Xencode

Local-first code-aware RAG assistant and agentic coding system for PHP/JS/TS/Vue projects.

Features

• Fast embedding with BGE-M3 ONNX model (local, no API needed)
• Smart chunking - extracts functions/classes, avoids vendor noise
• Hybrid search - vector + keyword matching
• Full context - stores complete code for LLM (no truncation)
• CLI spinners - animated progress with ETA
• Stale index warning - alerts if index is outdated
• Agentic coding - plan → retrieve → patch → diff → approve → apply
• Inline approval UI - keyboard-driven patch review (Enter/Esc/E/V)
• Multi-project workspaces - index and switch between multiple projects

Quick Start

# 1. Install
npm install

# 2. Start your LLM server (e.g., MLX, Ollama) on port 8080

# 3. Index your codebase
node src/app.js index ./my-project

# 4. Ask questions
node src/app.js ask "How does billing work?"

# 5. Generate and apply patches
node src/app.js agent "Add refund method to PaymentService"

Usage

# Index a codebase
node src/app.js index ./my-project

# List indexed projects
node src/app.js projects

# Switch between projects
node src/app.js use my-project

# Ask questions (auto-detects current project)
node src/app.js ask "Show me the Product model"
node src/app.js ask "How does refund payment work?"

# Generate patches with inline approval
node src/app.js agent "Add refund method to PaymentService"
node src/app.js agent "Create Laravel refund service" --review

# Re-index after code changes
node src/app.js index ./my-project

Configuration

# LLM endpoint (default: http://127.0.0.1:8080)
export LLM_URL=http://127.0.0.1:8080

# LLM models (optional — role-based routing)
export LLM_MODEL=""
export LLM_MODEL_PLANNER=""    # Fast model for intent parsing
export LLM_MODEL_CODER=""      # Best model for patch generation
export LLM_MODEL_REVIEWER=""   # Strict model for review

Tech Stack

• Transformers.js (local ONNX embedding: BGE-M3)
• better-sqlite3
• Node.js ESM
• OpenAI-compatible LLM API
• diff + chalk for colored diff output

Project Structure

src/
  app.js          - CLI entry point
  workspace.js    - Multi-project workspace manager
  embedder.js     - BGE-M3 embedding (256 char truncation)
  indexer.js      - Code scanning & chunking
  search.js       - Hybrid search (vector + keyword)
  context.js      - Result formatting
  llm.js          - LLM client (role-based routing)
  db.js           - Per-project SQLite storage
  ui.js           - Spinner animations
  agent/
    agent.js      - Orchestrator (pipeline)
    planner.js    - Intent parser
    coder.js      - Patch generator
    reviewer.js   - Validation + retry
    tool.js       - Patch application
    diff.js       - Unified diff output
    approval.js   - Inline approval UI (keypress)
.xencode/
  projects/       - Per-project databases
  current_project.json
docs/
  guide.md        - Full user guide
  architecture.md - System design
  performance.md  - Benchmarks
  troubleshooting.md - Common issues

Documentation

• User Guide - Step-by-step instructions
• Architecture - System design
• Performance - Benchmarks
• Troubleshooting - Common issues

License

MIT