quantara 0.1.3

A local-first vector database engine with pluggable indexes, collection management, metadata filtering, and vector similarity search.

Quantara

A lightweight, local-first vector database built in pure Python with support for collections, metadata filtering, persistence, pluggable indexes, and custom index registration.

Quantara is designed for developers, researchers, and students who want a simple yet extensible vector database that runs entirely on their local machine without requiring external services.

---

Features

Core Database

• Collection-based organization
• CRUD operations
• Batch document insertion
• Metadata support
• Metadata filtering during search
• Collection cloning
• Collection renaming
• Collection deletion
• Collection statistics

Vector Search

• Cosine Similarity
• Dot Product Similarity
• Euclidean Distance
• Top-K nearest neighbor search
• Configurable search metrics

Persistence

• Local .db storage
• Automatic persistence
• JSON import/export
• Database configuration persistence
• Index persistence and restoration

Indexing

• Pluggable indexing architecture
• Built-in Brute Force index
• Custom index registration API
• Collection-specific indexes
• Index save/load support

Developer Experience

• Type hints throughout the codebase
• Dataclass-based records
• Extensible architecture
• Local-first design
• No external database dependencies

---

Installation

pip install quantara

Or install from source:

git clone https://github.com/<your-username>/quantara.git

cd quantara

pip install -e .

---

Quick Start

import ollama

from quantara import Database

EMBEDDING_MODEL = "snowflake-arctic-embed:335m"

db = Database(
    "my_database",
    auto_persist=True
)

text = "I want to build AI agents."

embedding = ollama.embeddings(
    model=EMBEDDING_MODEL,
    prompt=text
)["embedding"]

db.insert_doc(
    name=text,
    vector=embedding,
    metadata={
        "category": "ai"
    }
)

query = "How can I build autonomous AI systems?"

query_embedding = ollama.embeddings(
    model=EMBEDDING_MODEL,
    prompt=query
)["embedding"]

results = db.search_doc(
    query_embedding,
    top_k=3
)

print(results)

---

Collections

Create a collection:

db.create_collection(
    "research"
)

Insert into a collection:

db.insert_doc(
    collection="research",
    name="Transformers Paper",
    vector=embedding,
    metadata={
        "author": "Vaswani"
    }
)

List collections:

db.list_collections()

Rename a collection:

db.rename_collection(
    "research",
    "papers"
)

Clone a collection:

db.clone_collection(
    "papers",
    "papers_backup"
)

Delete a collection:

db.delete_collection(
    "papers_backup"
)

---

Metadata Filtering

Search only documents matching specific metadata:

results = db.search_doc(
    query_embedding,
    collection="research",
    filters={
        "author": "Vaswani"
    }
)

---

Batch Insertion

db.batch_insert_docs(
    objects=[
        (
            "Document 1",
            embedding_1,
            {
                "type": "paper"
            }
        ),
        (
            "Document 2",
            embedding_2,
            {
                "type": "article"
            }
        )
    ],
    collection="research"
)

---

Similarity Metrics

Quantara supports multiple similarity metrics:

db.search_doc(
    query_embedding,
    metric="cosine"
)

db.search_doc(
    query_embedding,
    metric="dot"
)

db.search_doc(
    query_embedding,
    metric="euclidean"
)

---

Persistence

Persist database state:

db.persist_doc()

Export database:

db.export_to_json(
    "backup.json"
)

Import database:

db.import_from_json(
    "backup.json"
)

---

Indexes

Saving an Index

db.save_index(
    collection="research",
    path="research.index"
)

Loading an Index

db = Database(
    "my_database",
    index_paths={
        "research": "research.index"
    }
)

---

Creating Custom Indexes

Quantara provides a pluggable indexing architecture.

Create a custom index:

from quantara import BaseIndex


class MyIndex(BaseIndex):

    def build(self, vectors):
        ...

    def add(self, record_id, vector):
        ...

    def remove(self, record_id):
        ...

    def update(self, record_id, vector):
        ...

    def search(
        self,
        query_vector,
        top_k=3,
        **kwargs
    ):
        ...

    def clear(self):
        ...

    def size(self):
        ...

    def save(self, path):
        ...

    def load(self, path):
        ...

Register the index:

from quantara import register_index

register_index(
    "myindex",
    MyIndex
)

Use it in a collection:

db.create_collection(
    "research",
    index_type="myindex"
)

---

Statistics

Database statistics:

db.stats()

Collection statistics:

db.collection_stats(
    "research"
)

---

Architecture

Database
│
├── Collections
│   ├── Records
│   └── Indexes
│
├── Persistence Layer
│
├── Metadata Layer
│
└── Search Layer

---

Project Goals

Quantara aims to provide:

• A lightweight local vector database
• A simple developer experience
• Extensible indexing support
• Fast experimentation for AI and RAG applications
• An educational implementation of vector database concepts

---

Roadmap

Version 1.x

• Brute Force Index
• Metadata Filtering
• Collection Management
• Persistent Storage
• Custom Index Registration

Version 2.x

• HNSW Index
• Approximate Nearest Neighbor Search
• Additional Search Optimizations

Version 3.x

• Vector Quantization
• Compressed Vector Storage
• Advanced Indexing Strategies

---

License

MIT License

---

Author

Abhijeet Rajhans

Built as a learning-focused vector database project exploring vector search, indexing systems, persistence, and database architecture in Python.