benchmarks/run_router_eval.py

"""
Evaluation harness for RouterBench-inspired policies in InferRoute.

This script evaluates various routing strategies (Always Cloud, Always Local, Zero Router, 
Rule Router, KNN Router, MLP Router, Oracle Router, and Cascade Router) against the 
workload dataset. It sweeps parameters (mixture ratio p, willingness to pay lambda) 
to trace cost-quality curves.

Inspired by:
"ROUTERBENCH: A Benchmark for Multi-LLM Routing System" (withmartian/routerbench)
"""

import os
import sys
import json
import time
import asyncio
from typing import Any, List
from httpx import AsyncClient

# Add project root to path to resolve imports correctly
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))

from inferroute.main import app
from benchmarks.evaluate_quality import evaluate_response_quality

DATASET_PATH = os.path.join(os.path.dirname(__file__), "datasets", "workload.json")
RESULTS_DIR = os.path.join(os.path.dirname(__file__), "results")
RESULTS_PATH = os.path.join(RESULTS_DIR, "eval_results.json")

# Ensure results directory exists
os.makedirs(RESULTS_DIR, exist_ok=True)

# API Token for acme_corp tenant
HEADERS = {"Authorization": "Bearer sk-inferroute-demo"}

# Defined evaluation scenarios
SCENARIOS = [
    {
        "name": "always-openai",
        "payload_patch": {"model": "gpt-4o-mini", "routing": {"allow_local": False}}
    },
    {
        "name": "always-gemini",
        "payload_patch": {"model": "gemini-1.5-flash", "routing": {"allow_local": False}}
    },
    {
        "name": "always-vllm",
        "payload_patch": {"model": "meta-llama/Meta-Llama-3-8B-Instruct", "routing": {"allow_cloud": False}}
    },
    {
        "name": "always-ollama",
        "payload_patch": {"model": "llama3", "routing": {"allow_cloud": False}}
    },
    {
        "name": "rule-router",
        "payload_patch": {"model": "edge/auto", "routing": {"policy": "rule"}}
    },
    {
        "name": "oracle-router",
        "payload_patch": {"model": "edge/auto", "routing": {"policy": "oracle"}}
    }
]

# Sweep Zero Router (baseline mixture ratios p from 0.0 to 1.0)
for p in [0.0, 0.2, 0.4, 0.6, 0.8, 1.0]:
    SCENARIOS.append({
        "name": f"zero-router_p{p:.1f}",
        "payload_patch": {"model": "edge/auto", "routing": {"policy": "zero", "mixture_ratio": p}}
    })

# Sweep KNN Router (willingness to pay lambdas)
for lam in [0.0, 0.25, 0.5, 1.0, 2.0, 5.0]:
    SCENARIOS.append({
        "name": f"knn-router_l{lam:.2f}",
        "payload_patch": {"model": "edge/auto", "routing": {"policy": "knn", "lambda": lam}}
    })

# Sweep MLP Router (willingness to pay lambdas)
for lam in [0.0, 0.25, 0.5, 1.0, 2.0, 5.0]:
    SCENARIOS.append({
        "name": f"mlp-router_l{lam:.2f}",
        "payload_patch": {"model": "edge/auto", "routing": {"policy": "mlp", "lambda": lam}}
    })

# Sweep Cascade Router (acceptance thresholds tau from 0.0 to 1.0)
for tau in [0.0, 0.2, 0.4, 0.6, 0.8, 1.0]:
    SCENARIOS.append({
        "name": f"cascade-router_t{tau:.2f}",
        "payload_patch": {"model": "edge/auto", "routing": {"policy": "cascade", "acceptance_threshold": tau}}
    })

async def execute_request(client: AsyncClient, payload: dict) -> dict[str, Any]:
    """Issues POST completion to the gateway and measures latencies."""
    start_time = time.time()
    
    response = await client.post(
        "/v1/chat/completions",
        headers=HEADERS,
        json=payload,
        timeout=10.0
    )
    
    latency = (time.time() - start_time) * 1000.0
    
    if response.status_code != 200:
        return {
            "success": False,
            "error": response.text,
            "latency_ms": latency,
            "ttft_ms": latency,
            "cost_usd": 0.0,
            "content": "",
            "fallback_triggered": False
        }
        
    data = response.json()
    choices = data.get("choices", [])
    content = choices[0].get("message", {}).get("content", "") if choices else ""
    timing = data.get("timing", {})
    usage = data.get("usage", {})
    route = data.get("route", {})
    fallback_count = route.get("fallback_count", 0)
    
    return {
        "success": True,
        "content": content,
        "latency_ms": timing.get("latency_ms", latency),
        "ttft_ms": timing.get("ttft_ms", latency / 2.0),
        "cost_usd": usage.get("estimated_cost_usd", 0.0),
        "model_used": data.get("model", ""),
        "fallback_triggered": fallback_count > 0
    }

async def run_evaluation():
    print("=============================================================")
    print("Starting RouterBench-driven Routing Policy Sweeping Harness")
    print("=============================================================")
    
    # Load dataset
    with open(DATASET_PATH, "r", encoding="utf-8") as f:
        prompts = json.load(f)
        
    print(f"Loaded {len(prompts)} prompts across {len(set(p['category'] for p in prompts))} categories.")
    
    results = []
    
    from httpx import ASGITransport
    transport = ASGITransport(app=app)
    
    async with AsyncClient(transport=transport, base_url="http://testserver") as client:
        # Give lifespan a millisecond to boot databases
        await asyncio.sleep(0.5)
        
        for scenario in SCENARIOS:
            name = scenario["name"]
            print(f"\n---> Running scenario: {name}")
            
            for item in prompts:
                prompt_id = item["id"]
                category = item["category"]
                prompt_text = item["prompt"]
                requires_json = item["requires_json"]
                expected_keys = item.get("expected_keys")
                ref_keywords = item.get("reference_keywords")
                
                # Base request body
                payload = {
                    "messages": [{"role": "user", "content": prompt_text}],
                    "temperature": 0.2,
                    "max_output_tokens": 128
                }
                payload.update(scenario["payload_patch"])
                
                # Regular execution (cascade fallback is handled on server side)
                res = await execute_request(client, payload)
                fallback_triggered = res.get("fallback_triggered", False)
                
                if res["success"]:
                    q_score = evaluate_response_quality(
                        category, res["content"], requires_json, expected_keys, ref_keywords
                    )
                else:
                    q_score = 0.0
                
                # Check SLO compliance (target: p95 latency < 2000ms, success=True)
                slo_compliant = res["success"] and (res["latency_ms"] < 2000.0)
                
                results.append({
                    "scenario": name,
                    "prompt_id": prompt_id,
                    "category": category,
                    "success": res["success"],
                    "model_used": res.get("model_used", "unknown"),
                    "latency_ms": res["latency_ms"],
                    "ttft_ms": res["ttft_ms"],
                    "cost_usd": res["cost_usd"],
                    "quality_score": q_score,
                    "fallback_triggered": fallback_triggered,
                    "slo_compliant": slo_compliant
                })
                
                print(f"  ID={prompt_id:<14} Model={res.get('model_used', 'None'):<14} Latency={res['latency_ms']:>6.1f}ms Cost=${res['cost_usd']:.6f} Quality={q_score:.2f}")
 
    # Write evaluation results to file
    with open(RESULTS_PATH, "w", encoding="utf-8") as f:
        json.dump(results, f, indent=2)
        
    print(f"\nEvaluation finished. Results saved to {RESULTS_PATH}")

if __name__ == "__main__":
    asyncio.run(run_evaluation())