Local AI với Ollama: Privacy-first Coding Setup cho Developer

Vấn đề

Khi dùng cloud AI tools, có một câu hỏi luôn là:

"Code của tôi có thực sự private không?"

Với:

• Proprietary algorithms
• Healthcare/finance regulations
• Internal APIs và credentials
• Client code under NDA

Cloud AI có thể không phải option.

Nhưng bạn vẫn muốn AI assistance. Ollama là câu trả lời.

---

Ollama là gì?

Ollama là local runtime cho open-source AI models. Giống như Docker cho containers, nhưng cho AI models.

# Cài đặt
curl -fsSL https://ollama.com/install.sh | sh

# Pull và chạy model
ollama pull llama3
ollama pull mistral
ollama pull codellama

# Sử dụng
ollama run llama3 "Explain this code:"

Supported Models cho Coding

Model	Size	Best For
Llama3	8B/70B	General coding
Mistral	7B	Fast responses
Codellama	7B/13B	Code-optimized
Qwen2.5-Coder	7B	Coding-specific
DeepSeek-Coder	6.7B	Code generation

---

Setup: Privacy-first Coding Environment

1. Ollama + VS Code

# Cài Ollama
brew install ollama  # macOS
# Windows: winget install Ollama.Ollama

# Pull Codellama
ollama pull codellama:7b

# Install VS Code extension: "Ollama"
# (Search marketplace)

2. Ollama + Cline (BYOK Alternative)

# Cline đã hỗ trợ Ollama
# Setting > Cline > Advanced > Ollama API
# 
# Endpoint: http://localhost:11434
# Model: codellama:7b

3. Ollama + OpenCode

# OpenCode với Ollama
opencode --provider local --model codellama:7b

---

Performance: So sánh Local vs Cloud

Metric	Ollama (Codellama 7B)	Claude Code Pro
Setup	Local	Cloud
Privacy	100% (code never leaves)	Partial (data processing)
Cost	Hardware only	$20-200/tháng
Speed	Depends on hardware	Fast
Quality	Good (code-specialized)	Best (frontier models)
Context window	4K-16K (model dependent)	200K

Trade-off: Privacy vs Quality.

---

Khi nào Local AI đủ tốt

Đủ tốt cho:

• Code review — quick feedback, privacy maintained
• Documentation generation — không cần frontier model
• Boilerplate code — templates, standard patterns
• Debug assistance — explain errors, suggest fixes
• Learning — explain concepts, patterns

Không đủ cho:

• Complex refactors — frontier models better
• Architectural decisions — need deeper reasoning
• Production code generation — quality matters
• Long context tasks — limited window

---

Ollama trong CI/CD

# .github/workflows/ai-review.yml
name: Local AI Code Review

on: [pull_request]

jobs:
  review:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      
      - name: Start Ollama
        run: |
          docker run -d -p 11434:11434 \
            -v ollama:/root/.ollama \
            ollama/ollama:latest
      
      - name: Run Code Review
        run: |
          # Use Cline with Ollama backend
          CLINE_OLLAMA_ENDPOINT=http://localhost:11434 \
          npx cline review --diff ${{ github.event.pull_request.diff }}

---

Enterprise Use Cases

Ollama đặc biệt hữu ích cho:

1. Regulated industries — Healthcare, fintech, legal
2. Proprietary code — IP protection mandatory
3. Air-gapped environments — no internet access
4. Cost-sensitive teams — hardware cheaper than subscriptions

Case Study: Healthcare Startup

"Chúng tôi cần AI assist cho developers nhưng không thể gửi patient data ra ngoài. Ollama + Codellama cho phép developers get AI assistance locally. Code never leaves our network."

---

Setup Guide: macOS/Windows

macOS

# 1. Install
brew install ollama

# 2. Start service
brew services start ollama

# 3. Pull models
ollama pull codellama:7b
ollama pull llama3

# 4. API available at
# http://localhost:11434

Windows

# 1. Download từ ollama.com
# 2. Install
# 3. PowerShell:
ollama serve

# Another terminal:
ollama pull codellama:7b

---

So sánh: Ollama vs Tabby vs Continue.dev

Feature	Ollama	Tabby	Continue.dev
Models	100+	Code-specialized	Multiple
Setup	Self-contained	Self-hosted	Extension
Privacy	Full local	Full local	BYOK
Code quality	Good	Good	Varies
Ease of use	High	Medium	High

---

Góc nhìn từ team BKGlobal

Tại BKGlobal, chúng tôi recommend Ollama cho:

1. Projects with privacy requirements — client code, proprietary algorithms
2. Learning environments — students, training
3. Cost-sensitive teams — hardware investment vs subscription

Khi nào chọn Ollama:

• Privacy is paramount
• Budget doesn't allow Claude Code subscription
• Air-gapped or restricted environment

Khi nào chọn cloud AI:

• Maximum quality required
• Complex reasoning tasks
• Long context needs

Hybrid approach: Ollama for simple tasks, Claude Code for complex — balance privacy và quality.

---

Takeaway

Ollama enables privacy-first AI coding — code never leaves your machine, no subscription costs, good enough for most tasks.

Đáng để setup nếu:

• Bạn có privacy requirements
• Bạn muốn reduce subscription costs
• Bạn thoải mái với slightly lower quality cho simple tasks

Setup takes < 10 minutes, models download automatically.

---

Son Do — BKGlobal Tech Team

#BKGlobal #dotnet #architecture #1percentbetter