Semantic Grep

A local semantic search tool for your codebase. Instead of matching exact strings like regular grep, this uses a locally running LLM (Ollama with llama3.2) to understand what you mean and find relevant code/text.

How it works

1. Indexer scans your project files and splits them into chunks
2. Embedding generates vector embeddings for each chunk using Ollama
3. Search compares your natural language query against all chunks using cosine similarity and returns the best matches

Requirements

• Python 3.10+
• Ollama installed and running (ollama serve)
• llama3.2 model pulled (ollama pull llama3.2)
• requests library (pip install requests)

Usage

Search with a query directly:

python main.py "find the database connection logic" -d /path/to/project

Or run in interactive mode (no query argument):

python main.py -d /path/to/project

Options

• query — natural language search query (optional, starts interactive mode if omitted)
• -d, --dir — root directory to scan (default: current directory)
• -n, --results — number of results to show (default: 10)

Project structure

• config.py — settings (model name, allowed file types, chunk size, etc.)
• indexer.py — file walking and chunking
• embedding.py — Ollama embedding generation
• search.py — cosine similarity search and result formatting
• main.py — CLI entry point