n8n-bookmarks

Personal AI-powered bookmark hive mind. Send any URL (article, YouTube, tweet, Instagram) to a Telegram bot — it gets fetched, summarized by Gemini, classified into a workspace, stored in Supabase + an Obsidian vault, and retrievable by natural-language /ask queries.

What it does

• Ingest: forward any URL to the Telegram bot.
• Extract: real content is fetched — YouTube transcripts via yt-dlp, tweets via the syndication endpoint, articles via Jina Reader, Instagram via Jina fallback.
• Process: Gemini 2.5 Flash classifies into one of the predefined workspaces and produces a summary + key points + tags.
• Store twice: Supabase row (with text-embedding-004 embedding for semantic search) and an Obsidian markdown file with [[wikilinks]] so the graph view builds itself.
• Retrieve: /ask what did I save about X → embedding search → Gemini answers with cited source URLs.

Stack

Layer	Tool
Trigger	Telegram bot
Orchestrator	n8n (homelab, behind Cloudflare Tunnel)
AI	Gemini 2.5 Flash + text-embedding-004 (free tier)
Vector DB	Supabase Cloud (Postgres + pgvector)
Browsing / graph view	Obsidian vault (Syncthing-synced)

Repo layout

.
├── README.md                       you are here
├── .env.example                    secrets template
├── docs/                           architecture, setup, workspaces, prompt design
├── supabase/schema.sql             database schema + match_bookmarks RPC
├── prompts/                        source of truth for Gemini prompts
├── n8n/workflows/                  exported workflow JSON
├── scripts/                        yt-dlp + URL normalization helpers
└── obsidian-vault-template/        starter vault structure

Quick start

1. Follow docs/setup.md — provisions Supabase, Telegram bot, Gemini API key, n8n credentials.
2. Import the three workflows from n8n/workflows/ into your n8n instance.
3. Send a URL to your bot.

Required n8n container config

The ingest workflow writes Obsidian markdown files into /files/obsidian-vault/ inside the n8n container (host bind mount). n8n's restrictFileAccessTo defaults to ~/.n8n-files, which blocks the vault path with a misleading "is not writable" error.

Add to the n8n service env in docker-compose.yml:

environment:
  - N8N_RESTRICT_FILE_ACCESS_TO=/files

Then docker compose up -d to recreate the container. The workspace subdirs (AI & LLMs, Dev Tools, …) must also exist on the host beforehand — n8n's readWriteFile node does not create parents.

Deploying ingest workflow

n8n's MCP create_workflow_from_code has reproducible 500 bugs. Deploy via raw REST API instead:

# Update
curl -X PUT https://<your-n8n>/api/v1/workflows/<id> \
  -H "X-N8N-API-KEY: ..." -H "Content-Type: application/json" \
  --data @workflow.json
# Activate
curl -X POST https://<your-n8n>/api/v1/workflows/<id>/activate \
  -H "X-N8N-API-KEY: ..."

Body must include name, nodes, connections, settings only. Readonly fields (availableInMCP, binaryMode, staticData, versionId, etc.) cause 400 errors if passed back unchanged.

Build status

See docs/architecture.md for the full architecture and current build sprint.

Workspaces

See docs/workspaces.md for the taxonomy. Edit prompts/classify.md to change them.