Using Langchain and OpenAI functions to structure data

Working with LLMs is fun but sometimes too much free text is a pain to parse and process.

Fortunately, with Langchain and OpenAI functions we can structure data in a few lines of code.

Before OpenAI functions I was prompting stuff like “Please, give me the output in the following JSON structure” and praying to the gods that it would magically solve my problems.

Unfortunately I was being given replies with structured data prefixed with “Sure! Here’s your data:” or suffixed with “Hope this structure works for you. Enjoy!”.

Getting structured outputs of given information

We can use JsonSchema (Pydantic is awesome but it is still very buggy 🐛 for Langchain v0.0.281 at the time I wrote this article) to structure whatever we want. Let’s use the straightforward example from Langchain documentation to parse someone’s age:

from typing import Optional

from dotenv import load_dotenv
from langchain.chains.openai_functions import create_structured_output_chain
from langchain.chat_models import ChatOpenAI
from langchain.prompts import ChatPromptTemplate

load_dotenv()  # take environment variables from .env so we can load OPENAI_API_KEY

json_schema = {
    "title": "Person",
    "description": "Identifying information about a person.",
    "type": "object",
    "properties": {
        "name": {"title": "Name", "description": "The person's name", "type": "string"},
        "age": {"title": "Age", "description": "The person's age", "type": "integer"},
        "fav_food": {
            "title": "Fav Food",
            "description": "The person's favorite food",
            "type": "string",
        },
    },
    "required": ["name", "age"],
}

llm = ChatOpenAI(model="gpt-4", temperature=0)
prompt = ChatPromptTemplate.from_messages(
    [
        ("system", "You are extracting information in structured formats."),
        ("human", "Use the given format to extract information from the following input: {input}")
    ]
)

chain = create_structured_output_chain(json_schema, llm, prompt, verbose=True)
output = chain.run("DJ Quesadilla is 38 and loves pizza")

print(output)

Getting structured outputs of generated information

Sometimes we need to generate data on the fly and output it in the right format. Let’s take the previous example and generate a random list of 5 people with portuguese names and their favourite food.

We just need to add a new people_json_schema to create an array of people and pass it to the create_structured_output_chain:

from dotenv import load_dotenv
from langchain.chains.openai_functions import create_structured_output_chain
from langchain.chat_models import ChatOpenAI
from langchain.prompts import ChatPromptTemplate

load_dotenv()  # take environment variables from .env so we can load OPENAI_API_KEY

person_json_schema = {
    "title": "Person",
    "description": "Identifying information about a person.",
    "type": "object",
    "properties": {
        "name": {"title": "Name", "description": "The person's name", "type": "string"},
        "age": {"title": "Age", "description": "The person's age", "type": "integer"},
        "fav_food": {
            "title": "Fav Food",
            "description": "The person's favorite food",
            "type": "string",
        },
    },
    "required": ["name", "age"],
}

people_json_schema = {
    "title": "People",
    "description": "A list of people.",
    "type": "object",
    "properties": {
        "people": {
            "title": "People",
            "description": "A list of people",
            "type": "array",
            "items": person_json_schema
        }
    }
}

llm = ChatOpenAI(model="gpt-4", temperature=0)
prompt = ChatPromptTemplate.from_messages(
    [
        ("system", "You are extracting information in structured formats."),
        ("human", "Use the given format to generate {number} random person with {type} names.")
    ]
)

chain = create_structured_output_chain(people_json_schema, llm, prompt, verbose=True)
output = chain.run({
    "number": 5,
    "type": "portuguese"
})

print(output)

You can find the code in my github repo.

Happy generating!