WEBVTT
Kind: captions
Language: zh-Hant

00:00:00.000 --> 00:00:01.700
如果你用過AI家教

00:00:01.920 --> 00:00:03.660
大概會有一個熟悉的感覺

00:00:04.240 --> 00:00:07.780
它回答得很快可是它不一定真的記得你卡

00:00:07.780 --> 00:00:10.160
在哪裡今天的GitHub雷達

00:00:10.560 --> 00:00:12.360
我們來看DeepTutor

00:00:12.360 --> 00:00:15.520
它不是把大型語言模型包成聊天框而已

00:00:16.060 --> 00:00:19.340
它想做的是一個終身個人化輔導工作區

00:00:19.840 --> 00:00:21.860
DeepTutor的問題意識很直接

00:00:22.320 --> 00:00:23.840
學習不是一次問答

00:00:24.240 --> 00:00:27.800
學生會留下錯題筆記閱讀材料知識庫

00:00:27.800 --> 00:00:29.680
還有一堆長期目標

00:00:30.040 --> 00:00:32.420
如果這些東西散在不同工具裡

00:00:32.740 --> 00:00:35.380
AI很難知道下一步該怎麼幫你

00:00:35.780 --> 00:00:37.260
它可能答對這一題

00:00:37.540 --> 00:00:40.820
卻完全不知道你為什麼一直卡在同一類題

00:00:40.820 --> 00:00:44.600
型所以DeepTutor把Chat Quiz Research Visualize

00:00:45.180 --> 00:00:46.860
Solve和Mastery Path接

00:00:46.860 --> 00:00:48.580
在同一個agent loop上

00:00:48.980 --> 00:00:50.500
你換的是學習任務

00:00:50.740 --> 00:00:52.300
不是換一顆陌生的腦

00:00:52.300 --> 00:00:55.620
它也把Knowledge Center Book Co-Writer Question

00:00:55.620 --> 00:00:57.960
Bank和Memory放進同一個工作

00:00:57.960 --> 00:01:00.560
區上下文有機會跟著人走

00:01:00.900 --> 00:01:03.040
而不是每次對話都重開機

00:01:03.460 --> 00:01:04.640
Knowledge Center這塊

00:01:04.640 --> 00:01:07.440
很重DeepTutor不只做一種RAG

00:01:07.960 --> 00:01:11.933
而是讓你在LlamaIndex PageIndex GraphRAG

00:01:11.933 --> 00:01:12.747
LightRAG

00:01:12.747 --> 00:01:13.980
LightRAG Server

00:01:14.460 --> 00:01:15.920
甚至Obsidian之間選

00:01:16.320 --> 00:01:18.020
這對學習其實很重要

00:01:18.020 --> 00:01:20.500
課本論文筆記講義

00:01:20.500 --> 00:01:23.280
不一定適合同一種檢索方式

00:01:23.340 --> 00:01:26.980
硬套一個向量庫常常只是看起來很聰明

00:01:27.320 --> 00:01:29.260
更關鍵的是它的記憶設計

00:01:29.600 --> 00:01:33.260
README說DeepTutor有L1 L2 L3三

00:01:33.260 --> 00:01:36.160
層Memory不是藏在背後的黑箱向量

00:01:36.160 --> 00:01:38.300
庫L1保留事件trace

00:01:38.760 --> 00:01:41.080
L2整理各個surface的事實

00:01:41.420 --> 00:01:43.440
L3再做跨場景綜合

00:01:43.440 --> 00:01:47.490
理想狀態下你可以追回一個個人化判斷到

00:01:47.490 --> 00:01:51.500
底從哪裡來論文摘要把這件事說得更研究

00:01:51.500 --> 00:01:54.540
一點DeepTutor結合引用可追溯

00:01:54.540 --> 00:01:57.300
的解題輔導難度校準的出題

00:01:57.620 --> 00:01:59.240
以及動態learner memory

00:01:59.580 --> 00:02:02.420
也就是說它不只回答這題怎麼算

00:02:02.780 --> 00:02:06.660
還要判斷你現在該練哪種難度該補哪段概

00:02:06.660 --> 00:02:09.540
念論文裡還提出TutorBench

00:02:09.540 --> 00:02:13.040
用帶個人檔案的學生模擬器做互動評測

00:02:13.040 --> 00:02:15.040
這比只看單題答案

00:02:15.040 --> 00:02:19.420
更接近真實教學摘要報告personalized metrics平

00:02:19.420 --> 00:02:22.880
均提升10.8 general agentic reasoning在

00:02:22.880 --> 00:02:24.460
五個backbone models上

00:02:24.460 --> 00:02:28.100
提升29.4這些數字是論文報告

00:02:28.340 --> 00:02:30.220
不是我們本機重跑的結果

00:02:30.700 --> 00:02:33.420
回到產品面DeepTutor的介面比

00:02:33.420 --> 00:02:35.360
較像一個學習作業系統

00:02:35.360 --> 00:02:39.140
Chat是入口但旁邊還有Partners My Agents

00:02:39.820 --> 00:02:40.500
Co-Writer

00:02:40.500 --> 00:02:42.580
Book Knowledge Center Learning Space

00:02:42.580 --> 00:02:43.140
和Settings

00:02:43.710 --> 00:02:45.060
Partners很有意思

00:02:45.480 --> 00:02:48.800
它讓持久化的IM夥伴接在同一個brain上

00:02:49.140 --> 00:02:51.620
訊息從Web或即時通訊進來

00:02:51.620 --> 00:02:54.600
都可以變成同一套ChatOrchestrator的

00:02:54.600 --> 00:02:57.460
turn My Agents則把Claude Code Codex

00:02:58.015 --> 00:02:58.570
Gemini

00:02:58.570 --> 00:03:00.400
Kimi opencode或MiMo這

00:03:00.400 --> 00:03:02.240
類外部agent接進聊天

00:03:02.240 --> 00:03:06.820
對開發者來說學習研究和實作可以互相呼

00:03:06.820 --> 00:03:08.820
叫Book也不是普通PDF

00:03:09.500 --> 00:03:12.160
DeepTutor把材料編成互動式活

00:03:12.160 --> 00:03:17.360
書可以有章節測驗圖解動畫概念圖和Page Chat

00:03:17.820 --> 00:03:19.920
Co-Writer則像一張寫作台

00:03:20.320 --> 00:03:24.140
你選一段文字它可以根據知識庫或網路證

00:03:24.140 --> 00:03:28.120
據改寫然後用差異對照讓你接受或拒絕

00:03:28.120 --> 00:03:30.540
Visualize和Math Animator讓

00:03:30.540 --> 00:03:34.180
抽象概念變成圖表互動模組或動畫

00:03:34.600 --> 00:03:37.940
數學資料結構模型流程這類內容

00:03:38.260 --> 00:03:39.960
純文字有時真的不夠

00:03:40.340 --> 00:03:42.060
設定面也做得很集中

00:03:42.460 --> 00:03:46.740
模型embedding搜尋TTS STT Image Generation

00:03:47.340 --> 00:03:48.960
Video Generation文件

00:03:48.960 --> 00:03:50.540
解析工具和MCP

00:03:50.940 --> 00:03:52.480
都在Settings裡管理

00:03:52.820 --> 00:03:56.880
安裝路徑有四種你可以用PyPI裝完整

00:03:56.880 --> 00:03:57.920
本地Web app

00:03:58.260 --> 00:04:00.360
也可以source install Docker單

00:04:00.360 --> 00:04:02.020
容器或只裝CLI

00:04:02.550 --> 00:04:04.580
Docker版本有個實用細節

00:04:04.980 --> 00:04:06.880
通常只需要開前端3782

00:04:07.667 --> 00:04:09.460
Next.js middleware會

00:04:09.460 --> 00:04:12.300
在容器內把API和WebSocket轉

00:04:12.300 --> 00:04:13.860
到FastAPI後端

00:04:14.260 --> 00:04:16.840
CLI是給agent很好用的入口

00:04:17.240 --> 00:04:19.520
deeptutor run可以用JSON串

00:04:19.520 --> 00:04:22.160
接單次能力deeptutor chat則

00:04:22.160 --> 00:04:23.800
是人類用的REPL

00:04:23.800 --> 00:04:27.620
最新README也提到OpenAI Codex OAuth

00:04:27.620 --> 00:04:30.080
可以用自己的ChatGPT plan登

00:04:30.080 --> 00:04:31.900
入DeepTutor的模型設定

00:04:32.300 --> 00:04:34.820
對不想管理傳統API key的人

00:04:35.220 --> 00:04:38.140
這一點很吸引人但這裡要踩煞車

00:04:38.540 --> 00:04:40.040
DeepTutor功能很大

00:04:40.360 --> 00:04:42.040
也代表導入成本不低

00:04:42.380 --> 00:04:45.000
你要配置模型embedding文件解

00:04:45.000 --> 00:04:47.220
析知識庫引擎權限

00:04:47.520 --> 00:04:49.640
還可能要處理多用戶隔離

00:04:49.640 --> 00:04:53.220
如果資料來源亂索引錯模型不穩

00:04:53.560 --> 00:04:55.540
個人化反而會變成噪音

00:04:55.900 --> 00:04:57.740
RAG也不是單選題

00:04:58.100 --> 00:05:01.170
GraphRAG LightRAG PageIndex

00:05:01.170 --> 00:05:02.120
LlamaIndex

00:05:02.120 --> 00:05:04.580
各自適合不同資料和問題

00:05:04.980 --> 00:05:07.060
不是打開就一定比較聰明

00:05:07.400 --> 00:05:10.280
另外多用戶部署雖然支援隔離

00:05:10.600 --> 00:05:14.980
但管理者還是要負責模型授權知識庫grants工

00:05:14.980 --> 00:05:17.200
具策略和程式執行風險

00:05:17.200 --> 00:05:19.980
所以我會把DeepTutor看成一個

00:05:19.980 --> 00:05:23.300
開源教育工作台而不是零設定玩具

00:05:23.720 --> 00:05:27.740
它適合願意整理材料建立知識庫記錄學習

00:05:27.740 --> 00:05:29.760
狀態的人如果你是老師

00:05:30.180 --> 00:05:33.460
可以拿它設計題庫回饋和補救教學

00:05:33.860 --> 00:05:37.630
如果你是學生可以把課本筆記和問答串起

00:05:37.630 --> 00:05:39.420
來如果你是開發者

00:05:39.780 --> 00:05:42.660
它更像一個agent-native learning platform的

00:05:42.660 --> 00:05:44.620
參考實作不適合誰

00:05:44.940 --> 00:05:48.540
如果你只想要一個最輕量問一句答一句的

00:05:48.540 --> 00:05:50.130
chatbot DeepTutor可

00:05:50.130 --> 00:05:54.300
能太重但如果你在意的是長期學習脈絡引

00:05:54.300 --> 00:05:57.480
用來源工具協作和可追溯記憶

00:05:57.820 --> 00:05:59.500
那它就值得花時間拆

00:05:59.920 --> 00:06:01.180
一句話總結DeepTutor

00:06:01.670 --> 00:06:05.850
它想把AI家教從會回答推到看得見你怎

00:06:05.850 --> 00:06:09.860
麼學對我來說最有價值的不是某一個炫技

00:06:09.860 --> 00:06:14.180
功能而是知識記憶工具和互動都在同一條

00:06:14.180 --> 00:06:18.140
線上如果你想看更多這類開源工具的拆解

00:06:18.480 --> 00:06:20.160
記得訂閱GitHub雷達

00:06:20.480 --> 00:06:21.460
下支影片見