#vector-search #onnx #semantic #code #statistics

bin+lib vectordb-cli

A CLI tool for semantic code search

6 releases (stable)

new 1.6.0 Apr 19, 2025
1.5.0 Apr 18, 2025

#5 in #code

Download history 674/week @ 2025-04-11

674 downloads per month

MIT license

660KB
12K SLoC

vectordb-cli

A lightweight command-line tool for fast, local code search using semantic retrieval powered by ONNX models and Qdrant. Now with multi-repository and branch-aware indexing!

Note: This repository contains both the vectordb-cli command-line tool and the underlying vectordb_lib library.

Table of Contents

Features

  • Semantic Search: Finds relevant code chunks based on meaning using ONNX models.
  • Repository Management: Manage configurations for multiple Git repositories.
  • Branch-Aware Indexing: Track and sync specific branches within repositories.
  • Qdrant Backend: Utilizes a Qdrant vector database instance for scalable storage and efficient search.
  • Local or Remote Qdrant: Can connect to a local Dockerized Qdrant or a remote instance.
  • Simple Indexing (Default): Recursively indexes specified directories (can be used alongside repository management).
  • Configurable: Supports custom ONNX embedding models/tokenizers and Qdrant connection details via config file or environment variables.
  • Semantic Code Editing: Powerful code editing capabilities that leverage its semantic understanding of code:
    • Semantic element targeting - Edit entire classes, functions, or methods using semantic identifiers
    • Line-based precision edits - Make targeted changes to specific sections of code
    • Validation-first workflow - Validate edits before applying them to ensure safety
    • Both CLI and gRPC interfaces - Use from the command line or programmatically

Use Cases

  • Debugging Assistance: Use semantic search to find potentially related code sections when investigating bugs. Combine with LLMs by providing relevant code snippets found through queries for diagnosis, explanation, or generating flow charts.
  • Code Exploration & Understanding: Quickly locate definitions, implementations, or usages of functions, classes, or variables across large codebases or multiple repositories, even if you don't know the exact name.
  • Finding Examples: Locate examples of how a particular API, library function, or design pattern is used within your indexed code.
  • Onboarding: Help new team members find relevant code sections related to specific features or concepts they need to learn.
  • Automated Code Editing: Make precise semantic-aware edits to code without manual file editing:
    • Replace entire classes or functions using semantic targeting
    • Add methods to existing classes with line-based targeting
    • Validate edits before applying for safety and reliability
    • Build automated refactoring workflows using the gRPC API
  • Building AI Coding Tools: Integrate with the VectorDB server using the vectordb-client crate to build your own AI-powered development tools, agents, or custom workflows.
  • Documentation Search: Index and search through Markdown documentation alongside code (Note: Current Markdown parsing is basic but will be improved).
  • Refactoring & Auditing: Identify code locations potentially affected by refactoring or search for specific patterns related to security or best practices.

Supported Languages

The CLI uses tree-sitter for Abstract Syntax Tree (AST) parsing to extract meaningful code chunks (like functions, classes, structs) for indexing. This leads to more contextually relevant search results compared to simple line-based splitting. Here is the current status of language support:

Language Status Supported Elements
Rust ✅ Supported functions, structs, enums, impls, traits, mods, macros, use, extern crates, type aliases, unions, statics, consts
Ruby ✅ Supported modules, classes, methods, singleton_methods
Go ✅ Supported functions, methods, types (struct/interface), consts, vars
Python ✅ Supported functions, classes, top-level statements
JavaScript ✅ Supported functions, classes, methods, assignments
TypeScript ✅ Supported functions, classes, methods, interfaces, enums, types, assignments
Markdown ✅ Supported headings, code blocks, list items, paragraphs
YAML ✅ Supported documents
Other ✅ Supported Whole file chunk (fallback_chunk)

Files with unsupported extensions will automatically use the whole-file fallback mechanism.

Planned Languages:

Support for the following languages is planned for future releases:

  • Java (.java)
  • C# (.cs)
  • C++ (.cpp, .h, .hpp)
  • C (.c, .h)
  • PHP (.php)
  • Swift (.swift)
  • Kotlin (.kt, .kts)
  • HTML (.html)
  • CSS (.css)
  • JSON (.json)

Setup

For new users, the Local Quickstart Guide provides minimal steps to get up and running quickly.

Prerequisites

  • Rust: Required for building the project. Install from rustup.rs.
    curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
    # After installing rustup, source the Cargo environment script or restart your terminal
    source "$HOME/.cargo/env"
    
  • Git: Required for repository management features (repo add, repo sync, etc.).
  • Build Tools: Rust often requires a C linker and build tools.
    • Linux (Debian/Ubuntu):
      sudo apt-get update && sudo apt-get install build-essential git-lfs libssl-dev pkg-config
      
    • macOS: Install the Xcode Command Line Tools. If you don't have Xcode installed, running the following command in your terminal will prompt you to install them:
      xcode-select --install
      
      Install required packages using Homebrew:
      brew install git-lfs pkg-config
      
  • Qdrant: A Qdrant instance (v1.7.0 or later recommended) must be running and accessible. See Qdrant Setup.
  • ONNX Model Files: An ONNX embedding model and its corresponding tokenizer files are required. See Installation and Configuration.

Qdrant Setup

vectordb-cli requires a running Qdrant instance. Each managed repository will have its own collection in Qdrant, named repo_<repository_name>.

Option 1: Docker (Recommended for Local Use)

docker run -p 6333:6333 -p 6334:6334 \
    -v $(pwd)/qdrant_storage:/qdrant/storage:z \
    qdrant/qdrant:latest

This starts Qdrant with the default HTTP/REST port (6333, used for the web UI at http://localhost:6333/dashboard) and gRPC port (6334, used by vectordb-cli) mapped to your host. Data will be persisted in the qdrant_storage directory in your current working directory.

Option 2: Qdrant Cloud or Other Deployment

Follow the instructions for your chosen deployment method. You will need the URL (including http:// or https:// and the port, typically 6333 for gRPC) and potentially an API Key if required by your setup.

Environment Setup Guides

For specific environment configurations (GPU acceleration), refer to the guides in the docs/ directory:

Installation

  1. Clone the Repository:

    git clone https://gitlab.com/amulvany/vectordb-cli.git
    cd vectordb-cli
    
  2. Prepare ONNX Model & Tokenizer: Download or obtain your desired ONNX embedding model (.onnx file) and its tokenizer configuration (tokenizer.json and potentially other files like vocab.txt, merges.txt, etc., usually in a single directory). Place them in a known location. See Configuration for how to tell the tool where these are.

    Using the Example Model: This repository includes an example all-MiniLM-L6-v2 model in the onnx/ directory, managed via Git LFS. If you followed the prerequisites and installed Git LFS, Git should handle pulling the model files automatically when you clone or pull updates. If the .onnx file in onnx/model/ is small (a pointer file), you might need to run git lfs pull manually.

    Note: The tool dynamically detects the embedding dimension from the provided .onnx model.

  3. Build:

    • Standard (CPU):
      cargo build --release
      
    • With CUDA GPU Support (Linux): Ensure you have NVIDIA drivers, the CUDA toolkit, and cudnn installed (see docs/CUDA_SETUP.md). Then build with:
      cargo build --release --features ort/cuda
      
    • With Metal GPU Support (macOS): (See docs/MACOS_GPU_SETUP.md)
      cargo build --release --features ort/coreml # Or ort/metal if preferred/available
      
    • With Server Support: To build with gRPC server functionality:
      cargo build --release --features server
      
    • With Server and GPU Support: To combine server functionality with GPU acceleration:
      cargo build --release --features ort/cuda,server # Linux with NVIDIA GPU
      # OR
      cargo build --release --features ort/coreml,server # macOS with Metal GPU
      

    For a complete reference of all build options and feature flags, see Compilation Options.

  4. Understanding the Build Process (Linux/macOS):

    • The project uses a build script (build.rs) to simplify setup.
    • During the build, this script automatically finds the necessary ONNX Runtime libraries (downloaded by the ort crate to ~/.cache/ort.pyke.io/) including provider-specific libraries (like CUDA .so files or macOS .dylib files).
    • It copies these libraries into the final build output directory (target/release/lib/).
    • It sets the necessary RPATH ($ORIGIN/lib on Linux, @executable_path/lib on macOS) on the vectordb-cli executable.
    • This means you typically do not need to manually set LD_LIBRARY_PATH (Linux) or DYLD_LIBRARY_PATH (macOS).
  5. Install Binary (Optional): Symlink the compiled binary to a location in your PATH.

    # Example for Linux/macOS to set it up globally
    sudo ln -s $PWD/target/release/vectordb-cli /usr/local/bin
    

Configuration

vectordb-cli uses a hierarchical configuration system:

  1. Command-line Arguments: Highest priority (e.g., --onnx-model-path-arg, --onnx-tokenizer-dir-arg).
  2. Environment Variables: Second priority.
  3. Configuration File (config.toml): Lowest priority.

Environment Variables

  • QDRANT_URL: URL of the Qdrant gRPC endpoint (e.g., http://localhost:6334). Defaults to http://localhost:6334 if not set.
  • QDRANT_API_KEY: API key for Qdrant authentication (optional).
  • VECTORDB_ONNX_MODEL: Full path to the .onnx model file.
  • VECTORDB_ONNX_TOKENIZER_DIR: Full path to the directory containing the tokenizer.json file.

Configuration File (config.toml)

The tool looks for a config.toml file in the XDG configuration directory:

  • Linux/macOS: ~/.config/vectordb-cli/config.toml

Example config.toml:

# URL for the Qdrant gRPC endpoint
qdrant_url = "http://localhost:6334"

# --- Optional: Qdrant API Key ---
# api_key = "your_qdrant_api_key"

# --- Optional: ONNX Model Configuration ---
# These are only needed if not provided via args or env vars.

# Path to the ONNX model file
onnx_model_path = "/path/to/your/model.onnx"

# Path to the directory containing tokenizer.json
# Note: Key name is `onnx_tokenizer_path`
onnx_tokenizer_path = "/path/to/your/tokenizer_directory"

# --- Optional: Repository Storage Configuration ---
# Base path where all repositories will be stored
# If not provided, uses ~/.local/share/vectordb-cli/repositories
repositories_base_path = "/path/to/your/repositories"

# --- Repository Management ---
# The active repository (used by default for commands like sync, query)
# Set via `repo use <n>`
active_repository = "my-project"

# List of managed repositories
[[repositories]]
name = "my-project"
# Local path where the repository was cloned
local_path = "/home/user/dev/my-project"
# Branches tracked by `repo sync`
tracked_branches = ["main", "develop"]
# The branch currently checked out locally
active_branch = "main" # Updated automatically by `repo use-branch`
# Last commit hash synced for each tracked branch
# Updated automatically by `repo sync`
[repositories.last_synced_commits]
main = "a1b2c3d4e5f6..."
develop = "f6e5d4c3b2a1..."

[[repositories]]
name = "another-repo"
local_path = "/home/user/dev/another-repo"
tracked_branches = ["release-v1"]
active_branch = "release-v1"
[repositories.last_synced_commits]
release-v1 = "deadbeef..."

# ... other repositories ...

Note: You must provide the ONNX model and tokenizer paths via one of these methods (arguments, environment variables, or config file) for commands like index, query, and repo sync to work. The repositories section is managed automatically by the repo subcommands.

Usage (CLI)

This section focuses on the vectordb-cli command-line tool.

Global Options

These options can be used with most commands:

  • -m, --onnx-model <PATH>: Path to the ONNX model file (overrides config & env var).
  • -t, --onnx-tokenizer-dir <PATH>: Path to the ONNX tokenizer directory (overrides config & env var).

Simple Commands (simple)

These commands operate on a default, non-repository-specific Qdrant collection (vectordb-code-search).

simple index

Recursively indexes files in specified directories or specific files into the default collection.

vectordb-cli simple index <PATHS>... [-e <ext>] [--extension <ext>]
  • <PATHS>...: One or more file or directory paths to index.
  • -e <ext>, --extension <ext>: Optional: Filter by specific file extensions (without the dot, e.g., -e rs, -e py). If omitted, attempts to parse based on known extensions.

simple query

Performs a semantic search against the default collection.

vectordb-cli simple query "<query text>" [-l <limit>] [--lang <language>] [--type <element_type>]
  • <query text>: The natural language query.
  • -l <limit>, --limit <limit> (Optional): Max number of results (default: 10).
  • --lang <language> (Optional): Filter by language (e.g., rust, python).
  • --type <element_type> (Optional): Filter by code element type (e.g., function).

simple clear

Deletes the entire simple index collection (vectordb-code-search). This does not affect repository indices. Requires confirmation unless -y is provided.

vectordb-cli simple clear [-y]
  • -y: Confirm deletion without prompting.

Repository Management (repo)

This subcommand group manages configurations for Git repositories, allowing you to index and query specific branches within dedicated Qdrant collections (repo_<repository_name>).

repo add

Clones a Git repository locally (if not already present) and adds it to the managed list.

vectordb-cli repo add --url <repo-url> [--local-path <path>] [--name <repo-name>] [--branch <branch-name>] [--remote <remote_name>] [--ssh-key <path>] [--ssh-passphrase <passphrase>]
  • --url <repo-url>: The URL of the Git repository (HTTPS or SSH).
  • --local-path <path> (Optional): Local directory to clone into (defaults to <config_dir>/repos/<repo_name>).
  • --name <repo-name> (Optional): Name for the repository configuration (defaults to deriving from URL).
  • --branch <branch-name> (Optional): Initial branch to track (defaults to the repo's default).
  • --remote <remote_name> (Optional): Name for the Git remote (defaults to "origin").
  • --ssh-key <path> (Optional): Path to the SSH private key file for authentication.
  • --ssh-passphrase <passphrase> (Optional): Passphrase for the SSH key.

repo config

Configure repository management settings.

vectordb-cli repo config set-repo-base-path <path>
  • <path>: The directory path where all repositories will be stored by default.

This command sets the global repository storage location. New repositories added with repo add will be stored in this directory unless overridden with --local-path. Existing repositories will remain at their current locations.

repo list

Lists all configured repositories, their URLs, local paths, tracked branches, and detected indexed languages. Indicates the active repository with a *.

vectordb-cli repo list

Example Output:

Managed Repositories:
 * my-project (https://github.com/user/my-project.git) -> /home/user/.config/vectordb-cli/repos/my-project
     Default Branch: main
     Active Branch: main
     Tracked Branches: ["main", "develop"]
     Indexed Languages: rust, markdown
   another-repo (https://github.com/user/another.git) -> /home/user/.config/vectordb-cli/repos/another-repo
     Default Branch: main
     Active Branch: main
     Tracked Branches: ["main"]
     Indexed Languages: python

repo use

Sets a repository as the active one, used by default for other repo subcommands like query, sync, use-branch, clear, stats.

vectordb-cli repo use <name>
  • <name>: (Required) The name of the repository configuration to activate.

repo remove

Removes a repository configuration and its corresponding Qdrant collection (repo_<name>). This also removes the local clone by default.

vectordb-cli repo remove <name> [-y]
  • <name>: (Required) The name of the repository configuration to remove.
  • -y: Skip confirmation prompt.

This operation is irreversible and deletes the Qdrant data and local clone.

repo use-branch

Checks out a specific branch in the active repository locally and adds it to the list of tracked branches for syncing.

vectordb-cli repo use-branch <branch_name>
  • <branch_name>: (Required) The name of the branch to check out and track. Fetches from the configured remote if the branch isn't available locally.

repo sync

Fetches updates from the configured remote for the currently checked-out, tracked branch of the active repository (or specified repository). It calculates changes since the last sync and updates the Qdrant index accordingly (adding/modifying/deleting points).

vectordb-cli repo sync [-n <name>] [--name <name>] [-e <ext>,...] [--extensions <ext>,...] [--force]
  • -n <name>, --name <name> (Optional): Name of the repository to sync. Defaults to the active repository.
  • -e <ext>,..., --extensions <ext>,... (Optional): Specify file extensions to sync (without the dot, comma-separated or multiple flags: -e rs,py or -e rs -e py). If omitted, syncs files matching known parsers.
  • --force (Optional): Force a full re-index of the specified files for the branch, ignoring the last synced commit state.

repo clear

Clears the index (Qdrant collection repo_<repo_name>) for a specific repository without removing the repository configuration or local clone. Requires confirmation unless -y is provided.

vectordb-cli repo clear [-n <name>] [--name <name>] [-y]
  • -n <name>, --name <name> (Optional): The name of the repository index to clear. Defaults to the active repository.
  • -y: Confirm deletion without prompting.

This operation is irreversible.

repo query

Performs a semantic search across the indexed data for the active repository.

vectordb-cli repo query "<query text>" [-l <limit>] [--lang <language>] [--type <element_type>]
  • <query text>: The natural language query.
  • -l <limit>, --limit <limit> (Optional): Max number of results (default: 10).
  • --lang <language> (Optional): Filter by language (e.g., rust, python).
  • --type <element_type> (Optional): Filter by code element type (e.g., function).

Results display file paths (relative to the repository root), line numbers, scores, and the relevant code chunk.

repo stats

Displays statistics (like point count) about the Qdrant collection for the active repository.

vectordb-cli repo stats

Library integration via gRPC

To integrate semantic code search functionality into your own applications, use the vectordb-client crate to connect to VectorDB in server mode.

use vectordb_client::VectorDBClient;

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Connect to the server
    let client = VectorDBClient::default().await?;
    
    // Perform a semantic search
    let results = client.query(
        "implementing authentication middleware", 
        10,  // limit 
        None, // language filter (optional)
        None, // element type filter (optional)
    ).await?;
    
    // Process the results
    for result in results.results {
        println!("{}:{} (Score: {})", result.file_path, result.line_number, result.score);
        println!("{}", result.content);
        println!("---");
    }
    
    Ok(())
}

For detailed information on client usage, see the Server Mode Documentation.

Development

The project has 42% unit test coverage and thorough end-to-end testing for key features.

# Run tests without server features (faster, fewer dependencies)
cargo test

# Run tests including server functionality
cargo test --features server

# Run only ignored tests (many server tests are ignored as they require a running server)
cargo test --features server -- --ignored

# Run clippy
cargo clippy --all-targets -- -D warnings

# Format code
cargo fmt

Certain tests are conditionally compiled based on feature flags to allow for faster testing during development. Server-specific functionality is guarded behind the server feature flag.

Contributing

(Contribution guidelines)

License

MIT License

Server Mode

VectorDB-CLI can be run as a gRPC server, allowing you to integrate semantic code search into your own applications.

Note: Server functionality requires compiling with the server feature flag: cargo build --release --features server. See the Build section for details.

# Start the server with default settings (localhost:50051)
vectordb-cli server start

# Or with custom host and port
vectordb-cli server start --host 0.0.0.0 --port 8080

# With authentication
vectordb-cli server start --api-key your_secret_key

# With TLS
vectordb-cli server start --tls --tls-cert /path/to/cert.pem --tls-key /path/to/key.pem

For detailed information on server configuration, API usage, and client examples, see the Server Mode Documentation.

gRPC API

The server exposes a gRPC API that can be used by clients in any language. The API is defined in the proto/vectordb.proto file.

Client libraries:

Usage Examples

Code Editing

The edit feature allows you to make precise changes to your code with built-in validation:

# Example: Replace a class with semantic targeting
vectordb-cli edit apply --file src/my_app.py --element "class:Calculator" --content-file new_calculator.py

# Example: Add a method to a class with line-based targeting
vectordb-cli edit apply --file src/my_app.py --line-start 25 --line-end 25 \
  --content "    def multiply(self, x):\n        self.value *= x\n        return self.value"

# Example: Validate before applying an edit
vectordb-cli edit validate --file src/my_app.py --element "function:process_data" --content-file new_function.py

For more details and best practices, see the edit feature documentation.

Dependencies

~121MB
~3M SLoC