• 세종대왕

Stella's Thread House

• I'm a big fan of Jennie. If there are any copyright issues, I will remove the content.

Overview

This service was born from a simple question: instead of extracting information from AI and manually moving it into separate documents, could we document the AI responses themselves to manage knowledge permanently? And could we automatically link related documents by considering the relationships between the current document and others, making information easier to navigate?

How It Works

A question asked to an LLM can become a subtitle. For example, the message "Tell me about the modern history of our country" can be turned into a title like "Modern History of Our Country." Based on the generated response, hashtags are automatically created by AI, and documents can be linked (Related Documents) based on matching hashtags from other documents. Of course, hashtags can have homonyms, so the more hashtags two documents share, the closer the relationship. When the LLM responds using web search, the referenced documents (Reference Documents) are visualized, numbered, and cited within the response to show which sources were used. Based on this, a loose implementation of the Zettelkasten method has been applied.

Feature Details

Modes

There are two modes: Thread and Chat. Writing in Thread mode automatically generates hashtags and links documents together. If you want to ask casual questions lightly, use Chat mode. However, Chat mode uses the same base engine for AI responses but does not support related document linking via hashtags or editing of AI-generated messages.

Thread Mode

1. Reference Documents
   • When the AI performs a web search to answer a question, the referenced web pages are numbered, and images or other content from the search appear above them. Key images or YouTube videos related to the document can be pinned to the screen. The content editor also allows you to delete unnecessary images and reposition them.
2. User message → Subtitle conversion
   • Clicking on a user message allows you to edit it into a subtitle/title.
3. Right-side menu
   • Reorder and move sections up or down by subtitle
   • Delete a document set (question + answer) by subtitle
   • Manage AI-generated hashtags (add/delete)
   • Related Documents — browse documents linked by shared hashtags

Dashboard

• Browse recent posts and view hashtag generation info
• Visualize connections across all documents

How I Use It

Here are some ways I personally use this service:

1. Paste a YouTube link and request a summary, play it in PiP mode, and ask follow-up questions mid-video to study.
2. When reading a book, web-search for a summary of the book, then document questions that come up while reading.
3. Provide a web link on a specific topic and have the content organized.

Requirements

Tested primarily with the gemma4 and oss-20b models served via Ollama / LM Studio. The API Endpoint must end with /v1!

• ex) http://ollama:11434 → http://100.78.190.8:11434/v1

Type	Purpose
Local LLM	OpenAI-compatible endpoint URL
Tavily API Key	For web search

Tested AI Models

Built against the OpenAI GPT-compatible spec. Below is the list of models that have been tested.

• Note: every AI Endpoint must end with /v1!

Type	Model	Purpose
Ollama	Gemma4:26b	Reasoning Model
Ollama	Gemma4:26b	Vision Model
OpenAI	gpt-4o	Reasoning Model
OpenAI	gpt-4o	Vision Model
Anthropic	opus	Reasoning Model

Deploy

Quick Start

• You only need an admin account + AI endpoint + Tavily key.

.env

TH_ADMIN_EMAIL_ID=admin@example.com               # admin login email
TH_ADMIN_PASSWORD=changeme1234                    # admin login password

# Postgres
POSTGRES_DB=stella
POSTGRES_USER=stella
POSTGRES_PASSWORD=stella_pass@

docker-compose.yml

services:
  postgres:
    image: pgvector/pgvector:pg18
    environment:
      POSTGRES_DB: ${POSTGRES_DB:-stella}
      POSTGRES_USER: ${POSTGRES_USER:-stella}
      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD:-stella}
    volumes:
      - pgdata:/var/lib/postgresql/18/docker

  app:
    image: charles1031/stella-th:latest
    restart: unless-stopped
    ports:
      - "3100:3100"                  # web (frontend)
      - "4100:4100"                  # API (backend)
    environment:
      TH_ADMIN_EMAIL_ID: ${TH_ADMIN_EMAIL_ID}
      TH_ADMIN_PASSWORD: ${TH_ADMIN_PASSWORD}
      DATABASE_URL: postgres://${POSTGRES_USER:-stella}:${POSTGRES_PASSWORD:-stella}@postgres:5432/${POSTGRES_DB:-stella}
    depends_on:
      - postgres

volumes:
  pgdata:

docker compose up -d

Open http://localhost:3100.

Need custom ports, host volumes, healthchecks, or a reverse proxy? See the full setup below.

1. Create .env file

# ── Required ──────────────────────────────────────────
# Admin account created on first boot
TH_ADMIN_EMAIL_ID=admin@example.com
TH_ADMIN_PASSWORD=changeme1234

# Web (frontend) host port behind your reverse proxy. Change on port conflict.
TH_PORT=3100

# AI — configured per group: Reasoning (text answers) and Vision (image analysis).
# Each group has its own OpenAI-compatible endpoint (include /v1), API key, model, and max output tokens.
# These seed the DB on first boot; afterwards values set in Settings → AI take priority.
#   OpenAI cloud:  https://api.openai.com/v1        (e.g. model gpt-4o)
#   Anthropic:     https://api.anthropic.com/v1     (e.g. model claude-...)
#   Local runtime (Ollama/vLLM/LM Studio …): http://host.docker.internal:11434/v1
AI_REASONING_ENDPOINT=https://api.openai.com/v1
AI_REASONING_API_KEY=sk-xxxxxxxxxxxxxxxxxxxxx   # leave empty for local servers that need no auth
AI_REASONING_MODEL=gpt-4o
AI_REASONING_MAX_TOKENS=16384
AI_VISION_ENDPOINT=https://api.openai.com/v1    # can point to a different provider/model
AI_VISION_API_KEY=sk-xxxxxxxxxxxxxxxxxxxxx
AI_VISION_MODEL=gpt-4o                          # vision-capable model
AI_VISION_MAX_TOKENS=8192

# ── Optional ──────────────────────────────────────────
# Web search feature (disabled if not provided)
TAVILY_API_KEY=tvly-xxxxxxxxxxxxxxxxxxxxx

# DB credentials (can be omitted to use defaults)
# POSTGRES_USER=stella
# POSTGRES_PASSWORD=stella_dev_pass
# POSTGRES_DB=stella

2. Create docker-compose.yml

services:
  postgres:
    image: pgvector/pgvector:pg18   # PostgreSQL 18 + pgvector (required for the vector extension)
    container_name: stella-th-postgres
    restart: unless-stopped
    environment:
      POSTGRES_DB: ${POSTGRES_DB:-stella}
      POSTGRES_USER: ${POSTGRES_USER:-stella}
      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD:-stella_dev_pass}
    volumes:
      - postgres-data:/var/lib/postgresql/18/docker
    networks:
      - stella-net
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U ${POSTGRES_USER:-stella} -d ${POSTGRES_DB:-stella}"]
      interval: 5s
      timeout: 5s
      retries: 20
      start_period: 10s

  app:
    image: charles1031/stella-th:1.0.1   # Docker Hub — replace with the latest tag
    container_name: stella-th-app
    restart: unless-stopped
    ports:
      - "4100:4100"   # backend (NestJS)
      - "${TH_PORT:-3100}:3100"   # web (frontend) — change TH_PORT in .env on conflict
    env_file:
      - .env
    environment:
      DATABASE_URL: postgres://${POSTGRES_USER:-stella}:${POSTGRES_PASSWORD:-stella_dev_pass}@postgres:5432/${POSTGRES_DB:-stella}
      UPLOAD_DIR: /app/backend/uploads
      LOG_DIR: /app/backend/logs
    volumes:
      - attachments-data:/app/backend/uploads
      - logs-data:/app/backend/logs
    depends_on:
      postgres:
        condition: service_healthy
    networks:
      - stella-net

networks:
  stella-net:
    driver: bridge

volumes:
  postgres-data:
  attachments-data:
  logs-data:

3. Run

docker compose up -d

4. Access

• http://localhost:3100