Satya: An Offline-First, Low-Resource AI Learning Platform for Nepali Education

概述

Overview

Satya 是一个专为尼泊尔教育环境优化的本地优先学习平台。它利用检索增强生成技术和 Phi 1.5 语言模型，提供内容索引和查询功能，在离线和在线环境中功能完全一致。该系统专为在资源有限的硬件上运行而设计，确保无论基础设施如何，都能实现访问。

Satya is a local-first learning platform optimized for the Nepali educational context. It provides content indexing and query capabilities using Retrieval-Augmented Generation (RAG) and the Phi 1.5 language model, functioning identically in both offline and online environments. The system is engineered to run on hardware with limited resources, ensuring access regardless of infrastructure.

使命与愿景

Mission & Vision

我们的使命

Our Mission

通过让智能辅导惠及每一位学生，无论其地理位置、网络连接或硬件资源如何，来普及人工智能驱动的教育。

To democratize AI-powered education by making intelligent tutoring accessible to every student, regardless of their location, internet connectivity, or hardware resources.

Satya 通过提供一个自包含的AI辅导系统，解决了农村教育中的基础设施限制。它消除了对高速互联网或现代设备的需求，确保偏远地区的学生能够获得与网络发达地区学生相同的学习资源。

Satya addresses the infrastructure limitations in rural education by offering a self-contained AI tutoring system. It eliminates the need for high-speed internet or modern devices, ensuring students in remote areas have access to the same learning resources as those in connected environments.

教育鸿沟

The Educational Divide

Important
139万 尼泊尔中学生缺乏可靠的互联网接入。虽然 79.3% 的城市家庭拥有互联网，但只有 17.4% 的农村家庭拥有。

Important
1.39 million secondary students in Nepal lack reliable internet access. While 79.3% of urban households have internet, only 17.4% of rural households do.

农村学生的现实：

The Reality for Rural Students:

• 连接性差距 - 79.3% 城市家庭已连接 vs 仅 17.4% 农村家庭
• 设备访问 - 仅 3% 的农村儿童能同时使用电脑和互联网
• 硬件限制 - 学校依赖2015年生产的4GB+内存的电脑
• 学校基础设施 - 仅 12% 的公立学校拥有正常运行的IT连接
• 成本障碍 - 软件订阅预算为0美元 vs 每月20美元的云端工具

• Connectivity Gap - 79.3% urban households connected vs only 17.4% rural households
• Device Access - Only 3% of rural children have access to both a computer and internet
• Hardware Limitations - Schools rely on computers with 4GB+ RAM from 2015
• School Infrastructure - Only 12% of public schools have functioning IT connectivity
• Cost Barriers - $0 budget for software subscriptions vs $20/month cloud tools

结果： 系统性被排除在AI革命之外。现有的教育科技解决方案假设了农村课堂根本不具备的基础设施。

The Result: A systematic exclusion from the AI revolution. Existing ed-tech solutions assume infrastructure that simply doesn't exist in rural classrooms.

我们的解决方案：离线优先的AI教育

Our Solution: Offline-First AI Education

Satya 通过 彻底的易用性 打破这些障碍：

Satya breaks down these barriers through radical accessibility:

1. 离线优先架构一种软件设计理念，确保应用在无网络连接时仍能完全正常运行，网络连接仅作为可选功能而非必需条件。

1. Offline-First Architecture

• 无需互联网连接即可实现完整功能
• 一次性下载，终身离线使用
• 无云端依赖或订阅费用

• Complete functionality without internet connection
• One-time download, lifetime offline use
• No cloud dependencies or subscription fees

2. 低资源优化

2. Low-Resource Optimization

• 在4GB内存和仅CPU处理下运行
• 适用于农村学校常见的十年老旧硬件
• 针对第三代英特尔i3处理器优化

• Runs on 4GB RAM with CPU-only processing
• Works on decade-old hardware common in rural schools
• Optimized for 3rd gen Intel i3 processors

3. 智能RAG系统

3. Intelligent RAG System

• 本地向量数据库用于内容发现
• 同时搜索教科书和教师笔记
• 无需外部API即可提供上下文感知的答案

• Local vector database (ChromaDBAn open-source vector database designed for storing and querying embeddings.) for content discovery
• Searches textbooks and teacher notes simultaneously
• Context-aware answers without external APIs

4. 单一模型效率

4. Single Model Efficiency

• 微软Phi 1.5模型微软开发的小型语言模型，参数量约800MB，专为资源受限环境优化，在Satya平台中用于所有AI任务处理。处理所有AI任务
• 无需多个模型或复杂流水线
• 针对有限资源优化的快速推理

• Microsoft Phi 1.5 (800MB) handles all AI tasks
• No multiple models or complex pipelines
• Fast inference optimized for limited resources

5. 社区驱动的内容

5. Community-Driven Content

• 教师贡献本地课程材料
• 支持PDF、扫描文档、手写笔记
• 透明、协作的内容工作流

• Teachers contribute local curriculum materials
• Supports PDFs, scanned documents, handwritten notes
• Transparent, collaborative content workflow

影响与覆盖范围

Impact & Reach

目标受益者：

Target Beneficiaries:

• 主要： 139万+中学生
• 次要： 农村地区的公立学校
• 第三级： 基础设施有限的远程学习中心

• Primary: 1.39 million+ secondary students (Grades 8-12)
• Secondary: Public Schools in rural districts
• Tertiary: Remote learning centers with limited infrastructure

可衡量的成果：

Measurable Outcomes:

• 可访问性： 无需互联网即可24/7获得AI辅导
• 公平性： 农村和城市地区享有同等质量的教育
• 可负担性： 初始设置后零持续成本
• 可扩展性： 一名教师可为数千名学生准备内容
• 可持续性： 社区维护、开源平台

• Accessibility: AI tutoring available 24/7 without internet
• Equity: Same quality education in rural and urban areas
• Affordability: Zero ongoing costs after initial setup
• Scalability: One teacher can prepare content for thousands of students
• Sustainability: Community-maintained, open-source platform

设计理念

Design Philosophy

Note
Satya中的每一个技术决策都优先考虑 可访问性而非性能，简洁性而非功能，以及 离线能力而非云端便利性。

Note
Every technical decision in Satya prioritizes accessibility over performance, simplicity over features, and offline capability over cloud convenience.

核心原则：

Core Principles:

1. 离线优先 - 互联网是可选项，非必需
2. 资源意识 - 针对学生实际拥有的硬件进行优化
3. 赋能教育者 - 教师而非公司控制内容
4. 以学生为中心 - 学习体验重于技术复杂性
5. 社区驱动 - 透明、协作的开发

1. Offline-First - Internet is optional, not required
2. Resource-Conscious - Optimized for the hardware students actually have
3. Educator-Empowered - Teachers control content, not corporations
4. Student-Centered - Learning experience over technical complexity
5. Community-Driven - Transparent, collaborative development

为何这很重要

Why This Matters

教育是一项基本权利，而非特权。 人工智能驱动的学习应该惠及每一位学生，而不仅仅是那些位于网络发达城市中心的学生。

Education is a fundamental right, not a privilege. AI-powered learning should be accessible to every student, not just those in well-connected urban centers.

Satya 证明了 智能、个性化的教育并不需要昂贵的基础设施。通过周密的工程设计和社区协作，我们可以将AI辅导带给最需要的学生——那些目前被排除在AI革命之外的学生。

Satya proves that intelligent, personalized education doesn't require expensive infrastructure. With thoughtful engineering and community collaboration, we can deliver AI tutoring to the students who need it most—those currently excluded from the AI revolution.

这不仅关乎技术，更关乎教育公平。

This isn't just about technology. It's about educational justice.

核心特性

Key Features

面向学生的功能

Student-Facing Features

内容检索

Content Retrieval (RAG)

• 语义搜索基于语义理解而非关键词匹配的搜索技术，能理解查询意图和内容含义。 - ChromaDBAn open-source vector database designed for storing and querying embeddings.向量数据库检索相关内容
• 上下文处理 - 在生成答案前引用适当的学习材料
• 多源搜索 - 同时搜索教科书和教师笔记
• 过滤 - 应用学科感知约束
• 状态反馈 - 实时进度更新

• Semantic Search - ChromaDBAn open-source vector database designed for storing and querying embeddings. vector database retrieves relevant content
• Context Handling - References appropriate study materials before generation
• Multi-Source - Searches both textbooks and teacher notes
• Filtering - Applies subject-aware constraints
• Status Feedback - Real-time progress updates

Tip
RAG系统会自动查询教科书和笔记集合，以提供全面的答案。

Tip
The RAG system automatically queries both textbooks and notes collections for comprehensive answers.

学习辅助

Learning Assistance

• 响应生成 - 生成简洁的3-4句解释
• 令牌流式传输 - 低延迟字符显示
• 置信度指标 - 显示低置信度生成的警告
• 输入规范化 - 自动纠正大小写和格式

• Response Generation - Produces concise 3-4 sentence explanations (100-150 tokens)
• Token Streaming - Low-latency character display
• Confidence Metrics - Displays warnings for low-confidence generations (< 70%)
• Input Normalization - Auto-corrects case and formatting

视觉解释

Visual Explanations

• ASCII图表 - 从文本生成结构、流程和流程图
• 年级感知库 - 预建的适合年龄的图表库
• 自然触发 - 智能逻辑，仅在视觉上有帮助时显示图表
• 模式识别 - 从RAG内容中识别循环、层次结构和顺序步骤
• 零依赖 - 纯文本渲染，无需外部库

• ASCII Diagrams - Generates structural, process, and flowchart diagrams from text
• Grade-Aware Library - Pre-built library of age-appropriate diagrams (Grades 8-12)
• Natural Triggering - Intelligent logic to show diagrams only when visually helpful
• Pattern Recognition - Identifies cycles, hierarchies, and sequential steps from RAG content
• Zero-Dependency - Pure text rendering requiring no external libraries

用户界面

User Interfaces

• 命令行界面 - 带有进度指示器的丰富终端界面
• 图形用户界面 - 具有响应式设计的现代CustomTkinter界面
• 进度跟踪 - 详细的分析和可视化
• 导出/导入 - 保存和恢复学习进度

• Command-Line Interface (CLI) - Rich terminal interface with progress indicators
• Graphical User Interface (GUI) - Modern CustomTkinter interface with responsive design
• Progress Tracking - Detailed analytics and visualizations
• Export/Import - Save and restore learning progress

面向教师的功能

Teacher-Facing Features

内容管理

Content Management

• 通用摄取 - 单一脚本处理PDF、扫描文档、手写笔记
• 自动检测 - 自动检测内容类型并应用适当的处理
• OCR支持 - 扫描PDF使用Tesseract，手写笔记使用EasyOCR
• 智能分块 - 512个令牌，10%重叠，实现最佳检索
• 元数据提取 - 从文件夹结构自动检测年级和学科

• Universal Ingestion - Single script handles PDFs, scanned documents, handwritten notes
• Auto-Detection - Automatically detects content type and applies appropriate processing
• OCR Support - Tesseract for scanned PDFs, EasyOCR for handwritten notes
• Smart Chunking - 512 tokens with 10% overlap for optimal retrieval
• Metadata Extraction - Auto-detects grade and subject from folder structure

Note
使用 scripts/ingest_content.py 进行所有内容摄取。它取代了之前所有的摄取脚本。

Note
Use scripts/ingest_content.py for all content ingestion. It replaces all previous ingestion scripts.

系统架构

System Architecture

高层架构

High-Level Architecture

Important
架构已在2.0版本中更新。单一的Phi 1.5模型微软开发的小型语言模型，参数量约800MB，专为资源受限环境优化，在Satya平台中用于所有AI任务处理。取代了之前的多模型方法。

Important
Architecture has been updated in version 2.0. Single Phi 1.5 model replaces previous multi-model approach.

graph TB
    subgraph "Student Interface Layer"
        CLI[CLI Interface]
        GUI[GUI Interface]
    end
    
    subgraph "Application Layer"
        RAG[RAG Retrieval Engine]
        DS[Diagram Service]
        PM[Progress Manager]
    end
    
    subgraph "AI Layer"
        MH[Model Handler]
        PH[Phi 1.5 Handler]
    end
    
    subgraph "Data Layer"
        CDB[(ChromaDB)]
        DL[(Diagram Library)]
        PROG[Progress Data]
    end
    
    CLI --> RAG
    CLI --> MH
    CLI --> DS
    GUI --> RAG
    GUI --> MH
    GUI --> DS
    
    PM --> PROG
    RAG --> CDB
    MH --> PH
    PH --> CDB
    DS --> DL
    DS -.-> MH

组件架构

Component Architecture

1. 通用内容摄取

1. Universal Content Ingestion

实现 (scripts/ingest_content.py)

Implementation (scripts/ingest_content.py)

• 自动检测 - 识别文本PDF、扫描PDF或手写笔记
• 多格式支持 - PDF, TXT, MD, JSONL
• OCR模式 - 自动检测、强制或从不
• 智能处理 - 文本使用PyMuPDF，图像使用Tesseract/EasyOCR

• Auto-Detection - Identifies text PDFs, scanned PDFs, or handwritten notes
• Multi-Format Support - PDF, TXT, MD, JSONL
• OCR Modes - Auto-detect, force, or never
• Smart Processing - PyMuPDF for text, Tesseract/EasyOCR for images

处理流程：

Processing Flow:

Input Files (PDF/TXT/MD)
    ↓
Content Type Detection
    ↓
Extraction (PyMuPDF/Tesseract/EasyOCR)
    ↓
Smart Chunking (512 tokens, 10% overlap)
    ↓
Embedding Generation (all-MiniLM-L6-v2)
    ↓
ChromaDB Storage

技术规格与对比

Technical Specifications & Comparison

核心组件规格

Core Component Specifications

常见问题（FAQ）

Satya平台需要联网才能使用吗？

不需要。Satya采用离线优先架构一种软件设计理念，确保应用在无网络连接时仍能完全正常运行，网络连接仅作为可选功能而非必需条件。，一次下载即可终身离线使用，无需互联网连接或云端依赖，专门为网络条件差的农村地区设计。

Satya对电脑配置要求高吗？

要求很低。Satya专为4GB内存的旧硬件优化，可在仅CPU处理的条件下运行，兼容农村学校常见的十年老旧电脑（如第三代i3处理器）。

Satya如何解决农村教育资源不足的问题？

通过集成RAG技术和Phi 1.5模型微软开发的小型语言模型，参数量约800MB，专为资源受限环境优化，在Satya平台中用于所有AI任务处理。，Satya在本地提供智能辅导，可同时搜索教科书和教师笔记，让偏远地区学生获得与城市相同的学习资源。

Component	Specification/Model	Key Features	Resource Footprint
AI Model	Microsoft Phi 1.5	Single model for all tasks, 800MB size	CPU-only inference
Vector Database	ChromaDBAn open-source vector database designed for storing and querying embeddings.	Local storage, semantic search	Low memory footprint
Text Embedding Model	all-MiniLM-L6-v2	Generates vectors for content chunks