Skip to content

~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.

7.5 KB
Raw Blame
#!/usr/bin/env python
"""
Custom LLM multi-benchmark optimization example for Local Deep Research.

This script demonstrates how to run multi-benchmark optimization with custom LLM models.

Usage:
    # Run from project root with PDM
    cd /path/to/local-deep-research
    pdm run python examples/optimization/llm_multi_benchmark.py --model "your-model" --provider "your-provider"
"""

import argparse
import os
import sys
from datetime import datetime, UTC
from pathlib import Path
from typing import Any, Dict, Optional

from loguru import logger

# Import benchmark optimization functions
from local_deep_research.benchmarks.optimization.api import optimize_parameters


def setup_llm_config(
    model: Optional[str] = None,
    provider: Optional[str] = None,
    endpoint_url: Optional[str] = None,
    api_key: Optional[str] = None,
    temperature: float = 0.7,
) -> Dict[str, Any]:
    """
    Set up LLM configuration for benchmarks and optimization.

    Args:
        model: LLM model name
        provider: LLM provider
        endpoint_url: Custom endpoint URL for OpenRouter or other services
        api_key: API key for the service
        temperature: LLM temperature

    Returns:
        Dictionary with LLM configuration
    """
    config = {
        "model_name": model,
        "provider": provider,
        "temperature": temperature,
    }

    if endpoint_url:
        config["openai_endpoint_url"] = endpoint_url
        os.environ["OPENAI_ENDPOINT_URL"] = endpoint_url
        os.environ["LDR_LLM__OPENAI_ENDPOINT_URL"] = endpoint_url

    if api_key:
        # Set API key in environment
        if provider == "openai" or provider == "openai_endpoint":
            os.environ["OPENAI_API_KEY"] = api_key
            os.environ["LDR_LLM__OPENAI_API_KEY"] = api_key
            if provider == "openai_endpoint":
                os.environ["OPENAI_ENDPOINT_API_KEY"] = api_key
                os.environ["LDR_LLM__OPENAI_ENDPOINT_API_KEY"] = api_key
        elif provider == "anthropic":
            os.environ["ANTHROPIC_API_KEY"] = api_key
            os.environ["LDR_LLM__ANTHROPIC_API_KEY"] = api_key

        config["api_key"] = api_key

    # Set model and provider in environment
    if model:
        os.environ["LDR_LLM__MODEL"] = model
    if provider:
        os.environ["LDR_LLM__PROVIDER"] = provider

    return config


def main():
    """Run multi-benchmark optimization with custom LLM."""
    parser = argparse.ArgumentParser(
        description="Run multi-benchmark optimization with custom LLM"
    )

    # LLM configuration
    parser.add_argument("--model", help="LLM model name")
    parser.add_argument(
        "--provider", help="LLM provider (openai, anthropic, openai_endpoint)"
    )
    parser.add_argument(
        "--endpoint-url", help="Custom endpoint URL (for OpenRouter etc.)"
    )
    parser.add_argument("--api-key", help="API key for the LLM provider")
    parser.add_argument(
        "--temperature", type=float, default=0.7, help="Temperature for LLM"
    )

    # Optimization parameters
    parser.add_argument(
        "--mode",
        choices=["balanced", "speed", "quality"],
        default="balanced",
        help="Optimization mode",
    )
    parser.add_argument(
        "--trials", type=int, default=3, help="Number of trials (default: 3)"
    )
    parser.add_argument(
        "--output-dir", help="Output directory (default: auto-generated)"
    )

    args = parser.parse_args()

    # Create timestamp-based directory for results
    timestamp = datetime.now(UTC).strftime("%Y%m%d_%H%M%S")
    if args.output_dir:
        output_dir = args.output_dir
    else:
        output_dir = str(
            Path("examples")
            / "optimization"
            / "results"
            / f"llm_multi_benchmark_{timestamp}"
        )

    os.makedirs(output_dir, exist_ok=True)
    print(f"Results will be saved to: {output_dir}")

    # Set up LLM configuration
    setup_llm_config(
        model=args.model,
        provider=args.provider,
        endpoint_url=args.endpoint_url,
        api_key=args.api_key,
        temperature=args.temperature,
    )

    if args.model and args.provider:
        print(f"Using LLM: {args.model} via {args.provider}")
    else:
        print("Using default LLM configuration from environment or database")

    # Define a small parameter space for quick demonstration
    param_space = {
        "iterations": {
            "type": "int",
            "low": 1,
            "high": 2,
            "step": 1,
        },
        "questions_per_iteration": {
            "type": "int",
            "low": 1,
            "high": 2,
            "step": 1,
        },
        "search_strategy": {
            "type": "categorical",
            "choices": ["rapid", "source_based"],  # Limited choices for speed
        },
    }

    # Example query for running optimization
    query = "Recent developments in fusion energy research"

    # Define metrics weights based on mode
    if args.mode == "speed":
        metric_weights = {"speed": 0.8, "quality": 0.2}
    elif args.mode == "quality":
        metric_weights = {"quality": 0.9, "speed": 0.1}
    else:  # balanced
        metric_weights = {"quality": 0.5, "speed": 0.5}

    # Run optimization with multi-benchmark weights
    print(
        f"\n🔍 Running {args.mode}-focused optimization with SimpleQA and BrowseComp..."
    )
    try:
        # Run optimization with combined benchmark weights
        benchmark_weights = {
            "simpleqa": 0.7,
            "browsecomp": 0.3,
        }  # 70% SimpleQA, 30% BrowseComp

        params, score = optimize_parameters(
            query=query,
            param_space=param_space,
            output_dir=output_dir,
            n_trials=args.trials,
            model_name=args.model,
            provider=args.provider,
            openai_endpoint_url=args.endpoint_url,
            temperature=args.temperature,
            api_key=args.api_key,
            benchmark_weights=benchmark_weights,
            metric_weights=metric_weights,
            search_tool="searxng",
        )

        print("\n" + "=" * 50)
        print(f" OPTIMIZATION RESULTS - {args.mode.upper()} MODE ")
        print("=" * 50)
        print(f"SCORE: {score:.4f}")
        print("Benchmark weights: SimpleQA 70%, BrowseComp 30%")
        print(f"Metrics weights: {metric_weights}")

        if args.model and args.provider:
            print(f"LLM: {args.model} via {args.provider}")

        print("\nBest Parameters:")
        for param, value in params.items():
            print(f"  {param}: {value}")
        print("=" * 50 + "\n")

        # Save results to file
        import json

        with open(
            Path(output_dir) / "multi_benchmark_results.json",
            "w",
            encoding="utf-8",
        ) as f:
            json.dump(
                {
                    "timestamp": timestamp,
                    "mode": args.mode,
                    "model": args.model,
                    "provider": args.provider,
                    "n_trials": args.trials,
                    "benchmark_weights": benchmark_weights,
                    "metric_weights": metric_weights,
                    "best_parameters": params,
                    "best_score": float(score),
                },
                f,
                indent=2,
            )

        print(
            f"Results saved to {Path(output_dir) / 'multi_benchmark_results.json'}"
        )

    except Exception:
        logger.exception("Error running optimization")
        import traceback

        traceback.print_exc()
        return 1

    return 0


if __name__ == "__main__":
    sys.exit(main())