SteeringAPI

Step 1: Get Your API Key

After signing up, navigate to the API Keys page in your dashboard:

Go to API Keys in your dashboard
Click "Create API Key"
Give your key a descriptive name
Copy and securely store your API key

Important

Your API key will only be shown once. Store it securely and never commit it to version control.

Step 2: Install the SDK

Python

pip install vllm-sdk

For Python, install the vLLM SDK. For JavaScript/Node.js, you can use the native fetch API (no additional packages required).

Step 3: Make Your First Request

Basic Chat Completion

Start with a simple chat completion without steering:

import asyncio
from vllm_sdk import VLLMClient, ChatMessage

async def main():
    async with VLLMClient(api_key="<YOUR_API_KEY>") as client:
        response = await client.chat_completions(
            model="meta-llama/Llama-3.1-8B-Instruct",
            messages=[
                ChatMessage(role="user", content="Tell me about the ocean")
            ],
        )
        print(response.choices[0].message.content)

asyncio.run(main())

Step 4: Add Feature Steering

Now let's add steering to control the model's behavior. We'll use feature 99 which represents "pirate speech patterns":

import asyncio
from vllm_sdk import VLLMClient, ChatMessage, Variant

async def main():
    async with VLLMClient(api_key="<YOUR_API_KEY>") as client:
        variant = Variant("meta-llama/Llama-3.1-8B-Instruct")
        variant.add_intervention(feature_id=99, strength=0.5, mode="add")

        response = await client.chat_completions(
            model=variant,
            messages=[
                ChatMessage(role="user", content="Tell me about the ocean")
            ],
        )
        print(response.choices[0].message.content)
        # Output: "Arr, the ocean be a vast body of water..."

asyncio.run(main())

Understanding Steering Values

Positive values (0.1 - 1.0): Amplify the feature's effect
Negative values (-1.0 - 0): Suppress the feature's effect
Typical range: -0.5 to 0.5 for most use cases
Experiment: Start small and adjust based on results

Step 5: Discover Features

Use the Feature Search API to find relevant features for your use case:

import requests

response = requests.post(
    "https://api.goodfire.ai/v1/features/search",
    headers={"X-API-Key": "<YOUR_API_KEY>"},
    json={
        "query": "formal academic writing",
        "model_name": "meta-llama/Llama-3.1-8B-Instruct",
        "top_k": 10
    }
)

features = response.json()["data"]
for feature in features:
    print(f"ID: {feature['id']}, Label: {feature['label']}")

You can also browse features interactively in the Feature Search dashboard.

Step 6: Combine Multiple Features

You can steer on multiple features simultaneously for fine-grained control:

import asyncio
from vllm_sdk import VLLMClient, ChatMessage, Variant

async def main():
    async with VLLMClient(api_key="<YOUR_API_KEY>") as client:
        variant = Variant("meta-llama/Llama-3.1-8B-Instruct")
        variant.add_intervention(feature_id=1234, strength=0.3, mode="add")   # Increase technical detail
        variant.add_intervention(feature_id=5678, strength=-0.2, mode="add")  # Reduce jargon
        variant.add_intervention(feature_id=9012, strength=0.4, mode="add")   # Add enthusiasm

        response = await client.chat_completions(
            model=variant,
            messages=[
                ChatMessage(role="user", content="Explain quantum computing")
            ],
        )
        print(response.choices[0].message.content)

asyncio.run(main())