Chatwoot + Dify — the classic AI support combo

What this combo solves#

Four pain points cluster together for most SMB teams that want self-hosted AI support:

1. SaaS is expensive — Intercom AI seats break $20k/yr easily
2. Existing platforms are not “AI-native” — bolting LLMs onto traditional ticketing is painful
3. Channels are scattered — web, WhatsApp, email, WeChat live in silos
4. Vendor lock-in is unacceptable — teams want to swap models and KBs on demand

Chatwoot + Dify is the consensus 2026 “default starter” — sane licensing, mature components, ample docs.

Architecture#

Component roles#

Component	Role	Required
Chatwoot	Channels, agent desk, ticketing	Yes
Dify	AI brain — prompts, KB, workflow, model governance	Yes
PostgreSQL	Shared relational DB (separate databases recommended)	Yes
Redis	Queues, cache	Yes
Weaviate	Vector store (or pgvector / Milvus)	Yes
Sidekiq worker	Chatwoot async tasks (in the image)	Yes
n8n	Business orchestration (orders, CRM writes)	Optional
Postmark / SES	Email delivery	Strongly recommended

Hardware and cost#

Test / POC (single host)#

Resource	Count	Unit (mo)	Notes
4 vCPU / 8 GB / 80 GB VPS	1	$20-40	Hetzner CPX31 / DigitalOcean s-4vcpu-8gb
LLM tokens (~5k convs/mo)	—	$5-30	Lower with DeepSeek / Doubao
Domain + Postmark	—	$5-15
Total		$30-85 / mo

Small production (two hosts)#

Resource	Count	Unit (mo)
4C/8G Chatwoot node	1	$30
8C/16G Dify node	1	$60
Managed Postgres (small)	1	$25
Total		$115-150 / mo

Mid production (3 hosts + heavier LLM)#

$250-500/mo including LLM tokens.

Deployment (45-90 min)#

1. Server + DNS#

# Ubuntu 24.04
curl -fsSL https://get.docker.com | sh
sudo usermod -aG docker $USER

# DNS:
#   support.example.com  → server IP
#   dify.example.com     → server IP

2. Bring up Dify#

git clone https://github.com/langgenius/dify
cd dify/docker
cp .env.example .env

# Key env:
#   CONSOLE_WEB_URL=https://dify.example.com
#   APP_API_URL=https://dify.example.com
#   SECRET_KEY=$(openssl rand -base64 42)

docker compose up -d

Open https://dify.example.com, register admin, Settings → Model Providers → add LLM (DeepSeek to start) + Embedding (bge-m3 via SiliconFlow for Chinese).

3. Create a Dify chatbot#

• Type: Chat Assistant or Agent (Agent if tool calls needed)
• Upload KB docs (FAQ, product, refund policy, etc.)
• Chunking: Parent-Child (outperforms default)
• Prompt prefix with hard constraints:

  You are [Brand]'s support assistant.
  Rules:
  1. Answer only from "Reference"
  2. If insufficient, reply "No info, handing off to a human"
  3. Never invent amounts or dates

• Publish and copy the API key (app-xxxxxxxx)

4. Bring up Chatwoot#

git clone https://github.com/chatwoot/chatwoot
cd chatwoot
cp docker/.env.example .env
# Edit FRONTEND_URL, SMTP, SECRET_KEY_BASE
docker compose -f docker-compose.production.yaml \
  run --rm rails bundle exec rails db:chatwoot_prepare
docker compose -f docker-compose.production.yaml up -d

Open https://support.example.com, register, create Account / Inbox.

5. Wire Dify as Agent Bot#

Chatwoot → Settings → Agent Bots → New:

Field	Value
Name	Dify Bot
Outgoing URL	https://dify.example.com/v1/chat-messages
Auth Header	Authorization: Bearer app-xxxxxxxx

Set this bot as the default in the target Inbox.

6. Verify#

Send a test message via the widget:

• Chatwoot picks up the conversation
• Bot auto-replies in < 3 s
• Dify shows the call in Application Logs
• Low confidence (< 0.5) triggers a human handoff

Key tuning#

KB preparation#

• Clean scattered Notion / Lark exports first (drop TOCs, emoji titles, empty images)
• Chunk by H2 / H3 headings, not by character count
• Run a regression eval weekly (see KB preparation in practice)

Fallback design#

Situation	Behavior
Top-1 score < 0.5	Auto reply “handing off” + create ticket
Workflow error	Failure branch returns degraded answer + alert
LLM token quota hit	Dify auto-stops app + alert
Sensitive keyword (complaint, refund)	n8n hands to human + ping on-call

Human-in-the-loop#

Don’t have AI fully replace agents. The safe pattern is AI drafts → human reviews → one-click send:

1. Chatwoot forwards customer message to Dify
2. Dify writes its reply back as a draft (not sent)
3. Agent clicks send or edits then sends
4. Edits flow into the improvement queue for the prompt engineer’s Monday review

KPI targets#

Metric	Target	Observed (6 mo)
First response time (FRT)	< 3 s	1.2-2.4 s
AI deflection rate	45-70%	KB-dependent
24h return rate	< 10%	8%
CSAT (AI leg)	> 4.0 / 5	4.2-4.4
Cost per conversation	< $0.05	$0.02-0.04

Common pitfalls#

1. Email fails — SMTP must be Postmark / SES with MAILER_SENDER_EMAIL inside the verified domain
2. Dify returns zero hits — switched embedding without rebuilding the vector store; delete and re-ingest
3. Cloudflare Tunnel 502 — container bound to 127.0.0.1 instead of 0.0.0.0
4. AI goes off-topic — prompt missing constraints; KB contaminated by stale docs
5. Token bill explodes — workflow loop or bot traffic; set monthly cap per app + Cloudflare Turnstile

Evolution path#

Stage	Add	When
1. Start	Chatwoot + Dify	Day 1
2. Business integration	n8n (orders / CRM)	> 2k conv/mo
3. KB upgrade	RAGFlow (complex PDFs)	High doc complexity
4. Multi-LLM	Dify multi-model	Cost pressure
5. HA	Split DBs + Swarm / K8s	> 30k conv/mo
6. Local inference	Ollama + Qwen 14B/32B	Compliance kicks in

Related#

• Chatwoot deep review
• Dify workflow patterns
• Captain vs external Dify
• Cloudflare + VPS deploy guide