WuYa (无涯) - Theory-Grounded Academic Review System

English | 中文

WuYa (无涯, meaning "boundless") is a theory-driven multi-agent system that transforms classical philosophical principles into an operational architecture for academic paper evaluation and journal recommendation.

🌟 Key Features

• Retrieval-Evaluation Coupling: RAG serves as the core triggering mechanism for judgment, not just citation
• Two-Phase Routing: CUDOS gatekeeping + parallel expert evaluation (Innovation, Method, Evidence, Application)
• Hybrid Knowledge Strategy: Combines prompt-internalized "principles" (dao) with RAG-retrieved "evidence" (shu)
• LLM-as-Mapper with Discipline Priors: Cross-disciplinary paper localization calibrated by citation network analysis
• Self-Improving Frontier Discovery: Learns evaluation preferences from editor feedback
• DEA Efficiency Analysis: Data Envelopment Analysis for quantitative paper-to-journal matching
• CLI & Docker: One-command evaluation, containerized deployment

🚀 Quick Start

Installation

# Clone the repository
git clone https://github.com/wuya-team/wuya.git
cd wuya

# Install (with all optional dependencies)
pip install -e ".[all]"

# Or install core only
pip install -e .

Configure

# Copy environment template
cp .env.example .env

# Edit .env with your API keys
# OPENAI_API_KEY=sk-...
# WUYA_LLM_PROVIDER=openai

CLI Usage

# Evaluate a single paper
wuya evaluate paper.pdf --journal "Nature Machine Intelligence" -o report.md

# Batch evaluate papers in a directory
wuya batch ./papers/ --output-dir ./reports/

# Show current configuration
wuya config show

# Start API server
wuya serve --port 8000

# Show version
wuya --version

Python API

import asyncio
from wuya_agents.router import TwoPhaseRouter
from wuya_agents.subagents import (
    CUDOSSubAgent, InnovationSubAgent, MethodSubAgent,
    EvidenceSubAgent, ApplicationSubAgent
)
from wuya_agents.dea_subagent import DEASubAgent, DEAEngine
from wuya_agents.rag.client import RAGClientImpl
from wuya_agents.base import ParsedPaper
from tests.conftest import MockLLMClient, MockRAGClient  # For testing

async def main():
    # Initialize agents (use real clients in production)
    mock_llm = MockLLMClient()
    mock_rag = MockRAGClient()

    router = TwoPhaseRouter(
        cudos_agent=CUDOSSubAgent(llm_client=mock_llm, rag_client=mock_rag),
        innovation_agent=InnovationSubAgent(llm_client=mock_llm, rag_client=mock_rag),
        method_agent=MethodSubAgent(llm_client=mock_llm, rag_client=mock_rag),
        evidence_agent=EvidenceSubAgent(llm_client=mock_llm, rag_client=mock_rag),
        application_agent=ApplicationSubAgent(llm_client=mock_llm, rag_client=mock_rag),
        rag_client=mock_rag,
    )

    # Prepare paper
    paper = ParsedPaper(
        paper_id="paper_001",
        title="Your Paper Title",
        abstract="Paper abstract...",
        content="Full paper content...",
        authors=["Author Name"],
        keywords=["keyword1", "keyword2"],
        discipline="computer science",
    )

    # Run evaluation
    report = await router.route(
        paper,
        target_journal="Nature",
        reference_papers=[...]  # Optional: for DEA analysis
    )

    print(f"Overall Score: {report.overall_score:.2f}")
    print(f"Tier Estimate: {report.tier_estimate}")
    print(f"Status: {report.status}")

asyncio.run(main())

🐳 Docker

Quick Start with Docker Compose

# Start all services (WuYa + ChromaDB)
docker compose up -d

# Run evaluation
docker compose exec wuya wuya evaluate /app/data/paper.pdf

# View logs
docker compose logs -f wuya

Build and Run Manually

# Build the image
docker build -t wuya-agents .

# Run CLI commands
docker run --rm -it wuya-agents wuya --help
docker run --rm -it -v $(pwd)/papers:/app/data wuya-agents wuya evaluate /app/data/paper.pdf

# Start API server
docker run --rm -it -p 8000:8000 wuya-agents wuya serve

# Run tests
docker run --rm -it wuya-agents test

🏗️ Architecture

┌─────────────────────────────────────────────────────────────┐
│                        Router                               │
└──────────────┬──────────────────────────────────────────────┘
               │
    ┌──────────┴──────────┐
    ▼                     ▼
┌──────────┐      ┌─────────────────┐
│  CUDOS   │      │  Evaluation     │
│  Gate    │      │  Sub-agents     │
│ (Phase 1)│      │  (Phase 2)      │
└──────────┘      └────────┬────────┘
                           │
        ┌──────────────────┼──────────────────┐
        ▼                  ▼                  ▼
   ┌──────────┐     ┌──────────┐      ┌──────────┐
   │Innovation│     │  Method  │      │ Evidence │
   └──────────┘     └──────────┘      └──────────┘
        │                  │                  │
        └──────────────────┼──────────────────┘
                           ▼
              ┌─────────────────────┐
              │   Paper Localization │
              │  Path A: LLM-Mapper  │
              │  Path B: DEA Analysis│
              └─────────────────────┘

📁 Project Structure

wuya/
├── wuya_agents/           # Agent implementations
│   ├── __init__.py        # Package exports and version
│   ├── config.py          # Global configuration (pydantic-settings)
│   ├── cli.py             # CLI entry point (typer)
│   ├── base.py            # Base classes and data models
│   ├── router.py          # TwoPhaseRouter orchestration
│   ├── aggregator.py      # Result aggregation
│   ├── parser.py          # Paper parsing (PDF/text)
│   ├── dea_subagent.py    # DEA Sub-agent
│   ├── subagents/         # Evaluation sub-agents
│   │   ├── cudos.py       # CUDOS gate
│   │   ├── innovation.py  # Innovation evaluation
│   │   ├── method.py      # Methodology evaluation
│   │   ├── evidence.py    # Evidence evaluation
│   │   ├── application.py # Application relevance
│   │   └── frontier.py    # Frontier discovery
│   ├── llm_client/        # LLM client abstraction
│   │   ├── config.py      # LLM configuration
│   │   ├── client.py      # Unified client
│   │   ├── openai_provider.py
│   │   └── anthropic_provider.py
│   └── rag/               # RAG components
│       ├── client.py      # RAG client
│       ├── embedding.py   # Embedding providers
│       └── vector_store.py # Vector storage
├── tests/                 # Test suite (218 tests)
├── scripts/               # Utility scripts
├── docs/                  # Documentation & ADRs
├── .github/workflows/     # CI/CD (GitHub Actions)
├── Dockerfile             # Docker image
├── docker-compose.yml     # Docker Compose
├── pyproject.toml         # Package configuration
├── .env.example           # Environment template
├── CHANGELOG.md           # Version history
├── CONTRIBUTING.md        # Contribution guide
├── SECURITY.md            # Security policy
├── LICENSE                # MIT License
└── README.md

🔬 Testing

Run Tests

# Run all tests
pytest tests/ -v

# Run with coverage
pytest tests/ --cov=wuya_agents --cov-report=term-missing

# Run specific module
pytest tests/test_subagents/ -v

# Skip slow tests
pytest tests/ -v -m "not slow and not e2e"

Test Results

============================= test session starts ==============================
218 passed in 4.11s

Coverage Report

Module	Coverage
Total	61%
wuya_agents/subagents/*	80-92%
wuya_agents/router.py	88%
wuya_agents/rag/client.py	90%
wuya_agents/base.py	69%
wuya_agents/dea_subagent.py	68%

⚙️ Configuration

All configuration is managed through environment variables or .env files:

Variable	Default	Description
WUYA_ENVIRONMENT	development	Environment: development/testing/production
WUYA_LLM_PROVIDER	openai	LLM provider: openai/anthropic
OPENAI_API_KEY	—	OpenAI API key
ANTHROPIC_API_KEY	—	Anthropic API key
WUYA_LLM_TEMPERATURE	0.1	Sampling temperature
WUYA_RAG_ENABLED	true	Enable RAG retrieval
WUYA_RAG_VECTOR_STORE_TYPE	in_memory	Vector store: in_memory/chromadb
WUYA_EVAL_CUDOS_THRESHOLD	3.0	CUDOS gate threshold

See .env.example for the complete list.

📄 Research Paper

The theoretical foundations and system architecture are described in our research paper:

"Retrieval-Driven Self-Improvement: A Multi-Agent Architecture for Theory-Grounded Academic Review"

See paper/WuYa_Paper.pdf

🧠 Theoretical Foundations

Dimension	Primary Theorists	Key Concepts
Innovation	Kuhn, Schumpeter, Mokyr	Paradigm shift, Creative destruction
Method	Pearl, Campbell, Fisher	Causal ladder, Internal/External validity
Evidence	Popper, Lakatos	Falsification, Research programmes
Application	Bush, Mokyr	Innovation pipeline, TRL
CUDOS	Merton	Scientific norms (Communalism, Universalism, etc.)

🤝 Contributing

We welcome contributions! Please see CONTRIBUTING.md for details.

🔒 Security

Please see SECURITY.md for our security policy and vulnerability reporting guidelines.

📜 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• Philosophical foundations drawn from classical works in philosophy of science
• DEA methodology based on Charnes, Cooper, and Rhodes (1978)
• Self-improving agent design inspired by Reflexion and Voyager

---

中文介绍

WuYa (无涯) 是一个基于理论驱动的多智能体系统，将经典哲学原理转化为学术论文评估和期刊推荐的运营架构。

核心特性

• 检索-评估耦合机制: RAG 作为判断的核心触发机制，而非仅用于引用
• 两阶段路由: CUDOS 把关 + 并行专家评估（创新性、方法、证据、应用）
• 混合知识策略: 提示内化的"道"与 RAG 检索的"术"相结合
• LLM-as-Mapper: 基于引文网络分析的学科先验校准
• 自改进前沿发现: 从编辑反馈中学习评估偏好
• DEA 效率分析: 数据包络分析用于定量的论文-期刊匹配
• CLI 和 Docker: 一键评估，容器化部署

快速开始

# 克隆仓库
git clone https://github.com/wuya-team/wuya.git
cd wuya

# 安装
pip install -e ".[all]"

# 配置环境变量
cp .env.example .env
# 编辑 .env 填入 API 密钥

# 评估单篇论文
wuya evaluate paper.pdf --journal "Nature Machine Intelligence" -o report.md

# 批量评估
wuya batch ./papers/ --output-dir ./reports/

# Docker 一键启动
docker compose up -d

测试

# 运行所有测试
pytest tests/ -v

# 运行带覆盖率报告
pytest tests/ --cov=wuya_agents --cov-report=term-missing

测试结果: 218 个测试全部通过

理论基础

评估维度	主要理论家	核心概念
创新性	库恩、熊彼特、莫基尔	范式转换、创造性破坏
方法	珀尔、坎贝尔、费舍尔	因果阶梯、内部/外部效度
证据	波普尔、拉卡托斯	证伪主义、研究纲领
应用	布什、莫基尔	创新管道、技术就绪水平
CUDOS	默顿	科学规范（共有主义、普遍主义等）