feat: Add experiments framework and novelty-driven agent loop
- Add complete experiments directory with pilot study infrastructure - 5 experimental conditions (direct, expert-only, attribute-only, full-pipeline, random-perspective) - Human assessment tool with React frontend and FastAPI backend - AUT flexibility analysis with jump signal detection - Result visualization and metrics computation - Add novelty-driven agent loop module (experiments/novelty_loop/) - NoveltyDrivenTaskAgent with expert perspective perturbation - Three termination strategies: breakthrough, exhaust, coverage - Interactive CLI demo with colored output - Embedding-based novelty scoring - Add DDC knowledge domain classification data (en/zh) - Add CLAUDE.md project documentation - Update research report with experiment findings Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
214
experiments/conditions/c4_full_pipeline.py
Normal file
214
experiments/conditions/c4_full_pipeline.py
Normal file
@@ -0,0 +1,214 @@
|
||||
"""
|
||||
Condition 4: Full Pipeline (Attributes + Experts)
|
||||
|
||||
The complete novelty-seeking system:
|
||||
1. Attribute decomposition into categories
|
||||
2. Expert team generation
|
||||
3. Expert keyword generation for each attribute
|
||||
4. Description generation for each keyword
|
||||
"""
|
||||
|
||||
import sys
|
||||
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 app.services.expert_source_service import expert_source_service
|
||||
from app.prompts.attribute_prompt import get_step1_dynamic_attributes_prompt
|
||||
from app.prompts.expert_transformation_prompt import (
|
||||
get_expert_keyword_generation_prompt,
|
||||
get_single_description_prompt
|
||||
)
|
||||
from experiments.config import (
|
||||
MODEL, TEMPERATURE, FIXED_CATEGORIES, EXPERT_COUNT,
|
||||
EXPERT_SOURCE, KEYWORDS_PER_EXPERT, PROMPT_LANGUAGE
|
||||
)
|
||||
|
||||
|
||||
async def generate_ideas(
|
||||
query: str,
|
||||
model: str = None,
|
||||
temperature: float = None,
|
||||
categories: List[str] = None,
|
||||
expert_count: int = None,
|
||||
expert_source: str = None,
|
||||
keywords_per_expert: int = None,
|
||||
lang: str = None
|
||||
) -> Dict[str, Any]:
|
||||
"""
|
||||
Generate ideas using the full pipeline (C4).
|
||||
|
||||
Args:
|
||||
query: The object/concept to generate ideas for
|
||||
model: LLM model to use
|
||||
temperature: Generation temperature
|
||||
categories: Categories for attribute decomposition
|
||||
expert_count: Number of experts
|
||||
expert_source: Source of experts
|
||||
keywords_per_expert: Keywords each expert generates per attribute
|
||||
lang: Language for prompts
|
||||
|
||||
Returns:
|
||||
Dict with ideas and metadata
|
||||
"""
|
||||
model = model or MODEL
|
||||
temperature = temperature or TEMPERATURE
|
||||
categories = categories or FIXED_CATEGORIES
|
||||
expert_count = expert_count or EXPERT_COUNT
|
||||
expert_source = expert_source or EXPERT_SOURCE
|
||||
keywords_per_expert = keywords_per_expert or KEYWORDS_PER_EXPERT
|
||||
lang = lang or PROMPT_LANGUAGE
|
||||
|
||||
# Step 0: Get experts from curated source
|
||||
experts_data, actual_source = expert_source_service.get_experts(
|
||||
source=expert_source,
|
||||
count=expert_count,
|
||||
language=lang
|
||||
)
|
||||
|
||||
# Convert to expected format
|
||||
experts = [
|
||||
{
|
||||
"id": f"expert-{i}",
|
||||
"name": exp.get("name", "Expert"),
|
||||
"domain": exp.get("domain", ""),
|
||||
"perspective": exp.get("perspective", "")
|
||||
}
|
||||
for i, exp in enumerate(experts_data)
|
||||
]
|
||||
|
||||
# Step 1: Generate attributes
|
||||
category_defs = [
|
||||
{"name": cat, "description": f"Related {cat.lower()} of the object", "order": i}
|
||||
for i, cat in enumerate(categories)
|
||||
]
|
||||
|
||||
attr_prompt = get_step1_dynamic_attributes_prompt(
|
||||
query=query,
|
||||
categories=category_defs,
|
||||
lang=lang
|
||||
)
|
||||
|
||||
attr_response = await ollama_provider.generate(
|
||||
prompt=attr_prompt,
|
||||
model=model,
|
||||
temperature=temperature
|
||||
)
|
||||
|
||||
attributes_by_category = extract_json_from_response(attr_response)
|
||||
|
||||
# Step 2: Expert keyword generation for each category/attribute
|
||||
all_keywords = []
|
||||
|
||||
for category in categories:
|
||||
attrs = attributes_by_category.get(category, [])
|
||||
|
||||
for attr in attrs:
|
||||
# Generate keywords from all experts for this attribute
|
||||
keyword_prompt = get_expert_keyword_generation_prompt(
|
||||
category=category,
|
||||
attribute=attr,
|
||||
experts=experts,
|
||||
keywords_per_expert=keywords_per_expert,
|
||||
lang=lang
|
||||
)
|
||||
|
||||
keyword_response = await ollama_provider.generate(
|
||||
prompt=keyword_prompt,
|
||||
model=model,
|
||||
temperature=temperature
|
||||
)
|
||||
|
||||
keyword_result = extract_json_from_response(keyword_response)
|
||||
keywords = keyword_result.get("keywords", [])
|
||||
|
||||
for kw in keywords:
|
||||
all_keywords.append({
|
||||
"category": category,
|
||||
"attribute": attr,
|
||||
"keyword": kw.get("keyword", ""),
|
||||
"expert_id": kw.get("expert_id", ""),
|
||||
"expert_name": kw.get("expert_name", "")
|
||||
})
|
||||
|
||||
# Step 3: Generate descriptions for each keyword
|
||||
all_ideas = []
|
||||
|
||||
for kw_info in all_keywords:
|
||||
# Find expert details
|
||||
expert = next(
|
||||
(e for e in experts if e["id"] == kw_info["expert_id"]),
|
||||
{"name": kw_info["expert_name"], "domain": "", "id": kw_info["expert_id"]}
|
||||
)
|
||||
|
||||
desc_prompt = get_single_description_prompt(
|
||||
query=query,
|
||||
keyword=kw_info["keyword"],
|
||||
expert_id=expert["id"],
|
||||
expert_name=expert["name"],
|
||||
expert_domain=expert.get("domain", ""),
|
||||
lang=lang
|
||||
)
|
||||
|
||||
desc_response = await ollama_provider.generate(
|
||||
prompt=desc_prompt,
|
||||
model=model,
|
||||
temperature=temperature
|
||||
)
|
||||
|
||||
desc_result = extract_json_from_response(desc_response)
|
||||
description = desc_result.get("description", "")
|
||||
|
||||
all_ideas.append({
|
||||
"idea": description,
|
||||
"keyword": kw_info["keyword"],
|
||||
"category": kw_info["category"],
|
||||
"attribute": kw_info["attribute"],
|
||||
"expert_name": expert["name"],
|
||||
"expert_domain": expert.get("domain", "")
|
||||
})
|
||||
|
||||
return {
|
||||
"condition": "c4_full_pipeline",
|
||||
"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,
|
||||
"categories": categories,
|
||||
"attributes_by_category": attributes_by_category,
|
||||
"attribute_count": sum(len(v) for v in attributes_by_category.values()),
|
||||
"expert_count": expert_count,
|
||||
"expert_source": actual_source,
|
||||
"keywords_per_expert": keywords_per_expert,
|
||||
"total_keywords": len(all_keywords),
|
||||
"experts": [{"name": e["name"], "domain": e["domain"]} for e in experts],
|
||||
"mechanism": "full_pipeline_attributes_plus_experts"
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
# For testing
|
||||
if __name__ == "__main__":
|
||||
import asyncio
|
||||
|
||||
async def test():
|
||||
result = await generate_ideas("Chair")
|
||||
print(f"Generated {result['idea_count']} ideas using full pipeline:")
|
||||
print(f" Attributes: {result['metadata']['attribute_count']}")
|
||||
print(f" Experts: {result['metadata']['expert_count']}")
|
||||
print(f" Keywords: {result['metadata']['total_keywords']}")
|
||||
print("\nExperts used:")
|
||||
for exp in result['metadata']['experts']:
|
||||
print(f" - {exp['name']} ({exp['domain']})")
|
||||
print("\nSample ideas:")
|
||||
for i, item in enumerate(result['ideas_with_source'][:5], 1):
|
||||
print(f" {i}. [{item['expert_name']}] {item['keyword']}: {item['idea']}")
|
||||
|
||||
asyncio.run(test())
|
||||
Reference in New Issue
Block a user