lanchchain decorators is a layer on the top op LangChain that provides syntactic sugar 🍭 for writing custom langchain prompts and chains
Note: This is an unofficial addon to langchain library. It's not trying to compete, just to make using it easier. Lot's of ideas here are totally opinionated
Main principles and benefits:
- more
pythonicway of writing code - write multiline prompts that wont break your code flow with indentation
- making use of IDE in-built support for hinting, type checking and popup with docs to quickly peek in the function to see the prompt, parameters it consumes etc.
- leverage all the power of 🦜🔗 LangChain ecosystem
- adding support for optional parameters
- easily share parameters between the prompts by binding them to one class
Here is a simple example of a code written with LangChain Decorators ✨
@llm_prompt
def write_me_short_post(topic:str, platform:str="twitter", audience:str = "developers")->str:
"""
Write me a short header for my post about {topic} for {platform} platform.
It should be for {audience} audience.
(Max 15 words)
"""
return
# run it naturaly
write_me_short_post(topic="starwars")
# or
write_me_short_post(topic="starwars", platform="redit")pip install langchain_decoratorsGood idea on how to start is to review the examples here:
By default the prompt is is the whole function docs, unless you mark your prompt
We can specify what part of our docs is the prompt definition, by specifying a code block with language tag
@llm_prompt
def write_me_short_post(topic:str, platform:str="twitter", audience:str = "developers"):
"""
Here is a good way to write a prompt as part of a function docstring, with additional documentation for devs.
It needs to be a code block, marked as a `<prompt>` language
```<prompt>
Write me a short header for my post about {topic} for {platform} platform.
It should be for {audience} audience.
(Max 15 words)
```
Now only to code block above will be used as a prompt, and the rest of the docstring will be used as a description for developers.
(It has also a nice benefit that IDE (like VS code) will display the prompt properly (not trying to parse it as markdown, and thus not showing new lines properly))
"""
return For chat models is very useful to define prompt as a set of message templates... here is how to do it:
@llm_prompt
def simulate_conversation(human_input:str, agent_role:str="a pirate"):
"""
## System message
- note the `:system` sufix inside the <prompt:_role_> tag
```<prompt:system>
You are a {agent_role} hacker. You mus act like one.
You reply always in code, using python or javascript code block...
for example:
... do not reply with anything else.. just with code - respecting your role.
```
# human message
(we are using the real role that are enforced by the LLM - GPT supports system, assistant, user)
``` <prompt:user>
Helo, who are you
```
a reply:
``` <prompt:assistant>
\``` python <<- escaping inner code block with \ that should be part of the prompt
def hello():
print("Argh... hello you pesky pirate")
\```
```
we can also add some history using placeholder
```<prompt:placeholder>
{history}
```
```<prompt:user>
{human_input}
```
Now only to code block above will be used as a prompt, and the rest of the docstring will be used as a description for developers.
(It has also a nice benefit that IDE (like VS code) will display the prompt properly (not trying to parse it as markdown, and thus not showing new lines properly))
"""
passthe roles here are model native roles (assistant, user, system for chatGPT)
- you can define a whole sections of your prompt that should be optional
- if any input in the section is missing, the whole section wont be rendered
the syntax for this is as follows:
@llm_prompt
def prompt_with_optional_partials():
"""
this text will be rendered always, but
{? anything inside this block will be rendered only if all the {value}s parameters are not empty (None | "") ?}
you can also place it in between the words
this too will be rendered{? , but
this block will be rendered only if {this_value} and {this_value}
is not empty?} !
"""- llm_prompt decorator natively tries to detect the best output parser based on the output type. (if not set, it returns the raw string)
- list, dict and pydantic outputs are also supported natively (automaticaly)
# this code example is complete and should run as it is
from langchain_decorators import llm_prompt
@llm_prompt
def write_name_suggestions(company_business:str, count:int)->list:
""" Write me {count} good name suggestions for company that {company_business}
"""
pass
write_name_suggestions(company_business="sells cookies", count=5)-
currently supported only for the latest OpenAI chat models
-
all you need to do is annotate your function with @llm_function decorator.
-
This will parse the description for LLM (first coherent paragraph is considered as function description)
-
and aso parameter descriptions (Google, Numpy and Spihnx notations are supported for now)
-
by default the docstring format is automatically resolved, but setting the format of the docstring can speed things up a bit. -
auto(default): the format is automatically inferred from the docstring -google: the docstring is parsed as markdown (see Google docstring format) -numpy: the docstring is parsed as markdown (see Numpy docstring format) -sphinx: the docstring is parsed as sphinx format (see Sphinx docstring format)
To annotate an "enum" like argument, you can use this "typescript" like format: ["value_a" | "value_b"] ... if will be parsed out.
This text will be a part of a description too... if you dont want it, you can use this notation as a type notation.
Example:
Args:
message_type (["email" | "sms"]): type of a message / channel how to send the message
Then you pass these functions as arguments to and @llm_prompt (the argument must be named functions
here is how to use it:
from langchain.agents import load_tools
from langchian_decorators import llm_function, llm_prompt, GlobalSettings
@llm_function
def send_message(message:str, addressee:str=None, message_type:str="email"):
""" Use this if user asks to send some message
Args:
message (str): message text to send
addressee (str): email of the adressese... in format firstName.lastName@company.com
message_type (str, optional): enum: ["email"|"whatsapp"]
"""
if message_type=="email":
send_email(addressee, message)
elif message_type=="whatsapp":
send_whatsapp(addressee, message)
# load some other tools from langchain
list_of_other_tools = load_tools(
tool_names=[...],
llm=GlobalSettings.get_current_settings().default_llm)
@llm_prompt
def do_what_user_asks_for(user_input:str, functions:List[Union[Callable,BaseTool]]):
"""
``` <prompt:system>
Your role is to be a helpfull asistant.
```
``` <prompt:user>
{user_input}
```
"""
user_input="Yo, send an email to John Smith that I will be late for the meeting"
result = do_what_user_asks_for(
user_input=user_input,
functions=[send_message, *list_of_other_tools]
)
if result.is_function_call:
result.execute()
else:
print(result.output_text)If you use functions argument, the output will be always OutputWithFunctionCall
class OutputWithFunctionCall(BaseModel):
output_text:str
output:T
function_name:str =None
function_arguments:Union[Dict[str,Any],str,None]
function:Callable = None
function_async:Callable = None
@property
def is_function_call(self):
...
@property
def support_async(self):
...
@property
def support_sync(self):
...
async def execute_async(self):
"""Executes the function asynchronously."""
...
def execute(self):
""" Executes the function synchronously.
If the function is async, it will be executed in a event loop.
"""
...If you want to see how to schema has been build, you can use get_function_schema method that is added to the function by the decorator:
@llm_function
def my_func(arg1:str):
...
print(my_func.get_function_schema())Here we are just marking a function as a prompt with llm_prompt decorator, turning it effectively into a LLMChain. Instead of running it
Standard LLMchain takes much more init parameter than just inputs_variables and prompt... here is this implementation detail hidden in the decorator. Here is how it works:
- Using Global settings:
# define global settings for all prompty (if not set - chatGPT is the current default)
from langchain_decorators import GlobalSettings
GlobalSettings.define_settings(
default_llm=ChatOpenAI(temperature=0.0), this is default... can change it here globally
default_streaming_llm=ChatOpenAI(temperature=0.0,streaming=True), this is default... can change it here for all ... will be used for streaming
)- Using predefined prompt types
#You can change the default prompt types
from langchain_decorators import PromptTypes, PromptTypeSettings
PromptTypes.AGENT_REASONING.llm = ChatOpenAI()
# Or you can just define your own ones:
class MyCustomPromptTypes(PromptTypes):
GPT4=PromptTypeSettings(llm=ChatOpenAI(model="gpt-4"))
@llm_prompt(prompt_type=MyCustomPromptTypes.GPT4)
def write_a_complicated_code(app_idea:str)->str:
...- Define the settings directly in the decorator
from langchain.llms import OpenAI
@llm_prompt(
llm=OpenAI(temperature=0.7),
stop_tokens=["\nObservation"],
...
)
def creative_writer(book_title:str)->str:
...To pass any of these, just declare them in the function (or use kwargs to pass anything)
@llm_prompt()
async def write_me_short_post(topic:str, platform:str="twitter", memory:SimpleMemory = None):
"""
{history_key}
Write me a short header for my post about {topic} for {platform} platform.
It should be for {audience} audience.
(Max 15 words)
"""
pass
await write_me_short_post(topic="old movies")If we wan't to leverage streaming:
- we need to define prompt as async function
- turn on the streaming on the decorator, or we can define PromptType with streaming on
- capture the stream using StreamingContext
This way we just mark which prompt should be streamed, not needing to tinker with what LLM should we use, passing around the creating and distribute streaming handler into particular part of our chain... just turn the streaming on/off on prompt/prompt type...
The streaming will happen only if we call it in streaming context ... there we can define a simple function to handle the stream
# this code example is complete and should run as it is
from langchain_decorators import StreamingContext, llm_prompt
# this will mark the prompt for streaming (useful if we want stream just some prompts in our app... but don't want to pass distribute the callback handlers)
# note that only async functions can be streamed (will get an error if it's not)
@llm_prompt(capture_stream=True)
async def write_me_short_post(topic:str, platform:str="twitter", audience:str = "developers"):
"""
Write me a short header for my post about {topic} for {platform} platform.
It should be for {audience} audience.
(Max 15 words)
"""
pass
# just an arbitrary function to demonstrate the streaming... wil be some websockets code in the real world
tokens=[]
def capture_stream_func(new_token:str):
tokens.append(new_token)
# if we want to capture the stream, we need to wrap the execution into StreamingContext...
# this will allow us to capture the stream even if the prompt call is hidden inside higher level method
# only the prompts marked with capture_stream will be captured here
with StreamingContext(stream_to_stdout=True, callback=capture_stream_func):
result = await run_prompt()
print("Stream finished ... we can distinguish tokens thanks to alternating colors")
print("\nWe've captured",len(tokens),"tokens🎉\n")
print("Here is the result:")
print(result)for dict / pydantic you need to specify the formatting instructions... this can be tedious, that's why you can let the output parser generate you the instructions based on the model (pydantic)
from langchain_decorators import llm_prompt
from pydantic import BaseModel, Field
class TheOutputStructureWeExpect(BaseModel):
name:str = Field (description="The name of the company")
headline:str = Field( description="The description of the company (for landing page)")
employees:list[str] = Field(description="5-8 fake employee names with their positions")
@llm_prompt()
def fake_company_generator(company_business:str)->TheOutputStructureWeExpect:
""" Generate a fake company that {company_business}
{FORMAT_INSTRUCTIONS}
"""
return
company = fake_company_generator(company_business="sells cookies")
# print the result nicely formatted
print("Company name: ",company.name)
print("company headline: ",company.headline)
print("company employees: ",company.employees)from pydantic import BaseModel
from langchain_decorators import llm_prompt
class AssistantPersonality(BaseModel):
assistant_name:str
assistant_role:str
field:str
@property
def a_property(self):
return "whatever"
def hello_world(self, function_kwarg:str=None):
"""
We can reference any {field} or {a_property} inside our prompt... and combine it with {function_kwarg} in the method
"""
@llm_prompt
def introduce_your_self(self)->str:
"""
``` <prompt:system>
You are an assistant named {assistant_name}.
Your role is to act as {assistant_role}
```
```<prompt:user>
Introduce your self (in less than 20 words)
```
"""
personality = AssistantPersonality(assistant_name="John", assistant_role="a pirate")
print(personality.introduce_your_self(personality))- these and few more examples are also available in the examples notebook here
- including the ReAct Agent re-implementation using purely langchain decorators
feedback, contributions and PR are welcomed 🙏
- this project is dependant on langchain (obviously)
- as well as on promptwatch promptwatch, which make it easy to track and store to logs, track changes in prompts and compare them by running unit tests over the prompts...