AI-driven YouTube Video Summarizer
BibiGPTAI 音视频助理
博客首页

How to Build a Personal Video Knowledge Base with BibiGPT + Obsidian (2026 PKM Workflow, 3 Steps)

Hands-on methodology: use BibiGPT to process Bilibili / YouTube / podcast videos, auto-save into Obsidian, and build a sustainable personal video knowledge base in three steps. Includes note template and tag taxonomy.

BibiGPT Team

How to Build a Personal Video Knowledge Base with BibiGPT + Obsidian (2026 PKM Workflow, 3 Steps)

80-word answer: Use BibiGPT as the AI reader for video, and Obsidian as the long-term note vault. Enable "auto-save summary to Obsidian" in the BibiGPT desktop client, and every video you watch becomes an entry in your local Vault. Add a tag taxonomy and backlink structure, and within three months you have a searchable, questionable private knowledge network.

This is a methodology article for knowledge workers, learners, and content creators — not a tool review, but a battle-tested PKM (Personal Knowledge Management) workflow.

1. Why plain note apps fail at video knowledge bases

If you have tried to build a video notebook in Notion, Obsidian, or Evernote, you've hit these issues:

  • Captions aren't copy-pasteable: most platforms don't expose caption export.
  • Timestamps break: hand-typed "10:24 mentioned X" becomes unsearchable later.
  • Search misses: video keywords live in audio, not in your note text — full-text search finds nothing.
  • Entry friction: 20 minutes of manual note-taking per video → you stop doing it.

The solution is: let AI deconstruct video into searchable text first, then hand it to your note vault for long-term management. BibiGPT handles the former; Obsidian handles the latter.

2. Three steps: from zero to a growing PKM system

Step 1: configure Obsidian auto-save in the BibiGPT desktop client

Open BibiGPT desktop → Settings → Note Integrations → enter your Obsidian Vault local path (e.g. /Users/you/Obsidian Vault/BibiGPT) → toggle on "auto-save to Obsidian after summary completes".

Obsidian Vault local path settingsObsidian Vault local path settings

Once configured, every video BibiGPT processes (Bilibili, YouTube, Xiaohongshu, Xiaoyuzhou podcast, local file) auto-generates a structured note into your local Obsidian Vault. No more URL-scheme dependency — the flow is stable.

Step 2: design the note structure (Frontmatter + chapters + follow-up section)

Recommended template for each BibiGPT export:

---
source: https://bilibili.com/video/xxx
platform: bilibili
author: creator name
duration: 12:34
processed_at: 2026-04-24
tags: [learning, AI, product-design]
---

# Video Title

## AI Summary
(auto-filled by BibiGPT)

## Chapter Timestamps
- 00:00 - Intro
- 02:15 - Core argument
- 07:40 - Case study
- 11:20 - Conclusion

## Raw Captions (collapsible)
(auto-filled by BibiGPT)

## My Follow-Up Questions
(manually jot 2-3 questions to chase later via BibiGPT AI chat)

## Backlinks
(auto-generated by Obsidian)

The key is separating "AI-generated" from "human-added" sections. AI parts can be regenerated on top; human parts are never overwritten.

Step 3: tag taxonomy + collection chat form the knowledge network

Use three tag layers in Obsidian:

  • Layer 1 (topic): #AI #product-design #business
  • Layer 2 (source type): #video #interview #course
  • Layer 3 (processing state): #followed-up #deep-dive-later #summarized

Pair this with BibiGPT's Collections AI Chat — group same-topic videos into one Collection. Once your vault grows past 50 videos, you can cross-video query inside the Collection. That depth of retrieval doesn't exist in Obsidian alone.

Collections AI chatCollections AI chat

Obsidian stores; BibiGPT queries. That two-layer structure is the core of this methodology.

3. Advanced: make your vault "ask-able", not just "save-able"

Custom prompts for role-based re-processing

BibiGPT's Custom Prompt Summary lets you deconstruct the same video through different lenses: "explain this with Feynman technique", "extract all action items", "restructure by Minto pyramid principle".

Custom prompt summaryCustom prompt summary

Regenerated summaries auto-sync to Obsidian as new note cards — the backlink graph wires them together.

Dataview queries for topical views

Use the Dataview plugin in Obsidian:

TABLE platform, duration, processed_at
FROM #AI AND #video
WHERE processed_at > date("2026-01-01")
SORT processed_at DESC

Instant "AI videos I've watched in 2026" table — something you'd otherwise maintain by hand.

Cloud sync for team sharing

BibiGPT also exports to Lark, Yuque, Notion — perfect for team projects where "watch video → process via BibiGPT → share to Lark" becomes the new default.

4. Typical cadences (learner / creator / researcher)

Learner (student, professional upskilling)

30 min per day watching 1–2 knowledge videos → BibiGPT auto-saves to Obsidian → weekend Dataview review → Collections chat to chase concepts. In 3 months you have 50–100 structured video notes.

Content creator

Paste competitor / niche videos → re-process with a "hook formula" custom prompt → notes auto-save → when drafting, Dataview surfaces last 30 days' topic library. Content velocity compounds.

Researcher / PM

Pull every interview, talk, and course video for a theme (e.g. "AI Agents trend") into one Collection → Collections AI Chat lets you synthesize consensus and disagreements in 20 minutes. A compounding tool for research phases.

5. FAQ

Q1: Why not just use Notion?

Notion is cloud-first; search depends on the cloud. Obsidian offers local backlinks and Dataview queries that Notion doesn't. Use Obsidian for your personal long-term vault and Notion for cloud distribution — BibiGPT exports to both.

Q2: Must I use the desktop client for Obsidian auto-save?

Yes. Local Vault path access is a desktop-only capability. Web BibiGPT falls back to URL-scheme — less reliable. Download BibiGPT desktop.

Q3: What about local video files (meeting recordings)?

Desktop supports local upload and drag-and-drop. Meetings, recorded lectures — all auto-transcribed + summarized + saved.

Q4: Will Obsidian slow down with lots of notes?

Obsidian's local engine handles Markdown files generously; up to 5,000 is generally fine. Beyond that, shard by year or use Dataview aggregated views instead of real-time search.

Q5: Does this workflow support team collaboration?

Collaboration belongs in the cloud (Lark / Notion / Yuque). BibiGPT exports to all. Obsidian stays your private long-term vault. Public/private separation is a PKM golden rule.

Q6: Is it free?

Obsidian is free for personal use. BibiGPT has a free tier; Plus/Pro subscriptions unlock desktop features including Obsidian auto-save, custom prompt library, and more.

Try it: auto-save last week's videos into your vault

试试粘贴你的视频链接

支持 YouTube、B站、抖音、小红书等 30+ 平台

+30

See an ingested example — what does a Bilibili course video look like after processing?

看看 BibiGPT 的 AI 总结效果

Let's build GPT: from scratch, in code, spelled out

Let's build GPT: from scratch, in code, spelled out

Andrej Karpathy walks through building a tiny GPT in PyTorch — tokenizer, attention, transformer block, training loop.

Summary

Andrej Karpathy spends two hours rebuilding a tiny but architecturally faithful version of GPT in a single Jupyter notebook. He starts from a 1MB Shakespeare text file with a character-level tokenizer, derives self-attention from a humble running average, layers in queries/keys/values, scales up to multi-head attention, and stacks the canonical transformer block. By the end the model produces uncanny pseudo-Shakespeare and the audience has a complete mental map of pretraining, supervised fine-tuning, and RLHF — the three stages that turn a next-token predictor into ChatGPT.

Highlights

  • 🧱 Build the dumbest version first. A bigram baseline gives a working training loop and a loss number to beat before any attention is introduced.
  • 🧮 Self-attention rederived three times. Explicit loop → triangular matmul → softmax-weighted matmul makes the formula click instead of memorise.
  • 🎯 Queries, keys, values are just learned linear projections. Once you see them as that, the famous attention diagram stops being magical.
  • 🩺 Residuals + LayerNorm are what make depth trainable. Karpathy shows how each one earns its place in a transformer block.
  • 🌍 Pretraining is only stage one. The toy model is what we built; supervised fine-tuning and RLHF are what turn it into an assistant.

#GPT #Transformer #Attention #LLM #AndrejKarpathy

Questions

  1. Why start with character-level tokens instead of BPE?
    • To keep the vocabulary tiny (65 symbols) and the focus on the model. Production GPTs use BPE for efficiency, but the architecture is identical.
  2. Why scale dot-product attention by 1/√d_k?
    • It keeps the variance of the scores roughly constant as the head dimension grows, so the softmax does not collapse to a one-hot distribution.
  3. What separates the toy GPT from ChatGPT?
    • Scale (billions vs. tens of millions of parameters), data, and two extra training stages: supervised fine-tuning on conversation data and reinforcement learning from human feedback.

Key Terms

  • Bigram model: A baseline language model that predicts the next token using only the previous token, implemented as a single embedding lookup.
  • Self-attention: A mechanism where each token attends to all earlier tokens via softmax-weighted dot products of query and key projections.
  • LayerNorm (pre-norm): Normalisation applied before each sublayer in modern transformers; keeps activations well-conditioned and lets you train deeper.
  • RLHF: Reinforcement learning from human feedback — the alignment stage that nudges a pretrained model toward responses humans actually prefer.

想要总结你自己的视频？

BibiGPT 支持 YouTube、B站、抖音等 30+ 平台，一键获得 AI 智能总结

免费试用 BibiGPT

Further reading:

Core features: Obsidian Auto-Save, Custom Prompt Summary, Collections AI Chat.

BibiGPT Team