Deep Learning Orchestrator (Enhanced)

Automated deep learning workflow using NotebookLM as the core engine. Handles material collection, knowledge synthesis, and multi-format artifact generation with intelligent intent detection.

🎯 Quick Start

User provides a topic or materials:

"Help me deeply understand Kubernetes architecture"
"Research this article: https://..."
"Learn from this video: https://youtube.com/..."
"帮我学习 Rust 内存安全"
"总结一下这篇文章的核心观点"

Auto-executed workflow:

1. Detect learning intent from context and keywords
2. Collect materials (via jina-reader, bilibili-subtitle, ultimate-search, etc.)
3. Create NotebookLM notebook and upload sources
4. Generate all artifacts in parallel (report, podcast, slides, video, quiz, flashcards)
5. Download and send to user via Feishu + save to Obsidian

🏗️ Architecture

deep-learning 是 orchestrator，不是 all-in-one：

deep-learning (orchestrator)
    ↓
content-bridge (内容摄取统一入口)
    ↓
独立摄取 skills (weixin/bilibili/youtube/web/document)
    ↓
NotebookLM (知识合成引擎)
    ↓
多格式产物 (report/podcast/slides/quiz)

📦 Dependencies

• content-bridge: 内容摄取路由层
• bilibili-subtitle: B站字幕提取（通过 content-bridge 调用）
• notebooklm-py: NotebookLM CLI
• ultimate-search: 网络搜索（可选）

🔥 Trigger System

Explicit Triggers

• "帮我学习 X", "深入了解 X", "研究一下 X"
• "深度学习 X", "全面了解 X", "系统学习 X"

Implicit Triggers

• "总结一下", "听不懂", "解释一下"
• "教程", "入门", "指南"
• User shares long article/URL then asks questions

When detected: Ask "是否需要深度学习？我可以生成报告、播客、PPT、测试题等。"

📦 Material Collection

Supported sources:

Source Type	Handler	Auto-Detection
Web articles	jina-reader	https:// URLs
YouTube videos	anything-to-notebooklm	youtube.com, youtu.be
Bilibili videos	bilibili-subtitle	bilibili.com, BV*
X/Twitter posts	jina-reader	x.com, twitter.com
Reddit posts	jina-reader	reddit.com
Medium articles	jina-reader	medium.com
LinkedIn posts	jina-reader	linkedin.com
Documents (PDF/DOCX)	anything-to-notebooklm	File paths
Search keywords	ultimate-search + NotebookLM research	Plain text

Social media handling: All social platform links automatically processed via jina-reader for clean extraction.

🎨 Artifact Generation

Default Artifacts (Always Generated)

• 📄 Study Guide Report (5-15 min) - Comprehensive learning guide
• 🎙️ Audio Podcast (10-20 min) - Deep-dive audio discussion
• 📊 Slide Deck (15-30 min) - Detailed presentation (PDF)
• 🗺️ Mind Map (instant) - Visual knowledge structure
• ❓ Quiz (5-15 min) - Medium difficulty test questions

Optional Artifacts (Configurable)

• 🎬 Video Brief (10-20 min) - Narrated slideshow video (MP4)
• 📊 Infographic (5-10 min) - Visual summary (PNG)
• 🃏 Flashcards (5-10 min) - Memory aid cards (JSON)

Research Modes

Mode	Duration	Use Case
Fast Research	10-20 seconds	Quick multi-angle source collection
Deep Research	2-30 minutes	Comprehensive single-topic deep dive

Default: Deep Research (auto-selected when no materials provided)

🛠️ Scripts

# Main orchestrator
scripts/orchestrate.sh \
  --topic "Kubernetes 架构" \
  --materials "https://..." "https://..." \
  --research-mode deep \
  --artifacts "report audio slides quiz video"

# Config loader
source scripts/config_loader.sh
load_config

# Prompt selector
scripts/prompt_selector.sh --intent summarize
scripts/prompt_selector.sh --category analysis
scripts/prompt_selector.sh --random

📋 Configuration

Edit config/default.conf or set environment variables:

# Artifact generation
DEFAULT_ARTIFACTS="report audio slides mindmap quiz"
ENABLE_VIDEO=true
ENABLE_INFOGRAPHIC=false
ENABLE_FLASHCARDS=true

# Research mode
DEFAULT_RESEARCH_MODE="deep"  # fast | deep

# Obsidian integration
OBSIDIAN_VAULT_PATH="$HOME/obsidian-vault"
ENABLE_OBSIDIAN_INTEGRATION=true

# Timeouts (seconds)
ARTIFACT_TIMEOUT=1800
RESEARCH_TIMEOUT_DEEP=1800
RESEARCH_TIMEOUT_FAST=60

🧠 Prompt Templates

Located in config/prompts/:

Basic (基础提问)

• Summarize key points
• Explain in simple terms
• Relate to known concepts
• Provide practical examples
• Create teaching outline

Analysis (深度分析)

• Identify core controversies
• Compare different schools of thought
• Trace historical evolution
• Predict future trends
• Clarify common misconceptions

Practical (实用场景)

• Workplace application
• Personalized advice
• Learning path design
• Resource recommendations

Creative (创意生成)

• Training curriculum design
• Beginner-friendly adaptation
• Test question generation
• Analogy-based explanations

Usage:

# Select prompt by intent
./prompt_selector.sh --intent summarize
./prompt_selector.sh --intent compare
./prompt_selector.sh --intent teach

# Select by category
./prompt_selector.sh --category analysis

# Random prompt for inspiration
./prompt_selector.sh --random

⚡ Execution Pattern

Use subagents for long-running tasks:

• Source processing (30s-10min per source)
• Artifact generation (5-45min per artifact)
• Research (2-30min for deep mode)

Main conversation continues while subagents work in background.

📤 Output & Delivery

All artifacts are:

1. Downloaded to ~/.openclaw/workspace/deep-learning-output/<notebook-id>/
2. Sent to user via Feishu with formatted summary
3. Saved to Obsidian Inbox (if enabled)
4. Quiz questions sent as interactive message (if supported)

Feishu message template:

📚 深度学习完成：[主题]

✅ 已保存到 Obsidian Inbox
📂 Research Mode: deep

产物：
📄 学习报告
📊 PPT 讲义
🎙️ 音频播客
❓ 测试题
🎬 视频解说

🔗 NotebookLM: https://notebooklm.google.com/notebook/[id]

🔧 Error Handling

Source processing fails:

• Log warning, continue with successful sources
• Minimum 1 source required to proceed
• Retry once for transient failures

Artifact generation fails:

• Retry once after 5 minutes
• If still fails, skip that artifact and continue
• Report which artifacts succeeded/failed

Rate limiting:

• Wait 10 minutes, retry once
• If persistent, suggest manual retry later
• Provide notebook URL for manual access

📊 Time Estimates

Phase	Typical Duration
Material collection	1-5 min
Source processing	2-10 min
Fast Research	10-20 sec
Deep Research	2-30 min
Report generation	5-15 min
Podcast generation	10-20 min
Slide generation	15-30 min
Video generation	10-20 min
Quiz generation	5-15 min
Download & send	1-2 min

Total: 10-60 minutes (most work happens in parallel via subagents)

📝 Dependencies

Required skills:

• notebooklm - Core NotebookLM operations
• jina-reader - Web page extraction (social media support)

Optional skills (auto-detected):

• bilibili-subtitle - Bilibili video transcription
• ultimate-search - Web search enhancement
• anything-to-notebooklm - Multi-format document support

🎯 Best Practices

1. Always specify topic clearly - Better input = better output
2. Provide materials when available - More control over sources
3. Use Fast Research for quick overviews - Save time on simple topics
4. Use Deep Research for complex subjects - Comprehensive coverage
5. Enable video for visual learners - Great for presentations
6. Review quiz questions - Adjust difficulty if needed

🚀 Examples

Example 1: Learn from URL

User: "帮我学习这篇文章 https://example.com/kubernetes
Assistant: 🚀 Starting deep learning workflow...
[Collects article, creates notebook, generates artifacts]
✅ 深度学习完成：Kubernetes 架构

Example 2: Keyword Research

User: "我想深入了解 Rust 的所有权机制"
Assistant: 🔬 No materials provided, activating Deep Research mode...
[Searches for 15+ sources, creates comprehensive notebook]
✅ 深度学习完成：Rust 所有权机制

Example 3: Video Tutorial

User: "把这个 B 站视频做成学习材料 https://bilibili.com/video/BV1xx"
Assistant: 📥 Collecting materials...
[Extracts subtitles, generates video brief + slides + quiz]
✅ 深度学习完成：[Video topic]

Example 4: Quick Summary

User: "总结一下这个概念，太快了看不懂"
Assistant: ⚡ Fast Research mode activated...
[Quick 10-20 second research, generates summary report]
✅ 快速总结完成