csep

Cosine Similarity Embeddings Print

Version 0.3.0 - Embedding cache replaced with libSQL vector database. See MIGRATION_GUIDE.md for details.

╭─── ╭────┬───────╮
│    ╰──╮ ├─  ╭───╯  
╰───────╯ ╰── ╵

Like Grep (Global Regular Expression Print) takes a regular expression and prints all the lines that have a match in it, Csep (Cosine Similarity Embeddings Print) takes an input phrase and prints all the chunks that are similar to it.

The goal of this project is to give users command line access to semantic search in the same way that grep is used for regular expressions. This not only gives you a command line semantic search tool on any unix like system, but also allows you to use it in scripts and pipelines. If you combine it with a command line llm tool like chat-gipity or Ollama you could even potentially perform RAG in a simple unix shell script.

How it works

csep searches files by semantic meaning, so results can match the intent of a query even when the exact words are not present. Embeddings can take time to compute, so csep caches them per directory. The cache is built automatically on first run in a directory, or you can prepare it ahead of time:

csep cache --build

Cache Location: ~/.cache/csep/embeddings/{model_name}/ (see Storage Locations for details)

Usage

Default output mirrors rg: file heading followed by line:match entries, with exact query tokens highlighted.

# Basic search
csep "main entry point"

# Filter to markdown files
csep -g '*.md' "main entry point"

# vimgrep-compatible output
csep --vimgrep "main entry point"

Example output

Input:

csep "main entry point"

Output:

src/main.rs
22:#[tokio::main]
23:async fn main() {

Input:

csep -g '*.md' "main entry point"

Output:

README.md
12:Print) takes an input phrase and prints all the chunks that are similar to it.

Installation

You can then install csep from this source using:

cargo install --path .

Or you can pull whatever the latest published version is from crates.io with

cargo install csep

GPU Acceleration

To enable GPU-accelerated embeddings, build with the appropriate feature flag:

# NVIDIA GPU (CUDA)
cargo install --path . --features cuda

# Apple Silicon (Metal)
cargo install --path . --features metal

Use --no-gpu at runtime to force the CPU/ONNX backend even when built with GPU support.

Storage Locations

CSEP uses two storage locations:

• Embeddings Cache: ~/.cache/csep/embeddings/ - Stores computed file embeddings
• Models Cache: ~/.local/share/csep/models/ - Stores downloaded embedding models (fastembed only)

See STORAGE_LOCATIONS.md for detailed information about cache management and migration.