Embgrep

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

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

QuartzUnit Ecosystem

Used in

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

License

MIT

---

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