01_Source_Management.md

uid	202601190828
created	2026-01-15 09:01
updated	2026-01-19 08:28
tags	inbox programmatic-upgrade
status	draft
ai_processed	true
model	AutoAgent-Ralph-v1
title	01_Source_Management

제3강. 50개의 두뇌를 하나로: 대규모 소스 관리의 기술

👋 안녕하세요! 여러분의 AI 연구 파트너, 안티그래비티입니다.

지난 시간, 우리는 연구소를 세팅하고 첫 번째 가동을 마쳤습니다. 이제 본격적으로 연구소에 **'원재료'**를 채워 넣을 시간입니다.

NotebookLM이 다른 AI와 가장 다른 점이 뭘까요? 바로 **"내가 준 것만 먹고 자란다"**는 점입니다. ChatGPT는 전 세계 인터넷 데이터를 먹고 자랐지만, NotebookLM은 여러분이 넣어준 PDF, 여러분이 고른 유튜브 영상만을 신뢰합니다. 그래서 **"어떤 소스를, 어떻게 관리하느냐"**가 결과물의 퀄리티를 100% 좌우합니다.

이번 강의에서는 2025년 가장 핫했던 **'앵커 문서 전략(Anchor Document Strategy)'**을 포함해, 50개 이상의 방대한 소스를 프로처럼 다루는 비법을 알려드리겠습니다.

---

📚 1. 50개 소스, 어떻게 채울까? (The Big Container)

Module 1. Procurement of Raw Data: Sourcing for English Excellence

(Cleaning OCR and Standardizing CSAT Passages)

👋 Hello, Data Chemists!

In English education, the quality of your source is everything. If you upload a blurry OCR scan of a EBS workbook with messy headers and footers, your AI will produce messy results.

In this module, we learn how to "clean" English educational data to ensure 100% accuracy in logical analysis.

---

🧹 1. The OCR Cleanup: English Workbook Edition

Most English teachers work with scanned PDFs or captured images from workbooks. These are full of "noise":

• Header/Footer Noise: "2026 EBS Su-neung-teukgang Page 42".
• Question Numbering: "31. 다음 글의 주제로 가장 적절한 것은?".
• Vocab Glossaries: Small footnotes at the bottom of the page.

The Solution: "Focus Cropping"

Use a PDF tool (like Acrobat or PDFElement) to crop the margins. If you only want the AI to analyze the Paragraph Logic, remove the question stems and the footnotes before uploading.

---

🎭 2. Managing Bilingual Content

English classrooms in Korea are primarily bilingual. You need to manage how NotebookLM sees English vs. Korean text.

Best Practice: The Layered Sourcing

1. Level 1 (English Only): Upload the pure English paragraph as a .txt file for Logic Mapping.
2. Level 2 (Bilingual): Upload the version with Korean translations for "Explanation Generation" and "Vocabulary Mapping".
3. Level 3 (Logic Anchor): Upload our 00_Anchor_Comparative_Lens_EN.txt to guide the AI's "Logical Voice".

---

📊 3. CSAT Data Tables: Analyzing Mock Exams

Did you just get the results of a National Mock Exam (모의고사)? Don't just look at the scores.

1. Upload the PDF of the student score results.
2. Use the Data Table feature to convert the PDF chart into a CSV.
3. Ask: "Which specific question type (Inference, Blank completion, Order) had the lowest accuracy?"
4. Result: Instant personalized clinic data for your entire class.

---

✂️ 4. Segmentation by CSAT Question Type

Instead of one giant "EBS Workbook.pdf", split your sources by Question Type:

• Folder A: Blank Completion (빈칸 추론)
• Folder B: Paragraph Ordering (순서 배열)
• Folder C: Sentence Insertion (문장 삽입)

By isolating the sources, you prevent the AI from confusing the distinct logical patterns required for each type.

---

🧪 5. Today's Mini-Lab: Source Refinement

📌 Mission: "The Clean Link"

Step 1: Find a PDF passage that is "messy" (contains page numbers, logos, or multiple questions). Step 2: Create a cleaned-up version of just the body text in a .txt file. Step 3: Upload both to a new Notebook and ask: "Summarize the logic of the paragraph." Step 4: Compare the results. Notice how the cleaned version leads to a much more "elegant" and accurate summary.

Pro-tip: A clean source is the first step to becoming a "Prompt Grandmaster". 기대해주세요!