Replicate

This example goes over how to use LangChain to interact with Replicate models

import os
from langchain.llms import Replicate
from langchain import PromptTemplate, LLMChain

os.environ["REPLICATE_API_TOKEN"] = "YOUR REPLICATE API TOKEN"

Setup

To run this notebook, you’ll need to create a replicate account and install the replicate python client.

Calling a model

Find a model on the replicate explore page, and then paste in the model name and version in this format: model_name/version

For example, for this flan-t5 model, click on the API tab. The model name/version would be: daanelson/flan-t5:04e422a9b85baed86a4f24981d7f9953e20c5fd82f6103b74ebc431588e1cec8

Only the model param is required, but we can add other model params when initializing.

For example, if we were running stable diffusion and wanted to change the image dimensions:

Replicate(model="stability-ai/stable-diffusion:db21e45d3f7023abc2a46ee38a23973f6dce16bb082a930b0c49861f96d1e5bf", input={'image_dimensions': '512x512'})

Note that only the first output of a model will be returned.

llm = Replicate(model="daanelson/flan-t5:04e422a9b85baed86a4f24981d7f9953e20c5fd82f6103b74ebc431588e1cec8")

prompt = """
Answer the following yes/no question by reasoning step by step. 
Can a dog drive a car?
"""
llm(prompt)

'The legal driving age of dogs is 2. Cars are designed for humans to drive. Therefore, the final answer is yes.'

We can call any replicate model using this syntax. For example, we can call stable diffusion.

text2image = Replicate(model="stability-ai/stable-diffusion:db21e45d3f7023abc2a46ee38a23973f6dce16bb082a930b0c49861f96d1e5bf", 
                       input={'image_dimensions': '512x512'})

image_output = text2image("A cat riding a motorcycle by Picasso")
image_output

'https://replicate.delivery/pbxt/Cf07B1zqzFQLOSBQcKG7m9beE74wf7kuip5W9VxHJFembefKE/out-0.png'

The model spits out a URL. Let’s render it.

from PIL import Image
import requests
from io import BytesIO

response = requests.get(image_output)
img = Image.open(BytesIO(response.content))

img

Chaining Calls

The whole point of langchain is to… chain! Here’s an example of how do that.

from langchain.chains import SimpleSequentialChain

First, let’s define the LLM for this model as a flan-5, and text2image as a stable diffusion model.

llm = Replicate(model="daanelson/flan-t5:04e422a9b85baed86a4f24981d7f9953e20c5fd82f6103b74ebc431588e1cec8")
text2image = Replicate(model="stability-ai/stable-diffusion:db21e45d3f7023abc2a46ee38a23973f6dce16bb082a930b0c49861f96d1e5bf")

First prompt in the chain

prompt = PromptTemplate(
    input_variables=["product"],
    template="What is a good name for a company that makes {product}?",
)

chain = LLMChain(llm=llm, prompt=prompt)

Second prompt to get the logo for company description

second_prompt = PromptTemplate(
    input_variables=["company_name"],
    template="Write a description of a logo for this company: {company_name}",
)
chain_two = LLMChain(llm=llm, prompt=second_prompt)

Third prompt, let’s create the image based on the description output from prompt 2

third_prompt = PromptTemplate(
    input_variables=["company_logo_description"],
    template="{company_logo_description}",
)
chain_three = LLMChain(llm=text2image, prompt=third_prompt)

Now let’s run it!

# Run the chain specifying only the input variable for the first chain.
overall_chain = SimpleSequentialChain(chains=[chain, chain_two, chain_three], verbose=True)
catchphrase = overall_chain.run("colorful socks")
print(catchphrase)

> Entering new SimpleSequentialChain chain...
novelty socks
todd & co.
https://replicate.delivery/pbxt/BedAP1PPBwXFfkmeD7xDygXO4BcvApp1uvWOwUdHM4tcQfvCB/out-0.png

> Finished chain.
https://replicate.delivery/pbxt/BedAP1PPBwXFfkmeD7xDygXO4BcvApp1uvWOwUdHM4tcQfvCB/out-0.png

response = requests.get("https://replicate.delivery/pbxt/eq6foRJngThCAEBqse3nL3Km2MBfLnWQNd0Hy2SQRo2LuprCB/out-0.png")
img = Image.open(BytesIO(response.content))
img