ragrep 0.2.1

Local semantic code recall with mxbai embeddings and SQLite

RAGrep

RAGrep is a dead-simple local semantic recall tool for code and text files.

It uses:

• mxbai-embed-large embeddings in-process (no server)
• a single local SQLite database file: .ragrep.db

No ChromaDB. No remote API keys.

Install

pip install ragrep

Embedding Model Storage

RAGrep downloads the embedding model automatically on first use.

Default model directories:

• Linux: ~/.config/ragrep/models
• macOS: ~/Library/Application Support/ragrep/models
• Windows: %APPDATA%\\ragrep\\models

Override with:

• env var: RAGREP_MODEL_DIR
• CLI flag: --model-dir
• Python API: RAGrep(model_dir="...")

GPU Usage

RAGrep can use GPU for embeddings when available.

• Default behavior: RAGREP_DEVICE=auto (prefers cuda, then mps, then cpu)
• Override via env: RAGREP_DEVICE=cpu|cuda|mps|cuda:0
• Override via CLI: --device ...
• Override via Python API: RAGrep(embedding_device="...")
• Note: GPU usage requires a GPU-capable PyTorch build in your environment.
• Check runtime GPU support: ragrep --check-gpu (or ragrep --check-gpu --json)

CLI Usage

Recall is the default command.

# Implied recall (auto-indexes when needed)
ragrep "authentication middleware"

# Explicit recall (same behavior)
ragrep recall "authentication middleware"

# Build/update index manually
ragrep index .

# Show stats
ragrep stats

# Stats alias
ragrep --stats

When --path is omitted, auto-indexing uses the previously indexed root if one exists; otherwise it uses the current directory.

Indexing is incremental by default:

• new files are added
• modified files are re-embedded
• removed files are deleted from the index

CLI runs that index files print newly indexed file paths to stdout.

Useful flags:

ragrep "query text" --path . --limit 10 --db-path ./.ragrep.db
ragrep "query text" --model-dir ~/.config/ragrep/models --json
ragrep "query text" --device auto
ragrep index . --force

Python Usage

from ragrep import RAGrep

rag = RAGrep(
    db_path="./.ragrep.db",
    embedding_model="mxbai-embed-large",
    embedding_device="auto",
)
rag.index(".")

result = rag.recall("database transactions", limit=5)
for match in result["matches"]:
    print(match["score"], match["metadata"]["source"])

print(rag.stats())
rag.close()

Library methods:

• index(path=".", force=False)
• recall(query, limit=20, path=".", auto_index=True)
• stats()

Backwards-compatible aliases still available:

• RAGSystem (alias of RAGrep)
• dump(...) (alias of recall(..., auto_index=False) result list)

Local Database

RAGrep stores everything in one SQLite file (default ./.ragrep.db):

• indexed files and mtimes
• chunked source text
• embedding vectors
• index metadata (model, chunk settings, root path)

Development

pip install -e .[dev]
python -m unittest discover -s tests -p 'test_*.py'
python -m build
twine check dist/*

License

MIT