MCP Vector Search

🔍 CLI-first semantic code search with MCP integration

⚠️ Production Release (v2.5.56): Stable and actively maintained. LanceDB is now the default backend for better performance and stability.

A modern, fast, and intelligent code search tool that understands your codebase through semantic analysis and AST parsing. Built with Python, powered by LanceDB, and designed for developer productivity.

✨ Features

🚀 Core Capabilities

• Semantic Search: Find code by meaning, not just keywords

• AST-Aware Parsing: Understands code structure (functions, classes, methods)

• Multi-Language Support: 13 languages - Python, JavaScript, TypeScript, C#, Dart/Flutter, PHP, Ruby, Java, Go, Rust, HTML, and Markdown/Text (with extensible architecture)

• Knowledge Graph: Temporal knowledge graph with KuzuDB for entity extraction and relationship mapping (kg build, kg status, kg query)

• Interactive Visualization: D3.js-powered visualization with 5+ views (Treemap, Sunburst, Force Graph, Knowledge Graph, Heatmap)

• Development Narratives: Generate git history narratives with story command (markdown, JSON, HTML output)

• Real-time Indexing: File watching with automatic index updates

• Automatic Version Tracking: Smart reindexing on tool upgrades

• Local-First: Complete privacy with on-device processing

• Zero Configuration: Auto-detects project structure and languages

🛠️ Developer Experience

• CLI-First Design: Simple commands for immediate productivity

• Rich Output: Syntax highlighting, similarity scores, context

• Fast Performance: Sub-second search responses, efficient indexing with pipeline parallelism (37% faster); IVF-PQ vector index delivers 4.9x faster queries (3.4ms vs 16.7ms)

• Modern Architecture: Async-first, type-safe, modular design

• Semi-Automatic Reindexing: Multiple strategies without daemon processes

• 17 MCP Tools: Comprehensive MCP integration for AI assistants (search, analysis, documentation, KG, story generation)

• Chat Mode: LLM-powered code Q&A with iterative refinement (up to 30 queries), deep search, and KG query tools

• CodeT5+ Embeddings: Code-specific embeddings via index-code command (Salesforce/codet5p-110m-embedding)

🔧 Technical Features

• Vector Database: LanceDB (serverless, file-based)

• Embedding Models: Configurable sentence transformers with GPU acceleration

• Smart Reindexing: Search-triggered, Git hooks, scheduled tasks, and manual options

• Extensible Parsers: Plugin architecture for new languages

• Configuration Management: Project-specific settings

• Production Ready: Write buffering, auto-indexing, comprehensive error handling

• Performance: Apple Silicon M4 Max optimizations (2-4x speedup with MPS)

Related MCP server: ChunkHound

🚀 Quick Start

Installation

# Install from PyPI (recommended)
pip install mcp-vector-search

# Or with UV (faster)
uv pip install mcp-vector-search

# Or install from source
git clone https://github.com/bobmatnyc/mcp-vector-search.git
cd mcp-vector-search
uv sync && uv pip install -e .

Verify Installation:

# Check that all dependencies are installed correctly
mcp-vector-search doctor

# Should show all ✓ marks
# If you see missing dependencies, try:
pip install --upgrade mcp-vector-search

Zero-Config Setup (Recommended)

The fastest way to get started - completely hands-off, just one command:

# Smart zero-config setup (recommended)
mcp-vector-search setup

What setup does automatically:

• ✅ Detects your project's languages and file types

• ✅ Initializes semantic search with optimal settings

• ✅ Indexes your entire codebase

• ✅ Configures ALL installed MCP platforms (Claude Code, Cursor, etc.)

• ✅ Uses native Claude CLI integration (claude mcp add) when available

• ✅ Falls back to .mcp.json if Claude CLI not available

• ✅ Sets up file watching for auto-reindex

• ✅ Zero user input required!

Behind the scenes:

• Server name: mcp (for consistency with other MCP projects)

• Command: uv run python -m mcp_vector_search.mcp.server {PROJECT_ROOT}

• File watching: Enabled via MCP_ENABLE_FILE_WATCHING=true

• Integration method: Native claude mcp add (or .mcp.json fallback)

Example output:

🚀 Smart Setup for mcp-vector-search
🔍 Detecting project...
   ✅ Found 3 language(s): Python, JavaScript, TypeScript
   ✅ Detected 8 file type(s)
   ✅ Found 2 platform(s): claude-code, cursor
⚙️  Configuring...
   ✅ Embedding model: sentence-transformers/all-MiniLM-L6-v2
🚀 Initializing...
   ✅ Vector database created
   ✅ Configuration saved
🔍 Indexing codebase...
   ✅ Indexing completed in 12.3s
🔗 Configuring MCP integrations...
   ✅ Using Claude CLI for automatic setup
   ✅ Registered with Claude CLI
   ✅ Configured 2 platform(s)
🎉 Setup Complete!

Options:

# Force re-setup
mcp-vector-search setup --force

# Verbose output for debugging (shows Claude CLI commands)
mcp-vector-search setup --verbose

Advanced Setup Options

For more control over the installation process:

# Manual setup with MCP integration
mcp-vector-search install --with-mcp

# Custom file extensions
mcp-vector-search install --extensions .py,.js,.ts,.dart

# Skip automatic indexing
mcp-vector-search install --no-auto-index

# Just initialize (no indexing or MCP)
mcp-vector-search init

Add MCP Integration for AI Tools

Automatic (Recommended):

# One command sets up all detected platforms
mcp-vector-search setup

Manual Platform Installation:

# Add Claude Code integration (project-scoped)
mcp-vector-search install claude-code

# Add Cursor IDE integration (global)
mcp-vector-search install cursor

# See all available platforms
mcp-vector-search install list

Note: The setup command uses native claude mcp add when Claude CLI is available, providing better integration than manual .mcp.json creation.

Remove MCP Integrations

# Remove specific platform
mcp-vector-search uninstall claude-code

# Remove all integrations
mcp-vector-search uninstall --all

# List configured integrations
mcp-vector-search uninstall list

Basic Usage

# Search your code
mcp-vector-search search "authentication logic"
mcp-vector-search search "database connection setup"
mcp-vector-search search "error handling patterns"

# Index your codebase (if not done during setup)
mcp-vector-search index

# Index with code-specific embeddings (CodeT5+)
mcp-vector-search index-code

# Check project status
mcp-vector-search status

# Start file watching (auto-update index)
mcp-vector-search watch

# Interactive visualization (5+ views)
mcp-vector-search visualize

# Generate development narrative from git history
mcp-vector-search story

# Knowledge graph operations
mcp-vector-search kg build
mcp-vector-search kg status
mcp-vector-search kg query "find all Python functions"

# Chat mode with LLM
mcp-vector-search chat "explain the authentication flow"

# Code analysis
mcp-vector-search analyze complexity
mcp-vector-search analyze dead-code

Smart CLI with "Did You Mean" Suggestions

The CLI includes intelligent command suggestions for typos:

# Typos are automatically detected and corrected
$ mcp-vector-search serach "auth"
No such command 'serach'. Did you mean 'search'?

$ mcp-vector-search indx
No such command 'indx'. Did you mean 'index'?

See docs/guides/cli-usage.md for more details.

Versioning & Releasing

This project uses semantic versioning with an automated release workflow.

Quick Commands

• make version-show - Display current version

• make release-patch - Create patch release

• make publish - Publish to PyPI

See docs/development/versioning.md for complete documentation.

🔍 AI Code Review

Context-aware code review using your entire codebase as context — Not just diff analysis!

What Makes It Different

Traditional code review tools only see individual files or diffs. MCP Vector Search analyzes code with full codebase context by:

• 🔎 Semantic Search: Finding related patterns and similar implementations

• 🕸️ Knowledge Graph: Understanding dependencies and callers

• 🤖 LLM Analysis: Deep analysis with language-specific standards

• ⚡ Smart Caching: 5x speedup with intelligent result caching

Quick Examples

# Security review of your codebase
mvs analyze review security

# Review a pull request with full context
mvs analyze review-pr --baseline main --head feature-branch

# Review only changed files (fast!)
mvs analyze review security --changed-only --baseline main

# Run multiple review types at once
mvs analyze review --types security,quality,architecture

Review Types

PR Review with Context

The killer feature — review PRs using the entire codebase as context:

# Review PR with context-aware analysis
mvs analyze review-pr --baseline main --format github-json

# For each changed file, finds:
# ✓ Similar patterns in codebase (consistency checking)
# ✓ Callers and dependencies (impact analysis)
# ✓ Existing tests (coverage gaps)
# ✓ Language-specific idioms (12 languages supported)

Context Strategy:

Changed File → Vector Search (similar patterns)
            → Knowledge Graph (callers, deps)
            → Test Discovery (coverage)
            → LLM Analysis (with full context)
            → Actionable Comments

Multi-Language Support

12 languages with language-specific idioms, anti-patterns, and security checks:

Python • TypeScript • JavaScript • Java • C# • Ruby • Go • Rust • PHP • Swift • Kotlin • Scala

Each language has tailored standards:

• Python: PEP 8, type hints, context managers, SQL injection patterns

• TypeScript: Strict mode, no any, XSS patterns

• Java: SOLID principles, Optional over null, XXE patterns

• Ruby: Guard clauses, blocks, RuboCop standards

• Go: Error handling, goroutines, interfaces

Custom Instructions

Create .mcp-vector-search/review-instructions.yaml:

language_standards:
  python:
    - "Enforce type hints on all public functions"
    - "Use Pydantic for data validation"

scope_standards:
  src/auth:
    - "All auth functions must have audit logging"

custom_review_focus:
  security:
    - "Flag any hardcoded credentials"

Auto-Discovery

Automatically reads and applies standards from your existing config files:

• Python: pyproject.toml, .flake8, mypy.ini, ruff.toml

• TypeScript: tsconfig.json, .eslintrc.json

• Ruby: .rubocop.yml

• Java: checkstyle.xml, pom.xml

• +8 more languages

CI/CD Integration

# .github/workflows/code-review.yml
- name: Review PR
  run: |
    mvs analyze review-pr \
      --baseline ${{ github.base_ref }} \
      --format sarif \
      --output review.sarif

- name: Upload to Security tab
  uses: github/codeql-action/upload-sarif@v3
  with:
    sarif_file: review.sarif

Output Formats

• console: Rich, colored output for humans

• json: Machine-readable structured data

• sarif: GitHub Security tab integration

• markdown: Reports for documentation

• github-json: PR comments (summary + inline)

Performance

• Vector Search: <0.5s (find relevant code)

• KG Queries: <0.2s (relationships)

• LLM Analysis: 10-15s (deep analysis)

• Cache Hit: 5x speedup on repeat reviews

Smart Caching: Unchanged code chunks return cached findings instantly.

Learn More

📚 Complete Documentation — Architecture, examples, best practices

🚀 CI/CD Integration Guide — GitHub Actions, GitLab CI, pre-commit hooks

🌍 Multi-Language Support — 12 languages with standards

Privacy Policy Auditor

Audit codebases against their stated privacy policies using semantic code search and knowledge graph analysis.

Quick Start

# Install with auditor dependencies
pip install 'mcp-vector-search[auditor]'

# Run a privacy audit
mvs audit run --target /path/to/repo --policy /path/to/repo/PRIVACY.md

# Check for policy/code drift
mvs audit drift-check --target /path/to/repo --policy /path/to/repo/PRIVACY.md

# Verify a certification
mvs audit verify audits/<target>/latest/

# List audit history
mvs audit list

How It Works

1. Extract Claims — Parses privacy policies into testable assertions (hybrid text analysis + LLM)

2. Collect Evidence — Queries the codebase via vector search, hybrid search, and knowledge graph

3. Judge Verdicts — LLM evaluates each claim against evidence (PASS / FAIL / INSUFFICIENT / MANUAL_REVIEW)

4. Certify — Produces signed certification documents with per-claim verdicts and evidence

Features

• Dual output — Certification saved in both auditor repo and target repo

• Multiple LLM backends — OpenRouter and Anthropic (auto-detected)

• GitHub Actions — On-demand audit workflow + daily drift detection

• GPG signing — Optional cryptographic signatures for certification integrity

• Auto-issue creation — GitHub issues for claims needing review

• .audit-ignore.yml — Suppress specific claims with documented justifications

See docs/features/privacy-auditor.md for full documentation.

📖 Documentation

Commands

setup - Zero-Config Smart Setup (Recommended)

# One command to do everything (recommended)
mcp-vector-search setup

# What it does automatically:
# - Detects project languages and file types
# - Initializes semantic search
# - Indexes entire codebase
# - Configures all detected MCP platforms
# - Sets up file watching
# - Zero configuration needed!

# Force re-setup
mcp-vector-search setup --force

# Verbose output for debugging
mcp-vector-search setup --verbose

Key Features:

• Zero Configuration: No user input required

• Smart Detection: Automatically discovers languages and platforms

• Comprehensive: Handles init + index + MCP setup in one command

• Idempotent: Safe to run multiple times

• Fast: Timeout-protected scanning (won't hang on large projects)

• Team-Friendly: Commit .mcp.json to share configuration

When to use:

• ✅ First-time project setup

• ✅ Team onboarding

• ✅ Quick testing in new codebases

• ✅ Setting up multiple MCP platforms at once

install - Install Project and MCP Integrations (Advanced)

# Manual setup with more control
mcp-vector-search install

# Install with all MCP integrations
mcp-vector-search install --with-mcp

# Custom file extensions
mcp-vector-search install --extensions .py,.js,.ts

# Skip automatic indexing
mcp-vector-search install --no-auto-index

# Platform-specific MCP integration
mcp-vector-search install claude-code      # Project-scoped
mcp-vector-search install cursor           # Global
mcp-vector-search install windsurf         # Global
mcp-vector-search install vscode           # Global

# List available platforms
mcp-vector-search install list

When to use:

• Use install when you need fine-grained control over extensions, models, or MCP platforms

• Use setup for quick, zero-config onboarding (recommended)

uninstall - Remove MCP Integrations

# Remove specific platform
mcp-vector-search uninstall claude-code

# Remove all integrations
mcp-vector-search uninstall --all

# List configured integrations
mcp-vector-search uninstall list

# Skip backup creation
mcp-vector-search uninstall claude-code --no-backup

# Alias (same as uninstall)
mcp-vector-search remove claude-code

init - Initialize Project (Simple)

# Basic initialization (no indexing or MCP)
mcp-vector-search init

# Custom configuration
mcp-vector-search init --extensions .py,.js,.ts --embedding-model sentence-transformers/all-MiniLM-L6-v2

# Force re-initialization
mcp-vector-search init --force

Note: For most users, use setup instead of init. The init command is for advanced users who want manual control.

index - Index Codebase

# Index all files
mcp-vector-search index

# Index specific directory
mcp-vector-search index /path/to/code

# Force re-indexing
mcp-vector-search index --force

# Reindex entire project
mcp-vector-search index reindex

# Reindex entire project (explicit)
mcp-vector-search index reindex --all

# Reindex entire project without confirmation
mcp-vector-search index reindex --force

# Reindex specific file
mcp-vector-search index reindex path/to/file.py

search - Semantic Search

# Basic search
mcp-vector-search search "function that handles user authentication"

# Adjust similarity threshold
mcp-vector-search search "database queries" --threshold 0.7

# Limit results
mcp-vector-search search "error handling" --limit 10

# Search in specific context
mcp-vector-search search similar "path/to/function.py:25"

auto-index - Automatic Reindexing

# Setup all auto-indexing strategies
mcp-vector-search auto-index setup --method all

# Setup specific strategies
mcp-vector-search auto-index setup --method git-hooks
mcp-vector-search auto-index setup --method scheduled --interval 60

# Check for stale files and auto-reindex
mcp-vector-search auto-index check --auto-reindex --max-files 10

# View auto-indexing status
mcp-vector-search auto-index status

# Remove auto-indexing setup
mcp-vector-search auto-index teardown --method all

watch - File Watching

# Start watching for changes
mcp-vector-search watch

# Check watch status
mcp-vector-search watch status

# Enable/disable watching
mcp-vector-search watch enable
mcp-vector-search watch disable

status - Project Information

# Basic status
mcp-vector-search status

# Detailed information
mcp-vector-search status --verbose

config - Configuration Management

# View configuration
mcp-vector-search config show

# Update settings
mcp-vector-search config set similarity_threshold 0.8
mcp-vector-search config set embedding_model microsoft/codebert-base

# Configure indexing behavior
mcp-vector-search config set skip_dotfiles true    # Skip dotfiles (default)
mcp-vector-search config set respect_gitignore true # Respect .gitignore (default)

# Get specific setting
mcp-vector-search config get skip_dotfiles
mcp-vector-search config get respect_gitignore

# List available models
mcp-vector-search config models

# List all configuration keys
mcp-vector-search config list-keys

index-code - Code-Specific Embeddings

# Index with CodeT5+ embeddings (code-optimized)
mcp-vector-search index-code

# Feature-flagged via environment variable
export MCP_CODE_ENRICHMENT=true
mcp-vector-search index-code

visualize - Interactive D3.js Visualization

# Launch visualization server
mcp-vector-search visualize

# Start on custom port
mcp-vector-search visualize --port 8080

# Available views:
# - Treemap: Hierarchical view with size/complexity encoding
# - Sunburst: Radial hierarchical view
# - Force Graph: Network visualization of code relationships
# - Knowledge Graph: Entity and relationship visualization
# - Heatmap: Complexity and quality heatmap

story - Development Narrative Generation

# Generate development narrative from git history
mcp-vector-search story

# Output formats
mcp-vector-search story --format markdown
mcp-vector-search story --format json
mcp-vector-search story --format html

# Serve as HTTP endpoint
mcp-vector-search story --serve

# Extract-only mode (no LLM)
mcp-vector-search story --no-llm

# Custom LLM model
mcp-vector-search story --model gpt-4o

kg - Knowledge Graph Operations

# Build knowledge graph
mcp-vector-search kg build

# Check knowledge graph status
mcp-vector-search kg status

# Query knowledge graph
mcp-vector-search kg query "find all Python functions"
mcp-vector-search kg query "show classes in module auth"

# Browse document ontology (file-level document classification)
mcp-vector-search kg ontology
mcp-vector-search kg ontology --category guide       # filter by category
mcp-vector-search kg ontology --verbose              # include file paths

# Knowledge graph entities:
# - CodeFile, Function, Class, Person
# - ProgrammingLanguage, ProgrammingFramework
# - Document (file-level, with doc_category classification)
# - Topic (hierarchical taxonomy)

chat - LLM-Powered Code Q&A

# Ask questions about your codebase
mcp-vector-search chat "explain the authentication flow"
mcp-vector-search chat "how does error handling work?"

# Iterative refinement (up to 30 queries)
# Automatically uses deep search and KG query tools

# Advanced reasoning mode
mcp-vector-search chat "architectural patterns" --think

# Filter by files
mcp-vector-search chat "validation logic" --files "src/*.py"

analyze - Code Analysis

# Complexity analysis
mcp-vector-search analyze complexity

# Dead code detection
mcp-vector-search analyze dead-code

# Output formats
mcp-vector-search analyze complexity --json
mcp-vector-search analyze complexity --sarif
mcp-vector-search analyze complexity --output-format markdown

# CI/CD integration
mcp-vector-search analyze complexity --fail-on-smell

🚀 Performance Features

Search Optimizations

MCP Vector Search includes several query-time optimizations that are automatically enabled as your index grows.

IVF-PQ Index is built automatically after indexing more than 256 rows. It uses Inverted File with Product Quantization to partition vectors into clusters, so queries scan only a relevant subset rather than the full index. The index parameters adapt to your data: num_partitions = clamp(sqrt(N), 16, 512) and num_sub_vectors = dim // 4.

Two-stage retrieval improves precision on top of the IVF-PQ scan: the engine probes 20 IVF partitions (nprobes=20) and fetches 5x the requested candidates, then reranks them with exact cosine similarity (refine_factor=5). Applied to both the LanceDB and legacy vector backends.

Contextual chunking prepends a compact metadata header to each chunk before embedding, so the vector captures file, language, class, and function context rather than code text alone. Format: File: core/search.py | Lang: python | Class: Engine | Fn: search | Uses: lancedb. Based on Anthropic research showing 35-49% fewer retrieval failures.

See docs/performance/search-optimizations.md for technical details and benchmark methodology.

LanceDB Backend (Default in v2.1+)

LanceDB is now the default vector database for better performance and stability:

• Serverless Architecture: No separate server process needed

• Better Scaling: Superior performance for large codebases (>100k chunks)

• File-Based Storage: Simple directory-based persistence

• Fewer Corruption Issues: More stable than ChromaDB's HNSW indices

• Write Buffering: 2-4x faster indexing with accumulated batch writes

To use ChromaDB (legacy), set environment variable:

export MCP_VECTOR_SEARCH_BACKEND=chromadb

Migrate existing ChromaDB database:

mcp-vector-search migrate db chromadb-to-lancedb

See docs/LANCEDB_BACKEND.md for detailed documentation.

Apple Silicon M4 Max Optimizations

2-4x speedup on Apple Silicon with automatic hardware detection:

• MPS Backend: Metal Performance Shaders GPU acceleration for embeddings

• Intelligent Batch Sizing: Auto-detects GPU memory (384-512 for M4 Max with 128GB RAM)

• Multi-Core Optimization: Utilizes all 12 performance cores efficiently

• Zero Configuration: Automatically enabled on Apple Silicon Macs

Environment variables for tuning:

export MCP_VECTOR_SEARCH_MPS_BATCH_SIZE=512  # Override MPS batch size
export MCP_VECTOR_SEARCH_BATCH_SIZE=128      # Override all backends

Semi-Automatic Reindexing

Multiple strategies to keep your index up-to-date without daemon processes:

1. Search-Triggered: Automatically checks for stale files during searches

2. Git Hooks: Triggers reindexing after commits, merges, checkouts

3. Scheduled Tasks: System-level cron jobs or Windows tasks

4. Manual Checks: On-demand via CLI commands

5. Periodic Checker: In-process periodic checks for long-running apps

# Setup all strategies
mcp-vector-search auto-index setup --method all

# Check status
mcp-vector-search auto-index status

Configuration

Projects are configured via .mcp-vector-search/config.json:

{
  "project_root": "/path/to/project",
  "file_extensions": [".py", ".js", ".ts"],
  "embedding_model": "sentence-transformers/all-MiniLM-L6-v2",
  "similarity_threshold": 0.75,
  "languages": ["python", "javascript", "typescript"],
  "watch_files": true,
  "cache_embeddings": true,
  "skip_dotfiles": true,
  "respect_gitignore": true
}

Indexing Configuration Options

skip_dotfiles (default: true)

• Controls whether files and directories starting with "." are skipped during indexing

• Whitelisted directories are always indexed regardless of this setting:

  • .github/ - GitHub workflows and actions

  • .gitlab-ci/ - GitLab CI configuration

  • .circleci/ - CircleCI configuration

• When false: All dotfiles are indexed (subject to gitignore rules if respect_gitignore is true)

respect_gitignore (default: true)

• Controls whether .gitignore patterns are respected during indexing

• When false: Files in .gitignore are indexed (subject to skip_dotfiles if enabled)

force_include_patterns (default: [])

• Glob patterns to force-include files/directories even if they are gitignored

• Patterns support ** for recursive matching (e.g., repos/**/*.java matches all Java files in repos/ and subdirectories)

• Force-include patterns override .gitignore rules, allowing selective indexing of gitignored directories

• Example use case: Index specific file types in a gitignored repos/ directory

Example: Force-include Java files from gitignored directory

# Set force_include_patterns via JSON list
mcp-vector-search config set force_include_patterns '["repos/**/*.java", "repos/**/*.kt"]'

# Or add patterns one at a time (requires custom CLI command)
# This allows .gitignore to exclude repos/ from git, but mcp-vector-search still indexes Java/Kotlin files

Example config.json with force_include_patterns:

{
  "respect_gitignore": true,
  "force_include_patterns": [
    "repos/**/*.java",
    "repos/**/*.kt",
    "vendor/internal/**/*.go"
  ]
}

Configuration Use Cases

Default Behavior (Recommended for most projects):

# Skip dotfiles AND respect .gitignore
mcp-vector-search config set skip_dotfiles true
mcp-vector-search config set respect_gitignore true

Index Everything (Useful for deep code analysis):

# Index all files including dotfiles and gitignored files
mcp-vector-search config set skip_dotfiles false
mcp-vector-search config set respect_gitignore false

Index Dotfiles but Respect .gitignore:

# Index configuration files but skip build artifacts
mcp-vector-search config set skip_dotfiles false
mcp-vector-search config set respect_gitignore true

Skip Dotfiles but Ignore .gitignore:

# Useful when you want to index files in .gitignore but skip hidden config files
mcp-vector-search config set skip_dotfiles true
mcp-vector-search config set respect_gitignore false

Selective Gitignore Override with Force-Include Patterns:

# Index specific file types from gitignored directories
# Example: .gitignore excludes repos/, but you want to index Java/Kotlin files
mcp-vector-search config set respect_gitignore true
mcp-vector-search config set force_include_patterns '["repos/**/*.java", "repos/**/*.kt"]'

# This allows:
# - .gitignore to exclude repos/ from git (keeps your repo clean)
# - mcp-vector-search to index Java/Kotlin files in repos/ (semantic search)
# - Other files in repos/ (e.g., .class, .jar) remain excluded

🏗️ Architecture

Core Components

• Parser Registry: Extensible system for language-specific parsing

• Semantic Indexer: Efficient code chunking and embedding generation

• Vector Database: LanceDB for similarity search

• File Watcher: Real-time monitoring and incremental updates

• CLI Interface: Rich, user-friendly command-line experience

Supported Languages

MCP Vector Search supports 13 programming languages with full semantic search capabilities:

New Language Support

HTML Support (Unreleased):

• Semantic Extraction: Content from h1-h6, p, section, article, main, aside, nav, header, footer

• Intelligent Chunking: Based on heading hierarchy (h1-h6)

• Context Preservation: Maintains class and id attributes for searchability

• Script/Style Filtering: Ignores non-content elements

• Use Cases: Static sites, documentation, web templates, HTML fragments

Dart/Flutter Support (v0.4.15):

• Widget Detection: StatelessWidget, StatefulWidget recognition

• State Classes: Automatic parsing of _WidgetNameState patterns

• Async Support: Future and async function handling

• Dartdoc: Triple-slash comment extraction

• Tree-sitter AST: Fast, accurate parsing with regex fallback

PHP Support (v0.5.0):

• Class Detection: Classes, interfaces, traits

• Method Extraction: Public, private, protected, static methods

• Magic Methods: __construct, __get, __set, __call, etc.

• PHPDoc: Full comment extraction

• Laravel Patterns: Controllers, Models, Eloquent support

• Tree-sitter AST: Fast parsing with regex fallback

Ruby Support (v0.5.0):

• Module/Class Detection: Full namespace support (::)

• Method Extraction: Instance and class methods

• Special Syntax: Method names with ?, ! support

• Attribute Macros: attr_accessor, attr_reader, attr_writer

• RDoc: Comment extraction (# and =begin...=end)

• Rails Patterns: ActiveRecord, Controllers support

• Tree-sitter AST: Fast parsing with regex fallback

🤝 Contributing

We welcome contributions! Please see our Contributing Guide for details.

Development Setup

# Clone the repository
git clone https://github.com/bobmatnyc/mcp-vector-search.git
cd mcp-vector-search

# Install development environment (includes dependencies + editable install)
make dev

# Test CLI from source (recommended during development)
./scripts/dev-mcp version        # Shows [DEV] indicator
./scripts/dev-mcp search "test"  # No reinstall needed after code changes

# Run tests and quality checks
make test-unit           # Run unit tests
make quality            # Run linting and type checking
make fix                # Auto-fix formatting issues

# View all available targets
make help

For detailed development workflow and dev-mcp usage, see the Development section below.

Adding Language Support

1. Create a new parser in src/mcp_vector_search/parsers/

2. Extend the BaseParser class

3. Register the parser in parsers/registry.py

4. Add tests and documentation

📊 Performance

• Indexing Speed: ~1000 files/minute (typical Python project)

• Search Latency: 3.4ms median with IVF-PQ index (4.9x faster than without)

• Memory Usage: ~50MB baseline + ~1MB per 1000 code chunks

• Storage: ~1KB per code chunk (compressed embeddings)

⚠️ Known Limitations (Alpha)

• Tree-sitter Integration: Currently using regex fallback parsing (Tree-sitter setup needs improvement)

• Search Relevance: Embedding model may need tuning for code-specific queries

• Error Handling: Some edge cases may not be gracefully handled

• Documentation: API documentation is minimal

• Testing: Limited test coverage, needs real-world validation

🙏 Feedback Needed

We're actively seeking feedback on:

• Search Quality: How relevant are the search results for your codebase?

• Performance: How does indexing and search speed feel in practice?

• Usability: Is the CLI interface intuitive and helpful?

• Language Support: Which languages would you like to see added next?

• Features: What functionality is missing for your workflow?

Please open an issue or start a discussion to share your experience!

🔮 Roadmap

v2.5: Production (Current) ✅

• Core CLI interface

• Multi-language parsing (13 languages: Python, JavaScript, TypeScript, C#, Dart, PHP, Ruby, Java, Go, Rust, HTML, Markdown, Text)

• LanceDB default backend (ChromaDB legacy support)

• Apple Silicon optimizations (2-4x speedup with MPS)

• File watching and auto-reindexing

• MCP server implementation with 17 tools

• Advanced search modes (semantic, contextual, similar code)

• Code analysis tools (complexity, dead code detection, code smells)

• Interactive D3.js visualization (5+ views: Treemap, Sunburst, Force Graph, KG, Heatmap)

• Knowledge Graph with KuzuDB (entity extraction, relationship mapping)

• Development narrative generation (story command)

• Chat mode with LLM integration (iterative refinement, up to 30 queries)

• CodeT5+ code-specific embeddings

• Pipeline parallelism (37% faster indexing)

• Production-ready performance (write buffering, GPU acceleration, async pipeline)

• IVF-PQ vector index with two-stage retrieval (4.9x faster queries)

• Contextual chunking (metadata-enriched embeddings, 35-49% fewer retrieval failures)

• CodeRankEmbed model support (nomic-ai/CodeRankEmbed, 768d, 8K context)

• Document ontology with 23 categories (kg ontology command)

v2.6+: Enhancements 🔮

• Hybrid search (vector + keyword + BM25)

• Additional language support (more languages beyond 13)

• IDE extensions (VS Code, JetBrains)

• Team collaboration features

• Advanced code refactoring suggestions

• Real-time collaboration on knowledge graph

• Multi-project knowledge graph federation

🛠️ Development

Three-Stage Development Workflow

Stage A: Local Development & Testing

# Setup development environment
make dev

# Run development tests
make test-unit

# Run CLI from source (recommended during development)
./dev-mcp version        # Visual [DEV] indicator
./dev-mcp status         # Any command works
./dev-mcp search "auth"  # Immediate feedback on changes

# Run quality checks
make quality

# Alternative: use uv run directly
uv run mcp-vector-search version

Using the dev-mcp Development Helper

The ./dev-mcp script provides a streamlined way to run the CLI from source code during development, eliminating the need for repeated installations.

Key Features:

• Visual [DEV] Indicator: Shows [DEV] prefix to distinguish from installed version

• No Reinstall Required: Reflects code changes immediately

• Complete Argument Forwarding: Works with all CLI commands and options

• Verbose Mode: Debug output with --verbose flag

• Built-in Help: Script usage with --help

Usage Examples:

# Basic commands (note the [DEV] prefix in output)
./dev-mcp version
./dev-mcp status
./dev-mcp index
./dev-mcp search "authentication logic"

# With CLI options
./dev-mcp search "error handling" --limit 10
./dev-mcp index --force

# Script verbose mode (shows Python interpreter, paths)
./dev-mcp --verbose search "database"

# Script help (shows dev-mcp usage, not CLI help)
./dev-mcp --help

# CLI command help (forwards --help to the CLI)
./dev-mcp search --help
./dev-mcp index --help

When to Use:

• ./dev-mcp → Development workflow (runs from source code)

• mcp-vector-search → Production usage (runs installed version via pipx/pip)

Benefits:

• Instant Feedback: Changes to source code are reflected immediately

• No Build Step: Skip the reinstall cycle during active development

• Clear Context: Visual [DEV] indicator prevents confusion about which version is running

• Error Handling: Built-in checks for uv installation and project structure

Requirements:

• Must have uv installed (pip install uv)

• Must run from project root directory

• Requires pyproject.toml in current directory

Stage B: Local Deployment Testing

# Build and test clean deployment
./scripts/deploy-test.sh

# Test on other projects
cd ~/other-project
mcp-vector-search init && mcp-vector-search index

Stage C: PyPI Publication

# Publish to PyPI
./scripts/publish.sh

# Verify published version
pip install mcp-vector-search --upgrade

Quick Reference

./scripts/workflow.sh  # Show workflow overview

See DEVELOPMENT.md for detailed development instructions.

📚 Documentation

For comprehensive documentation, see docs/index.md - the complete documentation hub.

Getting Started

• Installation Guide - Complete installation instructions

• First Steps - Quick start tutorial

• Configuration - Basic configuration

User Guides

• Searching Guide - Master semantic code search

• Indexing Guide - Indexing strategies and optimization

• CLI Usage - Advanced CLI features

• MCP Integration - AI tool integration

• File Watching - Real-time index updates

Reference

• CLI Commands - Complete command reference

• Configuration Options - All configuration settings

• Features - Feature overview

• Architecture - System architecture

Development

• Contributing - How to contribute

• Testing - Testing guide

• Code Quality - Linting and formatting

• API Reference - Internal API docs

• Deployment - Release and deployment guide

Advanced

• Troubleshooting - Common issues and solutions

• Performance - Performance optimization

• Extending - Adding new features

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request. For major changes, please open an issue first to discuss what you would like to change.

📄 License

Elastic License 2.0 - see LICENSE file for details.

Note: This software may not be provided to third parties as a hosted or managed service.

🙏 Acknowledgments

• LanceDB for vector database

• Tree-sitter for parsing infrastructure

• Sentence Transformers for embeddings

• Typer for CLI framework

• Rich for beautiful terminal output

Built with ❤️ for developers who love efficient code search