Agnostic Engineering: The Model Swapper

In this lesson, we put "Abstraction" into practice. We will build a simple function that can return either an OpenAI model or an Anthropic model based on a string name. This is the Professional Way to build AI apps—never locking yourself into one vendor.

1. The Multi-Provider Script

Ensure you have both providers installed: pip install langchain-openai langchain-anthropic

from dotenv import load_dotenv
from langchain_openai import ChatOpenAI
from langchain_anthropic import ChatAnthropic

load_dotenv()

def get_model(provider="openai"):
    if provider == "openai":
        return ChatOpenAI(model="gpt-4o-mini")
    elif provider == "anthropic":
        return ChatAnthropic(model="claude-3-haiku-20240307")
    else:
        raise ValueError("Unsupported provider")

# Test switching
for p in ["openai", "anthropic"]:
    model = get_model(p)
    print(f"--- Testing {p} ---")
    print(model.invoke("Who are you?").content)

---

2. Why "Agnosticism" reduces Cost

During development, you might use Claude 3 Haiku or GPT-4o Mini (cheap). When you go to production for high-value tasks, you switch the config to GPT-4o or Claude 3.5 Sonnet (expensive). By using the LangChain abstraction, this change takes 2 seconds instead of 2 hours.

---

3. Standardizing Parameters

Even though providers use different names (e.g., Anthropic uses max_tokens and some older OpenAI models used max_tokens_to_sample), LangChain attempts to normalize these parameters.

# This parameter name will generally work across providers in LangChain
model = ChatOpenAI(model="...", temperature=0.7)

---

4. Visualizing the Switcher

graph TD
    Config[Config: provider='anthropic'] --> Logic[Factory Function]
    Logic -->|Choice| A[Anthropic Driver]
    Logic -->|Choice| O[OpenAI Driver]
    A --> API[External Service]
    O --> API

---

Key Takeaways

• Factory Functions are the best pattern for loading models.
• Provider modularity allows for rapid cost optimization.
• LangChain normalizes parameters like temperature and max tokens.
• Always test your app's logic with at least two different models to ensure it doesn't "break" on one model's quirks.