SAGE Libraries Package (sage-libs)

📋 Overview

sage-libs 是 SAGE 框架的算法库层，定位为 接口/注册表层 (Interface Layer)。

🧭 Governance / 团队协作制度

本仓库协作规范与质量门槛请参考：

• .github/copilot-instructions.md
• quickstart.sh
• hooks/pre-commit
• hooks/pre-push

核心设计原则：

• 📦 轻量级接口：定义抽象基类和工厂函数
• 🔌 可插拔实现：重型实现迁出为独立 PyPI 包 (isage-*)
• 🏗️ 注册表模式：通过 register_* / create_* 动态加载实现

🏗️ Architecture

sage-libs (Interface Layer)
    ├── agentic/interface/     → isage-agentic (Agent framework)
    ├── rag/interface/         → isage-rag (RAG toolkit)
    ├── finetune/interface/    → isage-finetune (Fine-tuning)
    ├── eval/interface/        → isage-eval (Evaluation)
    ├── privacy/interface/     → isage-privacy (Privacy/Unlearning)
    ├── safety/interface/      → isage-safety (Guardrails)
    ├── amms/interface/        → isage-amms (AMM algorithms)
    └── foundation/            → Built-in utilities (no external deps)

ANNS 已完全迁移到独立包 isage-anns（导入路径：sage_anns），不再由 sage-libs 承载。

📚 Five Core Domains

1. 🤖 Agentic (Agent Framework)

接口：sage.libs.agentic.interface

from sage.libs.agentic.interface import (
    BaseAgent, BasePlanner, BaseToolSelector,
    create_agent, create_planner, register_agent
)

# Register implementation (from isage-agentic)
register_agent("react", ReactAgent)

# Create via factory
agent = create_agent("react", tools=[...])

2. 📖 RAG (Retrieval-Augmented Generation)

接口：sage.libs.rag.interface

from sage.libs.rag.interface import (
    DocumentLoader, Retriever, RAGPipeline,
    create_loader, create_retriever, create_pipeline
)

loader = create_loader("pdf")
retriever = create_retriever("faiss", dimension=768)

3. 🔧 Fine-tuning

接口：sage.libs.finetune.interface

from sage.libs.finetune.interface import (
    FineTuner, TrainingStrategy, TrainingConfig, LoRAConfig,
    create_trainer, create_strategy
)

strategy = create_strategy("lora")
trainer = create_trainer("lora", model_name="gpt2")

4. 📊 Evaluation

接口：sage.libs.eval.interface

from sage.libs.eval.interface import (
    BaseMetric, BaseLLMJudge, MetricResult,
    create_metric, create_judge
)

metric = create_metric("accuracy")
judge = create_judge("faithfulness", model="gpt-4")

5. 🔒 Privacy & Safety

Privacy 接口：sage.libs.privacy.interface

Safety 接口：sage.libs.safety.interface

from sage.libs.privacy import create_unlearner, create_mechanism
from sage.libs.safety import create_guardrail, create_jailbreak_detector

unlearner = create_unlearner("sisa", num_shards=5)
guardrail = create_guardrail("llm", model="gpt-4")

注意：sage-libs 仅保留隐私接口与注册表；sage.libs.privacy.unlearning 中的具体 DP / unlearning 实现已迁移到 isage-privacy（导入路径：sage_privacy.*）。

📦 External Packages (isage-*)

🚀 Installation

Basic Installation

# From PyPI
pip install isage-libs

# Development install (in SAGE repo)
pip install -e packages/sage-libs

With Optional Extras

# All interfaces
pip install isage-libs[all]

# Specific domains
pip install isage-libs[agentic]    # Agent framework
pip install isage-libs[rag]         # RAG toolkit
pip install isage-libs[finetune]    # Fine-tuning
pip install isage-libs[eval]        # Evaluation
pip install isage-libs[privacy]     # Privacy/Unlearning
pip install isage-libs[safety]      # Safety/Guardrails

🏛️ Built-in Utilities

These modules are included directly (no external deps):

Foundation (foundation/)

from sage.libs.foundation import (
    text_utils,    # Text processing
    io_utils,      # File I/O helpers
    async_utils,   # Async utilities
)

DataOps (dataops/)

from sage.libs.dataops import (
    text_ops,      # Normalization, truncation
    table_ops,     # DataFrame operations
    json_ops,      # JSON processing
    sampling,      # Sampling strategies
)

Lightweight Safety (safety/)

from sage.libs.safety import (
    content_filter,  # Pattern-based filtering
    pii_scrubber,    # PII detection
    policy_check,    # Policy validation
)

📖 Usage Example

# 1. Define custom implementation
from sage.libs.agentic.interface import BaseAgent, AgentResult, register_agent

class MyAgent(BaseAgent):
    @property
    def name(self) -> str:
        return "my_agent"

    def run(self, task, context=None):
        # Implementation
        return AgentResult(success=True, output="Done")

# 2. Register implementation
register_agent("my_agent", MyAgent)

# 3. Use via factory
from sage.libs.agentic.interface import create_agent
agent = create_agent("my_agent")
result = agent.run("Hello")

📚 Documentation

• Architecture: ../../CHANGELOG.md
• API Reference: docs-public/docs_src/api-reference/sage-libs/
• Tutorials: examples/tutorials/L3-libs/

🔗 Related Packages

• SAGE - Main framework
• sage-benchmark - Evaluation benchmarks
• SageVDB - Vector database
• NeuroMem - Memory system

📄 License

Apache 2.0 License