hz/ebiz-dify-ai
1
0
Fork 0
mirror of http://112.124.100.131/huang.ze/ebiz-dify-ai.git synced 2025-12-09 19:06:51 +08:00
Code Issues Packages Projects Releases Wiki Activity

feat: advanced prompt backend (#1301)

Browse Source 
Co-authored-by: takatost <takatost@gmail.com>
This commit is contained in:
Garfield Dai
2023-10-12 23:13:10 +08:00
committed by GitHub
parent 2d1cb076c6
commit 42a5b3ec17
61 changed files with 767 additions and 581 deletions
Download Patch File Download Diff File Expand all files Collapse all files

79
api/core/prompt/advanced_prompt_templates.py Normal file
View File

@@ -0,0 +1,79 @@
CONTEXT = "Use the following context as your learned knowledge, inside <context></context> XML tags.\n\n<context>\n{{#context#}}\n</context>\n\nWhen answer to user:\n- If you don't know, just say that you don't know.\n- If you don't know when you are not sure, ask for clarification.\nAvoid mentioning that you obtained the information from the context.\nAnd answer according to the language of the user's question.\n"
BAICHUAN_CONTEXT = "用户在与一个客观的助手对话。助手会尊重找到的材料，给出全面专业的解释，但不会过度演绎。同时回答中不会暴露引用的材料：\n\n```\n{{#context#}}\n```\n\n"
CHAT_APP_COMPLETION_PROMPT_CONFIG = {
"completion_prompt_config": {
"prompt": {
"text": "{{#pre_prompt#}}\nHere is the chat histories between human and assistant, inside <histories></histories> XML tags.\n\n<histories>\n{{#histories#}}\n</histories>\n\n\nHuman: {{#query#}}\n\nAssistant: "
},
"conversation_histories_role": {
"user_prefix": "Human",
"assistant_prefix": "Assistant"
}
}
}
CHAT_APP_CHAT_PROMPT_CONFIG = {
"chat_prompt_config": {
"prompt": [{
"role": "system",
"text": "{{#pre_prompt#}}"
}]
}
}
COMPLETION_APP_CHAT_PROMPT_CONFIG = {
"chat_prompt_config": {
"prompt": [{
"role": "user",
"text": "{{#pre_prompt#}}"
}]
}
}
COMPLETION_APP_COMPLETION_PROMPT_CONFIG = {
"completion_prompt_config": {
"prompt": {
"text": "{{#pre_prompt#}}"
}
}
}
BAICHUAN_CHAT_APP_COMPLETION_PROMPT_CONFIG = {
"completion_prompt_config": {
"prompt": {
"text": "{{#pre_prompt#}}\n\n用户和助手的历史对话内容如下：\n```\n{{#histories#}}\n```\n\n\n\n用户：{{#query#}}"
},
"conversation_histories_role": {
"user_prefix": "用户",
"assistant_prefix": "助手"
}
}
}
BAICHUAN_CHAT_APP_CHAT_PROMPT_CONFIG = {
"chat_prompt_config": {
"prompt": [{
"role": "system",
"text": "{{#pre_prompt#}}"
}]
}
}
BAICHUAN_COMPLETION_APP_CHAT_PROMPT_CONFIG = {
"chat_prompt_config": {
"prompt": [{
"role": "user",
"text": "{{#pre_prompt#}}"
}]
}
}
BAICHUAN_COMPLETION_APP_COMPLETION_PROMPT_CONFIG = {
"completion_prompt_config": {
"prompt": {
"text": "{{#pre_prompt#}}"
}
}
}

38
api/core/prompt/prompt_builder.py
View File

@@ -1,38 +1,24 @@
import re
from langchain.schema import BaseMessage, SystemMessage, AIMessage, HumanMessage
from langchain.prompts import SystemMessagePromptTemplate, HumanMessagePromptTemplate, AIMessagePromptTemplate
from langchain.schema import BaseMessage
from core.prompt.prompt_template import JinjaPromptTemplate
from core.prompt.prompt_template import PromptTemplateParser
class PromptBuilder:
@classmethod
def parse_prompt(cls, prompt: str, inputs: dict) -> str:
prompt_template = PromptTemplateParser(prompt)
prompt_inputs = {k: inputs[k] for k in prompt_template.variable_keys if k in inputs}
prompt = prompt_template.format(prompt_inputs)
return prompt
@classmethod
def to_system_message(cls, prompt_content: str, inputs: dict) -> BaseMessage:
prompt_template = JinjaPromptTemplate.from_template(prompt_content)
system_prompt_template = SystemMessagePromptTemplate(prompt=prompt_template)
prompt_inputs = {k: inputs[k] for k in system_prompt_template.input_variables if k in inputs}
system_message = system_prompt_template.format(**prompt_inputs)
return system_message
return SystemMessage(content=cls.parse_prompt(prompt_content, inputs))
@classmethod
def to_ai_message(cls, prompt_content: str, inputs: dict) -> BaseMessage:
prompt_template = JinjaPromptTemplate.from_template(prompt_content)
ai_prompt_template = AIMessagePromptTemplate(prompt=prompt_template)
prompt_inputs = {k: inputs[k] for k in ai_prompt_template.input_variables if k in inputs}
ai_message = ai_prompt_template.format(**prompt_inputs)
return ai_message
return AIMessage(content=cls.parse_prompt(prompt_content, inputs))
@classmethod
def to_human_message(cls, prompt_content: str, inputs: dict) -> BaseMessage:
prompt_template = JinjaPromptTemplate.from_template(prompt_content)
human_prompt_template = HumanMessagePromptTemplate(prompt=prompt_template)
human_message = human_prompt_template.format(**inputs)
return human_message
@classmethod
def process_template(cls, template: str):
processed_template = re.sub(r'\{{2}(.+)\}{2}', r'{\1}', template)
# processed_template = re.sub(r'\{([a-zA-Z_]\w+?)\}', r'\1', template)
# processed_template = re.sub(r'\{\{([a-zA-Z_]\w+?)\}\}', r'{\1}', processed_template)
return processed_template
return HumanMessage(content=cls.parse_prompt(prompt_content, inputs))

106
api/core/prompt/prompt_template.py
View File

@@ -1,79 +1,39 @@
import re
from typing import Any
from jinja2 import Environment, meta
from langchain import PromptTemplate
from langchain.formatting import StrictFormatter
REGEX = re.compile(r"\{\{([a-zA-Z_][a-zA-Z0-9_]{1,29}|#histories#|#query#|#context#)\}\}")
class JinjaPromptTemplate(PromptTemplate):
template_format: str = "jinja2"
"""The format of the prompt template. Options are: 'f-string', 'jinja2'."""
class PromptTemplateParser:
"""
Rules:
1. Template variables must be enclosed in `{{}}`.
2. The template variable Key can only be: letters + numbers + underscore, with a maximum length of 16 characters,
and can only start with letters and underscores.
3. The template variable Key cannot contain new lines or spaces, and must comply with rule 2.
4. In addition to the above, 3 types of special template variable Keys are accepted:
`{{#histories#}}` `{{#query#}}` `{{#context#}}`. No other `{{##}}` template variables are allowed.
"""
def __init__(self, template: str):
self.template = template
self.variable_keys = self.extract()
def extract(self) -> list:
# Regular expression to match the template rules
return re.findall(REGEX, self.template)
def format(self, inputs: dict, remove_template_variables: bool = True) -> str:
def replacer(match):
key = match.group(1)
value = inputs.get(key, match.group(0)) # return original matched string if key not found
if remove_template_variables:
return PromptTemplateParser.remove_template_variables(value)
return value
return re.sub(REGEX, replacer, self.template)
@classmethod
def from_template(cls, template: str, **kwargs: Any) -> PromptTemplate:
"""Load a prompt template from a template."""
env = Environment()
template = template.replace("{{}}", "{}")
ast = env.parse(template)
input_variables = meta.find_undeclared_variables(ast)
if "partial_variables" in kwargs:
partial_variables = kwargs["partial_variables"]
input_variables = {
var for var in input_variables if var not in partial_variables
}
return cls(
input_variables=list(sorted(input_variables)), template=template, **kwargs
)
class OutLinePromptTemplate(PromptTemplate):
@classmethod
def from_template(cls, template: str, **kwargs: Any) -> PromptTemplate:
"""Load a prompt template from a template."""
input_variables = {
v for _, v, _, _ in OneLineFormatter().parse(template) if v is not None
}
return cls(
input_variables=list(sorted(input_variables)), template=template, **kwargs
)
def format(self, **kwargs: Any) -> str:
"""Format the prompt with the inputs.
Args:
kwargs: Any arguments to be passed to the prompt template.
Returns:
A formatted string.
Example:
.. code-block:: python
prompt.format(variable1="foo")
"""
kwargs = self._merge_partial_and_user_variables(**kwargs)
return OneLineFormatter().format(self.template, **kwargs)
class OneLineFormatter(StrictFormatter):
def parse(self, format_string):
last_end = 0
results = []
for match in re.finditer(r"{([a-zA-Z_]\w*)}", format_string):
field_name = match.group(1)
start, end = match.span()
literal_text = format_string[last_end:start]
last_end = end
results.append((literal_text, field_name, '', None))
remaining_literal_text = format_string[last_end:]
if remaining_literal_text:
results.append((remaining_literal_text, None, None, None))
return results
def remove_template_variables(cls, text: str):
return re.sub(REGEX, r'{\1}', text)

34
api/core/prompt/prompts.py
View File

@@ -61,36 +61,6 @@ User Input: yo, 你今天咋样？
User Input:
"""
CONVERSATION_SUMMARY_PROMPT = (
"Please generate a short summary of the following conversation.\n"
"If the following conversation communicating in English, you should only return an English summary.\n"
"If the following conversation communicating in Chinese, you should only return a Chinese summary.\n"
"[Conversation Start]\n"
"{context}\n"
"[Conversation End]\n\n"
"summary:"
)
INTRODUCTION_GENERATE_PROMPT = (
"I am designing a product for users to interact with an AI through dialogue. "
"The Prompt given to the AI before the conversation is:\n\n"
"```\n{prompt}\n```\n\n"
"Please generate a brief introduction of no more than 50 words that greets the user, based on this Prompt. "
"Do not reveal the developer's motivation or deep logic behind the Prompt, "
"but focus on building a relationship with the user:\n"
)
MORE_LIKE_THIS_GENERATE_PROMPT = (
"-----\n"
"{original_completion}\n"
"-----\n\n"
"Please use the above content as a sample for generating the result, "
"and include key information points related to the original sample in the result. "
"Try to rephrase this information in different ways and predict according to the rules below.\n\n"
"-----\n"
"{prompt}\n"
)
SUGGESTED_QUESTIONS_AFTER_ANSWER_INSTRUCTION_PROMPT = (
"Please help me predict the three most likely questions that human would ask, "
"and keeping each question under 20 characters.\n"
@@ -157,10 +127,10 @@ and fill in variables, with a welcome sentence, and keep TLDR.
```
<< MY INTENDED AUDIENCES >>
{audiences}
{{audiences}}
<< HOPING TO SOLVE >>
{hoping_to_solve}
{{hoping_to_solve}}
<< OUTPUT >>
"""