Second Brain — Technical Architecture

Personal knowledge management via Telegram + AI. Capture, organize, search, and surface insights. Runs on Mac Mini + Supabase + Ollama.

🛠️ Tech Stack

Component	Technology	Why
Interface	Telegram Bot (Deno)	Primary capture interface — zero friction
Backend	Deno 2 + Oak	API + AI pipeline
Database	Supabase (PostgreSQL + pgvector)	Knowledge base + vector embeddings for semantic search
Embeddings	Ollama (Mac Mini) + MiniMax API fallback	Local: nomic-embed-text for embeddings; API for complex queries
Queue	Upstash QStash	Weekly digest scheduling, async processing
Email capture	Gmail API	Forward + auto-archive
Export	Notion API, Obsidian (Markdown files)	Portable output

🗄️ Data Model

kb_items
id	uuid	Primary key
user_id	uuid (FK)	Ref users
content_type	enum('text','url','email','voice','image','file')
raw_content	text	Original content
summary	text	AI-generated summary
embedding	vector(768)	pgvector for semantic search
tags	uuid[]	Ref to tags
entities_json	jsonb	Extracted people, places, orgs
source_url	varchar(500)	If from URL
source_email_id	varchar(255)	Gmail message ID
is_action_item	boolean
action_status	enum('open','done','dismissed')
created_at	timestamp

tags
id	uuid	Primary key
user_id	uuid (FK)	Ref users
name	varchar(100)	Tag name
color	varchar(7)	Hex color

item_relations (cross-refs)
item_id_a	uuid (FK)	Ref kb_items
item_id_b	uuid (FK)	Ref kb_items
relation_type	varchar(50)	related_to, cites, contradicts
confidence	float	AI confidence 0–1

digests
id	uuid	Primary key
user_id	uuid (FK)	Ref users
week_start	date	Week this digest covers
content	text	AI-generated digest content
item_count	integer	Items surfaced
sent_at	timestamp

🤖 AI Pipeline

Capture → Process (async)

1. Receive item via Telegram/email. 2. QStash queues processing job. 3. Ollama generates summary + extracts entities + generates embedding (nomic-embed-text on Mac Mini). 4. Store in Supabase with pgvector.

Cross-Reference (weekly)

Weekly job: for each new item, find semantically similar items using vector search. Generate cross-reference relations if similarity > 0.8. User approves or dismisses suggested links.

Weekly Digest (Sunday)

Agent reviews week's captures. Groups by theme, identifies patterns ("you captured 5 things about AI agents"), surfaces forgotten relevant items, generates action item reminders. MiniMax API for digest generation.

Ask the Brain

Natural language query → embed query → semantic search in pgvector → fetch top 10 items → MiniMax synthesizes answer with citations. Fast path: QStash timeout 30s for sync answers.

⚡ Second Brain — Architecture

🛠️ Tech Stack

🗄️ Data Model

🤖 AI Pipeline