Arch-Router-1.5B

A specialized 1.5B parameter model for intelligent routing between multiple LLMs based on domain and action preferences.

Overview Arch-Router learns to map user queries to domain-action preferences, enabling smart routing decisions in multi-model environments. It analyzes incoming prompts to determine the best-suited model for each request using structured routing logic.

Key Features:

Fast, deterministic routing decisions (<200ms) Structured XML input → JSON output format Transparent, configurable routing logic Production-ready for low-latency applications

Credits

| Original Model: katanemo/Arch-Router-1.5B | | Paper: Preference-Aligned Routing (arXiv:2506.16655) |
| Project: Arch Gateway |

Usage
Define your routes with domain-action mappings:

ollama run arch-router:1.5b '
<routes>
[
  {"name": "medical", "description": "Physical health concerns, pain, symptoms, medication questions"},
  {"name": "emotional", "description": "Mental health, stress, anxiety, emotional support"},
  {"name": "activity", "description": "Exercise, movement, daily activities, energy management"}
]
</routes>

<conversation>
[{"role": "user", "content": "My chronic pain is flaring up today"}]
</conversation>

Output the route name in JSON format: {"route": "route_name"}

Output: {“route”: “medical”}

Python Integration

pythonimport json
import requests

def route_query(user_message, routes):
    prompt = f"""
<routes>
{json.dumps(routes)}
</routes>

<conversation>
[{{"role": "user", "content": "{user_message}"}}]
</conversation>

Output the route name in JSON format: {{"route": "route_name"}}
"""
    
    response = requests.post("http://localhost:11434/api/generate", json={
        "model": "arch-router:1.5b",
        "prompt": prompt,
        "stream": False,
        "options": {"temperature": 0.0}
    })
    
    result = response.json()["response"].strip()
    return json.loads(result)["route"]

# Example
routes = [
    {"name": "code", "description": "Programming, debugging, code generation"},
    {"name": "writing", "description": "Creative writing, editing, content creation"}
]

route = route_query("Help me fix this Python bug", routes)
print(route)  # "code"

License Original model license applies. See HuggingFace model card for details.