novelty-seeking/experiments/conditions/c5_random_perspective.py

"""
Condition 5: Random-Perspective Control

Uses random words as "perspectives" instead of domain experts.
Tests whether the benefit from expert perspectives comes from
domain knowledge or simply from any perspective shift.
"""

import sys
import json
import random
from pathlib import Path

# Add backend to path for imports
sys.path.insert(0, str(Path(__file__).parent.parent.parent / "backend"))

from typing import List, Dict, Any
from app.services.llm_service import ollama_provider, extract_json_from_response
from experiments.config import (
    MODEL, TEMPERATURE, EXPERT_COUNT, IDEAS_PER_EXPERT,
    PROMPT_LANGUAGE, RANDOM_SEED, DATA_DIR
)


def load_random_words() -> List[str]:
    """Load the random word pool from data file."""
    words_file = DATA_DIR / "random_words.json"
    with open(words_file, "r", encoding="utf-8") as f:
        data = json.load(f)
    return data.get("words", [])


def get_random_perspective_prompt(
    query: str,
    perspective_word: str,
    idea_count: int,
    lang: str = "en"
) -> str:
    """Generate prompt for random-perspective idea generation."""
    if lang == "en":
        return f"""/no_think
Generate {idea_count} creative and innovative ideas for "{query}" inspired by the concept of "{perspective_word}".

Requirements:
1. Each idea should draw inspiration from "{perspective_word}" - its qualities, characteristics, or associations
2. Think about how concepts related to "{perspective_word}" could improve or reimagine "{query}"
3. Ideas should be specific and actionable (15-30 words each)
4. Be creative in connecting "{perspective_word}" to "{query}"

Return JSON only:
{{"ideas": ["idea 1", "idea 2", "idea 3", ...]}}

Generate exactly {idea_count} ideas inspired by "{perspective_word}"."""
    else:
        return f"""/no_think
為「{query}」生成 {idea_count} 個創意點子，靈感來自「{perspective_word}」這個概念。

要求：
1. 每個點子要從「{perspective_word}」獲得靈感——它的特質、特徵或聯想
2. 思考與「{perspective_word}」相關的概念如何改進或重新想像「{query}」
3. 點子要具體可行（每個 15-30 字）
4. 創意地連接「{perspective_word}」和「{query}」

只回傳 JSON：
{{"ideas": ["點子1", "點子2", "點子3", ...]}}

生成正好 {idea_count} 個受「{perspective_word}」啟發的點子。"""


async def generate_ideas(
    query: str,
    model: str = None,
    temperature: float = None,
    word_count: int = None,
    ideas_per_word: int = None,
    lang: str = None,
    seed: int = None
) -> Dict[str, Any]:
    """
    Generate ideas using random word perspectives (C5 control).

    Args:
        query: The object/concept to generate ideas for
        model: LLM model to use
        temperature: Generation temperature
        word_count: Number of random words to use (matches expert count)
        ideas_per_word: Ideas to generate per word
        lang: Language for prompts
        seed: Random seed for reproducibility

    Returns:
        Dict with ideas and metadata
    """
    model = model or MODEL
    temperature = temperature or TEMPERATURE
    word_count = word_count or EXPERT_COUNT
    ideas_per_word = ideas_per_word or IDEAS_PER_EXPERT
    lang = lang or PROMPT_LANGUAGE
    seed = seed or RANDOM_SEED

    # Load word pool and sample random words
    word_pool = load_random_words()

    # Use seeded random for reproducibility
    # Create a unique seed per query to get different words for different queries
    # but same words for same query across runs
    query_seed = seed + hash(query) % 10000
    rng = random.Random(query_seed)
    selected_words = rng.sample(word_pool, min(word_count, len(word_pool)))

    all_ideas = []
    word_details = []

    for word in selected_words:
        prompt = get_random_perspective_prompt(
            query=query,
            perspective_word=word,
            idea_count=ideas_per_word,
            lang=lang
        )

        response = await ollama_provider.generate(
            prompt=prompt,
            model=model,
            temperature=temperature
        )

        result = extract_json_from_response(response)
        ideas = result.get("ideas", [])

        # Tag ideas with perspective word source
        for idea in ideas:
            all_ideas.append({
                "idea": idea,
                "perspective_word": word
            })

        word_details.append({
            "word": word,
            "ideas_generated": len(ideas)
        })

    return {
        "condition": "c5_random_perspective",
        "query": query,
        "ideas": [item["idea"] for item in all_ideas],
        "ideas_with_source": all_ideas,
        "idea_count": len(all_ideas),
        "metadata": {
            "model": model,
            "temperature": temperature,
            "prompt_language": lang,
            "word_count": word_count,
            "ideas_per_word": ideas_per_word,
            "random_seed": seed,
            "query_seed": query_seed,
            "selected_words": selected_words,
            "word_details": word_details,
            "word_pool_size": len(word_pool),
            "mechanism": "random_perspective_control"
        }
    }


# For testing
if __name__ == "__main__":
    import asyncio

    async def test():
        result = await generate_ideas("Chair")
        print(f"Generated {result['idea_count']} ideas from {len(result['metadata']['selected_words'])} random words:")
        print(f"  Words used: {', '.join(result['metadata']['selected_words'])}")
        print(f"  Seed: {result['metadata']['random_seed']}, Query seed: {result['metadata']['query_seed']}")
        print("\nSample ideas:")
        for i, item in enumerate(result['ideas_with_source'][:5], 1):
            print(f"  {i}. [{item['perspective_word']}] {item['idea']}")

    asyncio.run(test())