embgrep

한국어 문서 · llms.txt

Local semantic search — embedding-powered grep for files, zero external services.

  

Search your codebase and documentation by meaning, not just keywords. embgrep indexes files into local embeddings and lets you run semantic queries — no API keys, no cloud services, no vector database servers.

Features

• Local embeddings — Uses fastembed (ONNX Runtime), no API keys needed
• SQLite storage — Single-file index, no external vector DB
• Incremental indexing — Only re-indexes changed files (SHA-256 hash comparison)
• Smart chunking — Function-level splitting for code, heading-level for docs
• MCP native — 4-tool FastMCP server for LLM agent integration
• 15+ file types — .py, .js, .ts, .java, .go, .rs, .md, .txt, .yaml, .json, .toml, and more

Install

pip install embgrep              # core (fastembed + numpy)
pip install embgrep[cli]         # + click/rich CLI
pip install embgrep[mcp]         # + FastMCP server
pip install embgrep[all]         # everything

Quick Start

Python API

from embgrep import EmbGrep

eg = EmbGrep()

# Index a directory
eg.index("./my-project", patterns=["*.py", "*.md"])

# Semantic search
results = eg.search("database connection pooling", top_k=5)
for r in results:
    print(f"{r.file_path}:{r.line_start}-{r.line_end} (score: {r.score:.4f})")
    print(f"  {r.chunk_text[:80]}...")

# Incremental update (only changed files)
eg.update()

# Index statistics
status = eg.status()
print(f"{status.total_files} files, {status.total_chunks} chunks, {status.index_size_mb} MB")

eg.close()

CLI

# Index a project
embgrep index ./my-project --patterns "*.py,*.md"

# Search
embgrep search "error handling patterns"

# Filter by file type
embgrep search "async database query" --path-filter "%.py"

# Check status
embgrep status

# Update changed files
embgrep update

Convenience functions

import embgrep

embgrep.index("./src")
results = embgrep.search("authentication middleware")
status = embgrep.status()
embgrep.update()

MCP Server

Add to your Claude Desktop / MCP client configuration:

{
  "mcpServers": {
    "embgrep": {
      "command": "embgrep-mcp"
    }
  }
}

Or with uvx:

{
  "mcpServers": {
    "embgrep": {
      "command": "uvx",
      "args": ["--from", "embgrep[mcp]", "embgrep-mcp"]
    }
  }
}

MCP Tools

| Tool | Description | |------|-------------| | index_directory | Index files in a directory for semantic search | | semantic_search | Search indexed files using natural language | | index_status | Get current index statistics | | update_index | Incremental update — re-index changed files only |

How It Works

flowchart TD
    A["📁 Files"] --> B["Smart Chunking\ncode: function-level\ndocs: heading-level"]
    B --> C["fastembed\nlocal embeddings"]
    C --> D["SQLite\nvector index"]
    D --> E["🔍 Query"]
    E --> F["Cosine Similarity\nranked results"]
    F --> G["✅ Matches\nwith context"]

1. Chunking — Files are split into semantically meaningful chunks:

• Code files (.py, .js, .ts, etc.): split by function/class boundaries
• Documents (.md, .txt): split by headings or paragraph breaks
• Config files: fixed-size chunking

1. Embedding — Each chunk is converted to a 384-dimensional vector using BGE-small-en-v1.5 via ONNX Runtime (no PyTorch needed)

1. Storage — Embeddings are stored as BLOBs in a local SQLite database

1. Search — Query text is embedded and compared against all chunks using cosine similarity

Configuration

| Parameter | Default | Description | |-----------|---------|-------------| | db_path | ~/.local/share/embgrep/embgrep.db | SQLite database location | | model | BAAI/bge-small-en-v1.5 | fastembed model name | | max_chunk_size | 1000 chars | Maximum chunk size for fixed-size splitting | | top_k | 5 | Number of search results |

QuartzUnit Ecosystem

| Package | Description | |---------|-------------| | markgrab | HTML/YouTube/PDF/DOCX to LLM-ready markdown | | snapgrab | URL to screenshot + metadata | | docpick | OCR + LLM document structure extraction | | browsegrab | Local LLM browser agent | | feedkit | RSS feed collection + MCP | | embgrep | Local semantic search for files |

Used in

• newswatch — RSS news monitoring pipeline (feedkit → markgrab → embgrep → diffgrab)

License

MIT

<!-- mcp-name: io.github.QuartzUnit/embgrep -->

---

_{Part of the QuartzUnit ecosystem — composable Python libraries for data collection, extraction, search, and AI agent safety.}