedQuestionsAfterAnswerOutputParser()
def generate_suggested_questions_after_answer(
cls,
tenant_id: str,
histories: str,
*,
instruction_prompt: str | None = None,
model_config: object | None = None,
) -> Sequence[str]:
output_parser = SuggestedQuestionsAfterAnswerOutputParser(instruction_prompt=instruction_prompt)
format_instructions = output_parser.get_format_instructions()
prompt_template = PromptTemplateParser(template="{{histories}}\n{{format_instructions}}\nquestions:\n")
@ -133,10 +170,36 @@ class LLMGenerator:
try:
model_manager = ModelManager.for_tenant(tenant_id=tenant_id)
model_instance = model_manager.get_default_model_instance(
tenant_id=tenant_id,
model_type=ModelType.LLM,
)
configured_model = cast(dict[str, object], model_config) if isinstance(model_config, dict) else {}
provider = configured_model.get("provider")
model_name = configured_model.get("name")
use_configured_model = False
if isinstance(provider, str) and provider and isinstance(model_name, str) and model_name:
try:
model_instance = model_manager.get_model_instance(
tenant_id=tenant_id,
model_type=ModelType.LLM,
provider=provider,
model=model_name,
)
use_configured_model = True
except Exception:
logger.warning(
"Failed to use configured suggested-questions model %s/%s, fallback to default model",
provider,
model_name,
exc_info=True,
)
model_instance = model_manager.get_default_model_instance(
tenant_id=tenant_id,
model_type=ModelType.LLM,
)
else:
model_instance = model_manager.get_default_model_instance(
tenant_id=tenant_id,
model_type=ModelType.LLM,
)
except InvokeAuthorizationError:
return []
@ -145,19 +208,29 @@ class LLMGenerator:
questions: Sequence[str] = []
try:
configured_completion_params = configured_model.get("completion_params")
if use_configured_model and isinstance(configured_completion_params, dict):
model_parameters, stop = _normalize_completion_params(configured_completion_params)
elif use_configured_model:
model_parameters = {}
stop = []
else:
# Default-model generation keeps the built-in suggested-questions tuning.
model_parameters = {
"max_tokens": DEFAULT_SUGGESTED_QUESTIONS_MAX_TOKENS,
"temperature": DEFAULT_SUGGESTED_QUESTIONS_TEMPERATURE,
}
stop = []
response: LLMResult = model_instance.invoke_llm(
prompt_messages=list(prompt_messages),
model_parameters={
"max_tokens": SUGGESTED_QUESTIONS_MAX_TOKENS,
"temperature": SUGGESTED_QUESTIONS_TEMPERATURE,
},
model_parameters=model_parameters,
stop=stop,
stream=False,
)
text_content = response.message.get_text_content()
questions = output_parser.parse(text_content) if text_content else []
except InvokeError:
questions = []
except Exception:
logger.exception("Failed to generate suggested questions after answer")
questions = []

12
api/core/llm_generator/output_parser/suggested_questions_after_answer.py
View File

@ -3,17 +3,21 @@ import logging
import re
from collections.abc import Sequence
from core.llm_generator.prompts import SUGGESTED_QUESTIONS_AFTER_ANSWER_INSTRUCTION_PROMPT
from core.llm_generator.prompts import DEFAULT_SUGGESTED_QUESTIONS_AFTER_ANSWER_INSTRUCTION_PROMPT
logger = logging.getLogger(__name__)
class SuggestedQuestionsAfterAnswerOutputParser:
def __init__(self, instruction_prompt: str | None = None) -> None:
self._instruction_prompt = instruction_prompt or DEFAULT_SUGGESTED_QUESTIONS_AFTER_ANSWER_INSTRUCTION_PROMPT
def get_format_instructions(self) -> str:
return SUGGESTED_QUESTIONS_AFTER_ANSWER_INSTRUCTION_PROMPT
return self._instruction_prompt
def parse(self, text: str) -> Sequence[str]:
action_match = re.search(r"\[.*?\]", text.strip(), re.DOTALL)
stripped_text = text.strip()
action_match = re.search(r"\[.*?\]", stripped_text, re.DOTALL)
questions: list[str] = []
if action_match is not None:
try:
@ -23,4 +27,6 @@ class SuggestedQuestionsAfterAnswerOutputParser:
else:
if isinstance(json_obj, list):
questions = [question for question in json_obj if isinstance(question, str)]
elif stripped_text:
logger.warning("Failed to find suggested questions payload array in text: %r", stripped_text[:200])
return questions

15
api/core/llm_generator/prompts.py
View File

@ -1,5 +1,4 @@
# Written by YORKI MINAKO🤡, Edited by Xiaoyi, Edited by yasu-oh
import os
CONVERSATION_TITLE_PROMPT = """You are asked to generate a concise chat title by decomposing the user’s input into two parts: “Intention” and “Subject”.
@ -96,8 +95,8 @@ JAVASCRIPT_CODE_GENERATOR_PROMPT_TEMPLATE = (
)
# Default prompt for suggested questions (can be overridden by environment variable)
_DEFAULT_SUGGESTED_QUESTIONS_AFTER_ANSWER_PROMPT = (
# Default prompt and model parameters for suggested questions.
DEFAULT_SUGGESTED_QUESTIONS_AFTER_ANSWER_INSTRUCTION_PROMPT = (
"Please help me predict the three most likely questions that human would ask, "
"and keep each question under 20 characters.\n"
"MAKE SURE your output is the SAME language as the Assistant's latest response. "
@ -105,14 +104,8 @@ _DEFAULT_SUGGESTED_QUESTIONS_AFTER_ANSWER_PROMPT = (
'["question1","question2","question3"]\n'
)
# Environment variable override for suggested questions prompt
SUGGESTED_QUESTIONS_AFTER_ANSWER_INSTRUCTION_PROMPT = os.getenv(
"SUGGESTED_QUESTIONS_PROMPT", _DEFAULT_SUGGESTED_QUESTIONS_AFTER_ANSWER_PROMPT
)
# Configurable LLM parameters for suggested questions (can be overridden by environment variables)
SUGGESTED_QUESTIONS_MAX_TOKENS = int(os.getenv("SUGGESTED_QUESTIONS_MAX_TOKENS", "256"))
SUGGESTED_QUESTIONS_TEMPERATURE = float(os.getenv("SUGGESTED_QUESTIONS_TEMPERATURE", "0"))
DEFAULT_SUGGESTED_QUESTIONS_MAX_TOKENS = 256
DEFAULT_SUGGESTED_QUESTIONS_TEMPERATURE = 0.0
GENERATOR_QA_PROMPT = (
"<Task> The user will send a long text. Generate a Question and Answer pairs only using the knowledge"