Skip to main content

Instructor

Instructor is a library that helps you extract structured data from natural language using LLMs. By instrumenting Instructor with Agenta, you can monitor and debug your applications more effectively.

This guide shows you how to instrument Instructor when using OpenAI models, you can use the same approach for other LLM providers like Anthropic, Google, etc. You just need to use the correct instrumentation library for the LLM provider.

Installation​

Install the required packages:

pip install -U agenta openai opentelemetry-instrumentation-openai instructor

Configure Environment Variables​

import os

os.environ["AGENTA_API_KEY"] = "YOUR_AGENTA_API_KEY"
os.environ["AGENTA_HOST"] = "https://cloud.agenta.ai"

Code Example​

import agenta as ag
import openai
import instructor
from pydantic import BaseModel
from opentelemetry.instrumentation.openai import OpenAIInstrumentor

ag.init()

# Instrument OpenAI before creating the Instructor client
OpenAIInstrumentor().instrument()

class UserInfo(BaseModel):
    name: str
    age: int

@ag.instrument(spankind="WORKFLOW")
def instructor_workflow():
    # Create an Instructor client using the instrumented OpenAI model
    client = instructor.from_openai(openai.OpenAI())

    # Extract structured data from natural language
    user_info = client.chat.completions.create(
        model="gpt-3.5-turbo",
        response_model=UserInfo,
        messages=[{"role": "user", "content": "John Doe is 30 years old."}],
    )
    return user_info

user_info = instructor_workflow()
print(user_info)
warning

Order Matters: Call OpenAIInstrumentor().instrument() before creating the Instructor client with instructor.from_openai(openai.OpenAI()). Both modify the OpenAI library, so the order ensures proper instrumentation.

Explanation​

  • Initialize Agenta: ag.init() sets up the Agenta SDK.
  • Instrument OpenAI: OpenAIInstrumentor().instrument() instruments the OpenAI library for tracing. This must come before creating the Instructor client.
  • Instrument the Workflow: The @ag.instrument(spankind="WORKFLOW") decorator creates a parent span. This is optional, but it's a good practice to instrument the main function of your application.