VectorLiteDB

SQLite for Embeddings — A simple, embedded vector database that stores everything in a single file.

No server. No setup. Just vectors in a file.

Why?

Every vector database is either a cloud service (Pinecone), needs a server (Chroma), or doesn't persist (FAISS).

Sometimes you just want to store embeddings in a file and search them. Like SQLite does for relational data.

Quick Start

pip install vectorlitedb

from vectorlitedb import VectorLiteDB

# Create a database (just a file)
db = VectorLiteDB("my_vectors.db", dimension=384)

# Store vectors with metadata
db.insert(id="doc1", vector=embedding, metadata={"text": "Hello world"})
db.insert(id="doc2", vector=embedding2, metadata={"text": "Goodbye world"})

# Search (returns most similar vectors)
results = db.search(query=query_embedding, top_k=5)

# Results format
for result in results:
    print(f"ID: {result['id']}")
    print(f"Similarity: {result['similarity']}")
    print(f"Metadata: {result['metadata']}")

Your vectors are saved in my_vectors.db.

What This Does

• Stores vectors (embeddings) with metadata in a single file
• Searches for similar vectors using cosine/L2/dot distance
• Persists everything to disk automatically
• Filters results by metadata
• Works offline - no internet required

What This Doesn't Do (Yet)

• ❌ Generate embeddings (use OpenAI, etc.)
• ❌ Scale beyond ~100K vectors (uses brute force search)
• ❌ Handle concurrent writes
• ❌ Optimize for speed

Use Cases

• Local RAG: Store document embeddings for offline search
• Personal AI: Give your assistant persistent memory
• Prototyping: Test vector search ideas without infrastructure
• Edge devices: Semantic search on Raspberry Pi
• Privacy: Keep embeddings on your device

API Reference

Create/Open Database

db = VectorLiteDB("path/to/file.db", dimension=384, distance_metric="cosine")
# distance_metric: "cosine" (default), "l2", or "dot"

Insert Vectors

db.insert(id="unique_id", vector=[0.1, 0.2, ...], metadata={"key": "value"})

Search Vectors

results = db.search(
    query=[0.1, 0.2, ...], 
    top_k=5,
    filter=lambda meta: meta.get("type") == "document"  # optional
)

Get/Delete Vectors

vector, metadata = db.get("unique_id")
db.delete("unique_id")

Database Info

print(len(db))  # number of vectors
print(repr(db))  # database summary

Current Status

This is v0.1.0 - a working alpha focused on simplicity over performance.

What works:

• ✅ Basic CRUD operations
• ✅ File persistence
• ✅ Metadata filtering
• ✅ ~100ms search for 10K vectors

Planned:

• 🚧 Better search algorithms (IVF, HNSW)
• 🚧 Memory-mapped files for large datasets
• 🚧 Language bindings (JavaScript, Rust, Go)

Contributing

The codebase is intentionally small (~200 lines) and readable.

Perfect for:

• First-time contributors
• Algorithm implementations
• Performance optimizations
• Documentation improvements

See issues labeled good first issue.

Inspiration

• SQLite: Single-file, serverless, embedded
• FAISS: Fast similarity search

VectorLiteDB combines the best parts: embedded like SQLite, simple like a Python dict, persistent like a real database.

License

VectorLiteDB is licensed under the Apache 2.0 License.