DeepSeek + Dify Build AI Applications

Core Advantages of Dify

• Visual prompt orchestration, drag-and-drop workflow design, no coding required
• Built-in RAG engine, supports PDF / Word / TXT / Markdown / Web pages and more
• Agent: tool calling, code execution, conditional branching
• Conversation logs and analytics to continuously optimize application performance
• One-click API publishing, supports iframe embedding and shareable links
• Seamless integration with DeepSeek via OpenAI-API-compatible

What is Dify?

Dify, developed by the LangGenius team, is a full-lifecycle development platform for LLM applications. It provides a one-stop solution from Prompt orchestration, knowledge base management, Agent workflows to application deployment. You can think of Dify as the "WordPress for AI applications" - through a visual interface, you can quickly build various AI applications without writing code.

Environment Requirements

One-Click Deployment with Docker Compose

The following commands will clone the Dify repository and start all services via Docker Compose:

# Clone the Dify repository git clone https://github.com/langgenius/dify.git # Enter the Docker directory cd dify/docker # Copy the environment variable configuration file cp .env.example .env # Start all services (PostgreSQL, Redis, Weaviate, Nginx, Dify API, Dify Web) docker compose up -d # View service status docker compose ps

After deployment, open your browser and visit http://localhost:3000 to enter the Dify management interface. On first visit, you will need to set up the administrator email and password.

Configuration Steps

1. After logging into the Dify management interface, click the avatar in the top right corner and select "Settings"
2. In the left menu, click "Model Providers"
3. In the model provider list, find "OpenAI-API-compatible"
4. Click the "Add to Providers" button
5. Fill in the following configuration information

Method 1: Connect DeepSeek Official API

If you use the DeepSeek official API, configure as follows:

API Key acquisition address: platform.deepseek.com/api_keys. You can create one after registration; new users have free credits.

Method 2: Connect DeepSeek Deployed Locally with Ollama

If you run DeepSeek models locally using Ollama, configure as follows:

If Dify and Ollama run on the same machine, use localhost. If Ollama is on another machine, replace with the corresponding IP address. Ensure Ollama has set the OLLAMA_HOST=0.0.0.0:11434 environment variable.

Add Model

After configuration, on the "Model Providers" page, click on the DeepSeek provider you just added, then click the "Add Model" button:

Step 1: Create an Application

1. On the Dify homepage, click the "Create Application" button
2. Select the "Chat Assistant" application type
3. Enter an application name (e.g., "DeepSeek Smart Assistant")
4. Choose an application icon (optional)
5. Click "Create"

Step 2: Configure the Model

After entering the application editing page, in the "Model Settings" area on the right:

• In the "Model" dropdown, select the DeepSeek model you just added (deepseek-chat or deepseek-reasoner)
• Set model parameters:
  • Temperature: 0.7 (conversation) / 0.1 (code) / 0.9 (creative writing)
  • Max Tokens: Set according to needs; 2048 is usually sufficient
  • Context Window: It is recommended to keep the default value

Step 3: Set the System Prompt

The system prompt defines the AI's role and behavior. In the "Orchestrate" page, in the "Prompt" area, enter the following:

You are a professional AI assistant powered by the DeepSeek model. Your response style: - Be concise and to the point, without unnecessary verbosity - If the question involves code, provide complete runnable examples - If the question involves math or logic, show the reasoning process - If the user's description is unclear, proactively ask for details - Respond in Chinese; technical terms may include the English original Your limitations: - If you don't know, say you don't know; do not fabricate information - For professional advice involving medical, legal, etc., remind users to consult professionals

Step 4: Test the Chat

In the "Preview" area on the right side of the page, you can directly test the chat. Enter a question and observe whether DeepSeek's response meets expectations. If the effect is not ideal, you can adjust the Temperature parameter or optimize the system prompt.

For more tips on designing system prompts, please refer to the prompt engineering section in DeepSeek Usage Tutorial.

Upload Documents

In the Dify top navigation bar, click "Knowledge Base" to enter the knowledge base management page:

1. Click the "Create Knowledge Base" button
2. Fill in the knowledge base name (e.g., "Company Product Manual")
3. Click "Upload Documents" and select the files to upload

Supported Document Formats

Chunk Settings

After uploading documents, Dify automatically splits the document into multiple "chunks", each chunk is independently vectorized and stored for retrieval. Chunk settings directly affect retrieval performance:

Retrieval Settings

In the "Retrieval Settings" of the knowledge base, you can configure the following parameters:

For more on RAG architecture, please refer to the Dify integration section in the DeepSeek Deployment Tutorial.

Create a Knowledge Base Q&A Application

1. On the Dify homepage, click "Create Application" and select "Chat Assistant"
2. Fill in the application name (e.g., "Product Documentation Q&A")
3. Go to the application orchestration page, and in the "Context" area, click "Add"
4. Select the knowledge base you just created
5. In the model settings on the right, select the DeepSeek model

Configure System Prompt

A knowledge base Q&A application requires a special system prompt to let the AI know how to use the retrieved documents:

You are a professional knowledge base Q&A assistant. Your answers must be based on the provided document content. Answering rules: 1. Prioritize using information from the documents to answer questions 2. If the documents contain relevant information, directly cite and answer 3. If the documents do not contain relevant information, clearly tell the user "This content is not included in the documents" 4. When answering, indicate the source of information (e.g., "According to Chapter 3 of the product manual...") 5. Do not fabricate information not present in the documents Document content: {{#context#}} User question: {{#query#}} Please answer the user's question based on the above document content.

RAG Workflow

The complete workflow of knowledge base Q&A is as follows:

Testing and Optimization

Test a few questions in the preview area to check the answer quality:

• If the answers are not accurate enough, check the document chunking and retrieval settings of the knowledge base
• If the answers are irrelevant to the documents, the variable references in the system prompt may be incorrect
• If relevant content cannot be retrieved, try lowering the similarity threshold or increasing the Top K value
• Check the "Logs & Annotations" page to see the specific document chunks retrieved each time

Create an Agent Application

1. On the Dify homepage, click "Create Application"
2. Select the "Agent" application type
3. Fill in the application name (e.g., "Data Analysis Assistant")
4. Go to the workflow orchestration page

Workflow Node Introduction

Hands-on: Build a Data Analysis Agent

Below is a complete data analysis Agent workflow example:

LLM Node Configuration Example

In the LLM node, select the DeepSeek model and configure the following prompt:

You are a data analysis expert. Please analyze the user's data analysis needs and determine which of the following types it belongs to: 1. Descriptive analysis: The user wants to understand the basic situation of the data (e.g., "What is the sales amount?") 2. Trend analysis: The user wants to understand the trend of the data (e.g., "Changes in the last three months") 3. Comparative analysis: The user wants to compare different dimensions (e.g., "Comparison between Product A and Product B") User requirement: {{#query#}} Please output only the analysis type number (1, 2, or 3), and do not output any other content.

Code Node Example

In the code node, you can write Python code for data processing:

def main(data: str) -> dict: """Process input data and return analysis results""" import json # Parse input data records = json.loads(data) # Calculate basic statistical indicators values = [r["value"] for r in records] total = sum(values) avg = total / len(values) if values else 0 max_val = max(values) if values else 0 min_val = min(values) if values else 0 return { "total": total, "average": round(avg, 2), "max": max_val, "min": min_val, "count": len(values), "summary": f"Total {len(values)} records, sum {total}, average {avg:.2f}" }

Conditional Branch Configuration

The conditional branch node routes based on the output of the LLM node:

Method 1: API Publishing

Publish the application as a REST API, which can be called from any backend program:

1. In the application editing page, click the "Publish" button in the top right corner
2. Select "Access API"
3. Obtain the API key and interface documentation

# Call the Dify application API curl -X POST https://your-dify-host/v1/chat-messages \ -H "Authorization: Bearer app-xxxxxxxxxxxxx" \ -H "Content-Type: application/json" \ -d '{ "inputs": {}, "query": "Hello, please help me analyze the recent sales data", "response_mode": "streaming", "user": "user-001" }'

Method 2: iframe Embedding

Embed the application into your website so users can use it directly on the web page:

1. In the application editing page, click "Publish" in the top right corner
2. Select "Embed in Website"
3. Copy the generated iframe code

<!-- Paste the following code into your web page --> <iframe src="https://your-dify-host/chatbot/xxxxxxxxxxxxx" style="width: 100%; height: 600px; border: none; border-radius: 12px;" allow="microphone"> </iframe>

Method 3: Share Link

Generate a public share link that anyone can use to access your application:

1. In the application editing page, click "Publish" in the top right corner
2. Select "Run"
3. Copy the generated share link

Access Control Settings

Case 1: Intelligent Customer Service Bot

Case 2: Document Q&A Assistant

Case 3: Data Analysis Assistant

More Application Scenarios