turboquant-db

Lightweight embedded vector database built on turboquant-py. Drop-in replacement for ChromaDB with 16x vector compression.

turboquant-db stores vectors using TurboQuant's near-optimal quantization (1-4 bits per coordinate) and metadata in SQLite. It provides a ChromaDB-compatible API with collections, metadata filtering, and concurrent read/write support — all in a few hundred lines of Python with no dependencies beyond turboquant-py and the standard library.

Installation

pip install turboquant-db

Quick Start

import numpy as np
from turbodb import TurboDB

db = TurboDB("./my_db")
collection = db.create_collection("docs", dim=384)

collection.add(
    ids=["doc1", "doc2"],
    vectors=np.random.randn(2, 384),
    metadatas=[{"source": "wiki", "year": 2024}, {"source": "arxiv", "year": 2025}],
)

results = collection.query(vector=np.random.randn(384), k=5)
for r in results:
    print(f"{r.id}: {r.score:.3f} — {r.metadata}")

API Reference

TurboDB(path)

Open or create a database at the given directory path.

db = TurboDB("./my_db")

Methods:

Method	Description
create_collection(name, dim, metric, bit_width)	Create a new collection
get_collection(name)	Open an existing collection
get_or_create_collection(name, dim, metric, bit_width)	Get or create a collection
delete_collection(name)	Delete a collection and all its data
list_collections()	List all collection names

---

Collection

A named group of quantized vectors with metadata.

collection = db.create_collection("docs", dim=384, metric="cosine", bit_width=2)

Parameter	Type	Default	Description
name	str	required	Collection name
dim	int	required	Vector dimensionality
metric	str	"cosine"	Distance metric: "cosine", "ip", or "l2"
bit_width	int	2	Bits per coordinate (1-4). Lower = more compression, less accuracy

Methods:

add(ids, vectors, metadatas, documents)

Add vectors with string IDs, metadata dicts, and optional document text. IDs must be unique.

collection.add(
    ids=["doc1", "doc2"],
    vectors=np.random.randn(2, 384),
    metadatas=[{"source": "wiki"}, {"source": "arxiv"}],
    documents=["The quick brown fox...", "A study of neural networks..."],
)

The documents parameter is optional. When provided, document text is indexed for BM25 keyword search via hybrid_query().

upsert(ids, vectors, metadatas, documents)

Insert or replace vectors. If an ID already exists, its vector, metadata, and document are replaced.

collection.upsert(
    ids=["doc1"],
    vectors=new_vector,
    metadatas=[{"source": "updated"}],
    documents=["Updated document text."],
)

query(vector, k, where, format)

Search for the top-k most similar vectors, optionally filtering by metadata.

results = collection.query(vector=query_vec, k=10)
results[0].id        # "doc2"
results[0].score     # 0.934
results[0].metadata  # {"source": "arxiv"}

Returns a list of QueryResult objects sorted by descending score.

hybrid_query(text, vector, k, fusion, alpha, where, format)

Hybrid search combining BM25 keyword matching with vector similarity. Requires documents to have been passed to add() or upsert().

results = collection.hybrid_query(
    text="quick fox",
    vector=query_vec,
    k=10,
    fusion="rrf",        # "rrf", "weighted", or "dbsf"
    alpha=0.5,           # only used when fusion="weighted"
)

Parameter	Type	Default	Description
text	str	required	Query text for BM25 scoring
vector	array-like	None	Query vector for semantic scoring. If omitted, performs pure BM25
k	int	10	Number of results
fusion	str	"rrf"	Fusion strategy (see below)
alpha	float	0.5	Vector weight for "weighted" fusion. 1.0 = pure vector, 0.0 = pure BM25
where	dict	None	Metadata filter
format	str	None	"chroma" for ChromaDB format

Fusion strategies:

Strategy	Description	When to use
"rrf"	Reciprocal Rank Fusion (Cormack et al.) — combines rankings, ignores raw scores	Default. Zero tuning required, solid out-of-the-box
"weighted"	Convex combination with min-max normalization	Best accuracy when alpha is tuned per dataset
"dbsf"	Distribution-Based Score Fusion — normalizes via mean ± 3·stddev	Robust to score outliers without tuning

Per Kuzi et al. (ACM TOIS 2023), "weighted" outperforms "rrf" when alpha is tuned — even ~50 labeled query pairs suffice.

get(ids)

Retrieve metadata by IDs without performing a search.

items = collection.get(ids=["doc1", "doc2"])
# [{"id": "doc1", "position": 0, "metadata": {"source": "wiki"}}, ...]

delete(ids, where)

Delete vectors by IDs, metadata filter, or both.

collection.delete(ids=["doc1"])
collection.delete(where={"source": {"$eq": "wiki"}})

compact()

Rewrite storage to reclaim space from deleted vectors.

collection.compact()

count() / name / dim / metric

collection.count()   # number of live vectors
collection.name      # "docs"
collection.dim       # 384
collection.metric    # "cosine"

---

QueryResult

Frozen dataclass returned by query() and hybrid_query().

Attribute	Type	Description
id	str	Vector ID
score	float	Similarity score (higher = more similar for cosine/ip)
metadata	dict	Associated metadata
document	str | None	Document text, if stored

---

ChromaDB compatibility

Pass format="chroma" to get results in ChromaDB's column-oriented format:

results = collection.query(vector=query_vec, k=10, format="chroma")
results["ids"][0]        # ["doc2", "doc5", ...]
results["distances"][0]  # [0.934, 0.891, ...]
results["metadatas"][0]  # [{"source": "arxiv"}, ...]

This makes migration straightforward — change the import, update the constructor, and add format="chroma" to your query calls. Remove format="chroma" at your own pace.

Metadata Filtering

Filter syntax matches ChromaDB and Pinecone conventions:

# Comparison operators
collection.query(vector=v, k=10, where={"year": {"$eq": 2025}})
collection.query(vector=v, k=10, where={"year": {"$ne": 2024}})
collection.query(vector=v, k=10, where={"year": {"$gt": 2023}})
collection.query(vector=v, k=10, where={"year": {"$gte": 2024}})
collection.query(vector=v, k=10, where={"year": {"$lt": 2026}})
collection.query(vector=v, k=10, where={"year": {"$lte": 2025}})

# Set operators
collection.query(vector=v, k=10, where={"source": {"$in": ["wiki", "arxiv"]}})
collection.query(vector=v, k=10, where={"source": {"$nin": ["blog"]}})

# Logical combinators
collection.query(vector=v, k=10, where={
    "$and": [
        {"year": {"$gte": 2024}},
        {"source": {"$eq": "arxiv"}},
    ]
})

collection.query(vector=v, k=10, where={
    "$or": [
        {"source": {"$eq": "wiki"}},
        {"year": {"$gt": 2024}},
    ]
})

Multiple top-level fields are implicitly ANDed:

# Equivalent to $and
collection.query(vector=v, k=10, where={"year": {"$gte": 2024}, "source": {"$eq": "arxiv"}})

Distance Metrics

Metric	Description	Score interpretation
cosine (default)	Cosine similarity	1.0 = identical, 0.0 = orthogonal
ip	Inner product	Higher = more similar
l2	Squared L2 distance	Lower = more similar

All metrics use TurboQuant's inner-product quantizer under the hood. Cosine normalizes vectors on add; L2 is computed from stored norms and inner products.

Compression

turbo-db compresses vectors using TurboQuant's Lloyd-Max quantization with random orthogonal rotation:

Bit-width	Compression ratio	Use case
1	32x	Maximum compression, rough similarity
2 (default)	16x	Good balance of quality and size
3	10.7x	Higher accuracy
4	8x	Near-lossless similarity search

At the default 2-bit setting, a collection of 1M 384-dimensional vectors uses ~9.6 MB for vector data, compared to ~1.5 GB uncompressed.

Storage

Each database is a directory. Each collection is a subdirectory:

my_db/
├── docs/
│   ├── vectors/        # Quantized vectors (numpy arrays)
│   ├── metadata.db     # SQLite: IDs, metadata, positions
│   └── lock            # Write lock file
├── embeddings/
│   └── ...
└── turbodb.json        # Database config

Metadata is stored in SQLite with WAL mode for concurrent read/write access. Vector data uses turboquant-py's bit-packed numpy format.

Concurrency

• Multiple readers + one writer: SQLite WAL mode allows concurrent reads during writes
• Write serialization: File locking ensures one write operation at a time per collection
• Crash safety: Vectors are written before metadata is committed. On restart, orphaned vectors are automatically trimmed to match SQLite state

Migrating from ChromaDB

# Before (ChromaDB)
import chromadb
client = chromadb.PersistentClient(path="./db")
collection = client.create_collection("docs")
collection.add(ids=["a"], embeddings=[[1, 2, 3]], metadatas=[{"k": "v"}])
results = collection.query(query_embeddings=[[1, 2, 3]], n_results=5)

# After (turboquant-db)
from turbodb import TurboDB
db = TurboDB("./db")
collection = db.create_collection("docs", dim=3)
collection.add(ids=["a"], vectors=[[1, 2, 3]], metadatas=[{"k": "v"}])
results = collection.query(vector=[1, 2, 3], k=5)
# Or with Chroma-compat format:
results = collection.query(vector=[1, 2, 3], k=5, format="chroma")

Key differences:

• embeddings → vectors
• query_embeddings → vector (single vector, not nested list)
• n_results → k
• dim is required on create_collection
• Results are QueryResult objects by default (use format="chroma" for column dicts)

References

• TurboQuant: arXiv:2504.19874
• QJL: arXiv:2406.03482
• turboquant-py: github.com/msilverblatt/turboquant-py