🔎 probe (pb)

probe is a smart code search tool that works in the shell like grep, but implements full-text search. It runs locally on your machine with no servers or dependencies, using full-text search with embedding-based reranking to help you, or your AI assistant find what you're looking for.

Features

• Full-text search - Better relevance scoring than simple text matching
• Any text file - Specialized Java chunking, plus a fallback indexer so Python, JS, Markdown, YAML, Makefiles, and other text files are still searchable
• ML-powered reranking - Reranking language model boosts more relevant results
• Works offline - No servers, APIs, or other dependencies (after the reranker model is downloaded from HuggingFace)
• Respects gitignore - Skips ignored files and binary files automatically
• Fast updates - Only reindex changed files, the reindexing is performed on-the-fly
• Cross-platform - Works on Linux, macOS, and Windows
• Zero setup - Works out of the box in any directory

Quick Start

# Search in current directory
probe "snippet generation"

# Search in specific directory
probe -d /path/to/project "cli argument parsing"

# Rebuild index (only useful for development of probe itself)
probe rebuild

# Show index statistics
probe stats

Installation

Pre-built binaries

Download the latest release for your platform from the releases page.

From source

git clone https://github.com/nikita2206/probe
cd probe
cargo build --release

The binary will be available at target/release/probe.

Using Cargo

cargo install --git https://github.com/nikita2206/probe

Why probe?

Better than grep: Uses full-text search to rank results by relevance, not just pattern matching. Embedding-based reranking puts the most useful results first.

Perfect for AI agents: Great alternative to vector search for AI coding assistants. Works offline, no API calls needed.

Understands code: Language model reranking helps find related functions, similar patterns, and contextually relevant matches.

Just works: No setup, no servers, no dependencies. Works like grep but smarter.

Performance

probe combines the speed of full-text search with the intelligence of embedding-based reranking. The full-text search provides fast initial results (pulling at least 10 most-relevant results), while the embedding-based reranking ensures the most relevant matches appear at the top.

Usage

Basic Search

# Search for text in current directory
probe "error handling"

# Complex queries
probe "snippet generation"

Index Management

# Rebuild index from scratch (normally not needed, probe is able to reindex on-the-fly when searching files change)
probe rebuild

# Show index statistics and file counts
probe stats

Directory Selection

# Search in specific directory
probe -d /path/to/project "search term"

# Search with custom directory
probe --dir ~/code/my-project "function"

How it works

1. File Discovery: Scans directories using the same .gitignore logic as Git
2. Full-text Indexing: Creates a full-text search index using Tantivy
3. Embedding Reranking: Uses language model embeddings to boost contextually relevant results
4. Smart Results: Combines full-text search scores with semantic similarity for optimal ranking

The index is stored in a .probe/ directory in your project root and is automatically managed.

For details on search scoring, see SCORING_GUIDE.md. For query syntax, see Tantivy's query documentation. Architecture, indexing, and language-processor snapshots live in docs/.

Configuration

probe works with zero configuration, but you can customize behavior:

• Respects .gitignore files automatically
• Skips binary files and common non-text formats
• Excludes the .probe/ index directory from searches

For detailed configuration options including project settings (probe.yml) and custom reranking models, see CONFIG.md.

Building from Source

git clone https://github.com/nikita2206/probe
cd probe
cargo build --release

Development

# Run tests
cargo test

# Run with debug output
RUST_LOG=debug cargo run -- "search term"

# Check code formatting
cargo fmt -- --check

# Run linter
cargo clippy -- -D warnings

License

This project is dual-licensed:

• MIT License - Free for personal use, open source projects, and non-commercial use
• Probe Commercial License - Required for commercial agentic use (AI assistants, automated tools, commercial services)

Commercial Agentic Use includes integrating probe into:

• AI coding assistants or development tools
• Commercial code analysis services
• Automated AI workflows that are sold or monetized
• AI-powered platforms or products

For commercial agentic use, please contact the project maintainer for licensing terms.

See the LICENSE file for complete details.

Similar Tools

• ripgrep: Fast regex search, but results aren't ranked by relevance
• grep: Basic pattern matching, no relevance scoring
• Vector search: Requires embeddings API, slower, needs network
• The Silver Searcher (ag): Fast text search but simple substring matching

probe gives you smart search results without the complexity of vector databases or API dependencies.