A fast, SQLite-like embedded vector database with graph-based approximate nearest neighbor search
Open-source project from the team behind Pardus AI

PardusDB is designed for developers building local AI applications — RAG pipelines, semantic search, recommendation systems, or any project that needs lightweight, persistent vector storage without external dependencies.

While Pardus AI gives non-technical users a powerful no-code platform to ask questions of their CSV, JSON, and PDF data in plain English, PardusDB gives developers the same speed and privacy in an embeddable, fully open-source vector database.

• Single-file storage — Everything lives in one .pardus file, just like SQLite
• Multiple tables — Store different vector dimensions and metadata in the same database
• Familiar SQL-like syntax — CREATE, INSERT, SELECT, UPDATE, DELETE feel natural
• UNIQUE constraints — O(1) duplicate detection using HashSet
• GROUP BY with aggregates — O(n) hash aggregation with COUNT, SUM, AVG, MIN, MAX
• JOINs — O(n+m) hash join algorithm for INNER, LEFT, RIGHT joins
• Fast vector similarity search — Graph-based approximate nearest neighbor search
• Thread-safe — Safe concurrent reads in multi-threaded applications
• Full transactions — BEGIN/COMMIT/ROLLBACK for atomic operations
• Optional GPU acceleration — For large batch inserts and queries
• Zero external dependencies — Pure Rust, MIT licensed

git clone https://github.com/pardus-ai/pardusdb
cd pardusdb
./setup.sh

This will build PardusDB and install it as the pardusdb command, available system-wide.

git clone https://github.com/pardus-ai/pardusdb
cd pardusdb
cargo build --release

The binary will be at target/release/pardusdb.

pardusdb

╔═══════════════════════════════════════════════════════════════╗
║                    PardusDB REPL                      ║
║          Vector Database with SQL Interface           ║
╚═══════════════════════════════════════════════════════════════╝

pardusdb [memory]> .create mydb.pardus
Created and opened: mydb.pardus

pardusdb [mydb.pardus]> CREATE TABLE docs (embedding VECTOR(768), content TEXT);
Table 'docs' created

pardusdb [mydb.pardus]> INSERT INTO docs (embedding, content)
VALUES ([0.1, 0.2, 0.3, ...], 'Hello World');
Inserted row with id=1

pardusdb [mydb.pardus]> SELECT * FROM docs
WHERE embedding SIMILARITY [0.1, 0.2, 0.3, ...] LIMIT 5;

Found 1 similar rows:
  id=1, distance=0.0000, values=[Vector([...]), Text("Hello World")]

pardusdb [mydb.pardus]> quit
Saved to: mydb.pardus
Goodbye!

# Persistent file
pardusdb mydata.pardus

# In-memory only
pardusdb

CREATE TABLE documents (
    id INTEGER PRIMARY KEY,
    embedding VECTOR(768),
    title TEXT,
    category TEXT,
    score FLOAT
);

INSERT INTO documents (embedding, title, category, score)
VALUES ([0.1, 0.2, ...], 'Introduction to Rust', 'tutorial', 0.95);

SELECT * FROM documents WHERE category = 'tutorial' LIMIT 10;

UPDATE documents SET score = 0.99 WHERE id = 1;

DELETE FROM documents WHERE id = 1;

Ensure column values are unique with O(1) duplicate detection:

CREATE TABLE users (
    embedding VECTOR(128),
    id INTEGER PRIMARY KEY,
    email TEXT UNIQUE
);

-- This will fail - duplicate email
INSERT INTO users (embedding, id, email) VALUES ([0.1, ...], 1, '[email protected]');
INSERT INTO users (embedding, id, email) VALUES ([0.2, ...], 2, '[email protected]');
-- Error: Duplicate value for UNIQUE column 'email'

Group and aggregate data with O(n) hash aggregation:

-- Aggregate functions: COUNT, SUM, AVG, MIN, MAX
SELECT category, COUNT(*), AVG(score), SUM(amount)
FROM sales
GROUP BY category;

-- With HAVING clause for filtered groups
SELECT category, SUM(amount) as total
FROM sales
GROUP BY category
HAVING SUM(amount) > 1000;

Join tables with O(n+m) hash join algorithm:

-- INNER JOIN
SELECT * FROM orders
INNER JOIN users ON orders.user_id = users.id;

-- LEFT JOIN (include all left rows)
SELECT users.email, orders.product
FROM users
LEFT JOIN orders ON users.id = orders.user_id;

-- RIGHT JOIN (include all right rows)
SELECT * FROM users
RIGHT JOIN orders ON users.id = orders.user_id;

SELECT * FROM documents
WHERE embedding SIMILARITY [0.12, 0.24, ...]
LIMIT 10;

Results are automatically ordered by distance (closest first).

SHOW TABLES;
DROP TABLE documents;

Real-world benchmark comparing PardusDB against Neo4j 5.15 for vector similarity operations.

Test Configuration:

• Vector dimension: 128
• Number of vectors: 10,000
• Number of queries: 100
• Top-K: 10

PardusDB supports batch inserts for massive performance gains:

Run the benchmark yourself:

# Without Neo4j (PardusDB only)
cargo run --release --bin benchmark_neo4j

# With Neo4j comparison (requires Neo4j running)
docker run -d -p 7687:7687 -e NEO4J_AUTH=neo4j/password123 neo4j:5.15
cargo run --release --features neo4j --bin benchmark_neo4j

Accuracy comparison against brute-force exact search (ground truth).

PardusDB Results:

PardusDB vs Neo4j Accuracy Comparison:

Run accuracy benchmark:

# Without Neo4j (PardusDB only)
cargo run --release --bin benchmark_accuracy

# With Neo4j comparison (requires Neo4j running)
cargo run --release --features neo4j --bin benchmark_accuracy

Comparison against HelixDB, an open-source graph-vector database built in Rust.

Test Configuration:

• Vector dimension: 128
• Number of vectors: 10,000
• Number of queries: 100
• Top-K: 10

Run the benchmark yourself:

# Without HelixDB (PardusDB only)
cargo run --release --bin benchmark_helix

# With HelixDB comparison (requires HelixDB running)
curl -sSL "https://install.helix-db.com" | bash
mkdir helix_bench && cd helix_bench
helix init
# Add schema.hx and queries.hx for vectors
helix push dev
cargo run --release --features helix --bin benchmark_helix

A complete RAG example demonstrating PardusDB's features:

cargo run --example simple_rag --release

This shows:

• Creating tables with VECTOR columns
• Individual inserts with insert_direct()
• Batch inserts with insert_batch_direct()
• Similarity search with search_similar()

See examples/python/simple_rag.py — a RAG demo using Ollama for embeddings and PardusDB as the vector store.

cd examples/python
pip install requests
python simple_rag.py

The Pardus AI team built PardusDB because we believe private, local-first AI tools should be accessible to everyone — from individual developers to large teams.

PardusDB gives you the low-level building block for fast, private vector search, while Pardus AI delivers the high-level no-code experience for analysts, marketers, and business users who just want answers from their data.

If you enjoy working with PardusDB, we’d love for you to try Pardus AI — upload your spreadsheets or documents and ask questions in plain English. Free tier available, no credit card required.

MIT License — use it freely in personal and commercial projects.

⭐ Star us on GitHub if you find this useful!
🚀 Building something cool with PardusDB? Share it with us on X or Discord — we’d love to hear from you.

Pardus AI — https://pardusai.org/