Dify + Ollama + DeepSeek：本地部署完全指南

小二·2026-06-06 15:11

前言

💡 痛点：想用 AI 但不想数据上云？API 调用成本高？模型响应慢？隐私数据不能出内网？

🎯 解决方案 ：Dify + Ollama + DeepSeek 本地部署方案，完全离线运行，数据不出本地，成本为零。

组合优势：
#mermaid-svg-8tW6AW231FuFJIn5{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-8tW6AW231FuFJIn5 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-8tW6AW231FuFJIn5 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-8tW6AW231FuFJIn5 .error-icon{fill:#552222;}#mermaid-svg-8tW6AW231FuFJIn5 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-8tW6AW231FuFJIn5 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-8tW6AW231FuFJIn5 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-8tW6AW231FuFJIn5 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-8tW6AW231FuFJIn5 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-8tW6AW231FuFJIn5 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-8tW6AW231FuFJIn5 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-8tW6AW231FuFJIn5 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-8tW6AW231FuFJIn5 .marker.cross{stroke:#333333;}#mermaid-svg-8tW6AW231FuFJIn5 svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-8tW6AW231FuFJIn5 p{margin:0;}#mermaid-svg-8tW6AW231FuFJIn5 .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-8tW6AW231FuFJIn5 .cluster-label text{fill:#333;}#mermaid-svg-8tW6AW231FuFJIn5 .cluster-label span{color:#333;}#mermaid-svg-8tW6AW231FuFJIn5 .cluster-label span p{background-color:transparent;}#mermaid-svg-8tW6AW231FuFJIn5 .label text,#mermaid-svg-8tW6AW231FuFJIn5 span{fill:#333;color:#333;}#mermaid-svg-8tW6AW231FuFJIn5 .node rect,#mermaid-svg-8tW6AW231FuFJIn5 .node circle,#mermaid-svg-8tW6AW231FuFJIn5 .node ellipse,#mermaid-svg-8tW6AW231FuFJIn5 .node polygon,#mermaid-svg-8tW6AW231FuFJIn5 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-8tW6AW231FuFJIn5 .rough-node .label text,#mermaid-svg-8tW6AW231FuFJIn5 .node .label text,#mermaid-svg-8tW6AW231FuFJIn5 .image-shape .label,#mermaid-svg-8tW6AW231FuFJIn5 .icon-shape .label{text-anchor:middle;}#mermaid-svg-8tW6AW231FuFJIn5 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-8tW6AW231FuFJIn5 .rough-node .label,#mermaid-svg-8tW6AW231FuFJIn5 .node .label,#mermaid-svg-8tW6AW231FuFJIn5 .image-shape .label,#mermaid-svg-8tW6AW231FuFJIn5 .icon-shape .label{text-align:center;}#mermaid-svg-8tW6AW231FuFJIn5 .node.clickable{cursor:pointer;}#mermaid-svg-8tW6AW231FuFJIn5 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-8tW6AW231FuFJIn5 .arrowheadPath{fill:#333333;}#mermaid-svg-8tW6AW231FuFJIn5 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-8tW6AW231FuFJIn5 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-8tW6AW231FuFJIn5 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-8tW6AW231FuFJIn5 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-8tW6AW231FuFJIn5 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-8tW6AW231FuFJIn5 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-8tW6AW231FuFJIn5 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-8tW6AW231FuFJIn5 .cluster text{fill:#333;}#mermaid-svg-8tW6AW231FuFJIn5 .cluster span{color:#333;}#mermaid-svg-8tW6AW231FuFJIn5 div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-8tW6AW231FuFJIn5 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-8tW6AW231FuFJIn5 rect.text{fill:none;stroke-width:0;}#mermaid-svg-8tW6AW231FuFJIn5 .icon-shape,#mermaid-svg-8tW6AW231FuFJIn5 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-8tW6AW231FuFJIn5 .icon-shape p,#mermaid-svg-8tW6AW231FuFJIn5 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-8tW6AW231FuFJIn5 .icon-shape .label rect,#mermaid-svg-8tW6AW231FuFJIn5 .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-8tW6AW231FuFJIn5 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-8tW6AW231FuFJIn5 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-8tW6AW231FuFJIn5 :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} 🚀 本地 AI 部署方案
🎨 Dify

可视化编排
🦙 Ollama

本地推理
🧠 DeepSeek

国产大模型
✅ 可视化工作流
✅ 知识库管理
✅ API 服务
✅ 一键部署
✅ 模型管理
✅ GPU 加速
✅ 中文能力强
✅ 推理速度快
✅ 完全免费

适用场景：

  • 🏢 企业内网部署（数据不出域）
  • 🔒 隐私敏感应用（医疗、金融、法律）
  • 💰 成本敏感项目（零 API 费用）
  • 🚀 需要定制化 AI 应用

一、架构设计

1.1 整体架构

#mermaid-svg-lGdoXRI8Eb045o78{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-lGdoXRI8Eb045o78 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-lGdoXRI8Eb045o78 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-lGdoXRI8Eb045o78 .error-icon{fill:#552222;}#mermaid-svg-lGdoXRI8Eb045o78 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-lGdoXRI8Eb045o78 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-lGdoXRI8Eb045o78 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-lGdoXRI8Eb045o78 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-lGdoXRI8Eb045o78 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-lGdoXRI8Eb045o78 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-lGdoXRI8Eb045o78 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-lGdoXRI8Eb045o78 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-lGdoXRI8Eb045o78 .marker.cross{stroke:#333333;}#mermaid-svg-lGdoXRI8Eb045o78 svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-lGdoXRI8Eb045o78 p{margin:0;}#mermaid-svg-lGdoXRI8Eb045o78 .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-lGdoXRI8Eb045o78 .cluster-label text{fill:#333;}#mermaid-svg-lGdoXRI8Eb045o78 .cluster-label span{color:#333;}#mermaid-svg-lGdoXRI8Eb045o78 .cluster-label span p{background-color:transparent;}#mermaid-svg-lGdoXRI8Eb045o78 .label text,#mermaid-svg-lGdoXRI8Eb045o78 span{fill:#333;color:#333;}#mermaid-svg-lGdoXRI8Eb045o78 .node rect,#mermaid-svg-lGdoXRI8Eb045o78 .node circle,#mermaid-svg-lGdoXRI8Eb045o78 .node ellipse,#mermaid-svg-lGdoXRI8Eb045o78 .node polygon,#mermaid-svg-lGdoXRI8Eb045o78 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-lGdoXRI8Eb045o78 .rough-node .label text,#mermaid-svg-lGdoXRI8Eb045o78 .node .label text,#mermaid-svg-lGdoXRI8Eb045o78 .image-shape .label,#mermaid-svg-lGdoXRI8Eb045o78 .icon-shape .label{text-anchor:middle;}#mermaid-svg-lGdoXRI8Eb045o78 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-lGdoXRI8Eb045o78 .rough-node .label,#mermaid-svg-lGdoXRI8Eb045o78 .node .label,#mermaid-svg-lGdoXRI8Eb045o78 .image-shape .label,#mermaid-svg-lGdoXRI8Eb045o78 .icon-shape .label{text-align:center;}#mermaid-svg-lGdoXRI8Eb045o78 .node.clickable{cursor:pointer;}#mermaid-svg-lGdoXRI8Eb045o78 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-lGdoXRI8Eb045o78 .arrowheadPath{fill:#333333;}#mermaid-svg-lGdoXRI8Eb045o78 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-lGdoXRI8Eb045o78 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-lGdoXRI8Eb045o78 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-lGdoXRI8Eb045o78 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-lGdoXRI8Eb045o78 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-lGdoXRI8Eb045o78 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-lGdoXRI8Eb045o78 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-lGdoXRI8Eb045o78 .cluster text{fill:#333;}#mermaid-svg-lGdoXRI8Eb045o78 .cluster span{color:#333;}#mermaid-svg-lGdoXRI8Eb045o78 div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-lGdoXRI8Eb045o78 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-lGdoXRI8Eb045o78 rect.text{fill:none;stroke-width:0;}#mermaid-svg-lGdoXRI8Eb045o78 .icon-shape,#mermaid-svg-lGdoXRI8Eb045o78 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-lGdoXRI8Eb045o78 .icon-shape p,#mermaid-svg-lGdoXRI8Eb045o78 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-lGdoXRI8Eb045o78 .icon-shape .label rect,#mermaid-svg-lGdoXRI8Eb045o78 .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-lGdoXRI8Eb045o78 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-lGdoXRI8Eb045o78 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-lGdoXRI8Eb045o78 :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} 用户
Dify Web
Dify API Server
Ollama
DeepSeek 模型
知识库
向量数据库

Weaviate
Dify Worker

组件说明：

组件 作用 端口
Dify Web 前端界面 80/443
Dify API 后端服务 5001
Dify Worker 异步任务 -
Ollama 模型推理 11434
Weaviate 向量数据库 8080

1.2 技术选型对比

方案 优点 缺点 推荐场景
Dify + Ollama 完全本地、易部署 需要 GPU 资源 企业生产
Dify + 商用 API 无需 GPU、响应快 有 API 成本 快速验证
纯 Ollama 最轻量 无可视化界面 开发测试
vLLM + Open-WebUI 性能最优 配置复杂 高并发场景

二、环境准备

2.1 硬件要求

#mermaid-svg-udIpOzmkqvZPw7DA{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-udIpOzmkqvZPw7DA .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-udIpOzmkqvZPw7DA .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-udIpOzmkqvZPw7DA .error-icon{fill:#552222;}#mermaid-svg-udIpOzmkqvZPw7DA .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-udIpOzmkqvZPw7DA .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-udIpOzmkqvZPw7DA .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-udIpOzmkqvZPw7DA .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-udIpOzmkqvZPw7DA .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-udIpOzmkqvZPw7DA .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-udIpOzmkqvZPw7DA .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-udIpOzmkqvZPw7DA .marker{fill:#333333;stroke:#333333;}#mermaid-svg-udIpOzmkqvZPw7DA .marker.cross{stroke:#333333;}#mermaid-svg-udIpOzmkqvZPw7DA svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-udIpOzmkqvZPw7DA p{margin:0;}#mermaid-svg-udIpOzmkqvZPw7DA .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-udIpOzmkqvZPw7DA .cluster-label text{fill:#333;}#mermaid-svg-udIpOzmkqvZPw7DA .cluster-label span{color:#333;}#mermaid-svg-udIpOzmkqvZPw7DA .cluster-label span p{background-color:transparent;}#mermaid-svg-udIpOzmkqvZPw7DA .label text,#mermaid-svg-udIpOzmkqvZPw7DA span{fill:#333;color:#333;}#mermaid-svg-udIpOzmkqvZPw7DA .node rect,#mermaid-svg-udIpOzmkqvZPw7DA .node circle,#mermaid-svg-udIpOzmkqvZPw7DA .node ellipse,#mermaid-svg-udIpOzmkqvZPw7DA .node polygon,#mermaid-svg-udIpOzmkqvZPw7DA .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-udIpOzmkqvZPw7DA .rough-node .label text,#mermaid-svg-udIpOzmkqvZPw7DA .node .label text,#mermaid-svg-udIpOzmkqvZPw7DA .image-shape .label,#mermaid-svg-udIpOzmkqvZPw7DA .icon-shape .label{text-anchor:middle;}#mermaid-svg-udIpOzmkqvZPw7DA .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-udIpOzmkqvZPw7DA .rough-node .label,#mermaid-svg-udIpOzmkqvZPw7DA .node .label,#mermaid-svg-udIpOzmkqvZPw7DA .image-shape .label,#mermaid-svg-udIpOzmkqvZPw7DA .icon-shape .label{text-align:center;}#mermaid-svg-udIpOzmkqvZPw7DA .node.clickable{cursor:pointer;}#mermaid-svg-udIpOzmkqvZPw7DA .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-udIpOzmkqvZPw7DA .arrowheadPath{fill:#333333;}#mermaid-svg-udIpOzmkqvZPw7DA .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-udIpOzmkqvZPw7DA .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-udIpOzmkqvZPw7DA .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-udIpOzmkqvZPw7DA .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-udIpOzmkqvZPw7DA .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-udIpOzmkqvZPw7DA .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-udIpOzmkqvZPw7DA .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-udIpOzmkqvZPw7DA .cluster text{fill:#333;}#mermaid-svg-udIpOzmkqvZPw7DA .cluster span{color:#333;}#mermaid-svg-udIpOzmkqvZPw7DA div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-udIpOzmkqvZPw7DA .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-udIpOzmkqvZPw7DA rect.text{fill:none;stroke-width:0;}#mermaid-svg-udIpOzmkqvZPw7DA .icon-shape,#mermaid-svg-udIpOzmkqvZPw7DA .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-udIpOzmkqvZPw7DA .icon-shape p,#mermaid-svg-udIpOzmkqvZPw7DA .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-udIpOzmkqvZPw7DA .icon-shape .label rect,#mermaid-svg-udIpOzmkqvZPw7DA .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-udIpOzmkqvZPw7DA .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-udIpOzmkqvZPw7DA .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-udIpOzmkqvZPw7DA :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} 硬件配置
CPU
内存
GPU
存储
≥ 8 核
≥ 32GB
≥ RTX 3090

(24GB 显存)
≥ 100GB

最低配置 vs 推荐配置：

配置项 最低配置 推荐配置 说明
CPU 4 核 16 核 DeepSeek 需要
内存 16GB 64GB 加载模型需要
GPU 无需 RTX 3090 有 GPU 速度快 10 倍
存储 50GB 200GB 模型文件大
网络 10Mbps 100Mbps 下载模型用

2.2 软件环境

bash 复制代码 
# 系统要求
✅ Ubuntu 20.04 / 22.04  (推荐)
✅ CentOS 7+ 
✅ Debian 11+
❌ Windows (不推荐，可用 WSL2)

# 依赖软件
✅ Docker 20.10+
✅ Docker Compose v2+
✅ NVIDIA Driver 525+ (有 GPU 时)
✅ CUDA 12.1+ (有 GPU 时)

检查环境：

bash 复制代码 
#!/bin/bash
# check-env.sh - 环境检查脚本

echo "=== 系统信息 ==="
uname -a
cat /etc/os-release

echo -e "\n=== CPU 信息 ==="
lscpu | grep "CPU(s)"
free -h

echo -e "\n=== GPU 信息 ==="
if command -v nvidia-smi &> /dev/null; then
    nvidia-smi
else
    echo "未检测到 NVIDIA GPU"
fi

echo -e "\n=== Docker 信息 ==="
docker --version
docker compose version

echo -e "\n=== 磁盘空间 ==="
df -h /

echo -e "\n✅ 环境检查完成"

三、Ollama 安装与配置

3.1 Ollama 安装

bash 复制代码 
# 方法 1：官方脚本安装（推荐）
curl -fsSL https://ollama.com/install.sh | sh

# 方法 2：手动下载安装
# 访问 https://github.com/ollama/ollama/releases
wget https://github.com/ollama/ollama/releases/download/v0.1.31/ollama-linux-amd64
chmod +x ollama-linux-amd64
sudo mv ollama-linux-amd64 /usr/local/bin/ollama

# 验证安装
ollama --version

3.2 配置 Ollama 服务

bash 复制代码 
# 创建系统服务
sudo tee /etc/systemd/system/ollama.service <<EOF
[Unit]
Description=Ollama Service
After=network-online.target

[Service]
ExecStart=/usr/local/bin/ollama serve
User=root
Group=root
Restart=always
RestartSec=3
Environment="PATH=$PATH"
# 允许跨域访问（Dify 需要）
Environment="OLLAMA_ORIGINS=*"
# 监听所有网卡
Environment="OLLAMA_HOST=0.0.0.0:11434"
# GPU 层数（根据显存调整）
Environment="OLLAMA_NUM_GPU=999"

[Install]
WantedBy=default.target
EOF

# 启动服务
sudo systemctl daemon-reload
sudo systemctl enable ollama
sudo systemctl start ollama
sudo systemctl status ollama

关键配置说明：

环境变量 说明 推荐值
OLLAMA_ORIGINS 允许跨域的来源 * (生产环境应限制)
OLLAMA_HOST 监听地址 0.0.0.0:11434
OLLAMA_NUM_GPU GPU 加载层数 999 (全部)
OLLAMA_MAX_LOADED_MODELS 最多加载模型数 3

3.3 下载 DeepSeek 模型

bash 复制代码 
# 下载 DeepSeek 模型（需要一段时间）
# deepseek-coder:6.7b - 代码模型（推荐）
ollama pull deepseek-coder:6.7b

# deepseek-v2:16b - 通用模型（效果更好，需要更大显存）
ollama pull deepseek-v2:16b

# 查看已下载模型
ollama list

# 测试运行
ollama run deepseek-coder:6.7b "写一个 Python 快速排序"

模型选择指南：
#mermaid-svg-KPtUyKUpa38AwR1F{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-KPtUyKUpa38AwR1F .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-KPtUyKUpa38AwR1F .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-KPtUyKUpa38AwR1F .error-icon{fill:#552222;}#mermaid-svg-KPtUyKUpa38AwR1F .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-KPtUyKUpa38AwR1F .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-KPtUyKUpa38AwR1F .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-KPtUyKUpa38AwR1F .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-KPtUyKUpa38AwR1F .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-KPtUyKUpa38AwR1F .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-KPtUyKUpa38AwR1F .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-KPtUyKUpa38AwR1F .marker{fill:#333333;stroke:#333333;}#mermaid-svg-KPtUyKUpa38AwR1F .marker.cross{stroke:#333333;}#mermaid-svg-KPtUyKUpa38AwR1F svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-KPtUyKUpa38AwR1F p{margin:0;}#mermaid-svg-KPtUyKUpa38AwR1F .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-KPtUyKUpa38AwR1F .cluster-label text{fill:#333;}#mermaid-svg-KPtUyKUpa38AwR1F .cluster-label span{color:#333;}#mermaid-svg-KPtUyKUpa38AwR1F .cluster-label span p{background-color:transparent;}#mermaid-svg-KPtUyKUpa38AwR1F .label text,#mermaid-svg-KPtUyKUpa38AwR1F span{fill:#333;color:#333;}#mermaid-svg-KPtUyKUpa38AwR1F .node rect,#mermaid-svg-KPtUyKUpa38AwR1F .node circle,#mermaid-svg-KPtUyKUpa38AwR1F .node ellipse,#mermaid-svg-KPtUyKUpa38AwR1F .node polygon,#mermaid-svg-KPtUyKUpa38AwR1F .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-KPtUyKUpa38AwR1F .rough-node .label text,#mermaid-svg-KPtUyKUpa38AwR1F .node .label text,#mermaid-svg-KPtUyKUpa38AwR1F .image-shape .label,#mermaid-svg-KPtUyKUpa38AwR1F .icon-shape .label{text-anchor:middle;}#mermaid-svg-KPtUyKUpa38AwR1F .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-KPtUyKUpa38AwR1F .rough-node .label,#mermaid-svg-KPtUyKUpa38AwR1F .node .label,#mermaid-svg-KPtUyKUpa38AwR1F .image-shape .label,#mermaid-svg-KPtUyKUpa38AwR1F .icon-shape .label{text-align:center;}#mermaid-svg-KPtUyKUpa38AwR1F .node.clickable{cursor:pointer;}#mermaid-svg-KPtUyKUpa38AwR1F .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-KPtUyKUpa38AwR1F .arrowheadPath{fill:#333333;}#mermaid-svg-KPtUyKUpa38AwR1F .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-KPtUyKUpa38AwR1F .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-KPtUyKUpa38AwR1F .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-KPtUyKUpa38AwR1F .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-KPtUyKUpa38AwR1F .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-KPtUyKUpa38AwR1F .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-KPtUyKUpa38AwR1F .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-KPtUyKUpa38AwR1F .cluster text{fill:#333;}#mermaid-svg-KPtUyKUpa38AwR1F .cluster span{color:#333;}#mermaid-svg-KPtUyKUpa38AwR1F div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-KPtUyKUpa38AwR1F .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-KPtUyKUpa38AwR1F rect.text{fill:none;stroke-width:0;}#mermaid-svg-KPtUyKUpa38AwR1F .icon-shape,#mermaid-svg-KPtUyKUpa38AwR1F .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-KPtUyKUpa38AwR1F .icon-shape p,#mermaid-svg-KPtUyKUpa38AwR1F .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-KPtUyKUpa38AwR1F .icon-shape .label rect,#mermaid-svg-KPtUyKUpa38AwR1F .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-KPtUyKUpa38AwR1F .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-KPtUyKUpa38AwR1F .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-KPtUyKUpa38AwR1F :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} 代码生成
通用对话
轻量部署
选择模型
用途?
deepseek-coder
deepseek-v2
deepseek-coder:1.3b
6.7b: 推荐

需要 12GB 显存
33b: 效果最好

需要 48GB 显存
16b: 推荐

需要 32GB 显存
236b: 顶级效果

需要多卡

模型 显存需求 内存需求 推荐 GPU 适用场景
deepseek-coder:1.3b 2GB 4GB 无需 GPU 轻量部署
deepseek-coder:6.7b 12GB 16GB RTX 3060 通用开发
deepseek-coder:33b 48GB 64GB A100 企业生产
deepseek-v2:16b 32GB 48GB A100 高质量对话

四、Dify 部署

4.1 Docker Compose 部署（推荐）

bash 复制代码 
# 克隆 Dify 仓库
git clone https://github.com/langgenius/dify.git
cd dify/docker

# 复制环境变量文件
cp .env.example .env

# 编辑配置（关键配置如下）
vim .env

关键配置（.env 文件）：

bash 复制代码 
# ===== 基础配置 =====
# Dify 域名（本地部署可用 IP）
DIFY_WEB_URL=http://your-server-ip
DIFY_API_URL=http://your-server-ip/api

# ===== 模型配置 =====
# 启用 Ollama
OLLAMA_API_BASE_URL=http://host.docker.internal:11434

# ===== 向量数据库 =====
# 使用 Weaviate（默认）
VECTOR_STORE=weaviate

# ===== 存储配置 =====
# 存储路径
STORAGE_TYPE=local
STORAGE_LOCAL_PATH=./volumes/storage

# ===== 高级配置 =====
# 启用知识库
KNOWLEDGE_ENABLED=true
# 启用 Agent
AGENT_ENABLED=true
# 工作流并发数
MAX_WORKFLOW_RUNS=10

启动 Dify：

bash 复制代码 
# 启动所有服务
docker compose up -d

# 查看运行状态
docker compose ps

# 查看日志
docker compose logs -f

# 停止服务
docker compose down

4.2 配置 Ollama 作为模型供应商

DeepSeek Ollama Dify 用户 DeepSeek Ollama Dify 用户 #mermaid-svg-C4mKeyiqnCWfRvOk{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-C4mKeyiqnCWfRvOk .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-C4mKeyiqnCWfRvOk .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-C4mKeyiqnCWfRvOk .error-icon{fill:#552222;}#mermaid-svg-C4mKeyiqnCWfRvOk .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-C4mKeyiqnCWfRvOk .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-C4mKeyiqnCWfRvOk .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-C4mKeyiqnCWfRvOk .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-C4mKeyiqnCWfRvOk .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-C4mKeyiqnCWfRvOk .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-C4mKeyiqnCWfRvOk .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-C4mKeyiqnCWfRvOk .marker{fill:#333333;stroke:#333333;}#mermaid-svg-C4mKeyiqnCWfRvOk .marker.cross{stroke:#333333;}#mermaid-svg-C4mKeyiqnCWfRvOk svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-C4mKeyiqnCWfRvOk p{margin:0;}#mermaid-svg-C4mKeyiqnCWfRvOk .actor{stroke:hsl(259.6261682243, 59.7765363128%, 87.9019607843%);fill:#ECECFF;}#mermaid-svg-C4mKeyiqnCWfRvOk text.actor>tspan{fill:black;stroke:none;}#mermaid-svg-C4mKeyiqnCWfRvOk .actor-line{stroke:hsl(259.6261682243, 59.7765363128%, 87.9019607843%);}#mermaid-svg-C4mKeyiqnCWfRvOk .innerArc{stroke-width:1.5;stroke-dasharray:none;}#mermaid-svg-C4mKeyiqnCWfRvOk .messageLine0{stroke-width:1.5;stroke-dasharray:none;stroke:#333;}#mermaid-svg-C4mKeyiqnCWfRvOk .messageLine1{stroke-width:1.5;stroke-dasharray:2,2;stroke:#333;}#mermaid-svg-C4mKeyiqnCWfRvOk #arrowhead path{fill:#333;stroke:#333;}#mermaid-svg-C4mKeyiqnCWfRvOk .sequenceNumber{fill:white;}#mermaid-svg-C4mKeyiqnCWfRvOk #sequencenumber{fill:#333;}#mermaid-svg-C4mKeyiqnCWfRvOk #crosshead path{fill:#333;stroke:#333;}#mermaid-svg-C4mKeyiqnCWfRvOk .messageText{fill:#333;stroke:none;}#mermaid-svg-C4mKeyiqnCWfRvOk .labelBox{stroke:hsl(259.6261682243, 59.7765363128%, 87.9019607843%);fill:#ECECFF;}#mermaid-svg-C4mKeyiqnCWfRvOk .labelText,#mermaid-svg-C4mKeyiqnCWfRvOk .labelText>tspan{fill:black;stroke:none;}#mermaid-svg-C4mKeyiqnCWfRvOk .loopText,#mermaid-svg-C4mKeyiqnCWfRvOk .loopText>tspan{fill:black;stroke:none;}#mermaid-svg-C4mKeyiqnCWfRvOk .loopLine{stroke-width:2px;stroke-dasharray:2,2;stroke:hsl(259.6261682243, 59.7765363128%, 87.9019607843%);fill:hsl(259.6261682243, 59.7765363128%, 87.9019607843%);}#mermaid-svg-C4mKeyiqnCWfRvOk .note{stroke:#aaaa33;fill:#fff5ad;}#mermaid-svg-C4mKeyiqnCWfRvOk .noteText,#mermaid-svg-C4mKeyiqnCWfRvOk .noteText>tspan{fill:black;stroke:none;}#mermaid-svg-C4mKeyiqnCWfRvOk .activation0{fill:#f4f4f4;stroke:#666;}#mermaid-svg-C4mKeyiqnCWfRvOk .activation1{fill:#f4f4f4;stroke:#666;}#mermaid-svg-C4mKeyiqnCWfRvOk .activation2{fill:#f4f4f4;stroke:#666;}#mermaid-svg-C4mKeyiqnCWfRvOk .actorPopupMenu{position:absolute;}#mermaid-svg-C4mKeyiqnCWfRvOk .actorPopupMenuPanel{position:absolute;fill:#ECECFF;box-shadow:0px 8px 16px 0px rgba(0,0,0,0.2);filter:drop-shadow(3px 5px 2px rgb(0 0 0 / 0.4));}#mermaid-svg-C4mKeyiqnCWfRvOk .actor-man line{stroke:hsl(259.6261682243, 59.7765363128%, 87.9019607843%);fill:#ECECFF;}#mermaid-svg-C4mKeyiqnCWfRvOk .actor-man circle,#mermaid-svg-C4mKeyiqnCWfRvOk line{stroke:hsl(259.6261682243, 59.7765363128%, 87.9019607843%);fill:#ECECFF;stroke-width:2px;}#mermaid-svg-C4mKeyiqnCWfRvOk :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} 发送对话请求调用 Ollama API加载模型推理返回生成结果返回响应展示结果

在 Dify 中添加 Ollama 模型：

  1. 访问 http://your-server-ip
  2. 注册管理员账号
  3. 进入 设置 → 模型供应商
  4. 选择 Ollama
  5. 填写配置：
yaml 复制代码 
# Ollama 配置
Model Name: deepseek-coder:6.7b
Base URL: http://host.docker.internal:11434
Model Type: LLM
Context Size: 16384  # DeepSeek 支持 16K 上下文

Python 脚本自动添加（可选）：

python 复制代码 
"""
自动添加 Ollama 模型到 Dify
"""
import requests

DIFY_API_BASE = "http://localhost/api"
ADMIN_TOKEN = "your-admin-token"  # 从 Dify 后台获取

def add_ollama_model():
    """添加 Ollama 模型"""
    
    url = f"{DIFY_API_BASE}/admin/workspaces/current/models"
    
    headers = {
        "Authorization": f"Bearer {ADMIN_TOKEN}",
        "Content-Type": "application/json"
    }
    
    payload = {
        "provider": "ollama",
        "model_name": "deepseek-coder:6.7b",
        "model_type": "llm",
        "config": {
            "base_url": "http://host.docker.internal:11434",
            "context_size": 16384
        }
    }
    
    response = requests.post(url, json=payload, headers=headers)
    
    if response.status_code == 201:
        print("✅ 模型添加成功")
    else:
        print(f"❌ 添加失败: {response.text}")

if __name__ == "__main__":
    add_ollama_model()

五、知识库配置

5.1 创建知识库

#mermaid-svg-SmuuC1Cw7r4CasZy{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-SmuuC1Cw7r4CasZy .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-SmuuC1Cw7r4CasZy .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-SmuuC1Cw7r4CasZy .error-icon{fill:#552222;}#mermaid-svg-SmuuC1Cw7r4CasZy .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-SmuuC1Cw7r4CasZy .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-SmuuC1Cw7r4CasZy .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-SmuuC1Cw7r4CasZy .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-SmuuC1Cw7r4CasZy .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-SmuuC1Cw7r4CasZy .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-SmuuC1Cw7r4CasZy .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-SmuuC1Cw7r4CasZy .marker{fill:#333333;stroke:#333333;}#mermaid-svg-SmuuC1Cw7r4CasZy .marker.cross{stroke:#333333;}#mermaid-svg-SmuuC1Cw7r4CasZy svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-SmuuC1Cw7r4CasZy p{margin:0;}#mermaid-svg-SmuuC1Cw7r4CasZy .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-SmuuC1Cw7r4CasZy .cluster-label text{fill:#333;}#mermaid-svg-SmuuC1Cw7r4CasZy .cluster-label span{color:#333;}#mermaid-svg-SmuuC1Cw7r4CasZy .cluster-label span p{background-color:transparent;}#mermaid-svg-SmuuC1Cw7r4CasZy .label text,#mermaid-svg-SmuuC1Cw7r4CasZy span{fill:#333;color:#333;}#mermaid-svg-SmuuC1Cw7r4CasZy .node rect,#mermaid-svg-SmuuC1Cw7r4CasZy .node circle,#mermaid-svg-SmuuC1Cw7r4CasZy .node ellipse,#mermaid-svg-SmuuC1Cw7r4CasZy .node polygon,#mermaid-svg-SmuuC1Cw7r4CasZy .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-SmuuC1Cw7r4CasZy .rough-node .label text,#mermaid-svg-SmuuC1Cw7r4CasZy .node .label text,#mermaid-svg-SmuuC1Cw7r4CasZy .image-shape .label,#mermaid-svg-SmuuC1Cw7r4CasZy .icon-shape .label{text-anchor:middle;}#mermaid-svg-SmuuC1Cw7r4CasZy .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-SmuuC1Cw7r4CasZy .rough-node .label,#mermaid-svg-SmuuC1Cw7r4CasZy .node .label,#mermaid-svg-SmuuC1Cw7r4CasZy .image-shape .label,#mermaid-svg-SmuuC1Cw7r4CasZy .icon-shape .label{text-align:center;}#mermaid-svg-SmuuC1Cw7r4CasZy .node.clickable{cursor:pointer;}#mermaid-svg-SmuuC1Cw7r4CasZy .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-SmuuC1Cw7r4CasZy .arrowheadPath{fill:#333333;}#mermaid-svg-SmuuC1Cw7r4CasZy .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-SmuuC1Cw7r4CasZy .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-SmuuC1Cw7r4CasZy .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-SmuuC1Cw7r4CasZy .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-SmuuC1Cw7r4CasZy .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-SmuuC1Cw7r4CasZy .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-SmuuC1Cw7r4CasZy .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-SmuuC1Cw7r4CasZy .cluster text{fill:#333;}#mermaid-svg-SmuuC1Cw7r4CasZy .cluster span{color:#333;}#mermaid-svg-SmuuC1Cw7r4CasZy div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-SmuuC1Cw7r4CasZy .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-SmuuC1Cw7r4CasZy rect.text{fill:none;stroke-width:0;}#mermaid-svg-SmuuC1Cw7r4CasZy .icon-shape,#mermaid-svg-SmuuC1Cw7r4CasZy .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-SmuuC1Cw7r4CasZy .icon-shape p,#mermaid-svg-SmuuC1Cw7r4CasZy .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-SmuuC1Cw7r4CasZy .icon-shape .label rect,#mermaid-svg-SmuuC1Cw7r4CasZy .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-SmuuC1Cw7r4CasZy .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-SmuuC1Cw7r4CasZy .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-SmuuC1Cw7r4CasZy :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} 上传文档
文本拆分
向量化
存储到 Weaviate
检索测试
④ 拆分策略
父子分段
按段落
固定长度
④ Embedding 模型
Ollama Embeddings
本地模型

操作步骤：

  1. 进入 知识库 → 创建知识库
  2. 上传文档（支持 PDF、Word、Markdown、TXT）
  3. 配置分段策略：
python 复制代码 
# 分段策略配置（通过 API）
SEGMENT_CONFIG = {
    "mode": "parent_child",  # 父子分段（推荐）
    "parent_splitter": {
        "type": "paragraph",  # 按段落拆分
        "max_length": 1000     # 父块最大长度
    },
    "child_splitter": {
        "type": "token",       # 按 token 拆分
        "max_length": 200      # 子块最大长度
    },
    "overlap": 50             # 重叠长度
}
  1. 选择 Embedding 模型：
bash 复制代码 
# 下载 Embedding 模型（Ollama）
ollama pull nomic-embed-text  # 推荐，效果好

# 或使用本地模型
# 需要在 Dify 配置中指定模型路径

5.2 知识库检索测试

python 复制代码 
"""
测试知识库检索
"""
import requests

def test_knowledge_retrieval(query: str, dataset_id: str):
    """测试检索效果"""
    
    url = f"{DIFY_API_BASE}/datasets/{dataset_id}/retrieve"
    
    payload = {
        "query": query,
        "retrieval_model": {
            "top_k": 5,
            "score_threshold": 0.5
        }
    }
    
    response = requests.post(url, json=payload, headers=headers)
    
    if response.status_code == 200:
        results = response.json()["data"]["records"]
        print(f"查询：{query}")
        print(f"检索到 {len(results)} 条结果\n")
        
        for i, record in enumerate(results, 1):
            print(f"结果 {i}:")
            print(f"  内容：{record['segment']['content'][:100]}...")
            print(f"  得分：{record['score']:.3f}\n")
    else:
        print(f"❌ 检索失败: {response.text}")

# 使用
test_knowledge_retrieval(
    query="如何部署 Dify？",
    dataset_id="your-dataset-id"
)

六、构建 AI 应用

6.1 创建聊天助手

#mermaid-svg-g6cJB7BLkE5TkiFk{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-g6cJB7BLkE5TkiFk .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-g6cJB7BLkE5TkiFk .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-g6cJB7BLkE5TkiFk .error-icon{fill:#552222;}#mermaid-svg-g6cJB7BLkE5TkiFk .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-g6cJB7BLkE5TkiFk .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-g6cJB7BLkE5TkiFk .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-g6cJB7BLkE5TkiFk .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-g6cJB7BLkE5TkiFk .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-g6cJB7BLkE5TkiFk .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-g6cJB7BLkE5TkiFk .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-g6cJB7BLkE5TkiFk .marker{fill:#333333;stroke:#333333;}#mermaid-svg-g6cJB7BLkE5TkiFk .marker.cross{stroke:#333333;}#mermaid-svg-g6cJB7BLkE5TkiFk svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-g6cJB7BLkE5TkiFk p{margin:0;}#mermaid-svg-g6cJB7BLkE5TkiFk .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-g6cJB7BLkE5TkiFk .cluster-label text{fill:#333;}#mermaid-svg-g6cJB7BLkE5TkiFk .cluster-label span{color:#333;}#mermaid-svg-g6cJB7BLkE5TkiFk .cluster-label span p{background-color:transparent;}#mermaid-svg-g6cJB7BLkE5TkiFk .label text,#mermaid-svg-g6cJB7BLkE5TkiFk span{fill:#333;color:#333;}#mermaid-svg-g6cJB7BLkE5TkiFk .node rect,#mermaid-svg-g6cJB7BLkE5TkiFk .node circle,#mermaid-svg-g6cJB7BLkE5TkiFk .node ellipse,#mermaid-svg-g6cJB7BLkE5TkiFk .node polygon,#mermaid-svg-g6cJB7BLkE5TkiFk .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-g6cJB7BLkE5TkiFk .rough-node .label text,#mermaid-svg-g6cJB7BLkE5TkiFk .node .label text,#mermaid-svg-g6cJB7BLkE5TkiFk .image-shape .label,#mermaid-svg-g6cJB7BLkE5TkiFk .icon-shape .label{text-anchor:middle;}#mermaid-svg-g6cJB7BLkE5TkiFk .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-g6cJB7BLkE5TkiFk .rough-node .label,#mermaid-svg-g6cJB7BLkE5TkiFk .node .label,#mermaid-svg-g6cJB7BLkE5TkiFk .image-shape .label,#mermaid-svg-g6cJB7BLkE5TkiFk .icon-shape .label{text-align:center;}#mermaid-svg-g6cJB7BLkE5TkiFk .node.clickable{cursor:pointer;}#mermaid-svg-g6cJB7BLkE5TkiFk .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-g6cJB7BLkE5TkiFk .arrowheadPath{fill:#333333;}#mermaid-svg-g6cJB7BLkE5TkiFk .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-g6cJB7BLkE5TkiFk .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-g6cJB7BLkE5TkiFk .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-g6cJB7BLkE5TkiFk .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-g6cJB7BLkE5TkiFk .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-g6cJB7BLkE5TkiFk .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-g6cJB7BLkE5TkiFk .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-g6cJB7BLkE5TkiFk .cluster text{fill:#333;}#mermaid-svg-g6cJB7BLkE5TkiFk .cluster span{color:#333;}#mermaid-svg-g6cJB7BLkE5TkiFk div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-g6cJB7BLkE5TkiFk .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-g6cJB7BLkE5TkiFk rect.text{fill:none;stroke-width:0;}#mermaid-svg-g6cJB7BLkE5TkiFk .icon-shape,#mermaid-svg-g6cJB7BLkE5TkiFk .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-g6cJB7BLkE5TkiFk .icon-shape p,#mermaid-svg-g6cJB7BLkE5TkiFk .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-g6cJB7BLkE5TkiFk .icon-shape .label rect,#mermaid-svg-g6cJB7BLkE5TkiFk .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-g6cJB7BLkE5TkiFk .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-g6cJB7BLkE5TkiFk .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-g6cJB7BLkE5TkiFk :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} 创建工作流
添加 LLM 节点
配置 Prompt
添加知识库
测试发布
选择 deepseek-coder:6.7b
系统提示词
关联知识库

步骤：

  1. 创建应用 ：进入 应用 → 创建应用 → 聊天助手

  2. 配置提示词：

markdown 复制代码 
# 系统提示词模板

你是一个技术文档助手，专门回答关于 {domain} 的问题。

## 角色设定
- 你具有深厚的技术背景
- 回答要准确、详细、易懂
- 如果不确定，不要编造答案

## 回答规范
1. 先给出简洁的结论
2. 再提供详细的解释
3. 必要时提供代码示例
4. 引用知识库中的文档

## 注意事项
- 使用中文回答
- 代码要标注语言
- 复杂概念要用比喻
  1. 关联知识库 ：
    • 在 上下文 → 添加知识库 中选择已创建的知识库
    • 配置检索参数：
yaml 复制代码 
检索模式: hybrid  # 混合检索（向量 + 关键词）
Top K: 5
Score Threshold: 0.5
  1. 测试与发布 ：
    • 在右侧对话框中测试效果
    • 满意后点击 发布

6.2 创建工作流

#mermaid-svg-Cq5oGt1KBq7WwEhh{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-Cq5oGt1KBq7WwEhh .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-Cq5oGt1KBq7WwEhh .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-Cq5oGt1KBq7WwEhh .error-icon{fill:#552222;}#mermaid-svg-Cq5oGt1KBq7WwEhh .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-Cq5oGt1KBq7WwEhh .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-Cq5oGt1KBq7WwEhh .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-Cq5oGt1KBq7WwEhh .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-Cq5oGt1KBq7WwEhh .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-Cq5oGt1KBq7WwEhh .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-Cq5oGt1KBq7WwEhh .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-Cq5oGt1KBq7WwEhh .marker{fill:#333333;stroke:#333333;}#mermaid-svg-Cq5oGt1KBq7WwEhh .marker.cross{stroke:#333333;}#mermaid-svg-Cq5oGt1KBq7WwEhh svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-Cq5oGt1KBq7WwEhh p{margin:0;}#mermaid-svg-Cq5oGt1KBq7WwEhh .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-Cq5oGt1KBq7WwEhh .cluster-label text{fill:#333;}#mermaid-svg-Cq5oGt1KBq7WwEhh .cluster-label span{color:#333;}#mermaid-svg-Cq5oGt1KBq7WwEhh .cluster-label span p{background-color:transparent;}#mermaid-svg-Cq5oGt1KBq7WwEhh .label text,#mermaid-svg-Cq5oGt1KBq7WwEhh span{fill:#333;color:#333;}#mermaid-svg-Cq5oGt1KBq7WwEhh .node rect,#mermaid-svg-Cq5oGt1KBq7WwEhh .node circle,#mermaid-svg-Cq5oGt1KBq7WwEhh .node ellipse,#mermaid-svg-Cq5oGt1KBq7WwEhh .node polygon,#mermaid-svg-Cq5oGt1KBq7WwEhh .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-Cq5oGt1KBq7WwEhh .rough-node .label text,#mermaid-svg-Cq5oGt1KBq7WwEhh .node .label text,#mermaid-svg-Cq5oGt1KBq7WwEhh .image-shape .label,#mermaid-svg-Cq5oGt1KBq7WwEhh .icon-shape .label{text-anchor:middle;}#mermaid-svg-Cq5oGt1KBq7WwEhh .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-Cq5oGt1KBq7WwEhh .rough-node .label,#mermaid-svg-Cq5oGt1KBq7WwEhh .node .label,#mermaid-svg-Cq5oGt1KBq7WwEhh .image-shape .label,#mermaid-svg-Cq5oGt1KBq7WwEhh .icon-shape .label{text-align:center;}#mermaid-svg-Cq5oGt1KBq7WwEhh .node.clickable{cursor:pointer;}#mermaid-svg-Cq5oGt1KBq7WwEhh .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-Cq5oGt1KBq7WwEhh .arrowheadPath{fill:#333333;}#mermaid-svg-Cq5oGt1KBq7WwEhh .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-Cq5oGt1KBq7WwEhh .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-Cq5oGt1KBq7WwEhh .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-Cq5oGt1KBq7WwEhh .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-Cq5oGt1KBq7WwEhh .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-Cq5oGt1KBq7WwEhh .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-Cq5oGt1KBq7WwEhh .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-Cq5oGt1KBq7WwEhh .cluster text{fill:#333;}#mermaid-svg-Cq5oGt1KBq7WwEhh .cluster span{color:#333;}#mermaid-svg-Cq5oGt1KBq7WwEhh div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-Cq5oGt1KBq7WwEhh .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-Cq5oGt1KBq7WwEhh rect.text{fill:none;stroke-width:0;}#mermaid-svg-Cq5oGt1KBq7WwEhh .icon-shape,#mermaid-svg-Cq5oGt1KBq7WwEhh .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-Cq5oGt1KBq7WwEhh .icon-shape p,#mermaid-svg-Cq5oGt1KBq7WwEhh .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-Cq5oGt1KBq7WwEhh .icon-shape .label rect,#mermaid-svg-Cq5oGt1KBq7WwEhh .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-Cq5oGt1KBq7WwEhh .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-Cq5oGt1KBq7WwEhh .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-Cq5oGt1KBq7WwEhh :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} 技术问题
闲聊
代码生成
开始
用户输入
意图识别
知识库检索
直接对话
代码助手
LLM 推理
输出结果

创建工作流应用：

  1. 创建应用 ：应用 → 创建应用 → 工作流

  2. 添加节点：

python 复制代码 
# 节点 1：意图识别（LLM 节点）
INTENT_RECOGNITION_PROMPT = """
分析用户输入的意图，输出 JSON：

用户输入：{input}

输出格式：
{
    "intent": "tech_question | casual_chat | code_generation",
    "confidence": 0.0-1.0
}
"""

# 节点 2：知识库检索（知识库节点）
# 配置同上

# 节点 3：生成回答（LLM 节点）
ANSWER_GENERATION_PROMPT = """
基于以下上下文回答用户问题：

上下文：{context}
用户问题：{query}

要求：
1. 优先使用上下文中的信息
2. 如果上下文不足，可以基于你的知识回答
3. 标注信息来源
"""

  1. 连接节点：拖拽连接各节点

  2. 测试运行 ：点击 预览 进行测试

  3. 发布 ：点击 发布

七、API 集成

7.1 获取 API Key

python 复制代码 
"""
获取 Dify API Key（通过界面）
"""
# 1. 进入应用
# 2. 点击「发布」→「访问 API」
# 3. 创建 API Key
# 4. 复制 API Key（格式：app-xxxxxxxx）

7.2 调用 API

python 复制代码 
"""
调用 Dify API 进行对话
"""
import requests
from typing import Generator

class DifyClient:
    """Dify API 客户端"""
    
    def __init__(self, base_url: str, api_key: str):
        self.base_url = base_url.rstrip('/')
        self.api_key = api_key
        self.headers = {
            "Authorization": f"Bearer {api_key}",
            "Content-Type": "application/json"
        }
    
    def chat(
        self,
        query: str,
        conversation_id: str = None,
        user: str = "default-user"
    ) -> dict:
        """发送对话请求"""
        
        url = f"{self.base_url}/chat-messages"
        
        payload = {
            "inputs": {},
            "query": query,
            "response_mode": "blocking",  # blocking 或 streaming
            "user": user
        }
        
        if conversation_id:
            payload["conversation_id"] = conversation_id
        
        response = requests.post(
            url,
            json=payload,
            headers=self.headers
        )
        
        if response.status_code == 200:
            return response.json()
        else:
            raise Exception(f"API 调用失败: {response.text}")
    
    def chat_stream(
        self,
        query: str,
        user: str = "default-user"
    ) -> Generator[str, None, None]:
        """流式对话（SSE）"""
        
        url = f"{self.base_url}/chat-messages"
        
        payload = {
            "inputs": {},
            "query": query,
            "response_mode": "streaming",  # 流式模式
            "user": user
        }
        
        response = requests.post(
            url,
            json=payload,
            headers=self.headers,
            stream=True
        )
        
        for line in response.iter_lines():
            if line:
                line = line.decode('utf-8')
                if line.startswith('data: '):
                    data = line[6:]
                    if data != '[DONE]':
                        import json
                        chunk = json.loads(data)
                        yield chunk['answer']
    
    def get_conversations(self, user: str = "default-user") -> list:
        """获取会话列表"""
        
        url = f"{self.base_url}/conversations"
        params = {"user": user, "limit": 100}
        
        response = requests.get(
            url,
            params=params,
            headers=self.headers
        )
        
        if response.status_code == 200:
            return response.json()["data"]
        else:
            return []

# 使用示例
client = DifyClient(
    base_url="http://your-server-ip/v1",
    api_key="app-xxxxxxxxxxxxxxxx"
)

# 阻塞式调用
response = client.chat("如何在 Ubuntu 上安装 Docker？")
print(response["answer"])

# 流式调用
for chunk in client.chat_stream("写一个 Python 快速排序"):
    print(chunk, end="", flush=True)

7.3 集成到现有系统

python 复制代码 
"""
将 Dify 集成到 FastAPI 应用
"""
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel

app = FastAPI()
dify_client = DifyClient(
    base_url="http://localhost/v1",
    api_key="app-xxxxxxxx"
)

class ChatRequest(BaseModel):
    query: str
    user_id: str = "default"

@app.post("/api/chat")
async def chat(request: ChatRequest):
    """对话接口"""
    try:
        response = dify_client.chat(
            query=request.query,
            user=request.user_id
        )
        return {
            "success": True,
            "data": {
                "answer": response["answer"],
                "conversation_id": response["conversation_id"]
            }
        }
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))

@app.get("/api/conversations/{user_id}")
async def get_conversations(user_id: str):
    """获取会话列表"""
    conversations = dify_client.get_conversations(user=user_id)
    return {"success": True, "data": conversations}

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=8000)

八、性能优化

8.1 Ollama 性能优化

bash 复制代码 
# 1. 启用 GPU 加速（最关键）
# 确保 Ollama 能识别到 GPU
ollama run deepseek-coder:6.7b "测试"
# 看到 "gpu layers" 说明启用成功

# 2. 调整并发数
# 编辑 Ollama 服务配置
sudo vim /etc/systemd/system/ollama.service
# 添加：
Environment="OLLAMA_MAX_LOADED_MODELS=2"
Environment="OLLAMA_NUM_PARALLEL=2"

# 3. 使用更快的量化级别
# Q4_K_M: 推荐（速度最快，效果略降）
# Q5_K_M: 平衡
# Q8_0: 最高精度（慢）
ollama pull deepseek-coder:6.7b-q4_K_M

# 4. 预热模型（避免首次加载慢）
# 在服务器启动时预先加载
ollama run deepseek-coder:6.7b "warmup" &

性能对比：

配置 首 Token 延迟 生成速度 显存占用
CPU 模式 5-10s 5-10 tok/s 0GB
GPU (Q4) 0.5-1s 30-50 tok/s 8GB
GPU (Q8) 0.5-1s 20-30 tok/s 12GB
GPU (FP16) 0.3-0.5s 15-25 tok/s 24GB

8.2 Dify 性能优化

yaml 复制代码 
# docker-compose.yml 性能优化

# 1. 增加 Worker 并发数
services:
  api:
    environment:
      # 工作流并发数
      - MAX_WORKFLOW_RUNS=20
      # Celery 并发数
      - CELERY_WORKER_PREFETCH_MULTIPLIER=10
  
  worker:
    deploy:
      replicas: 3  # 增加 Worker 实例数
    environment:
      - CELERY_WORKER_CONCURRENCY=10

# 2. 配置 Redis 缓存
  redis:
    command: redis-server --maxmemory 2gb --maxmemory-policy allkeys-lru

# 3. 使用外部向量数据库（生产环境）
  weaviate:
    profiles:
      - local  # 默认不启动，使用外部

九、常见问题排查

9.1 Ollama 问题

问题 原因 解决方案
无法启动 端口占用 sudo lsof -i :11434 然后 kill
模型加载慢 首次下载 等待下载完成，后续会快
GPU 未启用 驱动问题 运行 nvidia-smi 检查
显存不足 模型太大 换用更小的模型或 Q4 量化
响应慢 并发过高 降低 OLLAMA_MAX_LOADED_MODELS

调试命令：

bash 复制代码 
# 查看 Ollama 日志
journalctl -u ollama -f

# 测试 API
curl http://localhost:11434/api/tags

# 查看模型信息
ollama show deepseek-coder:6.7b

# 性能测试
time ollama run deepseek-coder:6.7b "写一个快速排序"

9.2 Dify 问题

问题 原因 解决方案
无法访问 Web 端口占用 docker ps 查看端口映射
Ollama 连接失败 网络不通 检查 OLLAMA_API_BASE_URL
知识库检索慢 向量维度高 换用更快的 Embedding 模型
Worker 报错 依赖缺失 检查 Docker 日志
文件上传失败 存储权限 检查 volumes/storage 权限

调试命令：

bash 复制代码 
# 查看 Dify 日志
cd dify/docker
docker compose logs -f api
docker compose logs -f worker

# 进入容器调试
docker compose exec api bash

# 检查网络连接
docker compose exec api curl http://host.docker.internal:11434/api/tags

# 重启服务
docker compose restart

十、生产环境部署

10.1 安全配置

#mermaid-svg-muveVaLHFB6iH8uZ{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-muveVaLHFB6iH8uZ .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-muveVaLHFB6iH8uZ .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-muveVaLHFB6iH8uZ .error-icon{fill:#552222;}#mermaid-svg-muveVaLHFB6iH8uZ .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-muveVaLHFB6iH8uZ .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-muveVaLHFB6iH8uZ .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-muveVaLHFB6iH8uZ .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-muveVaLHFB6iH8uZ .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-muveVaLHFB6iH8uZ .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-muveVaLHFB6iH8uZ .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-muveVaLHFB6iH8uZ .marker{fill:#333333;stroke:#333333;}#mermaid-svg-muveVaLHFB6iH8uZ .marker.cross{stroke:#333333;}#mermaid-svg-muveVaLHFB6iH8uZ svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-muveVaLHFB6iH8uZ p{margin:0;}#mermaid-svg-muveVaLHFB6iH8uZ .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-muveVaLHFB6iH8uZ .cluster-label text{fill:#333;}#mermaid-svg-muveVaLHFB6iH8uZ .cluster-label span{color:#333;}#mermaid-svg-muveVaLHFB6iH8uZ .cluster-label span p{background-color:transparent;}#mermaid-svg-muveVaLHFB6iH8uZ .label text,#mermaid-svg-muveVaLHFB6iH8uZ span{fill:#333;color:#333;}#mermaid-svg-muveVaLHFB6iH8uZ .node rect,#mermaid-svg-muveVaLHFB6iH8uZ .node circle,#mermaid-svg-muveVaLHFB6iH8uZ .node ellipse,#mermaid-svg-muveVaLHFB6iH8uZ .node polygon,#mermaid-svg-muveVaLHFB6iH8uZ .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-muveVaLHFB6iH8uZ .rough-node .label text,#mermaid-svg-muveVaLHFB6iH8uZ .node .label text,#mermaid-svg-muveVaLHFB6iH8uZ .image-shape .label,#mermaid-svg-muveVaLHFB6iH8uZ .icon-shape .label{text-anchor:middle;}#mermaid-svg-muveVaLHFB6iH8uZ .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-muveVaLHFB6iH8uZ .rough-node .label,#mermaid-svg-muveVaLHFB6iH8uZ .node .label,#mermaid-svg-muveVaLHFB6iH8uZ .image-shape .label,#mermaid-svg-muveVaLHFB6iH8uZ .icon-shape .label{text-align:center;}#mermaid-svg-muveVaLHFB6iH8uZ .node.clickable{cursor:pointer;}#mermaid-svg-muveVaLHFB6iH8uZ .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-muveVaLHFB6iH8uZ .arrowheadPath{fill:#333333;}#mermaid-svg-muveVaLHFB6iH8uZ .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-muveVaLHFB6iH8uZ .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-muveVaLHFB6iH8uZ .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-muveVaLHFB6iH8uZ .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-muveVaLHFB6iH8uZ .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-muveVaLHFB6iH8uZ .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-muveVaLHFB6iH8uZ .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-muveVaLHFB6iH8uZ .cluster text{fill:#333;}#mermaid-svg-muveVaLHFB6iH8uZ .cluster span{color:#333;}#mermaid-svg-muveVaLHFB6iH8uZ div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-muveVaLHFB6iH8uZ .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-muveVaLHFB6iH8uZ rect.text{fill:none;stroke-width:0;}#mermaid-svg-muveVaLHFB6iH8uZ .icon-shape,#mermaid-svg-muveVaLHFB6iH8uZ .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-muveVaLHFB6iH8uZ .icon-shape p,#mermaid-svg-muveVaLHFB6iH8uZ .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-muveVaLHFB6iH8uZ .icon-shape .label rect,#mermaid-svg-muveVaLHFB6iH8uZ .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-muveVaLHFB6iH8uZ .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-muveVaLHFB6iH8uZ .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-muveVaLHFB6iH8uZ :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} 互联网
Nginx 反向代理
Dify Web
Dify API
内网
Ollama
Weaviate
HTTPS

SSL 证书
认证

JWT Token

Nginx 配置：

nginx 复制代码 
# /etc/nginx/sites-available/dify
server {
    listen 80;
    server_name your-domain.com;
    
    # 重定向到 HTTPS
    return 301 https://$server_name$request_uri;
}

server {
    listen 443 ssl http2;
    server_name your-domain.com;
    
    # SSL 证书
    ssl_certificate /path/to/cert.pem;
    ssl_certificate_key /path/to/key.pem;
    
    # 安全头
    add_header X-Frame-Options DENY;
    add_header X-Content-Type-Options nosniff;
    add_header X-XSS-Protection "1; mode=block";
    
    # Dify Web
    location / {
        proxy_pass http://localhost:80;
        proxy_set_header Host $host;
        proxy_set_header X-Real-IP $remote_addr;
        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
        proxy_set_header X-Forwarded-Proto $scheme;
    }
    
    # Dify API
    location /api {
        proxy_pass http://localhost:5001;
        proxy_set_header Host $host;
        proxy_set_header X-Real-IP $remote_addr;
        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
    }
    
    # 限制上传大小
    client_max_body_size 50M;
}

10.2 监控与告警

python 复制代码 
"""
简单的监控脚本
"""
import requests
import time
from datetime import datetime

def monitor_dify():
    """监控 Dify 服务状态"""
    
    services = {
        "Dify Web": "http://localhost",
        "Dify API": "http://localhost:5001/health",
        "Ollama": "http://localhost:11434/api/tags",
        "Weaviate": "http://localhost:8080/v1/meta"
    }
    
    while True:
        print(f"\n{'='*50}")
        print(f"监控时间: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}")
        print(f"{'='*50}\n")
        
        for name, url in services.items():
            try:
                start = time.time()
                response = requests.get(url, timeout=5)
                latency = (time.time() - start) * 1000
                
                if response.status_code == 200:
                    print(f"✅ {name}: 正常 ({latency:.0f}ms)")
                else:
                    print(f"⚠️  {name}: 异常 (状态码: {response.status_code})")
            except Exception as e:
                print(f"❌ {name}: 无法访问 ({e})")
        
        # 每 60 秒检查一次
        time.sleep(60)

if __name__ == "__main__":
    monitor_dify()

十一、总结

#mermaid-svg-6200mZbizdZj2xvR{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-6200mZbizdZj2xvR .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-6200mZbizdZj2xvR .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-6200mZbizdZj2xvR .error-icon{fill:#552222;}#mermaid-svg-6200mZbizdZj2xvR .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-6200mZbizdZj2xvR .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-6200mZbizdZj2xvR .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-6200mZbizdZj2xvR .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-6200mZbizdZj2xvR .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-6200mZbizdZj2xvR .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-6200mZbizdZj2xvR .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-6200mZbizdZj2xvR .marker{fill:#333333;stroke:#333333;}#mermaid-svg-6200mZbizdZj2xvR .marker.cross{stroke:#333333;}#mermaid-svg-6200mZbizdZj2xvR svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-6200mZbizdZj2xvR p{margin:0;}#mermaid-svg-6200mZbizdZj2xvR .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-6200mZbizdZj2xvR .cluster-label text{fill:#333;}#mermaid-svg-6200mZbizdZj2xvR .cluster-label span{color:#333;}#mermaid-svg-6200mZbizdZj2xvR .cluster-label span p{background-color:transparent;}#mermaid-svg-6200mZbizdZj2xvR .label text,#mermaid-svg-6200mZbizdZj2xvR span{fill:#333;color:#333;}#mermaid-svg-6200mZbizdZj2xvR .node rect,#mermaid-svg-6200mZbizdZj2xvR .node circle,#mermaid-svg-6200mZbizdZj2xvR .node ellipse,#mermaid-svg-6200mZbizdZj2xvR .node polygon,#mermaid-svg-6200mZbizdZj2xvR .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-6200mZbizdZj2xvR .rough-node .label text,#mermaid-svg-6200mZbizdZj2xvR .node .label text,#mermaid-svg-6200mZbizdZj2xvR .image-shape .label,#mermaid-svg-6200mZbizdZj2xvR .icon-shape .label{text-anchor:middle;}#mermaid-svg-6200mZbizdZj2xvR .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-6200mZbizdZj2xvR .rough-node .label,#mermaid-svg-6200mZbizdZj2xvR .node .label,#mermaid-svg-6200mZbizdZj2xvR .image-shape .label,#mermaid-svg-6200mZbizdZj2xvR .icon-shape .label{text-align:center;}#mermaid-svg-6200mZbizdZj2xvR .node.clickable{cursor:pointer;}#mermaid-svg-6200mZbizdZj2xvR .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-6200mZbizdZj2xvR .arrowheadPath{fill:#333333;}#mermaid-svg-6200mZbizdZj2xvR .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-6200mZbizdZj2xvR .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-6200mZbizdZj2xvR .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-6200mZbizdZj2xvR .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-6200mZbizdZj2xvR .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-6200mZbizdZj2xvR .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-6200mZbizdZj2xvR .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-6200mZbizdZj2xvR .cluster text{fill:#333;}#mermaid-svg-6200mZbizdZj2xvR .cluster span{color:#333;}#mermaid-svg-6200mZbizdZj2xvR div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-6200mZbizdZj2xvR .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-6200mZbizdZj2xvR rect.text{fill:none;stroke-width:0;}#mermaid-svg-6200mZbizdZj2xvR .icon-shape,#mermaid-svg-6200mZbizdZj2xvR .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-6200mZbizdZj2xvR .icon-shape p,#mermaid-svg-6200mZbizdZj2xvR .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-6200mZbizdZj2xvR .icon-shape .label rect,#mermaid-svg-6200mZbizdZj2xvR .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-6200mZbizdZj2xvR .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-6200mZbizdZj2xvR .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-6200mZbizdZj2xvR :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} 🎯 Dify + Ollama + DeepSeek

本地部署总结
✅ 优势
⚠️ 注意事项
🚀 下一步
数据完全本地
零 API 成本
完全可定制
离线可用
需要 GPU 资源
模型效果略逊商用
运维成本
添加更多模型
优化知识库
集成到业务系统
监控与告警

核心要点

要点 说明
部署难度 ⭐⭐⭐ (中等)
硬件成本 需要 GPU，约 2-5 万元
运维成本 需要专人维护
适用场景 数据敏感、成本敏感、定制需求
推荐指数 ⭐⭐⭐⭐ (企业内网部署推荐)

本文基于 Dify 0.6+、Ollama 0.1.31+、DeepSeek-Coder 6.7B 编写。不同版本配置可能略有差异，建议参考官方文档。如有问题欢迎评论区讨论！

相关推荐
企服AI产品测评局2 小时前
2026年Agent元年！深度解析实在Agent未来路线图：从自动化工具到全能数字员工的跃迁
运维·人工智能·ai·chatgpt·自动化
weixin_468466852 小时前
UNet 模型结构从零搭建与实战解析
人工智能·深度学习·算法·机器学习·ai·unet
小二·2 小时前
OpenAI API 实战指南
ai·openai·api
AI英德西牛仔2 小时前
Claude 导出 pdf 颜色不一样怎么办，选用 AI 导出鸭优化格式转换，多维度落地修正 PDF 色彩失真问题
javascript·人工智能·ai·chatgpt·pdf·deepseek·ai导出鸭
海绵宝宝de派小星2 小时前
MCP与A2A协议深度解析：Agent时代的“TCP/IP“如何诞生
arm开发·网络协议·tcp/ip·ai
ZFSS2 小时前
VS Code + Serp MCP：让 Copilot 实时上网查询
人工智能·ai·ai作画·copilot·ai编程·ai写作
装不满的克莱因瓶3 小时前
基于 Python 进行二维空间线性可分数据单/多层感知器实战
人工智能·python·深度学习·神经网络·ai·卷积
金融RPA机器人丨实在智能3 小时前
最终决定选择实在Agent的关键因素通常是什么？
人工智能·ai
土星云SaturnCloud3 小时前
边缘计算赋能烟草行业数字化转型
服务器·人工智能·ai·边缘计算
热门推荐
01GitHub 镜像站点022026 年 AI 编程工具终极横评：Cursor vs Claude Code vs Copilot vs Windsurf03【AI】2026 年具身智能模型和世界模型总结04Codex 下载安装指南：Windows 和 macOS 官方版下载05Codex 桌面端更新后 Chrome 插件和 Computer Use 不可用，怎么排查和修复06CC-Switch 下载、安装与使用配置指南【2026.5.29】07【踩坑记录 | 第一篇】微软商店无法使用时，如何手动安装 OpenAI Codex？附`.msix`文件系统错误解决方法08Codex 接入 DeepSeek API 完整配置文档09CC-Switch & Claude 基于 Linux 服务器安装使用指南10裂开！ChatGPT 居然开始要手机号验证，附详细解决方法