microvec 0.2.0

A lightweight, production-grade in-memory vector database

MicroVector

A lightweight, production-grade in-memory vector database for Python.

pip install microvector

No external services. No complex setup. Just numpy and your embeddings.

---

Features

• Fast similarity search — cosine, euclidean, and dot product metrics
• Rich results — every search result includes the document text and metadata, not just an index
• Batch operations — insert thousands of vectors in one call
• Metadata filtering — filter candidates before scoring with any Python predicate
• Safe persistence — JSON + numpy format (no pickle, no security risk)
• Full type hints — works great with mypy and IDEs
• Zero required dependencies beyond numpy

---

Quick Start

microvec stores and searches vectors — you bring the embeddings from any model you like.

With OpenAI

import numpy as np
from openai import OpenAI
from microvector import MicroVectorDB

client = OpenAI()  # uses OPENAI_API_KEY env var

def embed(text: str) -> np.ndarray:
    result = client.embeddings.create(input=text, model="text-embedding-3-small")
    return np.array(result.data[0].embedding, dtype=np.float32)

db = MicroVectorDB()  # dimension inferred from first insert
db.add_node(embed("Paris is the capital of France"), "Paris is the capital of France")
db.add_node(embed("The Eiffel Tower is in Paris"),   "The Eiffel Tower is in Paris")
db.add_node(embed("Python is a programming language"), "Python is a programming language")

results = db.search_top_k(embed("What city is the Eiffel Tower in?"), k=2)
for r in results:
    print(f"{r.score:.3f}  {r.document}")
# 0.921  The Eiffel Tower is in Paris
# 0.887  Paris is the capital of France

With a local model (no API key)

import numpy as np
from sentence_transformers import SentenceTransformer
from microvector import MicroVectorDB

model = SentenceTransformer("all-MiniLM-L6-v2")  # pip install sentence-transformers

def embed(text: str) -> np.ndarray:
    return model.encode(text, normalize_embeddings=True)

db = MicroVectorDB()  # dimension inferred from first insert
db.add_node(embed("Paris is the capital of France"), "Paris is the capital of France")

results = db.search_top_k(embed("What is the capital of France?"), k=1)
print(results[0].document)  # Paris is the capital of France

Save and reload

db.save("my_knowledge_base")
db = MicroVectorDB.load("my_knowledge_base")

---

Installation

Requires Python 3.9+ and numpy >= 1.21.

pip install microvector

Or from source:

git clone https://github.com/huolter/microVector
cd microVector
pip install -e ".[dev]"

---

API Reference

MicroVectorDB(dimension=None)

Create a new database. If dimension is omitted, it is inferred automatically from the first vector inserted and locked for all subsequent inserts.

db = MicroVectorDB()           # recommended — dimension inferred from first insert
db = MicroVectorDB(dimension=512)  # or set explicitly

---

add_node(vector, document, metadata=None, num_bits=None) → int

Add a single vector. Returns the assigned index.

idx = db.add_node(
    vector=np.array([...]),
    document="The text this vector represents",
    metadata={"source": "wikipedia", "date": "2024-01"},
    num_bits=8,   # optional: apply 8-bit scalar quantization
)

Parameter	Type	Description
vector	np.ndarray	1-D array matching the database dimension
document	str	Text or identifier associated with this vector
metadata	dict	Optional key-value pairs stored alongside the vector
num_bits	int	Optional quantization (1–16 bits). Reduces memory at the cost of precision.

---

add_nodes(vectors, documents, metadata=None) → list[int]

Batch insert. All inputs are validated before any insertion occurs.

import numpy as np

vectors = np.random.rand(1000, 512)
docs = [f"Document {i}" for i in range(1000)]
indices = db.add_nodes(vectors, docs)

---

search_top_k(query_vector, k, metric='cosine', filter_fn=None) → list[SearchResult]

Find the top-k most similar vectors. Returns results sorted by score descending.

results = db.search_top_k(query, k=5)
results = db.search_top_k(query, k=5, metric="euclidean")

# Filter before scoring — only consider documents tagged "news"
results = db.search_top_k(
    query, k=5,
    filter_fn=lambda node: node.metadata.get("type") == "news"
)

Metrics:

Metric	Range	Notes
cosine	[-1, 1]	Best for text embeddings. Direction-based, scale-invariant.
euclidean	(0, 1]	1 / (1 + dist). Identical vectors = 1.0.
dot	(-∞, +∞)	Fastest. Best for pre-normalized unit vectors.

SearchResult fields:

result.index     # int   — node's assigned index
result.document  # str   — the document text
result.score     # float — similarity score (higher = more similar)
result.metadata  # dict  — metadata stored with this node

---

search_by_threshold(query_vector, min_score, metric='cosine') → list[SearchResult]

Return all nodes with score >= min_score, sorted descending.

results = db.search_by_threshold(query, min_score=0.8)

---

get_node(index) → Node

Retrieve a node by index.

node = db.get_node(42)
print(node.document, node.metadata)

---

update_node(index, vector=None, document=None, metadata=None)

Update fields of an existing node. Omit fields to leave them unchanged.

db.update_node(42, document="Updated text")
db.update_node(42, metadata={"status": "reviewed"})

---

remove_node(index)

Remove a node. Its index is permanently retired (never reused).

db.remove_node(42)

---

save(path) / MicroVectorDB.load(path)

Persist to and restore from a .mvdb/ directory. Safe format: JSON + numpy binary.

db.save("my_db")              # creates my_db.mvdb/
db = MicroVectorDB.load("my_db")

The .mvdb/ directory contains:

• index.json — node metadata and documents (human-readable)
• vectors.npy — all vectors as a 2-D numpy array

---

Utility methods

len(db)           # number of nodes
42 in db          # check if index exists
repr(db)          # MicroVectorDB(dimension=512, nodes=1000, next_index=1000)
db.stats()        # {'count': 1000, 'dimension': 512, 'next_index': 1000, ...}

---

Exceptions

All exceptions inherit from MicroVectorError.

from microvector import (
    MicroVectorError,
    DimensionMismatchError,  # wrong vector dimension
    EmptyDatabaseError,      # search on empty database
    NodeNotFoundError,       # get/update/remove non-existent index
)

---

Design Notes

Why in-memory? MicroVector is designed for small-to-medium datasets (up to ~100k vectors) where the simplicity of pure-Python outweighs the need for a dedicated service. It starts instantly, needs no configuration, and is trivially embeddable in any Python application.

Why not pickle? Pickle can execute arbitrary code when loading untrusted files. MicroVector uses JSON + numpy binary format which is safe, portable, and inspectable with any text editor.

Index monotonicity. Deleted indexes are never reused. This prevents stale external references from silently pointing to new data.

---

Roadmap

• Approximate nearest neighbor (HNSW)
• Voronoi cell indexing
• FAISS optional backend for large datasets
• Async search support
• Benchmarks

---

License

MIT