Skip to content

setup-dify.md

Latest commit

 

History

History
234 lines (171 loc) · 6.22 KB

setup-dify.md

File metadata and controls

234 lines (171 loc) · 6.22 KB

Dify — AI Application Platform

Overview

Dify is an open-source platform for building AI applications with visual workflows, RAG pipelines, and agent orchestration. It provides a web UI for creating and managing AI apps without writing code.

  • URL: https://dify.<domain> (via Traefik)
  • Port (IP-only mode): 80 (via dify-nginx container)
  • Components: API server, background worker, web frontend, PostgreSQL, Redis
  • Default admin: created on first visit (setup wizard)

Architecture

Traefik (443) → dify-nginx (80)
                 ├── /console/api, /api, /v1, /files → dify-api (5001)
                 └── /                               → dify-web (3000)

dify-api ←→ dify-db (PostgreSQL 16)
dify-api ←→ dify-redis (Redis 7)
dify-worker ←→ dify-db, dify-redis (async tasks)

Configuration

Environment Variables

In .env:

DIFY_VERSION=latest
DIFY_SECRET_KEY=CHANGE-ME-dify-secret-key
DIFY_DB_PASSWORD=CHANGE-ME-dify-db-password
DIFY_REDIS_PASSWORD=CHANGE-ME-dify-redis-password

Generate secure values:

# Generate DIFY_SECRET_KEY (32+ chars)
openssl rand -base64 32

# Generate passwords
openssl rand -base64 16

First Launch

cd ~/ai-lab-server-setup
docker compose up -d

# Check all Dify containers are running
docker compose ps | grep dify

Open https://dify.<domain> (or http://<server-ip>:80 without Traefik). The setup wizard creates the admin account on first visit.

Connect to Ollama

Dify can use Ollama as a model provider for local LLM inference.

  1. Go to Settings → Model Providers → Ollama
  2. Add a new model:
    • Model Name: llama3.2
    • Base URL: http://ollama-compose:11434 (Docker Compose)
    • Or http://host.docker.internal:11434 (native Ollama on host)
  3. Click Save

If Ollama is installed natively (via setup.sh), use host.docker.internal since Dify runs in Docker but Ollama runs on the host.

For native Ollama, ensure it listens on all interfaces:

sudo systemctl edit ollama
# Add: Environment="OLLAMA_HOST=0.0.0.0"
sudo systemctl restart ollama

Connect to Qdrant

Dify supports Qdrant as a vector database for RAG (Knowledge Base).

  1. Go to Settings → Model Providers and add an Embedding Model (e.g., nomic-embed-text via Ollama)
  2. Create a Knowledge Base → choose Qdrant as vector store:
    • URL: http://qdrant-compose:6333 (Docker Compose)
    • Or http://host.docker.internal:6333 (native Qdrant on host)
    • API Key: value from QDRANT_API_KEY in .env

Build a RAG Application

  1. Create Knowledge Base:

    • Go to Knowledge → Create Knowledge
    • Upload documents (PDF, TXT, Markdown)
    • Dify chunks, embeds, and stores in Qdrant automatically

  2. Create App:

    • Go to Studio → Create App → Chat App
    • In Context, link your Knowledge Base
    • Set the Model to llama3.2 (Ollama)
    • Configure system prompt and retrieval settings

  3. Publish:

    • Click Publish to get a shareable URL or API endpoint

API Access

Dify provides an OpenAI-compatible API for each published app:

# Get API key from app settings → API Access
curl -X POST https://dify.<domain>/v1/chat-messages \
  -H "Authorization: Bearer app-YOUR-API-KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "inputs": {},
    "query": "What is RAG?",
    "user": "test-user",
    "response_mode": "blocking"
  }'

Python SDK

pip install dify-client
from dify_client import ChatClient

client = ChatClient(api_key="app-YOUR-API-KEY")
client.base_url = "https://dify.<domain>/v1"

response = client.create_chat_message(
    inputs={},
    query="Explain vector databases",
    user="test-user",
    response_mode="blocking",
)
print(response.json()["answer"])

Workflow Builder

Dify's visual workflow builder supports:

  • LLM nodes — call any configured model
  • Knowledge Retrieval — query RAG knowledge bases
  • Code nodes — run Python/JavaScript inline
  • HTTP Request — call external APIs
  • Conditional logic — branching and loops
  • Variable aggregation — collect and merge outputs

Example: RAG Workflow

Start → Knowledge Retrieval → LLM (with context) → Answer
  1. Create a Workflow App
  2. Add Knowledge Retrieval node → select your Knowledge Base
  3. Add LLM node → connect retrieval output as context
  4. Set system prompt: "Answer based on the provided context"

Without Domain (IP Only)

Expose dify-nginx port directly:

# In docker-compose.yml, add to dify-nginx:
ports:
  - "8080:80"

Access via http://<server-ip>:8080.

Backup and Restore

Backup

# Stop services
docker compose stop dify-api dify-worker dify-web dify-nginx

# Backup database
docker exec dify-db pg_dump -U dify dify > dify-backup-$(date +%Y%m%d).sql

# Backup storage (uploaded files)
docker cp dify-api:/app/api/storage ./dify-storage-backup/

docker compose start dify-api dify-worker dify-web dify-nginx

Restore

docker compose stop dify-api dify-worker dify-web dify-nginx
cat dify-backup.sql | docker exec -i dify-db psql -U dify dify
docker cp ./dify-storage-backup/. dify-api:/app/api/storage/
docker compose start dify-api dify-worker dify-web dify-nginx

Troubleshooting

Issue Solution
Setup wizard not loading Check all 5 containers: docker compose ps | grep dify
"Connection refused" to Ollama Use host.docker.internal:11434 for native Ollama
Slow document processing Check dify-worker logs: docker compose logs dify-worker
502 error via Traefik Verify dify-nginx is on traefik-public network
Database connection error Check dify-db health: docker compose ps dify-db
Redis connection error Verify DIFY_REDIS_PASSWORD matches in .env
File upload fails Check storage volume permissions and client_max_body_size

Check Logs

# API server
docker compose logs -f dify-api

# Background worker (document processing)
docker compose logs -f dify-worker

# Nginx routing
docker compose logs -f dify-nginx

# All Dify services
docker compose logs -f dify-api dify-worker dify-web dify-nginx dify-db dify-redis