[WIP] LangChain Agent

[Future-Outlier]

Tracking issue

https://github.com/flyteorg/flyte/issues/3936

Example

import os
from typing import Any, Union

from flytekit import workflow
from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts.prompt import PromptTemplate
from langchain_openai import ChatOpenAI

api_key = os.environ.get("OPENAI_API_KEY")

model = ChatOpenAI(model="gpt-3.5-turbo", 
                   openai_api_key=api_key, 
                   openai_organization="org-NayNG68kGnVXMJ8Ak4PMgQv7",)

prompt = PromptTemplate(
    input_variables=[
        "question",
    ],
    template="Question: {question}?",
)

output_parser = StrOutputParser()

@workflow
def wf(input: str) -> Union[str, Any]:
    message = prompt(input=input)
    o0 = model(input=message)
    o1 = output_parser(input=o0)
    return o1

Screen Shots

local execution remote execution

Details for discussion and review

LangChain LCEL

https://hackmd.io/@Future-Outlier/HykgW1P70

LangChain Agent Roadmap

https://hackmd.io/J3erYppETQe0OigP74-uJw?view

Potential Problems

We have to try to support cases of Any -> str when the input is str. Since we want to use StrOutputParser as the output.

[Future-Outlier]

how to setup langchain dev mode for debugger trace code

reference: https://python.langchain.com/docs/contributing/code

0. install langchain from github

git clone https://github.com/langchain-ai/langchain.git

1. cd langchain/libs

cd libs/community
poetry install --with lint,typing,test,test_integration

cd libs/core
poetry install --with test

cd libs/experimental
poetry install --with test

cd libs/partners
poetry add --optional openai

3. check your dependency

pip list | grep -i lang

langchain                                0.1.5                   /Users/future-outlier/code/langchain/libs/langchain
langchain-community                      0.0.19                  /Users/future-outlier/code/langchain/libs/community
langchain-core                           0.1.21                  /Users/future-outlier/code/langchain/libs/core
langchain-experimental                   0.0.50                  /Users/future-outlier/code/langchain/libs/experimental
langchain-openai                         0.0.5                   /Users/future-outlier/code/langchain/libs/partners/openai

[codecov[bot]]

Codecov Report

All modified and coverable lines are covered by tests :white_check_mark:

Project coverage is 83.48%. Comparing base (fa2aa0b) to head (0980ad0). Report is 3 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2191      +/-   ##
==========================================
- Coverage   83.89%   83.48%   -0.42%     
==========================================
  Files         342      324      -18     
  Lines       25483    24716     -767     
  Branches     3725     3516     -209     
==========================================
- Hits        21380    20634     -746     
+ Misses       3472     3450      -22     
- Partials      631      632       +1     

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.

[Future-Outlier]

Current thought:

1. ChatGPT-like implementation in LangChain is definitely doable and might be not that hard.
2. But 1 needs to support too much implementation
3. It is better to find a way to support Airflow-like implementation

[Future-Outlier]

Current Test Example

from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI
import os
from flytekit import task, workflow

# before
prompt = ChatPromptTemplate.from_template("{topic}")
# print(prompt)
# after
# prompt = ChatPromptTemplate(task_id="tmp")
# template = prompt.from_template("{topic}")

"""
<class 'tuple'> <class 'dict'>
args: (<class 'langchain_core.prompts.prompt.PromptTemplate'>,)
args len: 1
kwargs: {'input_variables': ['topic'], 'template': '{topic}', 'template_format': 'f-string', 'partial_variables': {}}
kwargs len: 4
"""
# model = ChatOpenAI(model="gpt-3.5-turbo", openai_api_key=os.environ.get('OPENAI_API_KEY'))
"""
<class 'tuple'> <class 'dict'>
args: (<class 'langchain_openai.chat_models.base.ChatOpenAI'>,)
args len: 1
kwargs: {'model': 'gpt-3.5-turbo', 'openai_api_key': 'xxx'}
kwargs len: 2
"""
# output_parser = StrOutputParser()
"""
<class 'tuple'> <class 'dict'>
args: (<class 'langchain_core.output_parsers.string.StrOutputParser'>,)
args len: 1
kwargs: {}
kwargs len: 
"""

@workflow
def wf(input: str) -> str:
    p = prompt(input=input)
    model = ChatOpenAI(model="gpt-3.5-turbo", openai_api_key=os.environ.get('OPENAI_API_KEY'))
    output = model(input=input)
    return output

if __name__ == "__main__":
    print(wf(input="hi"))

[Future-Outlier]

Have made some notes on Heptabase, if someone can help, I can send you my notes, thank you.

[Future-Outlier]

Currnet Work Version

from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI
import os
from typing import Any
from flytekit import workflow

api_key = os.environ.get('OPENAI_API_KEY')
model = ChatOpenAI(model="gpt-3.5-turbo", openai_api_key=api_key)
output_parser = StrOutputParser()

@workflow
def wf() -> str:
    o = model(input="hello")
    return output_parser(input=o)

if __name__ == '__main__':
    print(wf())

[Future-Outlier]

Update: maybe it will be doable if we monkey path the class, but not the method.

https://stackoverflow.com/questions/3765222/monkey-patch-python-class