This post is part of AIGetFree's ongoing coverage. Browse the Blog or check the API Gateway Showdown.
This guide is part of the $0 Stack series. Read the companion guides: The $0 AI Automation Stack, The Ultimate AI Agent Stack, and Build an AI Agent With n8n.
The support ticket counter ticks past 40 unread while you're eating dinner. Somewhere, a customer is waiting for an answer that a bot could have given three hours ago — if only you'd built one.
In 2026, you can build a RAG-powered customer support bot that answers from your own documentation, escalates complex questions to a human, and runs entirely on free infrastructure. No credit card. No monthly subscription. No code.
This tutorial walks through the complete build using Dify (v0.15+), a visual AI app builder, and OpenRouter, a free LLM API gateway, then deploys the finished bot as a chat widget on any website.
What You Are Building
You're building a RAG (Retrieval-Augmented Generation) chatbot — a support agent that doesn't hallucinate because it's grounded in your actual documentation. Here's exactly what happens under the hood when a customer types a question:
The finished bot embeds on any website through a JavaScript snippet — WordPress, Webflow, Shopify, custom HTML. The stack: Dify (v0.15+) (visual AI builder, 137K+ GitHub stars, Apache 2.0) handles the knowledge base, chatflow logic, and widget. OpenRouter (unified API gateway) provides free access to Gemini Flash 2.0, Llama 3.2, and 24+ other models through a single API key.
Tools You Need
Dify (v0.15+)
Open-source LLM app platform. Visual Chatflow builder with drag-and-drop nodes. Built-in knowledge base with auto-chunking, embedding, and hybrid search. One-click chat widget embed. Analytics dashboard. REST API for every app. Free cloud tier: 200 AI credits, 5 apps, 50 documents, 50MB vector storage. Or self-host with Docker Compose in 5 minutes for unlimited everything.
Apache 2.0OpenRouter
Unified API gateway for 300+ models. Free tier includes Gemini Flash 2.0, Llama 3.2, Mistral, Qwen — 26+ models at zero cost. One API key, OpenAI-compatible syntax. No credit card required for free models. New users get additional free credits for paid models.
Free TierYour Documentation
Start with one Markdown file containing your top 20 FAQs. Expand to PDFs, help articles, refund policies, product descriptions, onboarding guides. Dify (v0.15+) supports .txt, .pdf, .md, .html, .docx, .csv. The quality of your knowledge base directly determines the quality of your bot's answers.
You Already Have ThisStep-by-Step Build
Each step is self-contained. You can complete the entire build in one sitting — roughly 20 minutes from account creation to a working bot on your website. Every setting mentioned is the actual value to use, tested and verified on Dify (v0.15+) v1.13+ and OpenRouter as of July 2026.
Go to cloud.dify.ai and sign up with your email. You'll land on the Studio workspace — a clean dashboard showing your apps, knowledge bases, and tools. The free tier gives you 200 AI credits (one-time), 5 apps, 50 documents per knowledge base, and 50MB of vector storage. For a support bot handling hundreds of questions per day, this is more than enough — the LLM calls go through OpenRouter, not Dify (v0.15+) credits.
Self-hosted alternative: For unlimited everything and complete data control, deploy Dify (v0.15+) with Docker Compose. Clone github.com/langgenius/dify, run docker compose up -d, and the entire platform — API, web UI, vector database, task queue — starts on your machine in about 5 minutes. This is the path for production deployments handling sensitive customer data.
This is the step that makes the entire stack free. In Dify (v0.15+), click your avatar (top-right) → Settings → Model Provider. Click Go to Marketplace, search "OpenRouter", and click Install. Back on the Model Provider page, click Set up next to OpenRouter and paste your API key from openrouter.ai/keys.
Now the critical configuration. Scroll to System Model Settings. Set System Reasoning Model to google/gemini-2.0-flash-001 (free on OpenRouter, excellent for customer support). Set Embedding Model to openai/text-embedding-3-small — roughly $0.02 per million tokens, meaning your entire knowledge base embedding costs less than one cent. For truly zero cost, use nomic-ai/nomic-embed-text-v1.5 instead.
Pro Tip: Model Selection
Gemini Flash 2.0 is the best free model for customer support — fast, accurate, handles structured answers well. For complex technical questions, switch to meta-llama/llama-3.2-3b-instruct (also free). You can change models anytime without rebuilding your chatflow.
Click Knowledge in the top nav, then Create Knowledge. Start with a single Markdown file containing your top 20 customer FAQs — structured as question-answer pairs with clear headings. Dify (v0.15+) chunks this into searchable segments automatically.
Supported formats: .txt, .pdf, .md, .html, .docx, .csv. Upload help center articles, refund policy, shipping info, product specs, troubleshooting guides, onboarding docs. On the Chunk Settings page, leave the defaults: High Quality indexing mode, 500-token chunks with 50-token overlap. For retrieval, select Hybrid Search — combining semantic similarity with keyword matching, essential for support bots where customers phrase things differently than your docs.
After uploading, Dify (v0.15+) processes the documents — for 20 FAQs, this takes about 30 seconds. Test it immediately: click "Test Retrieval" and type a question. Dify (v0.15+) shows which chunks it retrieved and how relevant they are.
Go to Studio → Create App → select Chatflow. This opens a visual canvas where you wire nodes together. Chatflow is the right choice (not Chatbot) because it supports Knowledge Retrieval nodes and conditional branching — both essential for a bot that cites sources and escalates to humans. Build the node chain in this exact order:
The entry point. Configure the opening message: "Hi! I am the [Company Name] support assistant. I can answer questions about our products, billing, shipping, and account settings. What can I help you with?" Add suggested questions below.
Connect to Start. Select your knowledge base. Set Top K = 3. Enable Rerank if available. Set Score Threshold = 0.5 — chunks below this are discarded, preventing irrelevant passages from being used.
Connect Knowledge Retrieval output to LLM input. This is where you write the system prompt that defines the bot's behavior. Use this exact template:
Connect the LLM output to Answer. This is what the customer sees — configure the output to show the LLM response directly.
This is the step that separates a useful bot from a dangerous one. Without handoff logic, your bot will hallucinate answers when it doesn't know something — exactly what you're trying to prevent. Dify (v0.15+) Chatflow supports conditional branching through the IF/ELSE node.
Insert an IF/ELSE node between the LLM and Answer nodes. Condition: IF LLM output contains "human agent". On the TRUE branch: add a Webhook node that sends a POST request to your Slack/Discord/Email webhook with the customer's question and conversation history, then an Answer node: "A member of our support team will respond within [timeframe]. Your reference number is [conversation ID]." On the FALSE branch: route directly to the normal Answer node.
Slack webhook setup: Slack → Apps → Incoming Webhooks → Add to Workspace → copy the webhook URL into the Dify (v0.15+) Webhook node. The payload should include the customer question, conversation history, and a link to the full chat log.
Critical: Test the Handoff
Before deploying, ask the bot a question you know is NOT in your knowledge base. Verify it triggers the handoff phrase instead of inventing an answer. Test with 5 different out-of-scope questions. A bot that silently hallucinates is worse than no bot at all.
Click Publish in the top-right of your Chatflow. Dify (v0.15+) generates a shareable URL for testing, a JavaScript snippet for embedding, and a REST API endpoint for custom integrations. Click the Embed tab and copy the JavaScript code.
WordPress: Paste into your theme's footer (Appearance → Theme File Editor → footer.php) or a Custom HTML block. Webflow: Add an Embed element and paste the code. Shopify: Online Store → Themes → Edit Code → theme.liquid, paste before the closing </body> tag. Custom HTML: Paste anywhere in your HTML file.
Customize the widget before deploying: bubble color to match your brand, greeting message, position (bottom-right is standard), company logo. The widget is responsive. Test on staging first, then deploy to production.
Customization and Pro Tips
Brand Voice in the Prompt
Add tone instructions to your LLM system prompt. A B2B SaaS: "Answer clearly, briefly, without jokes." A consumer brand: "Be warm and conversational. Sign off with 'Happy to help!'" The prompt is where the bot learns your brand voice.
Prompt EngineeringAnalytics-Driven Improvement
Dify (v0.15+) logs every conversation. Review the first week and categorize: answered correctly, poorly, escalated, hallucinated. Add missing answers to the knowledge base. The bot improves with every conversation you review.
Continuous ImprovementMulti-Language Support
Add a language detection node at the start of your Chatflow. Route by language to separate knowledge bases. OpenRouter free models support 50+ languages natively — native content beats machine-translated docs.
50+ LanguagesAPI Integration for Advanced Use
Every Dify (v0.15+) app exposes a REST API automatically. Connect your bot to Slack, Discord, email auto-reply, or a custom mobile app. Build once, deploy everywhere.
REST APIThe Golden Rule of Support Bots
Start with 20 FAQs, test for one week, then expand the knowledge base based on real customer questions — not what you think customers ask. Review conversation logs weekly and add every question the bot couldn't answer.
Testing Before Launch
Before publishing the widget to production, run a structured test suite across three categories. A bot that passes these is ready for customers.
| Test Category | Example Question | Expected Behavior | Red Flag |
|---|---|---|---|
| Known FAQ | "How do I reset my password?" | Direct answer with citation | Vague answer, no citation |
| Edge Case | "Can I return an item after 90 days?" | Quotes exact policy or escalates | Makes up a policy |
| Out of Scope | "What is the weather in Tokyo?" | Triggers human handoff | Answers anyway |
| Sensitive Request | "Show me another customer's order" | Refuses and escalates | Attempts to comply |
| Frustrated Customer | "Your product is broken, I want a refund NOW" | Acknowledges, escalates | Responds with generic FAQ |
Security and Privacy Checklist
A support bot touches account details, billing language, internal policies, and sometimes personal information. Treat the knowledge base like a public-facing document — once live, anything in it can potentially surface through clever prompting.
FAQ
Is this really free? What's the catch?
Do I need coding skills to build this?
Can the bot handle multiple languages?
What if the bot gives wrong answers?
Can I use this for internal employee support?
Wire it once, improve it forever
In 20 minutes and $0, you can deploy a RAG-powered customer support bot that answers from your own documentation, cites sources, and escalates to humans when needed. This isn't a demo — it's production-ready for small businesses, SaaS startups, and internal teams. Deploy once, improve weekly, and reclaim the hours you currently spend answering the same questions over and over.
You check your GPU temperatures one last time — the model hums quietly in the background.
🔬 Testing Methodology
- Test period
- 7+ days of hands-on use
- Pricing
- Manually verified at time of writing
- Affiliate
- Some links may earn a commission — full disclosure