dd-vectordb

Unified Vector DB abstraction layer for Python.

Add semantic search to any project in minutes. Swap backends (in-memory, FAISS, ChromaDB, Qdrant) without changing your application code.

Supported Backends

Install

pip install dd-vectordb                  # InMemoryVectorDB only (numpy)
pip install "dd-vectordb[faiss]"         # + FAISS
pip install "dd-vectordb[chroma]"        # + ChromaDB
pip install "dd-vectordb[qdrant]"        # + Qdrant
pip install "dd-vectordb[all]"           # all adapters
pip install "dd-vectordb[dev]"           # dev tools

Quick Start

import numpy as np
from dd_vectordb import InMemoryVectorDB

# 1. Embed your texts (any encoder — OpenAI, sentence-transformers, Ollama, etc.)
texts = ["The quick brown fox", "Python programming", "Vector search rocks"]
embeddings = [np.random.rand(768).tolist() for _ in texts]  # replace with real embeddings

# 2. Add to the store
db = InMemoryVectorDB()
db.add_texts(texts=texts, embeddings=embeddings)

# 3. Search
query_vec = np.random.rand(768).tolist()  # replace with real query embedding
results = db.search(query_vec, k=2)
for r in results:
    print(f"#{r.rank}  score={r.score:.4f}  {r.document.text}")

API Reference

Core methods (all adapters)

Context manager

with FAISSVectorDB(dimension=768) as db:
    db.add_texts(texts, embeddings)
    results = db.search(query, k=5)
# close() called automatically

Pydantic models

from dd_vectordb import Document, SearchResult, CollectionInfo

doc = Document(id="1", text="hello", embedding=[0.1, 0.9], metadata={"src": "wiki"})
result: SearchResult  # .document, .score, .rank
info: CollectionInfo  # .name, .adapter, .count, .dimension, .metric

Examples

With FAISS

from dd_vectordb import FAISSVectorDB

db = FAISSVectorDB(dimension=768, metric="cosine")
db.add_texts(texts=["hello world"], embeddings=[[...768 floats...]])
results = db.search([...768 floats...], k=5)

# Persist to disk
db.save("my_index.faiss")
db2 = FAISSVectorDB.load("my_index.faiss")

With ChromaDB (persistent)

from dd_vectordb import ChromaVectorDB

db = ChromaVectorDB(collection_name="my_docs", persist_directory="./chroma_data")
db.add_texts(texts=["hello"], embeddings=[[0.1, 0.9]])
results = db.search([0.1, 0.9], k=1)

With Qdrant (in-memory)

from dd_vectordb import QdrantVectorDB

db = QdrantVectorDB(dimension=768, collection_name="docs")
db.add_texts(texts=["hello"], embeddings=[[...768 floats...]])
results = db.search([...768 floats...], k=5)

Metadata filtering

db.add_texts(
    texts=["wiki article", "blog post"],
    embeddings=[emb1, emb2],
    metadatas=[{"source": "wiki"}, {"source": "blog"}],
)
# Only search within wiki documents
results = db.search(query_vec, k=5, filter={"source": "wiki"})

Cookbooks

See cookbook/ for runnable examples:

• 01_in_memory_basics.py — full walkthrough with zero extra deps
• 02_faiss_basics.py — FAISS with save/load, metadata filtering

Running Tests

pip install -e ".[dev]"
python -m pytest

Tests use InMemoryVectorDB — no external server or extra install required.

Design

See docs/DESIGN.md for:

• Why pre-computed embeddings?
• Adapter comparison table
• Score normalisation convention
• How to add a new adapter

License

MIT