MooFile

A lightweight, embedded, single-file document store with a developer-friendly query API.
No server. No infrastructure. Just a file and a library.
🦀 Rust core available — 2-24× faster than pure Python.
🧠 On-device autoembedding — local embedding models for semantic search.

from moofile import Collection, count, mean

with Collection("mydata.bson", 
                indexes=["email", "age"],
                vector_indexes={"embedding": 1024},
                text_indexes=["content"],
                auto_embed={
                    "content": {
                        "model": "hf:jsonMartin/voyage-4-nano-gguf:voyage-4-nano-q8_0.gguf",
                        "target": "embedding",
                        "precision": "int8",
                    },
                }) as db:
    
    # Insert — auto-embeds content into embedding (int8, 1KB/doc)
    db.insert({
        "name": "Alice", 
        "email": "alice@example.com", 
        "age": 30,
        "content": "Machine learning and data science expert",
    })

    # Traditional query
    results = db.find({"age": {"$gt": 25}}).sort("age").to_list()
    
    # Vector similarity search (raw vector)
    similar = db.find({}).vector_search("embedding", query_vector, limit=5).to_list()
    
    # Semantic search — auto-embeds query text
    similar = db.find({}).semantic("content", "data science", limit=5).to_list()
    
    # BM25 text search
    text = db.find({}).text_search("content", "machine learning", limit=10).to_list()
    
    # Hybrid search — auto-embeds query vector from query text
    results = db.find({}).hybrid_search("content", "content", "data science", None, 10).to_list()

Why MooFile?

Target dataset size: megabytes to single-digit gigabytes.

Installation

pip install moofile

This installs the pure-Python version which works everywhere. See Native install below for the Rust-powered version.

Quick Start

from moofile import Collection

db = Collection("users.bson", 
                indexes=["email", "status"],
                text_indexes=["bio"],
                vector_indexes={"profile_vec": 128})

# Insert
alice = db.insert({"name": "Alice", "email": "a@ex.com", "age": 30, "status": "active"})
db.insert_many([...])

# Query
active = db.find({"status": "active"}).to_list()
young  = db.find({"age": {"$lt": 30}}).sort("age").to_list()
one    = db.find_one({"email": "alice@example.com"})

# Vector search
similar = db.find({}).vector_search("profile_vec", query_vector, limit=3).to_list()
for doc, score in similar:
    print(f"{doc['name']}: {score:.3f}")

# Text search
results = db.find({}).text_search("bio", "machine learning", limit=5).to_list()

# Update & Delete
db.update_one({"email": "a@ex.com"}, set={"age": 31})
db.update_many({"status": "trial"}, set={"status": "expired"})
db.delete_one({"email": "c@ex.com"})
db.delete_many({"status": "expired"})

With Autoembedding

from moofile import Collection

# Autoembedding: text in "abstract" is automatically embedded into
# "embedding" on insert, using a local GGUF model (downloaded on first use).
db = Collection("papers.bson",
    indexes=["year", "category"],
    vector_indexes={"embedding": 1024},
    auto_embed={
        "abstract": {
            "model": "hf:jsonMartin/voyage-4-nano-gguf:voyage-4-nano-q8_0.gguf",
            "target": "embedding",
            "dims": 1024,
            "precision": "int8",
        },
    })

# Insert — auto-embeds abstract → embedding (1 KB, int8 quantized)
db.insert({"title": "Quantum ML", "abstract": "Quantum computing for ML...", "year": 2025})

# Semantic search — query text is auto-embedded using the same model
results = db.find({"year": 2025}).semantic("abstract", "quantum algorithms", 5).to_list()
for doc, score in results:
    print(f"{doc['title']}: {score:.3f}")

# Hybrid search — auto-embeds query_text for the vector leg
results = db.find({}).hybrid_search("abstract", "abstract", "quantum", None, 10).to_list()

Native Install (Rust Core)

When the Rust native extension is installed, import moofile transparently uses it — same API, 2-24× faster.

From source (requires Rust)

# Install Rust: https://rustup.rs
curl --proto '=https' --tls v1.2 -sSf https://sh.rustup.rs | sh

# Build and install with native extension
pip install maturin
cd moofile
maturin develop --release

Prebuilt wheels

Coming soon — GitHub Actions CI will build platform wheels for:

In the meantime, pip install moofile always works (pure Python fallback).

CLI Tools

moosh users.bson --indexes email,age
moo2json users.bson users.json
moo2json --import users.json users.bson --indexes email
moo2mongo users.bson --uri mongodb://localhost/mydb --collection users
moo2sqlite users.bson users.db --table people

Full Documentation

• Specification — file format, architecture, design decisions
• API Reference — complete Python API, filter operators, aggregation
• bench_native.py — Python vs Rust head-to-head benchmark

Development

# Pure-Python tests (always run)
pytest tests/ -v

# Cross-implementation tests
pytest tests-cross/ -v

# Rust core tests
cd core && cargo test

# Rust benchmark
cd core && cargo run --example bench --release

# Python vs Rust benchmark
PYTHONPATH=. python bench_native.py

Project layout

moofile/
├── core/                    # Rust engine (cargo build)
│   ├── src/{lib,storage,index,query,text,cache,embed,errors}.rs
│   └── examples/bench.rs    # Pure-Rust benchmark
├── bindings/python/         # PyO3 binding (maturin build)
│   └── src/lib.rs
├── moofile/                 # Python package
│   ├── __init__.py          # Auto-detects Rust, falls back to Python
│   ├── _rust_adapter.py     # Adapts NativeCollection → Collection API
│   ├── collection.py        # Pure-Python reference implementation
│   ├── query.py, index.py, storage.py, ...
│   └── cli/                 # moosh, moo2json, moo2mongo, moo2sqlite
├── tests/                   # Python test suite
├── tests-cross/             # Cross-implementation validation
└── pyproject.toml

License

MIT — see LICENSE.