Skip to main content

Inline Code Completion

DevoxxGenie provides AI-powered inline code completion that suggests code as you type, appearing as ghost text directly in your editor. This feature uses Fill-in-the-Middle (FIM) models to intelligently complete code based on the context both before and after your cursor position.

What is Inline Completion?​

Unlike traditional code completion that only looks at the code before your cursor, DevoxxGenie's inline completion uses Fill-in-the-Middle (FIM) technology:

  • Prefix context: Code before your cursor (up to 4096 characters)
  • Suffix context: Code after your cursor (up to 1024 characters)
  • Smart completion: The model generates code that fits naturally between the prefix and suffix

This results in more contextually relevant suggestions that understand the broader structure of your code.

Requirements​

IntelliJ IDEA Version​

Inline completion requires IntelliJ IDEA 2024.3+. The feature uses the new debounced inline completion API available in these versions.

Supported Providers​

Inline completion is supported via two local LLM providers:

  • Ollama - Open-source, easy to use, wide model selection
  • LM Studio - Desktop app with GUI, OpenAI-compatible API

Choose the provider that best fits your workflow. Both require FIM-capable models.

Setup​

1. Choose and Configure Your Provider​

  1. Install Ollama from ollama.com
  2. Start Ollama (runs in background)
  3. Ensure Ollama URL is configured in Settings > DevoxxGenie > LLM Providers (default: http://localhost:11434)

2. Install a FIM Model​

Pull one of the recommended FIM models:

# Recommended: Lightweight and fast
ollama pull starcoder2:3b

# Alternative: Better quality, slightly slower
ollama pull qwen2.5-coder:7b

# Alternative: DeepSeek Coder (base version for FIM)
ollama pull deepseek-coder:6.7b-base

3. Enable Inline Completion​

  1. Open Settings > Tools > DevoxxGenie > Completion
  2. Select your provider from the "Fill-in-the-Middle Provider" dropdown:
    • None - Disable inline completion
    • Ollama - Use Ollama for completions
    • LM Studio - Use LM Studio for completions
  3. Select your FIM model from the dropdown (click Refresh Models if empty)
  4. Adjust performance settings if needed (see configuration)

4. Start Coding​

Once enabled, suggestions will appear automatically as you type. The ghost text appears in gray after your cursor.

Using Inline Completion​

Accepting Suggestions​

ActionShortcutDescription
Accept full suggestionTabInsert the entire completion
Accept next wordCtrl+Right Arrow (Windows/Linux) or Option+Right Arrow (Mac)Insert only the next word
Accept next lineCtrl+EnterInsert only the current line
DismissEscapeHide the suggestion

When Suggestions Appear​

Suggestions appear automatically when:

  • You're typing in a writable code editor
  • The file type is supported (not binary)
  • The file is under 500KB
  • No code completion popup is currently visible

Suggestions are not shown when:

  • The feature is disabled in settings (Provider set to "None")
  • No FIM model is configured
  • The editor is in viewer mode
  • A lookup/completion popup is active

Configuration​

Configure inline completion in Settings > Tools > DevoxxGenie > Completion:

SettingDescriptionDefaultRange
ProviderFIM provider: None, Ollama, or LM StudioNone-
Model nameThe FIM model to use-Available models
Max tokensMaximum tokens to generate6416-256
Timeout (ms)Request timeout in milliseconds50001000-30000
Debounce delay (ms)Delay after typing before requesting300100-2000
info

Provider URLs are configured in the main DevoxxGenie LLM Providers settings. Inline completion uses the same endpoints as the chat interface.

Tuning Recommendations​

For faster suggestions:

  • Reduce Debounce delay to 100-200ms
  • Reduce Max tokens to 32-48
  • Use a smaller model like starcoder2:3b

For better quality suggestions:

  • Increase Max tokens to 128-256
  • Use a larger model like qwen2.5-coder:7b
  • Increase Timeout if using a slower model

For slower machines:

  • Use starcoder2:3b (3 billion parameters)
  • Increase Debounce delay to 500-1000ms
  • Set Timeout to 10000ms or higher

Not all models support FIM. You need models specifically trained for Fill-in-the-Middle completion:

Lightweight (Fast, Good for Everyday Coding)​

ModelSizeBest ForOllamaLM Studio
starcoder2:3b / starcoder2-3b3BFast suggestions, general coding✅✅
qwen2.5-coder:1.5b / qwen2.5-coder-1.5b1.5BVery fast, lightweight tasks✅✅

Balanced (Quality vs Speed)​

ModelSizeBest ForOllamaLM Studio
qwen2.5-coder:7b / qwen2.5-coder-7b7BGood balance of quality and speed✅✅
deepseek-coder:6.7b-base / deepseek-coder-6.7b-base6.7BCode-specific training✅✅

Higher Quality (Slower but Better)​

ModelSizeBest ForOllamaLM Studio
qwen2.5-coder:14b / qwen2.5-coder-14b14BComplex code, larger context✅✅
tip

Start with starcoder2:3b for the best initial experience. It's fast enough for real-time suggestions while providing good completion quality.

Provider Comparison​

FeatureOllamaLM Studio
Setup ComplexitySimple CLIDesktop GUI
Model ManagementCommand lineVisual interface
Resource UsageLower overheadHigher (GUI)
Best ForDevelopers comfortable with CLIUsers preferring GUI
FIM SupportNative /api/generate with suffixOpenAI-compatible /v1/completions

How It Works​

Your Code:
----------------------------------------
public void calculateTotal() {
    double sum = 0;
    for (Item item : items) {
        sum += █
    }
    return sum;
}
----------------------------------------
         ↑
    Cursor position

Prefix sent to model:
  "public void calculateTotal() {\n    double sum = 0;\n    for (Item item : items) {\n        sum += "

Suffix sent to model:
  "\n    }\n    return sum;\n}"

Generated completion:
  "item.getPrice() * item.getQuantity();"

The model sees both the code before and after your cursor to generate contextually appropriate completions.

Troubleshooting​

No Suggestions Appearing​

  1. Verify the feature is enabled: Check Settings > DevoxxGenie > Completion (Provider should not be "None")
  2. Check your provider is running:
    • Ollama: Run ollama list in terminal
    • LM Studio: Check that the Local Server is started in the UI
  3. Check your model: Ensure you've selected a FIM-capable model
  4. Verify the URL: Check that the provider URL is correct in LLM Providers settings
  5. Check IntelliJ version: Requires 2024.3 or later

Slow Suggestions​

  1. Use a smaller model: Try starcoder2:3b instead of larger models
  2. Reduce max tokens: Lower to 32-48 for faster generation
  3. Increase debounce delay: Set to 500-1000ms to reduce request frequency
  4. Check system resources: Ensure your machine has enough RAM/CPU

Low Quality Suggestions​

  1. Use a larger model: Try qwen2.5-coder:7b or deepseek-coder:6.7b-base
  2. Increase max tokens: Allow the model to generate more context
  3. Ensure FIM model: Regular chat models don't work well for inline completion

Provider-Specific Issues​

Ollama:

  • Verify the model name is correct in settings
  • Check that Ollama is accessible at the configured URL
  • Try pulling the model again: ollama pull starcoder2:3b

LM Studio:

  • Ensure the Local Server is running (check the toggle in LM Studio UI)
  • Verify the model is loaded in LM Studio
  • Check that the URL ends with /v1 (e.g., http://localhost:1234/v1)
  • Try reloading the model in LM Studio

Best Practices​

  1. Start small: Begin with starcoder2:3b and upgrade if you need better quality
  2. Tune debounce delay: Find a balance between responsiveness and system load
  3. Use appropriate context: The model works best when there's clear surrounding code
  4. Accept word-by-word: Use partial acceptance (Ctrl+Right Arrow) for long suggestions
  5. Disable when not needed: Set Provider to "None" when doing non-coding tasks
  6. Keep models loaded: For LM Studio, keep the model loaded for faster first suggestions

Future Enhancements​

Planned improvements to inline completion include:

  • Support for additional providers
  • Multi-line completion improvements
  • Context-aware language detection
  • Integration with project-specific patterns