Skip to content

LangChain and LangSmith

Using Mirascope BasePrompt with LangChain

You may also want to use LangChain given it’s tight integration with LangSmith. For us, one issue we had when we first started using LangChain was that their invoke function had no type-safety or lint help. This means that calling invoke({"foox": "foo"}) was a difficult bug to catch. There’s so much functionality in LangChain, and we wanted to make using it more pleasant.

With Mirascope prompts, you can instantiate a ChatPromptTemplate from a Mirascope prompt template, and you can use the prompt’s model_dump method so you don’t have to worry about the invocation dictionary:

import os

from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI

from mirascope import BasePrompt

os.environ["OPENAI_API_KEY"] = "YOUR_API_KEY"

class JokePrompt(BasePrompt):
    prompt_template = "Tell me a short joke about {topic}"

    topic: str

joke_prompt = JokePrompt(topic="ice cream")
prompt = ChatPromptTemplate.from_template(joke_prompt.template())
# ^ instead of:
# prompt = ChatPromptTemplate.from_template("tell me a short joke about {topic}")

model = ChatOpenAI(model="gpt-4")
output_parser = StrOutputParser()
chain = prompt | model | output_parser

joke = chain.invoke(joke_prompt.model_dump())
# ^ instead of:
# joke = chain.invoke({"topic": "ice cream"})

Logging a LangSmith trace

You can use client wrappers (as mentioned in client wrappers) to integrate Mirascope with LangSmith. When using a wrapper, you can generate content as you would with a normal OpenAICall:

import os
from langsmith import wrappers
from mirascope.base import BaseConfig
from mirascope.openai import OpenAICall
os.environ["LANGCHAIN_TRACING_V2"] = "true"

class BookRecommender(OpenAICall):
    prompt_template = "Can you recommend some books on {topic}?"

    topic: str

    configuration = BaseConfig(client_wrappers[wrappers.wrap_openai])

response = BookRecommender(topic="sci-fi").call()

Now, if you log into LangSmith , you will be see your results have been traced. Of course, this integration works not just for call, but also for stream and extract.